diff --git "a/multilang/ggml-medium-encoder-openvino-int4/ggml-medium-encoder-openvino.xml" "b/multilang/ggml-medium-encoder-openvino-int4/ggml-medium-encoder-openvino.xml"
--- "a/multilang/ggml-medium-encoder-openvino-int4/ggml-medium-encoder-openvino.xml"
+++ "b/multilang/ggml-medium-encoder-openvino-int4/ggml-medium-encoder-openvino.xml"
@@ -2,12 +2,12 @@
 <net name="main_graph" version="11">
 	<layers>
 		<layer id="0" name="mel" type="Parameter" version="opset1">
-			<data shape="1,80,2000" element_type="f32" />
+			<data shape="1,80,3000" element_type="f32" />
 			<output>
 				<port id="0" precision="FP32" names="mel">
 					<dim>1</dim>
 					<dim>80</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -22,7 +22,7 @@
 			</output>
 		</layer>
 		<layer id="2" name="Convert_2377" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U8">
 					<dim>1024</dim>
@@ -31,7 +31,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>80</dim>
 					<dim>3</dim>
@@ -49,7 +49,7 @@
 			</output>
 		</layer>
 		<layer id="4" name="Convert_2380" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U8">
 					<dim>1024</dim>
@@ -58,29 +58,29 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>1</dim>
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="5" name="Subtract_2381" type="Subtract" version="opset1">
+		<layer id="5" name="onnx::Conv_2858/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>1024</dim>
 					<dim>80</dim>
 					<dim>3</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>1</dim>
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>1024</dim>
 					<dim>80</dim>
 					<dim>3</dim>
@@ -97,39 +97,39 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="7" name="onnx::Conv_2858/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
+		<layer id="7" name="onnx::Conv_2858/fq_weights_1" type="Multiply" version="opset1">
+			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>80</dim>
+					<dim>3</dim>
+				</port>
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>1</dim>
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>1024</dim>
-					<dim>1</dim>
-					<dim>1</dim>
+					<dim>80</dim>
+					<dim>3</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="8" name="onnx::Conv_2858/fq_weights_1" type="Multiply" version="opset1">
-			<data auto_broadcast="numpy" />
+		<layer id="8" name="onnx::Conv_2858/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>1024</dim>
 					<dim>80</dim>
 					<dim>3</dim>
 				</port>
-				<port id="1" precision="FP32">
-					<dim>1024</dim>
-					<dim>1</dim>
-					<dim>1</dim>
-				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>80</dim>
 					<dim>3</dim>
@@ -142,7 +142,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>80</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1024</dim>
@@ -154,7 +154,7 @@
 				<port id="2" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -193,7 +193,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</input>
 			<output>
@@ -298,7 +298,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
@@ -310,7 +310,7 @@
 				<port id="2" precision="FP32" names="/conv1/Conv_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -326,7 +326,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -334,7 +334,7 @@
 				<port id="2" precision="FP32" names="/Div_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -343,14 +343,14 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/Erf_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -366,7 +366,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -374,7 +374,7 @@
 				<port id="2" precision="FP32" names="/Add_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -384,19 +384,19 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/Mul_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -412,7 +412,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -420,7 +420,7 @@
 				<port id="2" precision="FP32" names="/Mul_1_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 			</output>
 		</layer>
@@ -435,7 +435,7 @@
 			</output>
 		</layer>
 		<layer id="30" name="Convert_2388" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U8">
 					<dim>1024</dim>
@@ -444,7 +444,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>1024</dim>
 					<dim>3</dim>
@@ -462,7 +462,7 @@
 			</output>
 		</layer>
 		<layer id="32" name="Convert_2391" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U8">
 					<dim>1024</dim>
@@ -471,29 +471,29 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>1</dim>
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="33" name="Subtract_2392" type="Subtract" version="opset1">
+		<layer id="33" name="onnx::Conv_2860/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>1024</dim>
 					<dim>1024</dim>
 					<dim>3</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>1</dim>
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>1024</dim>
 					<dim>1024</dim>
 					<dim>3</dim>
@@ -510,39 +510,39 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="35" name="onnx::Conv_2860/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
+		<layer id="35" name="onnx::Conv_2860/fq_weights_1" type="Multiply" version="opset1">
+			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+					<dim>3</dim>
+				</port>
+				<port id="1" precision="FP16">
 					<dim>1024</dim>
 					<dim>1</dim>
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>1024</dim>
-					<dim>1</dim>
-					<dim>1</dim>
+					<dim>1024</dim>
+					<dim>3</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="36" name="onnx::Conv_2860/fq_weights_1" type="Multiply" version="opset1">
-			<data auto_broadcast="numpy" />
+		<layer id="36" name="onnx::Conv_2860/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>1024</dim>
 					<dim>1024</dim>
 					<dim>3</dim>
 				</port>
-				<port id="1" precision="FP32">
-					<dim>1024</dim>
-					<dim>1</dim>
-					<dim>1</dim>
-				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 					<dim>3</dim>
@@ -555,7 +555,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>2000</dim>
+					<dim>3000</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1024</dim>
@@ -567,7 +567,7 @@
 				<port id="2" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -606,7 +606,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
@@ -711,7 +711,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
@@ -723,7 +723,7 @@
 				<port id="2" precision="FP32" names="/conv2/Conv_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -739,7 +739,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -747,7 +747,7 @@
 				<port id="2" precision="FP32" names="/Div_1_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -756,14 +756,14 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/Erf_1_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -779,7 +779,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -787,7 +787,7 @@
 				<port id="2" precision="FP32" names="/Add_1_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -797,19 +797,19 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/Mul_2_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -825,7 +825,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -833,7 +833,7 @@
 				<port id="2" precision="FP32" names="/Mul_3_output_0">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -850,7 +850,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>1024</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="I64">
 					<dim>3</dim>
@@ -859,16 +859,16 @@
 			<output>
 				<port id="2" precision="FP32" names="/Transpose_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="59" name="onnx::Add_2866" type="Const" version="opset1">
-			<data element_type="f32" shape="1000, 1024" offset="3405876" size="4096000" />
+			<data element_type="f32" shape="1500, 1024" offset="3405876" size="6144000" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2866">
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -878,18 +878,18 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/Add_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -899,14 +899,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -916,20 +916,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="63" name="onnx::Add_2891" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="7501876" size="4096" />
+			<data element_type="f32" shape="1024" offset="9549876" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2891">
 					<dim>1024</dim>
@@ -937,7 +937,7 @@
 			</output>
 		</layer>
 		<layer id="64" name="onnx::Add_2868" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="7505972" size="4096" />
+			<data element_type="f32" shape="1024" offset="9553972" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2868">
 					<dim>1024</dim>
@@ -945,13 +945,13 @@
 			</output>
 		</layer>
 		<layer id="65" name="Constant_424" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="66" name="Constant_423" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -980,7 +980,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -990,13 +990,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="70" name="blocks.0.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="7510084" size="4096" />
+			<data element_type="f32" shape="1024" offset="9558084" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.0.attn_ln.weight">
 					<dim>1024</dim>
@@ -1008,7 +1008,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -1018,13 +1018,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="72" name="blocks.0.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="7514180" size="4096" />
+			<data element_type="f32" shape="1024" offset="9562180" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.0.attn_ln.bias">
 					<dim>1024</dim>
@@ -1036,7 +1036,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -1046,7 +1046,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -1056,20 +1056,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="75" name="onnx::MatMul_2869" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="7518276" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="9566276" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -1079,7 +1079,7 @@
 			</output>
 		</layer>
 		<layer id="76" name="Convert_2420" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -1088,7 +1088,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1096,7 +1096,7 @@
 			</output>
 		</layer>
 		<layer id="77" name="onnx::MatMul_2869/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -1104,32 +1104,32 @@
 			</output>
 		</layer>
 		<layer id="78" name="Convert_2422" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="79" name="Subtract_2423" type="Subtract" version="opset1">
+		<layer id="79" name="onnx::MatMul_2869/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1137,7 +1137,7 @@
 			</output>
 		</layer>
 		<layer id="80" name="onnx::MatMul_2869/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="8042565" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="10090565" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -1146,57 +1146,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="81" name="onnx::MatMul_2869/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="82" name="onnx::MatMul_2869/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="81" name="onnx::MatMul_2869/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="83" name="Constant_2428" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="82" name="Constant_2427" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="84" name="Reshape_2429" type="Reshape" version="opset1">
+		<layer id="83" name="Reshape_2428" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1206,7 +1189,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="84" name="onnx::MatMul_2869/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -1217,7 +1215,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -1228,7 +1226,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -1241,20 +1239,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="87" name="/blocks.0/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.0/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -1266,7 +1264,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -1276,14 +1274,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="89" name="Constant_440" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -1294,7 +1292,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -1306,19 +1304,19 @@
 				<port id="2" precision="FP32" names="/blocks.0/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="91" name="/blocks.0/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.0/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="92" name="/blocks.0/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.0/attn/Constant_1_output_0" />
 			</output>
@@ -1339,7 +1337,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -1348,13 +1346,13 @@
 				<port id="2" precision="FP32" names="/blocks.0/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="95" name="onnx::MatMul_2871" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="8059037" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="10107037" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -1364,7 +1362,7 @@
 			</output>
 		</layer>
 		<layer id="96" name="Convert_2409" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -1373,7 +1371,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1381,7 +1379,7 @@
 			</output>
 		</layer>
 		<layer id="97" name="onnx::MatMul_2871/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10631325" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -1389,32 +1387,32 @@
 			</output>
 		</layer>
 		<layer id="98" name="Convert_2411" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="99" name="Subtract_2412" type="Subtract" version="opset1">
+		<layer id="99" name="onnx::MatMul_2871/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1422,7 +1420,7 @@
 			</output>
 		</layer>
 		<layer id="100" name="onnx::MatMul_2871/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="8583325" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="10631326" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -1431,57 +1429,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="101" name="onnx::MatMul_2871/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="102" name="onnx::MatMul_2871/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="101" name="onnx::MatMul_2871/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="103" name="Constant_2417" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="102" name="Constant_2416" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="104" name="Reshape_2418" type="Reshape" version="opset1">
+		<layer id="103" name="Reshape_2417" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1491,7 +1472,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="104" name="onnx::MatMul_2871/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -1502,7 +1498,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -1513,13 +1509,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="106" name="/blocks.0/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.0/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -1531,7 +1527,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -1541,14 +1537,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="108" name="Constant_445" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -1559,7 +1555,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -1572,7 +1568,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -1583,7 +1579,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -1592,7 +1588,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -1602,22 +1598,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -1627,16 +1623,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -1646,16 +1642,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -1665,21 +1661,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="115" name="onnx::Add_2873" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="8599741" size="4096" />
+			<data element_type="f32" shape="1024" offset="10647742" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2873">
 					<dim>1024</dim>
@@ -1687,7 +1683,7 @@
 			</output>
 		</layer>
 		<layer id="116" name="onnx::MatMul_2874" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="8603837" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="10651838" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -1697,7 +1693,7 @@
 			</output>
 		</layer>
 		<layer id="117" name="Convert_2398" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -1706,7 +1702,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1714,7 +1710,7 @@
 			</output>
 		</layer>
 		<layer id="118" name="onnx::MatMul_2874/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -1722,32 +1718,32 @@
 			</output>
 		</layer>
 		<layer id="119" name="Convert_2400" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="120" name="Subtract_2401" type="Subtract" version="opset1">
+		<layer id="120" name="onnx::MatMul_2874/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1755,7 +1751,7 @@
 			</output>
 		</layer>
 		<layer id="121" name="onnx::MatMul_2874/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="9128125" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="11176126" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -1764,57 +1760,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="122" name="onnx::MatMul_2874/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="123" name="onnx::MatMul_2874/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="122" name="onnx::MatMul_2874/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="124" name="Constant_2406" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="123" name="Constant_2405" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="125" name="Reshape_2407" type="Reshape" version="opset1">
+		<layer id="124" name="Reshape_2406" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -1824,7 +1803,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="125" name="onnx::MatMul_2874/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -1835,7 +1829,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -1846,7 +1840,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -1859,20 +1853,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="128" name="/blocks.0/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.0/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -1884,7 +1878,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -1894,14 +1888,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="130" name="Constant_450" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -1912,7 +1906,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -1924,7 +1918,7 @@
 				<port id="2" precision="FP32" names="/blocks.0/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -1935,13 +1929,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -1949,13 +1943,13 @@
 				<port id="2" precision="FP32" names="/blocks.0/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="133" name="Constant_457" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -1967,7 +1961,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -1977,7 +1971,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -1988,7 +1982,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -2000,7 +1994,7 @@
 			</output>
 		</layer>
 		<layer id="136" name="/blocks.0/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.0/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -2053,7 +2047,7 @@
 			</output>
 		</layer>
 		<layer id="141" name="/blocks.0/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.0/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -2085,7 +2079,7 @@
 			</output>
 		</layer>
 		<layer id="143" name="/blocks.0/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.0/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -2113,7 +2107,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -2124,13 +2118,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="146" name="onnx::MatMul_2892" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="9144509" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="11192510" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -2140,7 +2134,7 @@
 			</output>
 		</layer>
 		<layer id="147" name="Convert_2431" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -2149,7 +2143,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -2157,7 +2151,7 @@
 			</output>
 		</layer>
 		<layer id="148" name="onnx::MatMul_2892/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -2165,32 +2159,32 @@
 			</output>
 		</layer>
 		<layer id="149" name="Convert_2433" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="150" name="Subtract_2434" type="Subtract" version="opset1">
+		<layer id="150" name="onnx::MatMul_2892/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -2198,7 +2192,7 @@
 			</output>
 		</layer>
 		<layer id="151" name="onnx::MatMul_2892/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="9668797" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="11716799" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -2207,57 +2201,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="152" name="onnx::MatMul_2892/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="153" name="onnx::MatMul_2892/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="152" name="onnx::MatMul_2892/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="154" name="Constant_2439" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="153" name="Constant_2438" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="155" name="Reshape_2440" type="Reshape" version="opset1">
+		<layer id="154" name="Reshape_2439" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -2267,7 +2244,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="155" name="onnx::MatMul_2892/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -2278,7 +2270,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -2289,7 +2281,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -2302,14 +2294,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -2319,19 +2311,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -2341,20 +2333,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="160" name="onnx::Add_2897" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="9685181" size="4096" />
+			<data element_type="f32" shape="1024" offset="11733183" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2897">
 					<dim>1024</dim>
@@ -2362,7 +2354,7 @@
 			</output>
 		</layer>
 		<layer id="161" name="onnx::Add_2894" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="9689277" size="16384" />
+			<data element_type="f32" shape="4096" offset="11737279" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2894">
 					<dim>4096</dim>
@@ -2370,13 +2362,13 @@
 			</output>
 		</layer>
 		<layer id="162" name="Constant_483" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="163" name="Constant_482" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -2405,7 +2397,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -2415,13 +2407,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="167" name="blocks.0.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="9705661" size="4096" />
+			<data element_type="f32" shape="1024" offset="11753663" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.0.mlp_ln.weight">
 					<dim>1024</dim>
@@ -2433,7 +2425,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -2443,13 +2435,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="169" name="blocks.0.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="9709757" size="4096" />
+			<data element_type="f32" shape="1024" offset="11757759" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.0.mlp_ln.bias">
 					<dim>1024</dim>
@@ -2461,7 +2453,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -2471,7 +2463,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -2481,20 +2473,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="172" name="onnx::MatMul_2895" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="9713853" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="11761855" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -2504,7 +2496,7 @@
 			</output>
 		</layer>
 		<layer id="173" name="Convert_2442" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -2513,7 +2505,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -2521,7 +2513,7 @@
 			</output>
 		</layer>
 		<layer id="174" name="onnx::MatMul_2895/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -2529,32 +2521,32 @@
 			</output>
 		</layer>
 		<layer id="175" name="Convert_2444" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="176" name="Subtract_2445" type="Subtract" version="opset1">
+		<layer id="176" name="onnx::MatMul_2895/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -2562,7 +2554,7 @@
 			</output>
 		</layer>
 		<layer id="177" name="onnx::MatMul_2895/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="11811005" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="13859007" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -2571,57 +2563,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="178" name="onnx::MatMul_2895/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="179" name="onnx::MatMul_2895/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="178" name="onnx::MatMul_2895/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="180" name="Constant_2450" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="179" name="Constant_2449" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="181" name="Reshape_2451" type="Reshape" version="opset1">
+		<layer id="180" name="Reshape_2450" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -2631,7 +2606,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="181" name="onnx::MatMul_2895/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -2642,7 +2632,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -2653,7 +2643,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -2666,14 +2656,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -2689,7 +2679,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -2697,7 +2687,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -2706,14 +2696,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.0/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -2729,7 +2719,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -2737,7 +2727,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -2747,19 +2737,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -2775,7 +2765,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -2783,13 +2773,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="192" name="onnx::MatMul_2898" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="11876557" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="13924559" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -2799,7 +2789,7 @@
 			</output>
 		</layer>
 		<layer id="193" name="Convert_2453" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -2808,7 +2798,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -2816,7 +2806,7 @@
 			</output>
 		</layer>
 		<layer id="194" name="onnx::MatMul_2898/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -2824,32 +2814,32 @@
 			</output>
 		</layer>
 		<layer id="195" name="Convert_2455" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="196" name="Subtract_2456" type="Subtract" version="opset1">
+		<layer id="196" name="onnx::MatMul_2898/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -2857,7 +2847,7 @@
 			</output>
 		</layer>
 		<layer id="197" name="onnx::MatMul_2898/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="13973709" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="16021711" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -2866,57 +2856,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="198" name="onnx::MatMul_2898/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="199" name="onnx::MatMul_2898/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="198" name="onnx::MatMul_2898/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="200" name="Constant_2461" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="199" name="Constant_2460" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="201" name="Reshape_2462" type="Reshape" version="opset1">
+		<layer id="200" name="Reshape_2461" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -2926,7 +2899,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="201" name="onnx::MatMul_2898/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -2937,7 +2925,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -2948,7 +2936,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -2961,14 +2949,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -2978,19 +2966,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.0/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -3000,20 +2988,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="206" name="onnx::Add_2923" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="14039261" size="4096" />
+			<data element_type="f32" shape="1024" offset="16087263" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2923">
 					<dim>1024</dim>
@@ -3021,7 +3009,7 @@
 			</output>
 		</layer>
 		<layer id="207" name="onnx::Add_2900" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="14043357" size="4096" />
+			<data element_type="f32" shape="1024" offset="16091359" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2900">
 					<dim>1024</dim>
@@ -3029,13 +3017,13 @@
 			</output>
 		</layer>
 		<layer id="208" name="Constant_505" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="209" name="Constant_504" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -3064,7 +3052,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -3074,13 +3062,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="213" name="blocks.1.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="14047453" size="4096" />
+			<data element_type="f32" shape="1024" offset="16095455" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.1.attn_ln.weight">
 					<dim>1024</dim>
@@ -3092,7 +3080,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -3102,13 +3090,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="215" name="blocks.1.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="14051549" size="4096" />
+			<data element_type="f32" shape="1024" offset="16099551" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.1.attn_ln.bias">
 					<dim>1024</dim>
@@ -3120,7 +3108,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -3130,7 +3118,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -3140,20 +3128,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="218" name="onnx::MatMul_2901" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="14055645" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="16103647" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -3163,7 +3151,7 @@
 			</output>
 		</layer>
 		<layer id="219" name="Convert_2486" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -3172,7 +3160,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3180,7 +3168,7 @@
 			</output>
 		</layer>
 		<layer id="220" name="onnx::MatMul_2901/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -3188,32 +3176,32 @@
 			</output>
 		</layer>
 		<layer id="221" name="Convert_2488" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="222" name="Subtract_2489" type="Subtract" version="opset1">
+		<layer id="222" name="onnx::MatMul_2901/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3221,7 +3209,7 @@
 			</output>
 		</layer>
 		<layer id="223" name="onnx::MatMul_2901/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="14579933" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="16627935" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -3230,57 +3218,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="224" name="onnx::MatMul_2901/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="225" name="onnx::MatMul_2901/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="224" name="onnx::MatMul_2901/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="226" name="Constant_2494" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="225" name="Constant_2493" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="227" name="Reshape_2495" type="Reshape" version="opset1">
+		<layer id="226" name="Reshape_2494" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3290,7 +3261,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="227" name="onnx::MatMul_2901/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -3301,7 +3287,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -3312,7 +3298,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -3325,20 +3311,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="230" name="/blocks.1/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.1/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -3350,7 +3336,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -3360,14 +3346,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="232" name="Constant_521" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -3378,7 +3364,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -3390,19 +3376,19 @@
 				<port id="2" precision="FP32" names="/blocks.1/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="234" name="/blocks.1/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.1/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="235" name="/blocks.1/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.1/attn/Constant_1_output_0" />
 			</output>
@@ -3423,7 +3409,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -3432,13 +3418,13 @@
 				<port id="2" precision="FP32" names="/blocks.1/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="238" name="onnx::MatMul_2903" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="14596317" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="16644319" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -3448,7 +3434,7 @@
 			</output>
 		</layer>
 		<layer id="239" name="Convert_2475" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -3457,7 +3443,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3465,7 +3451,7 @@
 			</output>
 		</layer>
 		<layer id="240" name="onnx::MatMul_2903/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="17168607" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -3473,32 +3459,32 @@
 			</output>
 		</layer>
 		<layer id="241" name="Convert_2477" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="242" name="Subtract_2478" type="Subtract" version="opset1">
+		<layer id="242" name="onnx::MatMul_2903/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3506,7 +3492,7 @@
 			</output>
 		</layer>
 		<layer id="243" name="onnx::MatMul_2903/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="15120605" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="17168608" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -3515,57 +3501,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="244" name="onnx::MatMul_2903/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="245" name="onnx::MatMul_2903/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="244" name="onnx::MatMul_2903/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="246" name="Constant_2483" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="245" name="Constant_2482" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="247" name="Reshape_2484" type="Reshape" version="opset1">
+		<layer id="246" name="Reshape_2483" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3575,7 +3544,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="247" name="onnx::MatMul_2903/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -3586,7 +3570,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -3597,13 +3581,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="249" name="/blocks.1/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.1/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -3615,7 +3599,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -3625,14 +3609,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="251" name="Constant_526" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -3643,7 +3627,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -3656,7 +3640,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -3667,7 +3651,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -3676,7 +3660,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -3686,22 +3670,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -3711,16 +3695,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -3730,16 +3714,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -3749,21 +3733,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="258" name="onnx::Add_2905" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="15136989" size="4096" />
+			<data element_type="f32" shape="1024" offset="17184992" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2905">
 					<dim>1024</dim>
@@ -3771,7 +3755,7 @@
 			</output>
 		</layer>
 		<layer id="259" name="onnx::MatMul_2906" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="15141085" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="17189088" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -3781,7 +3765,7 @@
 			</output>
 		</layer>
 		<layer id="260" name="Convert_2464" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -3790,7 +3774,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3798,7 +3782,7 @@
 			</output>
 		</layer>
 		<layer id="261" name="onnx::MatMul_2906/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -3806,32 +3790,32 @@
 			</output>
 		</layer>
 		<layer id="262" name="Convert_2466" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="263" name="Subtract_2467" type="Subtract" version="opset1">
+		<layer id="263" name="onnx::MatMul_2906/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3839,7 +3823,7 @@
 			</output>
 		</layer>
 		<layer id="264" name="onnx::MatMul_2906/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="15665373" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="17713376" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -3848,57 +3832,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="265" name="onnx::MatMul_2906/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="266" name="onnx::MatMul_2906/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="265" name="onnx::MatMul_2906/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="267" name="Constant_2472" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="266" name="Constant_2471" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="268" name="Reshape_2473" type="Reshape" version="opset1">
+		<layer id="267" name="Reshape_2472" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -3908,7 +3875,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="268" name="onnx::MatMul_2906/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -3919,7 +3901,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -3930,7 +3912,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -3943,20 +3925,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="271" name="/blocks.1/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.1/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -3968,7 +3950,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -3978,14 +3960,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="273" name="Constant_531" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -3996,7 +3978,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -4008,7 +3990,7 @@
 				<port id="2" precision="FP32" names="/blocks.1/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -4019,13 +4001,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -4033,13 +4015,13 @@
 				<port id="2" precision="FP32" names="/blocks.1/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="276" name="Constant_538" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -4051,7 +4033,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -4061,7 +4043,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -4072,7 +4054,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -4084,7 +4066,7 @@
 			</output>
 		</layer>
 		<layer id="279" name="/blocks.1/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.1/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -4137,7 +4119,7 @@
 			</output>
 		</layer>
 		<layer id="284" name="/blocks.1/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.1/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -4169,7 +4151,7 @@
 			</output>
 		</layer>
 		<layer id="286" name="/blocks.1/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.1/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -4197,7 +4179,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -4208,13 +4190,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="289" name="onnx::MatMul_2924" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="15681757" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="17729760" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -4224,7 +4206,7 @@
 			</output>
 		</layer>
 		<layer id="290" name="Convert_2497" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -4233,7 +4215,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -4241,7 +4223,7 @@
 			</output>
 		</layer>
 		<layer id="291" name="onnx::MatMul_2924/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -4249,32 +4231,32 @@
 			</output>
 		</layer>
 		<layer id="292" name="Convert_2499" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="293" name="Subtract_2500" type="Subtract" version="opset1">
+		<layer id="293" name="onnx::MatMul_2924/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -4282,7 +4264,7 @@
 			</output>
 		</layer>
 		<layer id="294" name="onnx::MatMul_2924/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="16206045" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="18254048" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -4291,57 +4273,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="295" name="onnx::MatMul_2924/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="296" name="onnx::MatMul_2924/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="295" name="onnx::MatMul_2924/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="297" name="Constant_2505" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="296" name="Constant_2504" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="298" name="Reshape_2506" type="Reshape" version="opset1">
+		<layer id="297" name="Reshape_2505" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -4351,7 +4316,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="298" name="onnx::MatMul_2924/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -4362,7 +4342,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -4373,7 +4353,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -4386,14 +4366,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -4403,19 +4383,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -4425,20 +4405,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="303" name="onnx::Add_2929" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="16222429" size="4096" />
+			<data element_type="f32" shape="1024" offset="18270432" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2929">
 					<dim>1024</dim>
@@ -4446,7 +4426,7 @@
 			</output>
 		</layer>
 		<layer id="304" name="onnx::Add_2926" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="16226525" size="16384" />
+			<data element_type="f32" shape="4096" offset="18274528" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2926">
 					<dim>4096</dim>
@@ -4454,13 +4434,13 @@
 			</output>
 		</layer>
 		<layer id="305" name="Constant_564" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="306" name="Constant_563" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -4489,7 +4469,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -4499,13 +4479,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="310" name="blocks.1.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="16242909" size="4096" />
+			<data element_type="f32" shape="1024" offset="18290912" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.1.mlp_ln.weight">
 					<dim>1024</dim>
@@ -4517,7 +4497,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -4527,13 +4507,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="312" name="blocks.1.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="16247005" size="4096" />
+			<data element_type="f32" shape="1024" offset="18295008" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.1.mlp_ln.bias">
 					<dim>1024</dim>
@@ -4545,7 +4525,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -4555,7 +4535,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -4565,20 +4545,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="315" name="onnx::MatMul_2927" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="16251101" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="18299104" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -4588,7 +4568,7 @@
 			</output>
 		</layer>
 		<layer id="316" name="Convert_2508" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -4597,7 +4577,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -4605,7 +4585,7 @@
 			</output>
 		</layer>
 		<layer id="317" name="onnx::MatMul_2927/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -4613,32 +4593,32 @@
 			</output>
 		</layer>
 		<layer id="318" name="Convert_2510" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="319" name="Subtract_2511" type="Subtract" version="opset1">
+		<layer id="319" name="onnx::MatMul_2927/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -4646,7 +4626,7 @@
 			</output>
 		</layer>
 		<layer id="320" name="onnx::MatMul_2927/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="18348253" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="20396256" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -4655,57 +4635,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="321" name="onnx::MatMul_2927/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="322" name="onnx::MatMul_2927/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="321" name="onnx::MatMul_2927/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="323" name="Constant_2516" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="322" name="Constant_2515" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="324" name="Reshape_2517" type="Reshape" version="opset1">
+		<layer id="323" name="Reshape_2516" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -4715,7 +4678,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="324" name="onnx::MatMul_2927/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -4726,7 +4704,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -4737,7 +4715,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -4750,14 +4728,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -4773,7 +4751,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -4781,7 +4759,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -4790,14 +4768,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.1/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -4813,7 +4791,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -4821,7 +4799,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -4831,19 +4809,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -4859,7 +4837,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -4867,13 +4845,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="335" name="onnx::MatMul_2930" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="18413789" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="20461792" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -4883,7 +4861,7 @@
 			</output>
 		</layer>
 		<layer id="336" name="Convert_2519" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -4892,7 +4870,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -4900,7 +4878,7 @@
 			</output>
 		</layer>
 		<layer id="337" name="onnx::MatMul_2930/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -4908,32 +4886,32 @@
 			</output>
 		</layer>
 		<layer id="338" name="Convert_2521" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="339" name="Subtract_2522" type="Subtract" version="opset1">
+		<layer id="339" name="onnx::MatMul_2930/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -4941,7 +4919,7 @@
 			</output>
 		</layer>
 		<layer id="340" name="onnx::MatMul_2930/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="20510941" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="22558944" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -4950,57 +4928,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="341" name="onnx::MatMul_2930/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="342" name="onnx::MatMul_2930/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="341" name="onnx::MatMul_2930/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="343" name="Constant_2527" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="342" name="Constant_2526" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="344" name="Reshape_2528" type="Reshape" version="opset1">
+		<layer id="343" name="Reshape_2527" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5010,7 +4971,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="344" name="onnx::MatMul_2930/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -5021,7 +4997,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -5032,7 +5008,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -5045,14 +5021,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -5062,19 +5038,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.1/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -5084,20 +5060,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="349" name="onnx::Add_2955" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="20576477" size="4096" />
+			<data element_type="f32" shape="1024" offset="22624480" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2955">
 					<dim>1024</dim>
@@ -5105,7 +5081,7 @@
 			</output>
 		</layer>
 		<layer id="350" name="onnx::Add_2932" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="20580573" size="4096" />
+			<data element_type="f32" shape="1024" offset="22628576" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2932">
 					<dim>1024</dim>
@@ -5113,13 +5089,13 @@
 			</output>
 		</layer>
 		<layer id="351" name="Constant_586" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="352" name="Constant_585" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -5148,7 +5124,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -5158,13 +5134,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="356" name="blocks.2.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="20584669" size="4096" />
+			<data element_type="f32" shape="1024" offset="22632672" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.2.attn_ln.weight">
 					<dim>1024</dim>
@@ -5176,7 +5152,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -5186,13 +5162,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="358" name="blocks.2.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="20588765" size="4096" />
+			<data element_type="f32" shape="1024" offset="22636768" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.2.attn_ln.bias">
 					<dim>1024</dim>
@@ -5204,7 +5180,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -5214,7 +5190,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -5224,20 +5200,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="361" name="onnx::MatMul_2933" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="20592861" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="22640864" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -5247,7 +5223,7 @@
 			</output>
 		</layer>
 		<layer id="362" name="Convert_2552" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -5256,7 +5232,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5264,7 +5240,7 @@
 			</output>
 		</layer>
 		<layer id="363" name="onnx::MatMul_2933/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="23165152" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -5272,32 +5248,32 @@
 			</output>
 		</layer>
 		<layer id="364" name="Convert_2554" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="365" name="Subtract_2555" type="Subtract" version="opset1">
+		<layer id="365" name="onnx::MatMul_2933/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5305,7 +5281,7 @@
 			</output>
 		</layer>
 		<layer id="366" name="onnx::MatMul_2933/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="21117149" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="23165153" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -5314,57 +5290,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="367" name="onnx::MatMul_2933/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="368" name="onnx::MatMul_2933/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="367" name="onnx::MatMul_2933/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="369" name="Constant_2560" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="368" name="Constant_2559" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="370" name="Reshape_2561" type="Reshape" version="opset1">
+		<layer id="369" name="Reshape_2560" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5374,7 +5333,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="370" name="onnx::MatMul_2933/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -5385,7 +5359,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -5396,7 +5370,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -5409,20 +5383,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="373" name="/blocks.2/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.2/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -5434,7 +5408,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -5444,14 +5418,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="375" name="Constant_602" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -5462,7 +5436,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -5474,19 +5448,19 @@
 				<port id="2" precision="FP32" names="/blocks.2/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="377" name="/blocks.2/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.2/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="378" name="/blocks.2/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.2/attn/Constant_1_output_0" />
 			</output>
@@ -5507,7 +5481,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -5516,13 +5490,13 @@
 				<port id="2" precision="FP32" names="/blocks.2/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="381" name="onnx::MatMul_2935" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="21133533" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="23181537" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -5532,7 +5506,7 @@
 			</output>
 		</layer>
 		<layer id="382" name="Convert_2541" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -5541,7 +5515,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5549,7 +5523,7 @@
 			</output>
 		</layer>
 		<layer id="383" name="onnx::MatMul_2935/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10631325" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -5557,32 +5531,32 @@
 			</output>
 		</layer>
 		<layer id="384" name="Convert_2543" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="385" name="Subtract_2544" type="Subtract" version="opset1">
+		<layer id="385" name="onnx::MatMul_2935/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5590,7 +5564,7 @@
 			</output>
 		</layer>
 		<layer id="386" name="onnx::MatMul_2935/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="21657821" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="23705825" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -5599,57 +5573,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="387" name="onnx::MatMul_2935/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="388" name="onnx::MatMul_2935/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="387" name="onnx::MatMul_2935/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="389" name="Constant_2549" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="388" name="Constant_2548" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="390" name="Reshape_2550" type="Reshape" version="opset1">
+		<layer id="389" name="Reshape_2549" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5659,7 +5616,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="390" name="onnx::MatMul_2935/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -5670,7 +5642,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -5681,13 +5653,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="392" name="/blocks.2/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.2/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -5699,7 +5671,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -5709,14 +5681,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="394" name="Constant_607" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -5727,7 +5699,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -5740,7 +5712,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -5751,7 +5723,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -5760,7 +5732,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -5770,22 +5742,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -5795,16 +5767,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -5814,16 +5786,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -5833,21 +5805,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="401" name="onnx::Add_2937" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="21674205" size="4096" />
+			<data element_type="f32" shape="1024" offset="23722209" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2937">
 					<dim>1024</dim>
@@ -5855,7 +5827,7 @@
 			</output>
 		</layer>
 		<layer id="402" name="onnx::MatMul_2938" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="21678301" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="23726305" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -5865,7 +5837,7 @@
 			</output>
 		</layer>
 		<layer id="403" name="Convert_2530" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -5874,7 +5846,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5882,7 +5854,7 @@
 			</output>
 		</layer>
 		<layer id="404" name="onnx::MatMul_2938/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -5890,32 +5862,32 @@
 			</output>
 		</layer>
 		<layer id="405" name="Convert_2532" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="406" name="Subtract_2533" type="Subtract" version="opset1">
+		<layer id="406" name="onnx::MatMul_2938/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5923,7 +5895,7 @@
 			</output>
 		</layer>
 		<layer id="407" name="onnx::MatMul_2938/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="22202589" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="24250593" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -5932,57 +5904,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="408" name="onnx::MatMul_2938/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="409" name="onnx::MatMul_2938/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="408" name="onnx::MatMul_2938/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="410" name="Constant_2538" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="409" name="Constant_2537" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="411" name="Reshape_2539" type="Reshape" version="opset1">
+		<layer id="410" name="Reshape_2538" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -5992,7 +5947,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="411" name="onnx::MatMul_2938/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -6003,7 +5973,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -6014,7 +5984,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -6027,20 +5997,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="414" name="/blocks.2/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.2/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -6052,7 +6022,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -6062,14 +6032,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="416" name="Constant_612" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -6080,7 +6050,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -6092,7 +6062,7 @@
 				<port id="2" precision="FP32" names="/blocks.2/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -6103,13 +6073,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -6117,13 +6087,13 @@
 				<port id="2" precision="FP32" names="/blocks.2/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="419" name="Constant_619" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -6135,7 +6105,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -6145,7 +6115,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -6156,7 +6126,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -6168,7 +6138,7 @@
 			</output>
 		</layer>
 		<layer id="422" name="/blocks.2/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.2/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -6221,7 +6191,7 @@
 			</output>
 		</layer>
 		<layer id="427" name="/blocks.2/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.2/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -6253,7 +6223,7 @@
 			</output>
 		</layer>
 		<layer id="429" name="/blocks.2/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.2/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -6281,7 +6251,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -6292,13 +6262,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="432" name="onnx::MatMul_2956" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="22218973" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="24266977" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -6308,7 +6278,7 @@
 			</output>
 		</layer>
 		<layer id="433" name="Convert_2563" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -6317,7 +6287,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -6325,7 +6295,7 @@
 			</output>
 		</layer>
 		<layer id="434" name="onnx::MatMul_2956/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -6333,32 +6303,32 @@
 			</output>
 		</layer>
 		<layer id="435" name="Convert_2565" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="436" name="Subtract_2566" type="Subtract" version="opset1">
+		<layer id="436" name="onnx::MatMul_2956/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -6366,7 +6336,7 @@
 			</output>
 		</layer>
 		<layer id="437" name="onnx::MatMul_2956/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="22743261" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="24791266" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -6375,57 +6345,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="438" name="onnx::MatMul_2956/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="439" name="onnx::MatMul_2956/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="438" name="onnx::MatMul_2956/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="440" name="Constant_2571" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="439" name="Constant_2570" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="441" name="Reshape_2572" type="Reshape" version="opset1">
+		<layer id="440" name="Reshape_2571" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -6435,7 +6388,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="441" name="onnx::MatMul_2956/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -6446,7 +6414,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -6457,7 +6425,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -6470,14 +6438,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -6487,19 +6455,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -6509,20 +6477,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="446" name="onnx::Add_2961" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="22759645" size="4096" />
+			<data element_type="f32" shape="1024" offset="24807650" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2961">
 					<dim>1024</dim>
@@ -6530,7 +6498,7 @@
 			</output>
 		</layer>
 		<layer id="447" name="onnx::Add_2958" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="22763741" size="16384" />
+			<data element_type="f32" shape="4096" offset="24811746" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2958">
 					<dim>4096</dim>
@@ -6538,13 +6506,13 @@
 			</output>
 		</layer>
 		<layer id="448" name="Constant_645" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="449" name="Constant_644" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -6573,7 +6541,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -6583,13 +6551,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="453" name="blocks.2.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="22780125" size="4096" />
+			<data element_type="f32" shape="1024" offset="24828130" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.2.mlp_ln.weight">
 					<dim>1024</dim>
@@ -6601,7 +6569,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -6611,13 +6579,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="455" name="blocks.2.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="22784221" size="4096" />
+			<data element_type="f32" shape="1024" offset="24832226" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.2.mlp_ln.bias">
 					<dim>1024</dim>
@@ -6629,7 +6597,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -6639,7 +6607,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -6649,20 +6617,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="458" name="onnx::MatMul_2959" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="22788317" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="24836322" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -6672,7 +6640,7 @@
 			</output>
 		</layer>
 		<layer id="459" name="Convert_2574" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -6681,7 +6649,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -6689,7 +6657,7 @@
 			</output>
 		</layer>
 		<layer id="460" name="onnx::MatMul_2959/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="26933474" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -6697,32 +6665,32 @@
 			</output>
 		</layer>
 		<layer id="461" name="Convert_2576" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="462" name="Subtract_2577" type="Subtract" version="opset1">
+		<layer id="462" name="onnx::MatMul_2959/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -6730,7 +6698,7 @@
 			</output>
 		</layer>
 		<layer id="463" name="onnx::MatMul_2959/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="24885469" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="26933475" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -6739,57 +6707,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="464" name="onnx::MatMul_2959/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="465" name="onnx::MatMul_2959/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="464" name="onnx::MatMul_2959/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="466" name="Constant_2582" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="465" name="Constant_2581" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="467" name="Reshape_2583" type="Reshape" version="opset1">
+		<layer id="466" name="Reshape_2582" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -6799,7 +6750,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="467" name="onnx::MatMul_2959/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -6810,7 +6776,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -6821,7 +6787,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -6834,14 +6800,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -6857,7 +6823,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -6865,7 +6831,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -6874,14 +6840,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.2/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -6897,7 +6863,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -6905,7 +6871,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -6915,19 +6881,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -6943,7 +6909,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -6951,13 +6917,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="478" name="onnx::MatMul_2962" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="24951005" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="26999011" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -6967,7 +6933,7 @@
 			</output>
 		</layer>
 		<layer id="479" name="Convert_2585" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -6976,7 +6942,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -6984,7 +6950,7 @@
 			</output>
 		</layer>
 		<layer id="480" name="onnx::MatMul_2962/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="26933474" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -6992,32 +6958,32 @@
 			</output>
 		</layer>
 		<layer id="481" name="Convert_2587" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="482" name="Subtract_2588" type="Subtract" version="opset1">
+		<layer id="482" name="onnx::MatMul_2962/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7025,7 +6991,7 @@
 			</output>
 		</layer>
 		<layer id="483" name="onnx::MatMul_2962/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="27048157" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="29096163" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -7034,57 +7000,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="484" name="onnx::MatMul_2962/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="485" name="onnx::MatMul_2962/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="484" name="onnx::MatMul_2962/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="486" name="Constant_2593" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="485" name="Constant_2592" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="487" name="Reshape_2594" type="Reshape" version="opset1">
+		<layer id="486" name="Reshape_2593" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7094,7 +7043,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="487" name="onnx::MatMul_2962/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -7105,7 +7069,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -7116,7 +7080,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -7129,14 +7093,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -7146,19 +7110,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.2/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -7168,20 +7132,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="492" name="onnx::Add_2987" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="27113693" size="4096" />
+			<data element_type="f32" shape="1024" offset="29161699" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2987">
 					<dim>1024</dim>
@@ -7189,7 +7153,7 @@
 			</output>
 		</layer>
 		<layer id="493" name="onnx::Add_2964" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="27117789" size="4096" />
+			<data element_type="f32" shape="1024" offset="29165795" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2964">
 					<dim>1024</dim>
@@ -7197,13 +7161,13 @@
 			</output>
 		</layer>
 		<layer id="494" name="Constant_667" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="495" name="Constant_666" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -7232,7 +7196,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -7242,13 +7206,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="499" name="blocks.3.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="27121885" size="4096" />
+			<data element_type="f32" shape="1024" offset="29169891" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.3.attn_ln.weight">
 					<dim>1024</dim>
@@ -7260,7 +7224,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -7270,13 +7234,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="501" name="blocks.3.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="27125981" size="4096" />
+			<data element_type="f32" shape="1024" offset="29173987" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.3.attn_ln.bias">
 					<dim>1024</dim>
@@ -7288,7 +7252,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -7298,7 +7262,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -7308,20 +7272,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="504" name="onnx::MatMul_2965" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="27130077" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="29178083" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -7331,7 +7295,7 @@
 			</output>
 		</layer>
 		<layer id="505" name="Convert_2618" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -7340,7 +7304,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7348,7 +7312,7 @@
 			</output>
 		</layer>
 		<layer id="506" name="onnx::MatMul_2965/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -7356,32 +7320,32 @@
 			</output>
 		</layer>
 		<layer id="507" name="Convert_2620" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="508" name="Subtract_2621" type="Subtract" version="opset1">
+		<layer id="508" name="onnx::MatMul_2965/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7389,7 +7353,7 @@
 			</output>
 		</layer>
 		<layer id="509" name="onnx::MatMul_2965/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="27654365" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="29702371" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -7398,57 +7362,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="510" name="onnx::MatMul_2965/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="511" name="onnx::MatMul_2965/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="510" name="onnx::MatMul_2965/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="512" name="Constant_2626" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="511" name="Constant_2625" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="513" name="Reshape_2627" type="Reshape" version="opset1">
+		<layer id="512" name="Reshape_2626" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7458,7 +7405,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="513" name="onnx::MatMul_2965/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -7469,7 +7431,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -7480,7 +7442,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -7493,20 +7455,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="516" name="/blocks.3/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.3/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -7518,7 +7480,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -7528,14 +7490,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="518" name="Constant_683" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -7546,7 +7508,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -7558,19 +7520,19 @@
 				<port id="2" precision="FP32" names="/blocks.3/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="520" name="/blocks.3/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.3/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="521" name="/blocks.3/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.3/attn/Constant_1_output_0" />
 			</output>
@@ -7591,7 +7553,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -7600,13 +7562,13 @@
 				<port id="2" precision="FP32" names="/blocks.3/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="524" name="onnx::MatMul_2967" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="27670749" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="29718755" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -7616,7 +7578,7 @@
 			</output>
 		</layer>
 		<layer id="525" name="Convert_2607" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -7625,7 +7587,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7633,7 +7595,7 @@
 			</output>
 		</layer>
 		<layer id="526" name="onnx::MatMul_2967/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -7641,32 +7603,32 @@
 			</output>
 		</layer>
 		<layer id="527" name="Convert_2609" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="528" name="Subtract_2610" type="Subtract" version="opset1">
+		<layer id="528" name="onnx::MatMul_2967/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7674,7 +7636,7 @@
 			</output>
 		</layer>
 		<layer id="529" name="onnx::MatMul_2967/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="28195037" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="30243043" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -7683,57 +7645,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="530" name="onnx::MatMul_2967/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="531" name="onnx::MatMul_2967/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="530" name="onnx::MatMul_2967/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="532" name="Constant_2615" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="531" name="Constant_2614" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="533" name="Reshape_2616" type="Reshape" version="opset1">
+		<layer id="532" name="Reshape_2615" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7743,7 +7688,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="533" name="onnx::MatMul_2967/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -7754,7 +7714,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -7765,13 +7725,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="535" name="/blocks.3/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.3/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -7783,7 +7743,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -7793,14 +7753,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="537" name="Constant_688" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -7811,7 +7771,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -7824,7 +7784,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -7835,7 +7795,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -7844,7 +7804,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -7854,22 +7814,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -7879,16 +7839,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -7898,16 +7858,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -7917,21 +7877,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="544" name="onnx::Add_2969" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="28211421" size="4096" />
+			<data element_type="f32" shape="1024" offset="30259427" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2969">
 					<dim>1024</dim>
@@ -7939,7 +7899,7 @@
 			</output>
 		</layer>
 		<layer id="545" name="onnx::MatMul_2970" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="28215517" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="30263523" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -7949,7 +7909,7 @@
 			</output>
 		</layer>
 		<layer id="546" name="Convert_2596" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -7958,7 +7918,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -7966,7 +7926,7 @@
 			</output>
 		</layer>
 		<layer id="547" name="onnx::MatMul_2970/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -7974,32 +7934,32 @@
 			</output>
 		</layer>
 		<layer id="548" name="Convert_2598" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="549" name="Subtract_2599" type="Subtract" version="opset1">
+		<layer id="549" name="onnx::MatMul_2970/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -8007,7 +7967,7 @@
 			</output>
 		</layer>
 		<layer id="550" name="onnx::MatMul_2970/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="28739805" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="30787811" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -8016,57 +7976,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="551" name="onnx::MatMul_2970/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="552" name="onnx::MatMul_2970/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="551" name="onnx::MatMul_2970/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="553" name="Constant_2604" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="552" name="Constant_2603" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="554" name="Reshape_2605" type="Reshape" version="opset1">
+		<layer id="553" name="Reshape_2604" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -8076,7 +8019,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="554" name="onnx::MatMul_2970/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -8087,7 +8045,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -8098,7 +8056,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -8111,20 +8069,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="557" name="/blocks.3/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.3/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -8136,7 +8094,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -8146,14 +8104,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="559" name="Constant_693" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -8164,7 +8122,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -8176,7 +8134,7 @@
 				<port id="2" precision="FP32" names="/blocks.3/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -8187,13 +8145,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -8201,13 +8159,13 @@
 				<port id="2" precision="FP32" names="/blocks.3/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="562" name="Constant_700" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -8219,7 +8177,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -8229,7 +8187,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -8240,7 +8198,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -8252,7 +8210,7 @@
 			</output>
 		</layer>
 		<layer id="565" name="/blocks.3/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.3/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -8305,7 +8263,7 @@
 			</output>
 		</layer>
 		<layer id="570" name="/blocks.3/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.3/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -8337,7 +8295,7 @@
 			</output>
 		</layer>
 		<layer id="572" name="/blocks.3/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.3/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -8365,7 +8323,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -8376,13 +8334,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="575" name="onnx::MatMul_2988" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="28756189" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="30804195" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -8392,7 +8350,7 @@
 			</output>
 		</layer>
 		<layer id="576" name="Convert_2629" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -8401,7 +8359,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -8409,7 +8367,7 @@
 			</output>
 		</layer>
 		<layer id="577" name="onnx::MatMul_2988/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="31328483" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -8417,32 +8375,32 @@
 			</output>
 		</layer>
 		<layer id="578" name="Convert_2631" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="579" name="Subtract_2632" type="Subtract" version="opset1">
+		<layer id="579" name="onnx::MatMul_2988/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -8450,7 +8408,7 @@
 			</output>
 		</layer>
 		<layer id="580" name="onnx::MatMul_2988/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="29280477" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="31328484" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -8459,57 +8417,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="581" name="onnx::MatMul_2988/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="582" name="onnx::MatMul_2988/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="581" name="onnx::MatMul_2988/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="583" name="Constant_2637" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="582" name="Constant_2636" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="584" name="Reshape_2638" type="Reshape" version="opset1">
+		<layer id="583" name="Reshape_2637" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -8519,7 +8460,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="584" name="onnx::MatMul_2988/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -8530,7 +8486,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -8541,7 +8497,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -8554,14 +8510,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -8571,19 +8527,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -8593,20 +8549,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="589" name="onnx::Add_2993" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="29296861" size="4096" />
+			<data element_type="f32" shape="1024" offset="31344868" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2993">
 					<dim>1024</dim>
@@ -8614,7 +8570,7 @@
 			</output>
 		</layer>
 		<layer id="590" name="onnx::Add_2990" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="29300957" size="16384" />
+			<data element_type="f32" shape="4096" offset="31348964" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2990">
 					<dim>4096</dim>
@@ -8622,13 +8578,13 @@
 			</output>
 		</layer>
 		<layer id="591" name="Constant_726" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="592" name="Constant_725" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -8657,7 +8613,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -8667,13 +8623,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="596" name="blocks.3.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="29317341" size="4096" />
+			<data element_type="f32" shape="1024" offset="31365348" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.3.mlp_ln.weight">
 					<dim>1024</dim>
@@ -8685,7 +8641,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -8695,13 +8651,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="598" name="blocks.3.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="29321437" size="4096" />
+			<data element_type="f32" shape="1024" offset="31369444" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.3.mlp_ln.bias">
 					<dim>1024</dim>
@@ -8713,7 +8669,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -8723,7 +8679,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -8733,20 +8689,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="601" name="onnx::MatMul_2991" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="29325533" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="31373540" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -8756,7 +8712,7 @@
 			</output>
 		</layer>
 		<layer id="602" name="Convert_2640" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -8765,7 +8721,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -8773,7 +8729,7 @@
 			</output>
 		</layer>
 		<layer id="603" name="onnx::MatMul_2991/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="33470692" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -8781,32 +8737,32 @@
 			</output>
 		</layer>
 		<layer id="604" name="Convert_2642" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="605" name="Subtract_2643" type="Subtract" version="opset1">
+		<layer id="605" name="onnx::MatMul_2991/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -8814,7 +8770,7 @@
 			</output>
 		</layer>
 		<layer id="606" name="onnx::MatMul_2991/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="31422685" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="33470693" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -8823,57 +8779,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="607" name="onnx::MatMul_2991/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="608" name="onnx::MatMul_2991/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="607" name="onnx::MatMul_2991/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="609" name="Constant_2648" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="608" name="Constant_2647" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="610" name="Reshape_2649" type="Reshape" version="opset1">
+		<layer id="609" name="Reshape_2648" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -8883,7 +8822,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="610" name="onnx::MatMul_2991/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -8894,7 +8848,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -8905,7 +8859,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -8918,14 +8872,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -8941,7 +8895,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -8949,7 +8903,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -8958,14 +8912,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.3/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -8981,7 +8935,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -8989,7 +8943,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -8999,19 +8953,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -9027,7 +8981,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -9035,13 +8989,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="621" name="onnx::MatMul_2994" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="31488221" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="33536229" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -9051,7 +9005,7 @@
 			</output>
 		</layer>
 		<layer id="622" name="Convert_2651" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -9060,7 +9014,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9068,7 +9022,7 @@
 			</output>
 		</layer>
 		<layer id="623" name="onnx::MatMul_2994/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -9076,32 +9030,32 @@
 			</output>
 		</layer>
 		<layer id="624" name="Convert_2653" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="625" name="Subtract_2654" type="Subtract" version="opset1">
+		<layer id="625" name="onnx::MatMul_2994/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9109,7 +9063,7 @@
 			</output>
 		</layer>
 		<layer id="626" name="onnx::MatMul_2994/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="33585373" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="35633381" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -9118,57 +9072,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="627" name="onnx::MatMul_2994/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="628" name="onnx::MatMul_2994/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="627" name="onnx::MatMul_2994/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="629" name="Constant_2659" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="628" name="Constant_2658" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="630" name="Reshape_2660" type="Reshape" version="opset1">
+		<layer id="629" name="Reshape_2659" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9178,7 +9115,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="630" name="onnx::MatMul_2994/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -9189,7 +9141,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -9200,7 +9152,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -9213,14 +9165,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -9230,19 +9182,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.3/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -9252,20 +9204,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="635" name="onnx::Add_3019" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="33650909" size="4096" />
+			<data element_type="f32" shape="1024" offset="35698917" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3019">
 					<dim>1024</dim>
@@ -9273,7 +9225,7 @@
 			</output>
 		</layer>
 		<layer id="636" name="onnx::Add_2996" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="33655005" size="4096" />
+			<data element_type="f32" shape="1024" offset="35703013" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_2996">
 					<dim>1024</dim>
@@ -9281,13 +9233,13 @@
 			</output>
 		</layer>
 		<layer id="637" name="Constant_748" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="638" name="Constant_747" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -9316,7 +9268,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -9326,13 +9278,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="642" name="blocks.4.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="33659101" size="4096" />
+			<data element_type="f32" shape="1024" offset="35707109" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.4.attn_ln.weight">
 					<dim>1024</dim>
@@ -9344,7 +9296,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -9354,13 +9306,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="644" name="blocks.4.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="33663197" size="4096" />
+			<data element_type="f32" shape="1024" offset="35711205" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.4.attn_ln.bias">
 					<dim>1024</dim>
@@ -9372,7 +9324,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -9382,7 +9334,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -9392,20 +9344,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="647" name="onnx::MatMul_2997" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="33667293" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="35715301" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -9415,7 +9367,7 @@
 			</output>
 		</layer>
 		<layer id="648" name="Convert_2684" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -9424,7 +9376,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9432,7 +9384,7 @@
 			</output>
 		</layer>
 		<layer id="649" name="onnx::MatMul_2997/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -9440,32 +9392,32 @@
 			</output>
 		</layer>
 		<layer id="650" name="Convert_2686" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="651" name="Subtract_2687" type="Subtract" version="opset1">
+		<layer id="651" name="onnx::MatMul_2997/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9473,7 +9425,7 @@
 			</output>
 		</layer>
 		<layer id="652" name="onnx::MatMul_2997/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="34191581" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="36239589" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -9482,57 +9434,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="653" name="onnx::MatMul_2997/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="654" name="onnx::MatMul_2997/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="653" name="onnx::MatMul_2997/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="655" name="Constant_2692" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="654" name="Constant_2691" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="656" name="Reshape_2693" type="Reshape" version="opset1">
+		<layer id="655" name="Reshape_2692" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9542,7 +9477,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="656" name="onnx::MatMul_2997/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -9553,7 +9503,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -9564,7 +9514,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -9577,20 +9527,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="659" name="/blocks.4/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.4/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -9602,7 +9552,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -9612,14 +9562,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="661" name="Constant_764" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -9630,7 +9580,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -9642,19 +9592,19 @@
 				<port id="2" precision="FP32" names="/blocks.4/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="663" name="/blocks.4/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.4/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="664" name="/blocks.4/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.4/attn/Constant_1_output_0" />
 			</output>
@@ -9675,7 +9625,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -9684,13 +9634,13 @@
 				<port id="2" precision="FP32" names="/blocks.4/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="667" name="onnx::MatMul_2999" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="34207965" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="36255973" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -9700,7 +9650,7 @@
 			</output>
 		</layer>
 		<layer id="668" name="Convert_2673" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -9709,7 +9659,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9717,7 +9667,7 @@
 			</output>
 		</layer>
 		<layer id="669" name="onnx::MatMul_2999/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -9725,32 +9675,32 @@
 			</output>
 		</layer>
 		<layer id="670" name="Convert_2675" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="671" name="Subtract_2676" type="Subtract" version="opset1">
+		<layer id="671" name="onnx::MatMul_2999/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9758,7 +9708,7 @@
 			</output>
 		</layer>
 		<layer id="672" name="onnx::MatMul_2999/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="34732253" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="36780261" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -9767,57 +9717,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="673" name="onnx::MatMul_2999/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="674" name="onnx::MatMul_2999/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="673" name="onnx::MatMul_2999/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="675" name="Constant_2681" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="674" name="Constant_2680" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="676" name="Reshape_2682" type="Reshape" version="opset1">
+		<layer id="675" name="Reshape_2681" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -9827,7 +9760,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="676" name="onnx::MatMul_2999/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -9838,7 +9786,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -9849,13 +9797,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="678" name="/blocks.4/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.4/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -9867,7 +9815,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -9877,14 +9825,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="680" name="Constant_769" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -9895,7 +9843,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -9908,7 +9856,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -9919,7 +9867,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -9928,7 +9876,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -9938,22 +9886,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -9963,16 +9911,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -9982,16 +9930,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -10001,21 +9949,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="687" name="onnx::Add_3001" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="34748637" size="4096" />
+			<data element_type="f32" shape="1024" offset="36796645" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3001">
 					<dim>1024</dim>
@@ -10023,7 +9971,7 @@
 			</output>
 		</layer>
 		<layer id="688" name="onnx::MatMul_3002" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="34752733" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="36800741" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -10033,7 +9981,7 @@
 			</output>
 		</layer>
 		<layer id="689" name="Convert_2662" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -10042,7 +9990,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -10050,7 +9998,7 @@
 			</output>
 		</layer>
 		<layer id="690" name="onnx::MatMul_3002/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -10058,32 +10006,32 @@
 			</output>
 		</layer>
 		<layer id="691" name="Convert_2664" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="692" name="Subtract_2665" type="Subtract" version="opset1">
+		<layer id="692" name="onnx::MatMul_3002/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -10091,7 +10039,7 @@
 			</output>
 		</layer>
 		<layer id="693" name="onnx::MatMul_3002/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="35277021" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="37325029" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -10100,57 +10048,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="694" name="onnx::MatMul_3002/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="695" name="onnx::MatMul_3002/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="694" name="onnx::MatMul_3002/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="696" name="Constant_2670" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="695" name="Constant_2669" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="697" name="Reshape_2671" type="Reshape" version="opset1">
+		<layer id="696" name="Reshape_2670" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -10160,7 +10091,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="697" name="onnx::MatMul_3002/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -10171,7 +10117,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -10182,7 +10128,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -10195,20 +10141,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="700" name="/blocks.4/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.4/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -10220,7 +10166,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -10230,14 +10176,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="702" name="Constant_774" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -10248,7 +10194,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -10260,7 +10206,7 @@
 				<port id="2" precision="FP32" names="/blocks.4/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -10271,13 +10217,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -10285,13 +10231,13 @@
 				<port id="2" precision="FP32" names="/blocks.4/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="705" name="Constant_781" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -10303,7 +10249,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -10313,7 +10259,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -10324,7 +10270,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -10336,7 +10282,7 @@
 			</output>
 		</layer>
 		<layer id="708" name="/blocks.4/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.4/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -10389,7 +10335,7 @@
 			</output>
 		</layer>
 		<layer id="713" name="/blocks.4/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.4/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -10421,7 +10367,7 @@
 			</output>
 		</layer>
 		<layer id="715" name="/blocks.4/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.4/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -10449,7 +10395,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -10460,13 +10406,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="718" name="onnx::MatMul_3020" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="35293405" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="37341413" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -10476,7 +10422,7 @@
 			</output>
 		</layer>
 		<layer id="719" name="Convert_2695" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -10485,7 +10431,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -10493,7 +10439,7 @@
 			</output>
 		</layer>
 		<layer id="720" name="onnx::MatMul_3020/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -10501,32 +10447,32 @@
 			</output>
 		</layer>
 		<layer id="721" name="Convert_2697" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="722" name="Subtract_2698" type="Subtract" version="opset1">
+		<layer id="722" name="onnx::MatMul_3020/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -10534,7 +10480,7 @@
 			</output>
 		</layer>
 		<layer id="723" name="onnx::MatMul_3020/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="35817693" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="37865701" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -10543,57 +10489,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="724" name="onnx::MatMul_3020/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="725" name="onnx::MatMul_3020/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="724" name="onnx::MatMul_3020/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="726" name="Constant_2703" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="725" name="Constant_2702" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="727" name="Reshape_2704" type="Reshape" version="opset1">
+		<layer id="726" name="Reshape_2703" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -10603,7 +10532,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="727" name="onnx::MatMul_3020/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -10614,7 +10558,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -10625,7 +10569,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -10638,14 +10582,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -10655,19 +10599,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -10677,20 +10621,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="732" name="onnx::Add_3025" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="35834077" size="4096" />
+			<data element_type="f32" shape="1024" offset="37882085" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3025">
 					<dim>1024</dim>
@@ -10698,7 +10642,7 @@
 			</output>
 		</layer>
 		<layer id="733" name="onnx::Add_3022" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="35838173" size="16384" />
+			<data element_type="f32" shape="4096" offset="37886181" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3022">
 					<dim>4096</dim>
@@ -10706,13 +10650,13 @@
 			</output>
 		</layer>
 		<layer id="734" name="Constant_807" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="735" name="Constant_806" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -10741,7 +10685,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -10751,13 +10695,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="739" name="blocks.4.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="35854557" size="4096" />
+			<data element_type="f32" shape="1024" offset="37902565" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.4.mlp_ln.weight">
 					<dim>1024</dim>
@@ -10769,7 +10713,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -10779,13 +10723,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="741" name="blocks.4.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="35858653" size="4096" />
+			<data element_type="f32" shape="1024" offset="37906661" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.4.mlp_ln.bias">
 					<dim>1024</dim>
@@ -10797,7 +10741,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -10807,7 +10751,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -10817,20 +10761,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="744" name="onnx::MatMul_3023" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="35862749" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="37910757" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -10840,7 +10784,7 @@
 			</output>
 		</layer>
 		<layer id="745" name="Convert_2706" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -10849,7 +10793,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -10857,7 +10801,7 @@
 			</output>
 		</layer>
 		<layer id="746" name="onnx::MatMul_3023/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="40007909" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -10865,32 +10809,32 @@
 			</output>
 		</layer>
 		<layer id="747" name="Convert_2708" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="748" name="Subtract_2709" type="Subtract" version="opset1">
+		<layer id="748" name="onnx::MatMul_3023/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -10898,7 +10842,7 @@
 			</output>
 		</layer>
 		<layer id="749" name="onnx::MatMul_3023/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="37959901" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="40007910" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -10907,57 +10851,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="750" name="onnx::MatMul_3023/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="751" name="onnx::MatMul_3023/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="750" name="onnx::MatMul_3023/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="752" name="Constant_2714" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="751" name="Constant_2713" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="753" name="Reshape_2715" type="Reshape" version="opset1">
+		<layer id="752" name="Reshape_2714" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -10967,7 +10894,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="753" name="onnx::MatMul_3023/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -10978,7 +10920,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -10989,7 +10931,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -11002,14 +10944,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -11025,7 +10967,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -11033,7 +10975,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -11042,14 +10984,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.4/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -11065,7 +11007,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -11073,7 +11015,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -11083,19 +11025,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -11111,7 +11053,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -11119,13 +11061,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="764" name="onnx::MatMul_3026" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="38025437" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="40073446" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -11135,7 +11077,7 @@
 			</output>
 		</layer>
 		<layer id="765" name="Convert_2717" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -11144,7 +11086,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11152,7 +11094,7 @@
 			</output>
 		</layer>
 		<layer id="766" name="onnx::MatMul_3026/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -11160,32 +11102,32 @@
 			</output>
 		</layer>
 		<layer id="767" name="Convert_2719" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="768" name="Subtract_2720" type="Subtract" version="opset1">
+		<layer id="768" name="onnx::MatMul_3026/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11193,7 +11135,7 @@
 			</output>
 		</layer>
 		<layer id="769" name="onnx::MatMul_3026/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="40122589" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="42170598" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -11202,57 +11144,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="770" name="onnx::MatMul_3026/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="771" name="onnx::MatMul_3026/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="770" name="onnx::MatMul_3026/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="772" name="Constant_2725" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="771" name="Constant_2724" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="773" name="Reshape_2726" type="Reshape" version="opset1">
+		<layer id="772" name="Reshape_2725" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11262,7 +11187,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="773" name="onnx::MatMul_3026/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -11273,7 +11213,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -11284,7 +11224,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -11297,14 +11237,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -11314,19 +11254,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.4/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -11336,20 +11276,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="778" name="onnx::Add_3051" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="40188125" size="4096" />
+			<data element_type="f32" shape="1024" offset="42236134" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3051">
 					<dim>1024</dim>
@@ -11357,7 +11297,7 @@
 			</output>
 		</layer>
 		<layer id="779" name="onnx::Add_3028" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="40192221" size="4096" />
+			<data element_type="f32" shape="1024" offset="42240230" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3028">
 					<dim>1024</dim>
@@ -11365,13 +11305,13 @@
 			</output>
 		</layer>
 		<layer id="780" name="Constant_829" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="781" name="Constant_828" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -11400,7 +11340,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -11410,13 +11350,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="785" name="blocks.5.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="40196317" size="4096" />
+			<data element_type="f32" shape="1024" offset="42244326" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.5.attn_ln.weight">
 					<dim>1024</dim>
@@ -11428,7 +11368,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -11438,13 +11378,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="787" name="blocks.5.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="40200413" size="4096" />
+			<data element_type="f32" shape="1024" offset="42248422" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.5.attn_ln.bias">
 					<dim>1024</dim>
@@ -11456,7 +11396,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -11466,7 +11406,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -11476,20 +11416,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="790" name="onnx::MatMul_3029" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="40204509" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="42252518" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -11499,7 +11439,7 @@
 			</output>
 		</layer>
 		<layer id="791" name="Convert_2750" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -11508,7 +11448,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11516,7 +11456,7 @@
 			</output>
 		</layer>
 		<layer id="792" name="onnx::MatMul_3029/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -11524,32 +11464,32 @@
 			</output>
 		</layer>
 		<layer id="793" name="Convert_2752" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="794" name="Subtract_2753" type="Subtract" version="opset1">
+		<layer id="794" name="onnx::MatMul_3029/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11557,7 +11497,7 @@
 			</output>
 		</layer>
 		<layer id="795" name="onnx::MatMul_3029/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="40728797" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="42776806" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -11566,57 +11506,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="796" name="onnx::MatMul_3029/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="797" name="onnx::MatMul_3029/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="796" name="onnx::MatMul_3029/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="798" name="Constant_2758" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="797" name="Constant_2757" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="799" name="Reshape_2759" type="Reshape" version="opset1">
+		<layer id="798" name="Reshape_2758" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11626,7 +11549,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="799" name="onnx::MatMul_3029/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -11637,7 +11575,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -11648,7 +11586,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -11661,20 +11599,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="802" name="/blocks.5/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.5/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -11686,7 +11624,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -11696,14 +11634,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="804" name="Constant_845" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -11714,7 +11652,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -11726,19 +11664,19 @@
 				<port id="2" precision="FP32" names="/blocks.5/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="806" name="/blocks.5/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.5/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="807" name="/blocks.5/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.5/attn/Constant_1_output_0" />
 			</output>
@@ -11759,7 +11697,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -11768,13 +11706,13 @@
 				<port id="2" precision="FP32" names="/blocks.5/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="810" name="onnx::MatMul_3031" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="40745181" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="42793190" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -11784,7 +11722,7 @@
 			</output>
 		</layer>
 		<layer id="811" name="Convert_2739" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -11793,7 +11731,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11801,7 +11739,7 @@
 			</output>
 		</layer>
 		<layer id="812" name="onnx::MatMul_3031/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -11809,32 +11747,32 @@
 			</output>
 		</layer>
 		<layer id="813" name="Convert_2741" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="814" name="Subtract_2742" type="Subtract" version="opset1">
+		<layer id="814" name="onnx::MatMul_3031/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11842,7 +11780,7 @@
 			</output>
 		</layer>
 		<layer id="815" name="onnx::MatMul_3031/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="41269469" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="43317478" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -11851,57 +11789,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="816" name="onnx::MatMul_3031/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="817" name="onnx::MatMul_3031/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="816" name="onnx::MatMul_3031/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="818" name="Constant_2747" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="817" name="Constant_2746" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="819" name="Reshape_2748" type="Reshape" version="opset1">
+		<layer id="818" name="Reshape_2747" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -11911,7 +11832,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="819" name="onnx::MatMul_3031/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -11922,7 +11858,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -11933,13 +11869,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="821" name="/blocks.5/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.5/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -11951,7 +11887,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -11961,14 +11897,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="823" name="Constant_850" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -11979,7 +11915,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -11992,7 +11928,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -12003,7 +11939,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -12012,7 +11948,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -12022,22 +11958,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -12047,16 +11983,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -12066,16 +12002,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -12085,21 +12021,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="830" name="onnx::Add_3033" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="41285853" size="4096" />
+			<data element_type="f32" shape="1024" offset="43333862" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3033">
 					<dim>1024</dim>
@@ -12107,7 +12043,7 @@
 			</output>
 		</layer>
 		<layer id="831" name="onnx::MatMul_3034" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="41289949" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="43337958" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -12117,7 +12053,7 @@
 			</output>
 		</layer>
 		<layer id="832" name="Convert_2728" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -12126,7 +12062,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -12134,7 +12070,7 @@
 			</output>
 		</layer>
 		<layer id="833" name="onnx::MatMul_3034/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10631325" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -12142,32 +12078,32 @@
 			</output>
 		</layer>
 		<layer id="834" name="Convert_2730" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="835" name="Subtract_2731" type="Subtract" version="opset1">
+		<layer id="835" name="onnx::MatMul_3034/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -12175,7 +12111,7 @@
 			</output>
 		</layer>
 		<layer id="836" name="onnx::MatMul_3034/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="41814237" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="43862246" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -12184,57 +12120,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="837" name="onnx::MatMul_3034/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="838" name="onnx::MatMul_3034/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="837" name="onnx::MatMul_3034/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="839" name="Constant_2736" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="838" name="Constant_2735" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="840" name="Reshape_2737" type="Reshape" version="opset1">
+		<layer id="839" name="Reshape_2736" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -12244,7 +12163,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="840" name="onnx::MatMul_3034/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -12255,7 +12189,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -12266,7 +12200,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -12279,20 +12213,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="843" name="/blocks.5/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.5/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -12304,7 +12238,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -12314,14 +12248,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="845" name="Constant_855" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -12332,7 +12266,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -12344,7 +12278,7 @@
 				<port id="2" precision="FP32" names="/blocks.5/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -12355,13 +12289,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -12369,13 +12303,13 @@
 				<port id="2" precision="FP32" names="/blocks.5/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="848" name="Constant_862" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -12387,7 +12321,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -12397,7 +12331,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -12408,7 +12342,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -12420,7 +12354,7 @@
 			</output>
 		</layer>
 		<layer id="851" name="/blocks.5/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.5/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -12473,7 +12407,7 @@
 			</output>
 		</layer>
 		<layer id="856" name="/blocks.5/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.5/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -12505,7 +12439,7 @@
 			</output>
 		</layer>
 		<layer id="858" name="/blocks.5/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.5/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -12533,7 +12467,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -12544,13 +12478,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="861" name="onnx::MatMul_3052" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="41830621" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="43878630" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -12560,7 +12494,7 @@
 			</output>
 		</layer>
 		<layer id="862" name="Convert_2761" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -12569,7 +12503,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -12577,7 +12511,7 @@
 			</output>
 		</layer>
 		<layer id="863" name="onnx::MatMul_3052/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -12585,32 +12519,32 @@
 			</output>
 		</layer>
 		<layer id="864" name="Convert_2763" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="865" name="Subtract_2764" type="Subtract" version="opset1">
+		<layer id="865" name="onnx::MatMul_3052/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -12618,7 +12552,7 @@
 			</output>
 		</layer>
 		<layer id="866" name="onnx::MatMul_3052/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="42354909" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="44402918" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -12627,57 +12561,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="867" name="onnx::MatMul_3052/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="868" name="onnx::MatMul_3052/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="867" name="onnx::MatMul_3052/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="869" name="Constant_2769" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="868" name="Constant_2768" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="870" name="Reshape_2770" type="Reshape" version="opset1">
+		<layer id="869" name="Reshape_2769" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -12687,7 +12604,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="870" name="onnx::MatMul_3052/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -12698,7 +12630,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -12709,7 +12641,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -12722,14 +12654,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -12739,19 +12671,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -12761,20 +12693,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="875" name="onnx::Add_3057" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="42371293" size="4096" />
+			<data element_type="f32" shape="1024" offset="44419302" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3057">
 					<dim>1024</dim>
@@ -12782,7 +12714,7 @@
 			</output>
 		</layer>
 		<layer id="876" name="onnx::Add_3054" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="42375389" size="16384" />
+			<data element_type="f32" shape="4096" offset="44423398" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3054">
 					<dim>4096</dim>
@@ -12790,13 +12722,13 @@
 			</output>
 		</layer>
 		<layer id="877" name="Constant_888" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="878" name="Constant_887" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -12825,7 +12757,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -12835,13 +12767,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="882" name="blocks.5.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="42391773" size="4096" />
+			<data element_type="f32" shape="1024" offset="44439782" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.5.mlp_ln.weight">
 					<dim>1024</dim>
@@ -12853,7 +12785,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -12863,13 +12795,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="884" name="blocks.5.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="42395869" size="4096" />
+			<data element_type="f32" shape="1024" offset="44443878" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.5.mlp_ln.bias">
 					<dim>1024</dim>
@@ -12881,7 +12813,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -12891,7 +12823,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -12901,20 +12833,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="887" name="onnx::MatMul_3055" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="42399965" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="44447974" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -12924,7 +12856,7 @@
 			</output>
 		</layer>
 		<layer id="888" name="Convert_2772" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -12933,7 +12865,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -12941,7 +12873,7 @@
 			</output>
 		</layer>
 		<layer id="889" name="onnx::MatMul_3055/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -12949,32 +12881,32 @@
 			</output>
 		</layer>
 		<layer id="890" name="Convert_2774" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="891" name="Subtract_2775" type="Subtract" version="opset1">
+		<layer id="891" name="onnx::MatMul_3055/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -12982,7 +12914,7 @@
 			</output>
 		</layer>
 		<layer id="892" name="onnx::MatMul_3055/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="44497117" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="46545126" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -12991,57 +12923,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="893" name="onnx::MatMul_3055/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="894" name="onnx::MatMul_3055/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="893" name="onnx::MatMul_3055/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="895" name="Constant_2780" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="894" name="Constant_2779" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="896" name="Reshape_2781" type="Reshape" version="opset1">
+		<layer id="895" name="Reshape_2780" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -13051,7 +12966,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="896" name="onnx::MatMul_3055/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -13062,7 +12992,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -13073,7 +13003,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -13086,14 +13016,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -13109,7 +13039,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -13117,7 +13047,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -13126,14 +13056,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.5/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -13149,7 +13079,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -13157,7 +13087,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -13167,19 +13097,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -13195,7 +13125,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -13203,13 +13133,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="907" name="onnx::MatMul_3058" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="44562653" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="46610662" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -13219,7 +13149,7 @@
 			</output>
 		</layer>
 		<layer id="908" name="Convert_2783" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -13228,7 +13158,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13236,7 +13166,7 @@
 			</output>
 		</layer>
 		<layer id="909" name="onnx::MatMul_3058/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -13244,32 +13174,32 @@
 			</output>
 		</layer>
 		<layer id="910" name="Convert_2785" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="911" name="Subtract_2786" type="Subtract" version="opset1">
+		<layer id="911" name="onnx::MatMul_3058/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13277,7 +13207,7 @@
 			</output>
 		</layer>
 		<layer id="912" name="onnx::MatMul_3058/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="46659805" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="48707814" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -13286,57 +13216,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="913" name="onnx::MatMul_3058/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="914" name="onnx::MatMul_3058/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="913" name="onnx::MatMul_3058/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="915" name="Constant_2791" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="914" name="Constant_2790" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="916" name="Reshape_2792" type="Reshape" version="opset1">
+		<layer id="915" name="Reshape_2791" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13346,7 +13259,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="916" name="onnx::MatMul_3058/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -13357,7 +13285,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -13368,7 +13296,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -13381,14 +13309,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -13398,19 +13326,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.5/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -13420,20 +13348,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="921" name="onnx::Add_3083" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="46725341" size="4096" />
+			<data element_type="f32" shape="1024" offset="48773350" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3083">
 					<dim>1024</dim>
@@ -13441,7 +13369,7 @@
 			</output>
 		</layer>
 		<layer id="922" name="onnx::Add_3060" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="46729437" size="4096" />
+			<data element_type="f32" shape="1024" offset="48777446" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3060">
 					<dim>1024</dim>
@@ -13449,13 +13377,13 @@
 			</output>
 		</layer>
 		<layer id="923" name="Constant_910" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="924" name="Constant_909" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -13484,7 +13412,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -13494,13 +13422,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="928" name="blocks.6.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="46733533" size="4096" />
+			<data element_type="f32" shape="1024" offset="48781542" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.6.attn_ln.weight">
 					<dim>1024</dim>
@@ -13512,7 +13440,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -13522,13 +13450,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="930" name="blocks.6.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="46737629" size="4096" />
+			<data element_type="f32" shape="1024" offset="48785638" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.6.attn_ln.bias">
 					<dim>1024</dim>
@@ -13540,7 +13468,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -13550,7 +13478,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -13560,20 +13488,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="933" name="onnx::MatMul_3061" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="46741725" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="48789734" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -13583,7 +13511,7 @@
 			</output>
 		</layer>
 		<layer id="934" name="Convert_2816" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -13592,7 +13520,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13600,7 +13528,7 @@
 			</output>
 		</layer>
 		<layer id="935" name="onnx::MatMul_3061/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="17168607" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -13608,32 +13536,32 @@
 			</output>
 		</layer>
 		<layer id="936" name="Convert_2818" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="937" name="Subtract_2819" type="Subtract" version="opset1">
+		<layer id="937" name="onnx::MatMul_3061/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13641,7 +13569,7 @@
 			</output>
 		</layer>
 		<layer id="938" name="onnx::MatMul_3061/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="47266013" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="49314022" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -13650,57 +13578,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="939" name="onnx::MatMul_3061/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="940" name="onnx::MatMul_3061/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="939" name="onnx::MatMul_3061/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="941" name="Constant_2824" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="940" name="Constant_2823" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="942" name="Reshape_2825" type="Reshape" version="opset1">
+		<layer id="941" name="Reshape_2824" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13710,7 +13621,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="942" name="onnx::MatMul_3061/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -13721,7 +13647,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -13732,7 +13658,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -13745,20 +13671,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="945" name="/blocks.6/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.6/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -13770,7 +13696,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -13780,14 +13706,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="947" name="Constant_926" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -13798,7 +13724,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -13810,19 +13736,19 @@
 				<port id="2" precision="FP32" names="/blocks.6/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="949" name="/blocks.6/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.6/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="950" name="/blocks.6/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.6/attn/Constant_1_output_0" />
 			</output>
@@ -13843,7 +13769,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -13852,13 +13778,13 @@
 				<port id="2" precision="FP32" names="/blocks.6/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="953" name="onnx::MatMul_3063" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="47282397" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="49330406" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -13868,7 +13794,7 @@
 			</output>
 		</layer>
 		<layer id="954" name="Convert_2805" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -13877,7 +13803,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13885,7 +13811,7 @@
 			</output>
 		</layer>
 		<layer id="955" name="onnx::MatMul_3063/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -13893,32 +13819,32 @@
 			</output>
 		</layer>
 		<layer id="956" name="Convert_2807" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="957" name="Subtract_2808" type="Subtract" version="opset1">
+		<layer id="957" name="onnx::MatMul_3063/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13926,7 +13852,7 @@
 			</output>
 		</layer>
 		<layer id="958" name="onnx::MatMul_3063/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="47806685" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="49854694" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -13935,57 +13861,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="959" name="onnx::MatMul_3063/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="960" name="onnx::MatMul_3063/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="959" name="onnx::MatMul_3063/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="961" name="Constant_2813" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="960" name="Constant_2812" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="962" name="Reshape_2814" type="Reshape" version="opset1">
+		<layer id="961" name="Reshape_2813" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -13995,7 +13904,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="962" name="onnx::MatMul_3063/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -14006,7 +13930,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -14017,13 +13941,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="964" name="/blocks.6/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.6/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -14035,7 +13959,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -14045,14 +13969,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="966" name="Constant_931" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -14063,7 +13987,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -14076,7 +14000,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -14087,7 +14011,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -14096,7 +14020,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -14106,22 +14030,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -14131,16 +14055,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -14150,16 +14074,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -14169,21 +14093,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="973" name="onnx::Add_3065" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="47823069" size="4096" />
+			<data element_type="f32" shape="1024" offset="49871078" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3065">
 					<dim>1024</dim>
@@ -14191,7 +14115,7 @@
 			</output>
 		</layer>
 		<layer id="974" name="onnx::MatMul_3066" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="47827165" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="49875174" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -14201,7 +14125,7 @@
 			</output>
 		</layer>
 		<layer id="975" name="Convert_2794" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -14210,7 +14134,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -14218,7 +14142,7 @@
 			</output>
 		</layer>
 		<layer id="976" name="onnx::MatMul_3066/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="50399462" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -14226,32 +14150,32 @@
 			</output>
 		</layer>
 		<layer id="977" name="Convert_2796" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="978" name="Subtract_2797" type="Subtract" version="opset1">
+		<layer id="978" name="onnx::MatMul_3066/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -14259,7 +14183,7 @@
 			</output>
 		</layer>
 		<layer id="979" name="onnx::MatMul_3066/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="48351453" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="50399463" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -14268,57 +14192,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="980" name="onnx::MatMul_3066/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="981" name="onnx::MatMul_3066/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="980" name="onnx::MatMul_3066/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="982" name="Constant_2802" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="981" name="Constant_2801" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="983" name="Reshape_2803" type="Reshape" version="opset1">
+		<layer id="982" name="Reshape_2802" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -14328,7 +14235,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="983" name="onnx::MatMul_3066/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -14339,7 +14261,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -14350,7 +14272,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -14363,20 +14285,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="986" name="/blocks.6/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.6/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -14388,7 +14310,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -14398,14 +14320,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="988" name="Constant_936" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -14416,7 +14338,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -14428,7 +14350,7 @@
 				<port id="2" precision="FP32" names="/blocks.6/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -14439,13 +14361,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -14453,13 +14375,13 @@
 				<port id="2" precision="FP32" names="/blocks.6/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="991" name="Constant_943" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -14471,7 +14393,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -14481,7 +14403,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -14492,7 +14414,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -14504,7 +14426,7 @@
 			</output>
 		</layer>
 		<layer id="994" name="/blocks.6/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.6/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -14557,7 +14479,7 @@
 			</output>
 		</layer>
 		<layer id="999" name="/blocks.6/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.6/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -14589,7 +14511,7 @@
 			</output>
 		</layer>
 		<layer id="1001" name="/blocks.6/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.6/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -14617,7 +14539,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -14628,13 +14550,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1004" name="onnx::MatMul_3084" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="48367837" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="50415847" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -14644,7 +14566,7 @@
 			</output>
 		</layer>
 		<layer id="1005" name="Convert_2827" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -14653,7 +14575,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -14661,7 +14583,7 @@
 			</output>
 		</layer>
 		<layer id="1006" name="onnx::MatMul_3084/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -14669,32 +14591,32 @@
 			</output>
 		</layer>
 		<layer id="1007" name="Convert_2829" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1008" name="Subtract_2830" type="Subtract" version="opset1">
+		<layer id="1008" name="onnx::MatMul_3084/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -14702,7 +14624,7 @@
 			</output>
 		</layer>
 		<layer id="1009" name="onnx::MatMul_3084/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="48892125" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="50940135" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -14711,57 +14633,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1010" name="onnx::MatMul_3084/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1011" name="onnx::MatMul_3084/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1010" name="onnx::MatMul_3084/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1012" name="Constant_2835" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1011" name="Constant_2834" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1013" name="Reshape_2836" type="Reshape" version="opset1">
+		<layer id="1012" name="Reshape_2835" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -14771,7 +14676,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1013" name="onnx::MatMul_3084/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -14782,7 +14702,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -14793,7 +14713,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -14806,14 +14726,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -14823,19 +14743,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -14845,20 +14765,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1018" name="onnx::Add_3089" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="48908509" size="4096" />
+			<data element_type="f32" shape="1024" offset="50956519" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3089">
 					<dim>1024</dim>
@@ -14866,7 +14786,7 @@
 			</output>
 		</layer>
 		<layer id="1019" name="onnx::Add_3086" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="48912605" size="16384" />
+			<data element_type="f32" shape="4096" offset="50960615" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3086">
 					<dim>4096</dim>
@@ -14874,13 +14794,13 @@
 			</output>
 		</layer>
 		<layer id="1020" name="Constant_969" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1021" name="Constant_968" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -14909,7 +14829,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -14919,13 +14839,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1025" name="blocks.6.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="48928989" size="4096" />
+			<data element_type="f32" shape="1024" offset="50976999" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.6.mlp_ln.weight">
 					<dim>1024</dim>
@@ -14937,7 +14857,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -14947,13 +14867,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1027" name="blocks.6.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="48933085" size="4096" />
+			<data element_type="f32" shape="1024" offset="50981095" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.6.mlp_ln.bias">
 					<dim>1024</dim>
@@ -14965,7 +14885,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -14975,7 +14895,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -14985,20 +14905,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1030" name="onnx::MatMul_3087" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="48937181" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="50985191" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -15008,7 +14928,7 @@
 			</output>
 		</layer>
 		<layer id="1031" name="Convert_2838" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -15017,7 +14937,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -15025,7 +14945,7 @@
 			</output>
 		</layer>
 		<layer id="1032" name="onnx::MatMul_3087/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="40007909" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -15033,32 +14953,32 @@
 			</output>
 		</layer>
 		<layer id="1033" name="Convert_2840" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1034" name="Subtract_2841" type="Subtract" version="opset1">
+		<layer id="1034" name="onnx::MatMul_3087/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -15066,7 +14986,7 @@
 			</output>
 		</layer>
 		<layer id="1035" name="onnx::MatMul_3087/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="51034333" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="53082343" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -15075,57 +14995,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1036" name="onnx::MatMul_3087/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1037" name="onnx::MatMul_3087/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1036" name="onnx::MatMul_3087/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1038" name="Constant_2846" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="1037" name="Constant_2845" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1039" name="Reshape_2847" type="Reshape" version="opset1">
+		<layer id="1038" name="Reshape_2846" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -15135,7 +15038,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1039" name="onnx::MatMul_3087/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -15146,7 +15064,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -15157,7 +15075,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -15170,14 +15088,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -15193,7 +15111,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -15201,7 +15119,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -15210,14 +15128,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.6/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -15233,7 +15151,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -15241,7 +15159,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -15251,19 +15169,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -15279,7 +15197,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -15287,13 +15205,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1050" name="onnx::MatMul_3090" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="51099869" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="53147879" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -15303,7 +15221,7 @@
 			</output>
 		</layer>
 		<layer id="1051" name="Convert_2849" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -15312,7 +15230,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -15320,7 +15238,7 @@
 			</output>
 		</layer>
 		<layer id="1052" name="onnx::MatMul_3090/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="26933474" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -15328,32 +15246,32 @@
 			</output>
 		</layer>
 		<layer id="1053" name="Convert_2851" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1054" name="Subtract_2852" type="Subtract" version="opset1">
+		<layer id="1054" name="onnx::MatMul_3090/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -15361,7 +15279,7 @@
 			</output>
 		</layer>
 		<layer id="1055" name="onnx::MatMul_3090/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="53197021" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="55245031" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -15370,57 +15288,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1056" name="onnx::MatMul_3090/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1057" name="onnx::MatMul_3090/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1056" name="onnx::MatMul_3090/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1058" name="Constant_2857" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="1057" name="Constant_2856" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1059" name="Reshape_2858" type="Reshape" version="opset1">
+		<layer id="1058" name="Reshape_2857" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -15430,7 +15331,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1059" name="onnx::MatMul_3090/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -15441,7 +15357,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -15452,7 +15368,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -15465,14 +15381,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -15482,19 +15398,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.6/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -15504,20 +15420,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1064" name="onnx::Add_3115" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="53262557" size="4096" />
+			<data element_type="f32" shape="1024" offset="55310567" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3115">
 					<dim>1024</dim>
@@ -15525,7 +15441,7 @@
 			</output>
 		</layer>
 		<layer id="1065" name="onnx::Add_3092" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="53266653" size="4096" />
+			<data element_type="f32" shape="1024" offset="55314663" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3092">
 					<dim>1024</dim>
@@ -15533,13 +15449,13 @@
 			</output>
 		</layer>
 		<layer id="1066" name="Constant_991" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1067" name="Constant_990" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -15568,7 +15484,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -15578,13 +15494,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1071" name="blocks.7.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="53270749" size="4096" />
+			<data element_type="f32" shape="1024" offset="55318759" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.7.attn_ln.weight">
 					<dim>1024</dim>
@@ -15596,7 +15512,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -15606,13 +15522,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1073" name="blocks.7.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="53274845" size="4096" />
+			<data element_type="f32" shape="1024" offset="55322855" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.7.attn_ln.bias">
 					<dim>1024</dim>
@@ -15624,7 +15540,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -15634,7 +15550,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -15644,20 +15560,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1076" name="onnx::MatMul_3093" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="53278941" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="55326951" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -15667,7 +15583,7 @@
 			</output>
 		</layer>
 		<layer id="1077" name="Convert_2882" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -15676,7 +15592,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -15684,7 +15600,7 @@
 			</output>
 		</layer>
 		<layer id="1078" name="onnx::MatMul_3093/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -15692,32 +15608,32 @@
 			</output>
 		</layer>
 		<layer id="1079" name="Convert_2884" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1080" name="Subtract_2885" type="Subtract" version="opset1">
+		<layer id="1080" name="onnx::MatMul_3093/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -15725,7 +15641,7 @@
 			</output>
 		</layer>
 		<layer id="1081" name="onnx::MatMul_3093/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="53803229" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="55851239" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -15734,57 +15650,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1082" name="onnx::MatMul_3093/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1083" name="onnx::MatMul_3093/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1082" name="onnx::MatMul_3093/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1084" name="Constant_2890" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1083" name="Constant_2889" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1085" name="Reshape_2891" type="Reshape" version="opset1">
+		<layer id="1084" name="Reshape_2890" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -15794,7 +15693,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1085" name="onnx::MatMul_3093/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -15805,7 +15719,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -15816,7 +15730,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -15829,20 +15743,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1088" name="/blocks.7/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.7/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -15854,7 +15768,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -15864,14 +15778,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1090" name="Constant_1007" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -15882,7 +15796,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -15894,19 +15808,19 @@
 				<port id="2" precision="FP32" names="/blocks.7/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1092" name="/blocks.7/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.7/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="1093" name="/blocks.7/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.7/attn/Constant_1_output_0" />
 			</output>
@@ -15927,7 +15841,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -15936,13 +15850,13 @@
 				<port id="2" precision="FP32" names="/blocks.7/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1096" name="onnx::MatMul_3095" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="53819613" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="55867623" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -15952,7 +15866,7 @@
 			</output>
 		</layer>
 		<layer id="1097" name="Convert_2871" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -15961,7 +15875,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -15969,7 +15883,7 @@
 			</output>
 		</layer>
 		<layer id="1098" name="onnx::MatMul_3095/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="17168607" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -15977,32 +15891,32 @@
 			</output>
 		</layer>
 		<layer id="1099" name="Convert_2873" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1100" name="Subtract_2874" type="Subtract" version="opset1">
+		<layer id="1100" name="onnx::MatMul_3095/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16010,7 +15924,7 @@
 			</output>
 		</layer>
 		<layer id="1101" name="onnx::MatMul_3095/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="54343901" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="56391911" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -16019,57 +15933,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1102" name="onnx::MatMul_3095/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1103" name="onnx::MatMul_3095/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1102" name="onnx::MatMul_3095/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1104" name="Constant_2879" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1103" name="Constant_2878" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1105" name="Reshape_2880" type="Reshape" version="opset1">
+		<layer id="1104" name="Reshape_2879" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16079,7 +15976,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1105" name="onnx::MatMul_3095/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -16090,7 +16002,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -16101,13 +16013,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1107" name="/blocks.7/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.7/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -16119,7 +16031,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -16129,14 +16041,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1109" name="Constant_1012" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -16147,7 +16059,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -16160,7 +16072,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -16171,7 +16083,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -16180,7 +16092,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -16190,22 +16102,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -16215,16 +16127,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -16234,16 +16146,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -16253,21 +16165,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1116" name="onnx::Add_3097" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="54360285" size="4096" />
+			<data element_type="f32" shape="1024" offset="56408295" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3097">
 					<dim>1024</dim>
@@ -16275,7 +16187,7 @@
 			</output>
 		</layer>
 		<layer id="1117" name="onnx::MatMul_3098" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="54364381" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="56412391" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -16285,7 +16197,7 @@
 			</output>
 		</layer>
 		<layer id="1118" name="Convert_2860" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -16294,7 +16206,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16302,7 +16214,7 @@
 			</output>
 		</layer>
 		<layer id="1119" name="onnx::MatMul_3098/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -16310,32 +16222,32 @@
 			</output>
 		</layer>
 		<layer id="1120" name="Convert_2862" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1121" name="Subtract_2863" type="Subtract" version="opset1">
+		<layer id="1121" name="onnx::MatMul_3098/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16343,7 +16255,7 @@
 			</output>
 		</layer>
 		<layer id="1122" name="onnx::MatMul_3098/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="54888669" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="56936679" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -16352,57 +16264,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1123" name="onnx::MatMul_3098/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1124" name="onnx::MatMul_3098/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1123" name="onnx::MatMul_3098/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1125" name="Constant_2868" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1124" name="Constant_2867" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1126" name="Reshape_2869" type="Reshape" version="opset1">
+		<layer id="1125" name="Reshape_2868" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16412,7 +16307,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1126" name="onnx::MatMul_3098/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -16423,7 +16333,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -16434,7 +16344,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -16447,20 +16357,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1129" name="/blocks.7/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.7/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -16472,7 +16382,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -16482,14 +16392,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1131" name="Constant_1017" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -16500,7 +16410,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -16512,7 +16422,7 @@
 				<port id="2" precision="FP32" names="/blocks.7/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -16523,13 +16433,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -16537,13 +16447,13 @@
 				<port id="2" precision="FP32" names="/blocks.7/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1134" name="Constant_1024" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -16555,7 +16465,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -16565,7 +16475,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -16576,7 +16486,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -16588,7 +16498,7 @@
 			</output>
 		</layer>
 		<layer id="1137" name="/blocks.7/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.7/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -16641,7 +16551,7 @@
 			</output>
 		</layer>
 		<layer id="1142" name="/blocks.7/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.7/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -16673,7 +16583,7 @@
 			</output>
 		</layer>
 		<layer id="1144" name="/blocks.7/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.7/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -16701,7 +16611,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -16712,13 +16622,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1147" name="onnx::MatMul_3116" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="54905053" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="56953063" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -16728,7 +16638,7 @@
 			</output>
 		</layer>
 		<layer id="1148" name="Convert_2893" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -16737,7 +16647,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16745,7 +16655,7 @@
 			</output>
 		</layer>
 		<layer id="1149" name="onnx::MatMul_3116/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="50399462" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -16753,32 +16663,32 @@
 			</output>
 		</layer>
 		<layer id="1150" name="Convert_2895" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1151" name="Subtract_2896" type="Subtract" version="opset1">
+		<layer id="1151" name="onnx::MatMul_3116/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16786,7 +16696,7 @@
 			</output>
 		</layer>
 		<layer id="1152" name="onnx::MatMul_3116/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="55429341" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="57477351" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -16795,57 +16705,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1153" name="onnx::MatMul_3116/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1154" name="onnx::MatMul_3116/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1153" name="onnx::MatMul_3116/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1155" name="Constant_2901" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1154" name="Constant_2900" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1156" name="Reshape_2902" type="Reshape" version="opset1">
+		<layer id="1155" name="Reshape_2901" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -16855,7 +16748,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1156" name="onnx::MatMul_3116/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -16866,7 +16774,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -16877,7 +16785,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -16890,14 +16798,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -16907,19 +16815,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -16929,20 +16837,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1161" name="onnx::Add_3121" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="55445725" size="4096" />
+			<data element_type="f32" shape="1024" offset="57493735" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3121">
 					<dim>1024</dim>
@@ -16950,7 +16858,7 @@
 			</output>
 		</layer>
 		<layer id="1162" name="onnx::Add_3118" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="55449821" size="16384" />
+			<data element_type="f32" shape="4096" offset="57497831" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3118">
 					<dim>4096</dim>
@@ -16958,13 +16866,13 @@
 			</output>
 		</layer>
 		<layer id="1163" name="Constant_1050" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1164" name="Constant_1049" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -16993,7 +16901,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -17003,13 +16911,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1168" name="blocks.7.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="55466205" size="4096" />
+			<data element_type="f32" shape="1024" offset="57514215" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.7.mlp_ln.weight">
 					<dim>1024</dim>
@@ -17021,7 +16929,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -17031,13 +16939,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1170" name="blocks.7.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="55470301" size="4096" />
+			<data element_type="f32" shape="1024" offset="57518311" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.7.mlp_ln.bias">
 					<dim>1024</dim>
@@ -17049,7 +16957,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -17059,7 +16967,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -17069,20 +16977,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1173" name="onnx::MatMul_3119" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="55474397" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="57522407" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -17092,7 +17000,7 @@
 			</output>
 		</layer>
 		<layer id="1174" name="Convert_2904" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -17101,7 +17009,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -17109,7 +17017,7 @@
 			</output>
 		</layer>
 		<layer id="1175" name="onnx::MatMul_3119/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -17117,32 +17025,32 @@
 			</output>
 		</layer>
 		<layer id="1176" name="Convert_2906" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1177" name="Subtract_2907" type="Subtract" version="opset1">
+		<layer id="1177" name="onnx::MatMul_3119/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -17150,7 +17058,7 @@
 			</output>
 		</layer>
 		<layer id="1178" name="onnx::MatMul_3119/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="57571549" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="59619559" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -17159,57 +17067,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1179" name="onnx::MatMul_3119/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1180" name="onnx::MatMul_3119/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1179" name="onnx::MatMul_3119/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1181" name="Constant_2912" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="1180" name="Constant_2911" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1182" name="Reshape_2913" type="Reshape" version="opset1">
+		<layer id="1181" name="Reshape_2912" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -17219,7 +17110,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1182" name="onnx::MatMul_3119/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -17230,7 +17136,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -17241,7 +17147,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -17254,14 +17160,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -17277,7 +17183,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -17285,7 +17191,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -17294,14 +17200,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.7/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -17317,7 +17223,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -17325,7 +17231,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -17335,19 +17241,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -17363,7 +17269,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -17371,13 +17277,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1193" name="onnx::MatMul_3122" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="57637085" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="59685095" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -17387,7 +17293,7 @@
 			</output>
 		</layer>
 		<layer id="1194" name="Convert_2915" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -17396,7 +17302,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -17404,7 +17310,7 @@
 			</output>
 		</layer>
 		<layer id="1195" name="onnx::MatMul_3122/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -17412,32 +17318,32 @@
 			</output>
 		</layer>
 		<layer id="1196" name="Convert_2917" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1197" name="Subtract_2918" type="Subtract" version="opset1">
+		<layer id="1197" name="onnx::MatMul_3122/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -17445,7 +17351,7 @@
 			</output>
 		</layer>
 		<layer id="1198" name="onnx::MatMul_3122/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="59734237" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="61782247" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -17454,57 +17360,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1199" name="onnx::MatMul_3122/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1200" name="onnx::MatMul_3122/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1199" name="onnx::MatMul_3122/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1201" name="Constant_2923" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="1200" name="Constant_2922" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1202" name="Reshape_2924" type="Reshape" version="opset1">
+		<layer id="1201" name="Reshape_2923" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -17514,7 +17403,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1202" name="onnx::MatMul_3122/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -17525,7 +17429,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -17536,7 +17440,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -17549,14 +17453,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -17566,19 +17470,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.7/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -17588,20 +17492,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1207" name="onnx::Add_3147" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="59799773" size="4096" />
+			<data element_type="f32" shape="1024" offset="61847783" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3147">
 					<dim>1024</dim>
@@ -17609,7 +17513,7 @@
 			</output>
 		</layer>
 		<layer id="1208" name="onnx::Add_3124" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="59803869" size="4096" />
+			<data element_type="f32" shape="1024" offset="61851879" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3124">
 					<dim>1024</dim>
@@ -17617,13 +17521,13 @@
 			</output>
 		</layer>
 		<layer id="1209" name="Constant_1072" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1210" name="Constant_1071" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -17652,7 +17556,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -17662,13 +17566,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1214" name="blocks.8.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="59807965" size="4096" />
+			<data element_type="f32" shape="1024" offset="61855975" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.8.attn_ln.weight">
 					<dim>1024</dim>
@@ -17680,7 +17584,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -17690,13 +17594,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1216" name="blocks.8.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="59812061" size="4096" />
+			<data element_type="f32" shape="1024" offset="61860071" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.8.attn_ln.bias">
 					<dim>1024</dim>
@@ -17708,7 +17612,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -17718,7 +17622,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -17728,20 +17632,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1219" name="onnx::MatMul_3125" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="59816157" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="61864167" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -17751,7 +17655,7 @@
 			</output>
 		</layer>
 		<layer id="1220" name="Convert_2948" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -17760,7 +17664,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -17768,7 +17672,7 @@
 			</output>
 		</layer>
 		<layer id="1221" name="onnx::MatMul_3125/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -17776,32 +17680,32 @@
 			</output>
 		</layer>
 		<layer id="1222" name="Convert_2950" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1223" name="Subtract_2951" type="Subtract" version="opset1">
+		<layer id="1223" name="onnx::MatMul_3125/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -17809,7 +17713,7 @@
 			</output>
 		</layer>
 		<layer id="1224" name="onnx::MatMul_3125/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="60340445" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="62388455" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -17818,57 +17722,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1225" name="onnx::MatMul_3125/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1226" name="onnx::MatMul_3125/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1225" name="onnx::MatMul_3125/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1227" name="Constant_2956" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1226" name="Constant_2955" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1228" name="Reshape_2957" type="Reshape" version="opset1">
+		<layer id="1227" name="Reshape_2956" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -17878,7 +17765,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1228" name="onnx::MatMul_3125/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -17889,7 +17791,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -17900,7 +17802,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -17913,20 +17815,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1231" name="/blocks.8/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.8/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -17938,7 +17840,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -17948,14 +17850,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1233" name="Constant_1088" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -17966,7 +17868,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -17978,19 +17880,19 @@
 				<port id="2" precision="FP32" names="/blocks.8/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1235" name="/blocks.8/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.8/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="1236" name="/blocks.8/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.8/attn/Constant_1_output_0" />
 			</output>
@@ -18011,7 +17913,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -18020,13 +17922,13 @@
 				<port id="2" precision="FP32" names="/blocks.8/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1239" name="onnx::MatMul_3127" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="60356829" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="62404839" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -18036,7 +17938,7 @@
 			</output>
 		</layer>
 		<layer id="1240" name="Convert_2937" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -18045,7 +17947,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18053,7 +17955,7 @@
 			</output>
 		</layer>
 		<layer id="1241" name="onnx::MatMul_3127/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -18061,32 +17963,32 @@
 			</output>
 		</layer>
 		<layer id="1242" name="Convert_2939" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1243" name="Subtract_2940" type="Subtract" version="opset1">
+		<layer id="1243" name="onnx::MatMul_3127/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18094,7 +17996,7 @@
 			</output>
 		</layer>
 		<layer id="1244" name="onnx::MatMul_3127/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="60881117" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="62929127" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -18103,57 +18005,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1245" name="onnx::MatMul_3127/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1246" name="onnx::MatMul_3127/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1245" name="onnx::MatMul_3127/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1247" name="Constant_2945" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1246" name="Constant_2944" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1248" name="Reshape_2946" type="Reshape" version="opset1">
+		<layer id="1247" name="Reshape_2945" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18163,7 +18048,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1248" name="onnx::MatMul_3127/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -18174,7 +18074,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -18185,13 +18085,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1250" name="/blocks.8/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.8/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -18203,7 +18103,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -18213,14 +18113,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1252" name="Constant_1093" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -18231,7 +18131,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -18244,7 +18144,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -18255,7 +18155,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -18264,7 +18164,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -18274,22 +18174,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -18299,16 +18199,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -18318,16 +18218,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -18337,21 +18237,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1259" name="onnx::Add_3129" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="60897501" size="4096" />
+			<data element_type="f32" shape="1024" offset="62945511" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3129">
 					<dim>1024</dim>
@@ -18359,7 +18259,7 @@
 			</output>
 		</layer>
 		<layer id="1260" name="onnx::MatMul_3130" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="60901597" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="62949607" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -18369,7 +18269,7 @@
 			</output>
 		</layer>
 		<layer id="1261" name="Convert_2926" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -18378,7 +18278,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18386,7 +18286,7 @@
 			</output>
 		</layer>
 		<layer id="1262" name="onnx::MatMul_3130/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="33470692" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -18394,32 +18294,32 @@
 			</output>
 		</layer>
 		<layer id="1263" name="Convert_2928" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1264" name="Subtract_2929" type="Subtract" version="opset1">
+		<layer id="1264" name="onnx::MatMul_3130/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18427,7 +18327,7 @@
 			</output>
 		</layer>
 		<layer id="1265" name="onnx::MatMul_3130/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="61425885" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="63473895" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -18436,57 +18336,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1266" name="onnx::MatMul_3130/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1267" name="onnx::MatMul_3130/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1266" name="onnx::MatMul_3130/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1268" name="Constant_2934" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1267" name="Constant_2933" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1269" name="Reshape_2935" type="Reshape" version="opset1">
+		<layer id="1268" name="Reshape_2934" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18496,7 +18379,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1269" name="onnx::MatMul_3130/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -18507,7 +18405,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -18518,7 +18416,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -18531,20 +18429,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1272" name="/blocks.8/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.8/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -18556,7 +18454,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -18566,14 +18464,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1274" name="Constant_1098" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -18584,7 +18482,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -18596,7 +18494,7 @@
 				<port id="2" precision="FP32" names="/blocks.8/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -18607,13 +18505,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -18621,13 +18519,13 @@
 				<port id="2" precision="FP32" names="/blocks.8/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1277" name="Constant_1105" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -18639,7 +18537,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -18649,7 +18547,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -18660,7 +18558,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -18672,7 +18570,7 @@
 			</output>
 		</layer>
 		<layer id="1280" name="/blocks.8/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.8/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -18725,7 +18623,7 @@
 			</output>
 		</layer>
 		<layer id="1285" name="/blocks.8/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.8/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -18757,7 +18655,7 @@
 			</output>
 		</layer>
 		<layer id="1287" name="/blocks.8/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.8/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -18785,7 +18683,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -18796,13 +18694,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1290" name="onnx::MatMul_3148" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="61442269" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="63490279" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -18812,7 +18710,7 @@
 			</output>
 		</layer>
 		<layer id="1291" name="Convert_2959" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -18821,7 +18719,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18829,7 +18727,7 @@
 			</output>
 		</layer>
 		<layer id="1292" name="onnx::MatMul_3148/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -18837,32 +18735,32 @@
 			</output>
 		</layer>
 		<layer id="1293" name="Convert_2961" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1294" name="Subtract_2962" type="Subtract" version="opset1">
+		<layer id="1294" name="onnx::MatMul_3148/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18870,7 +18768,7 @@
 			</output>
 		</layer>
 		<layer id="1295" name="onnx::MatMul_3148/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="61966557" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="64014567" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -18879,57 +18777,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1296" name="onnx::MatMul_3148/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1297" name="onnx::MatMul_3148/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1296" name="onnx::MatMul_3148/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1298" name="Constant_2967" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1297" name="Constant_2966" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1299" name="Reshape_2968" type="Reshape" version="opset1">
+		<layer id="1298" name="Reshape_2967" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -18939,7 +18820,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1299" name="onnx::MatMul_3148/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -18950,7 +18846,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -18961,7 +18857,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -18974,14 +18870,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -18991,19 +18887,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -19013,20 +18909,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1304" name="onnx::Add_3153" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="61982941" size="4096" />
+			<data element_type="f32" shape="1024" offset="64030951" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3153">
 					<dim>1024</dim>
@@ -19034,7 +18930,7 @@
 			</output>
 		</layer>
 		<layer id="1305" name="onnx::Add_3150" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="61987037" size="16384" />
+			<data element_type="f32" shape="4096" offset="64035047" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3150">
 					<dim>4096</dim>
@@ -19042,13 +18938,13 @@
 			</output>
 		</layer>
 		<layer id="1306" name="Constant_1131" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1307" name="Constant_1130" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -19077,7 +18973,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -19087,13 +18983,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1311" name="blocks.8.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="62003421" size="4096" />
+			<data element_type="f32" shape="1024" offset="64051431" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.8.mlp_ln.weight">
 					<dim>1024</dim>
@@ -19105,7 +19001,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -19115,13 +19011,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1313" name="blocks.8.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="62007517" size="4096" />
+			<data element_type="f32" shape="1024" offset="64055527" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.8.mlp_ln.bias">
 					<dim>1024</dim>
@@ -19133,7 +19029,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -19143,7 +19039,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -19153,20 +19049,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1316" name="onnx::MatMul_3151" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="62011613" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="64059623" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -19176,7 +19072,7 @@
 			</output>
 		</layer>
 		<layer id="1317" name="Convert_2970" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -19185,7 +19081,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -19193,7 +19089,7 @@
 			</output>
 		</layer>
 		<layer id="1318" name="onnx::MatMul_3151/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="31328483" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -19201,32 +19097,32 @@
 			</output>
 		</layer>
 		<layer id="1319" name="Convert_2972" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1320" name="Subtract_2973" type="Subtract" version="opset1">
+		<layer id="1320" name="onnx::MatMul_3151/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -19234,7 +19130,7 @@
 			</output>
 		</layer>
 		<layer id="1321" name="onnx::MatMul_3151/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="64108765" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="66156775" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -19243,57 +19139,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1322" name="onnx::MatMul_3151/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1323" name="onnx::MatMul_3151/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1322" name="onnx::MatMul_3151/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1324" name="Constant_2978" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="1323" name="Constant_2977" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1325" name="Reshape_2979" type="Reshape" version="opset1">
+		<layer id="1324" name="Reshape_2978" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -19303,7 +19182,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1325" name="onnx::MatMul_3151/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -19314,7 +19208,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -19325,7 +19219,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -19338,14 +19232,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -19361,7 +19255,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -19369,7 +19263,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -19378,14 +19272,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.8/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -19401,7 +19295,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -19409,7 +19303,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -19419,19 +19313,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -19447,7 +19341,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -19455,13 +19349,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1336" name="onnx::MatMul_3154" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="64174301" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="66222311" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -19471,7 +19365,7 @@
 			</output>
 		</layer>
 		<layer id="1337" name="Convert_2981" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -19480,7 +19374,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -19488,7 +19382,7 @@
 			</output>
 		</layer>
 		<layer id="1338" name="onnx::MatMul_3154/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="40007909" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -19496,32 +19390,32 @@
 			</output>
 		</layer>
 		<layer id="1339" name="Convert_2983" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1340" name="Subtract_2984" type="Subtract" version="opset1">
+		<layer id="1340" name="onnx::MatMul_3154/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -19529,7 +19423,7 @@
 			</output>
 		</layer>
 		<layer id="1341" name="onnx::MatMul_3154/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="66271453" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="68319463" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -19538,57 +19432,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1342" name="onnx::MatMul_3154/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1343" name="onnx::MatMul_3154/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1342" name="onnx::MatMul_3154/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1344" name="Constant_2989" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="1343" name="Constant_2988" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1345" name="Reshape_2990" type="Reshape" version="opset1">
+		<layer id="1344" name="Reshape_2989" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -19598,7 +19475,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1345" name="onnx::MatMul_3154/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -19609,7 +19501,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -19620,7 +19512,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -19633,14 +19525,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -19650,19 +19542,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.8/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -19672,20 +19564,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1350" name="onnx::Add_3179" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="66336989" size="4096" />
+			<data element_type="f32" shape="1024" offset="68384999" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3179">
 					<dim>1024</dim>
@@ -19693,7 +19585,7 @@
 			</output>
 		</layer>
 		<layer id="1351" name="onnx::Add_3156" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="66341085" size="4096" />
+			<data element_type="f32" shape="1024" offset="68389095" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3156">
 					<dim>1024</dim>
@@ -19701,13 +19593,13 @@
 			</output>
 		</layer>
 		<layer id="1352" name="Constant_1153" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1353" name="Constant_1152" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -19736,7 +19628,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -19746,13 +19638,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1357" name="blocks.9.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="66345181" size="4096" />
+			<data element_type="f32" shape="1024" offset="68393191" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.9.attn_ln.weight">
 					<dim>1024</dim>
@@ -19764,7 +19656,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -19774,13 +19666,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1359" name="blocks.9.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="66349277" size="4096" />
+			<data element_type="f32" shape="1024" offset="68397287" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.9.attn_ln.bias">
 					<dim>1024</dim>
@@ -19792,7 +19684,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -19802,7 +19694,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -19812,20 +19704,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1362" name="onnx::MatMul_3157" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="66353373" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="68401383" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -19835,7 +19727,7 @@
 			</output>
 		</layer>
 		<layer id="1363" name="Convert_3014" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -19844,7 +19736,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -19852,7 +19744,7 @@
 			</output>
 		</layer>
 		<layer id="1364" name="onnx::MatMul_3157/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -19860,32 +19752,32 @@
 			</output>
 		</layer>
 		<layer id="1365" name="Convert_3016" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1366" name="Subtract_3017" type="Subtract" version="opset1">
+		<layer id="1366" name="onnx::MatMul_3157/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -19893,7 +19785,7 @@
 			</output>
 		</layer>
 		<layer id="1367" name="onnx::MatMul_3157/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="66877661" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="68925671" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -19902,57 +19794,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1368" name="onnx::MatMul_3157/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1369" name="onnx::MatMul_3157/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1368" name="onnx::MatMul_3157/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1370" name="Constant_3022" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1369" name="Constant_3021" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1371" name="Reshape_3023" type="Reshape" version="opset1">
+		<layer id="1370" name="Reshape_3022" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -19962,7 +19837,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1371" name="onnx::MatMul_3157/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -19973,7 +19863,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -19984,7 +19874,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -19997,20 +19887,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1374" name="/blocks.9/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.9/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -20022,7 +19912,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -20032,14 +19922,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1376" name="Constant_1169" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -20050,7 +19940,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -20062,19 +19952,19 @@
 				<port id="2" precision="FP32" names="/blocks.9/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1378" name="/blocks.9/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.9/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="1379" name="/blocks.9/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.9/attn/Constant_1_output_0" />
 			</output>
@@ -20095,7 +19985,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -20104,13 +19994,13 @@
 				<port id="2" precision="FP32" names="/blocks.9/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1382" name="onnx::MatMul_3159" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="66894045" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="68942055" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -20120,7 +20010,7 @@
 			</output>
 		</layer>
 		<layer id="1383" name="Convert_3003" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -20129,7 +20019,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20137,7 +20027,7 @@
 			</output>
 		</layer>
 		<layer id="1384" name="onnx::MatMul_3159/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -20145,32 +20035,32 @@
 			</output>
 		</layer>
 		<layer id="1385" name="Convert_3005" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1386" name="Subtract_3006" type="Subtract" version="opset1">
+		<layer id="1386" name="onnx::MatMul_3159/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20178,7 +20068,7 @@
 			</output>
 		</layer>
 		<layer id="1387" name="onnx::MatMul_3159/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="67418333" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="69466343" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -20187,57 +20077,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1388" name="onnx::MatMul_3159/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1389" name="onnx::MatMul_3159/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1388" name="onnx::MatMul_3159/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1390" name="Constant_3011" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1389" name="Constant_3010" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1391" name="Reshape_3012" type="Reshape" version="opset1">
+		<layer id="1390" name="Reshape_3011" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20247,7 +20120,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1391" name="onnx::MatMul_3159/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -20258,7 +20146,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -20269,13 +20157,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1393" name="/blocks.9/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.9/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -20287,7 +20175,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -20297,14 +20185,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1395" name="Constant_1174" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -20315,7 +20203,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -20328,7 +20216,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -20339,7 +20227,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -20348,7 +20236,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -20358,22 +20246,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -20383,16 +20271,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -20402,16 +20290,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -20421,21 +20309,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1402" name="onnx::Add_3161" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="67434717" size="4096" />
+			<data element_type="f32" shape="1024" offset="69482727" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3161">
 					<dim>1024</dim>
@@ -20443,7 +20331,7 @@
 			</output>
 		</layer>
 		<layer id="1403" name="onnx::MatMul_3162" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="67438813" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="69486823" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -20453,7 +20341,7 @@
 			</output>
 		</layer>
 		<layer id="1404" name="Convert_2992" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -20462,7 +20350,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20470,7 +20358,7 @@
 			</output>
 		</layer>
 		<layer id="1405" name="onnx::MatMul_3162/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -20478,32 +20366,32 @@
 			</output>
 		</layer>
 		<layer id="1406" name="Convert_2994" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1407" name="Subtract_2995" type="Subtract" version="opset1">
+		<layer id="1407" name="onnx::MatMul_3162/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20511,7 +20399,7 @@
 			</output>
 		</layer>
 		<layer id="1408" name="onnx::MatMul_3162/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="67963101" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="70011111" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -20520,57 +20408,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1409" name="onnx::MatMul_3162/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1410" name="onnx::MatMul_3162/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1409" name="onnx::MatMul_3162/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1411" name="Constant_3000" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1410" name="Constant_2999" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1412" name="Reshape_3001" type="Reshape" version="opset1">
+		<layer id="1411" name="Reshape_3000" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20580,7 +20451,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1412" name="onnx::MatMul_3162/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -20591,7 +20477,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -20602,7 +20488,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -20615,20 +20501,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1415" name="/blocks.9/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.9/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -20640,7 +20526,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -20650,14 +20536,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1417" name="Constant_1179" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -20668,7 +20554,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -20680,7 +20566,7 @@
 				<port id="2" precision="FP32" names="/blocks.9/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -20691,13 +20577,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -20705,13 +20591,13 @@
 				<port id="2" precision="FP32" names="/blocks.9/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1420" name="Constant_1186" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -20723,7 +20609,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -20733,7 +20619,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -20744,7 +20630,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -20756,7 +20642,7 @@
 			</output>
 		</layer>
 		<layer id="1423" name="/blocks.9/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.9/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -20809,7 +20695,7 @@
 			</output>
 		</layer>
 		<layer id="1428" name="/blocks.9/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.9/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -20841,7 +20727,7 @@
 			</output>
 		</layer>
 		<layer id="1430" name="/blocks.9/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.9/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -20869,7 +20755,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -20880,13 +20766,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1433" name="onnx::MatMul_3180" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="67979485" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="70027495" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -20896,7 +20782,7 @@
 			</output>
 		</layer>
 		<layer id="1434" name="Convert_3025" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -20905,7 +20791,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20913,7 +20799,7 @@
 			</output>
 		</layer>
 		<layer id="1435" name="onnx::MatMul_3180/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -20921,32 +20807,32 @@
 			</output>
 		</layer>
 		<layer id="1436" name="Convert_3027" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1437" name="Subtract_3028" type="Subtract" version="opset1">
+		<layer id="1437" name="onnx::MatMul_3180/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -20954,7 +20840,7 @@
 			</output>
 		</layer>
 		<layer id="1438" name="onnx::MatMul_3180/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="68503773" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="70551783" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -20963,57 +20849,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1439" name="onnx::MatMul_3180/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1440" name="onnx::MatMul_3180/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1439" name="onnx::MatMul_3180/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1441" name="Constant_3033" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1440" name="Constant_3032" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1442" name="Reshape_3034" type="Reshape" version="opset1">
+		<layer id="1441" name="Reshape_3033" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -21023,7 +20892,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1442" name="onnx::MatMul_3180/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -21034,7 +20918,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -21045,7 +20929,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21058,14 +20942,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21075,19 +20959,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21097,20 +20981,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1447" name="onnx::Add_3185" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="68520157" size="4096" />
+			<data element_type="f32" shape="1024" offset="70568167" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3185">
 					<dim>1024</dim>
@@ -21118,7 +21002,7 @@
 			</output>
 		</layer>
 		<layer id="1448" name="onnx::Add_3182" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="68524253" size="16384" />
+			<data element_type="f32" shape="4096" offset="70572263" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3182">
 					<dim>4096</dim>
@@ -21126,13 +21010,13 @@
 			</output>
 		</layer>
 		<layer id="1449" name="Constant_1212" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1450" name="Constant_1211" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -21161,7 +21045,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -21171,13 +21055,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1454" name="blocks.9.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="68540637" size="4096" />
+			<data element_type="f32" shape="1024" offset="70588647" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.9.mlp_ln.weight">
 					<dim>1024</dim>
@@ -21189,7 +21073,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -21199,13 +21083,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1456" name="blocks.9.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="68544733" size="4096" />
+			<data element_type="f32" shape="1024" offset="70592743" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.9.mlp_ln.bias">
 					<dim>1024</dim>
@@ -21217,7 +21101,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -21227,7 +21111,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21237,20 +21121,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1459" name="onnx::MatMul_3183" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="68548829" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="70596839" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -21260,7 +21144,7 @@
 			</output>
 		</layer>
 		<layer id="1460" name="Convert_3036" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -21269,7 +21153,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -21277,7 +21161,7 @@
 			</output>
 		</layer>
 		<layer id="1461" name="onnx::MatMul_3183/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="31328483" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -21285,32 +21169,32 @@
 			</output>
 		</layer>
 		<layer id="1462" name="Convert_3038" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1463" name="Subtract_3039" type="Subtract" version="opset1">
+		<layer id="1463" name="onnx::MatMul_3183/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -21318,7 +21202,7 @@
 			</output>
 		</layer>
 		<layer id="1464" name="onnx::MatMul_3183/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="70645981" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="72693991" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -21327,57 +21211,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1465" name="onnx::MatMul_3183/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1466" name="onnx::MatMul_3183/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1465" name="onnx::MatMul_3183/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1467" name="Constant_3044" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="1466" name="Constant_3043" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1468" name="Reshape_3045" type="Reshape" version="opset1">
+		<layer id="1467" name="Reshape_3044" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -21387,7 +21254,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1468" name="onnx::MatMul_3183/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -21398,7 +21280,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -21409,7 +21291,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -21422,14 +21304,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -21445,7 +21327,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -21453,7 +21335,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -21462,14 +21344,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.9/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -21485,7 +21367,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -21493,7 +21375,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -21503,19 +21385,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -21531,7 +21413,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -21539,13 +21421,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1479" name="onnx::MatMul_3186" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="70711517" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="72759527" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -21555,7 +21437,7 @@
 			</output>
 		</layer>
 		<layer id="1480" name="Convert_3047" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -21564,7 +21446,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -21572,7 +21454,7 @@
 			</output>
 		</layer>
 		<layer id="1481" name="onnx::MatMul_3186/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="74856679" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -21580,32 +21462,32 @@
 			</output>
 		</layer>
 		<layer id="1482" name="Convert_3049" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1483" name="Subtract_3050" type="Subtract" version="opset1">
+		<layer id="1483" name="onnx::MatMul_3186/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -21613,7 +21495,7 @@
 			</output>
 		</layer>
 		<layer id="1484" name="onnx::MatMul_3186/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="72808669" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="74856680" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -21622,57 +21504,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1485" name="onnx::MatMul_3186/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1486" name="onnx::MatMul_3186/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1485" name="onnx::MatMul_3186/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1487" name="Constant_3055" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="1486" name="Constant_3054" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1488" name="Reshape_3056" type="Reshape" version="opset1">
+		<layer id="1487" name="Reshape_3055" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -21682,7 +21547,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1488" name="onnx::MatMul_3186/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -21693,7 +21573,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -21704,7 +21584,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21717,14 +21597,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21734,19 +21614,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.9/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21756,20 +21636,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1493" name="onnx::Add_3211" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="72874205" size="4096" />
+			<data element_type="f32" shape="1024" offset="74922216" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3211">
 					<dim>1024</dim>
@@ -21777,7 +21657,7 @@
 			</output>
 		</layer>
 		<layer id="1494" name="onnx::Add_3188" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="72878301" size="4096" />
+			<data element_type="f32" shape="1024" offset="74926312" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3188">
 					<dim>1024</dim>
@@ -21785,13 +21665,13 @@
 			</output>
 		</layer>
 		<layer id="1495" name="Constant_1234" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1496" name="Constant_1233" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -21820,7 +21700,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -21830,13 +21710,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1500" name="blocks.10.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="72882397" size="4096" />
+			<data element_type="f32" shape="1024" offset="74930408" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.10.attn_ln.weight">
 					<dim>1024</dim>
@@ -21848,7 +21728,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -21858,13 +21738,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1502" name="blocks.10.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="72886493" size="4096" />
+			<data element_type="f32" shape="1024" offset="74934504" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.10.attn_ln.bias">
 					<dim>1024</dim>
@@ -21876,7 +21756,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -21886,7 +21766,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -21896,20 +21776,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1505" name="onnx::MatMul_3189" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="72890589" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="74938600" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -21919,7 +21799,7 @@
 			</output>
 		</layer>
 		<layer id="1506" name="Convert_3080" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -21928,7 +21808,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -21936,7 +21816,7 @@
 			</output>
 		</layer>
 		<layer id="1507" name="onnx::MatMul_3189/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="31328483" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -21944,32 +21824,32 @@
 			</output>
 		</layer>
 		<layer id="1508" name="Convert_3082" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1509" name="Subtract_3083" type="Subtract" version="opset1">
+		<layer id="1509" name="onnx::MatMul_3189/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -21977,7 +21857,7 @@
 			</output>
 		</layer>
 		<layer id="1510" name="onnx::MatMul_3189/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="73414877" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="75462888" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -21986,57 +21866,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1511" name="onnx::MatMul_3189/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1512" name="onnx::MatMul_3189/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1511" name="onnx::MatMul_3189/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1513" name="Constant_3088" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1512" name="Constant_3087" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1514" name="Reshape_3089" type="Reshape" version="opset1">
+		<layer id="1513" name="Reshape_3088" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22046,7 +21909,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1514" name="onnx::MatMul_3189/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -22057,7 +21935,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -22068,7 +21946,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -22081,20 +21959,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1517" name="/blocks.10/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.10/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -22106,7 +21984,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -22116,14 +21994,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1519" name="Constant_1250" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -22134,7 +22012,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -22146,19 +22024,19 @@
 				<port id="2" precision="FP32" names="/blocks.10/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1521" name="/blocks.10/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.10/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="1522" name="/blocks.10/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.10/attn/Constant_1_output_0" />
 			</output>
@@ -22179,7 +22057,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -22188,13 +22066,13 @@
 				<port id="2" precision="FP32" names="/blocks.10/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1525" name="onnx::MatMul_3191" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="73431261" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="75479272" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -22204,7 +22082,7 @@
 			</output>
 		</layer>
 		<layer id="1526" name="Convert_3069" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -22213,7 +22091,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22221,7 +22099,7 @@
 			</output>
 		</layer>
 		<layer id="1527" name="onnx::MatMul_3191/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="76003560" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -22229,32 +22107,32 @@
 			</output>
 		</layer>
 		<layer id="1528" name="Convert_3071" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1529" name="Subtract_3072" type="Subtract" version="opset1">
+		<layer id="1529" name="onnx::MatMul_3191/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22262,7 +22140,7 @@
 			</output>
 		</layer>
 		<layer id="1530" name="onnx::MatMul_3191/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="73955549" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="76003561" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -22271,57 +22149,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1531" name="onnx::MatMul_3191/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1532" name="onnx::MatMul_3191/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1531" name="onnx::MatMul_3191/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1533" name="Constant_3077" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1532" name="Constant_3076" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1534" name="Reshape_3078" type="Reshape" version="opset1">
+		<layer id="1533" name="Reshape_3077" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22331,7 +22192,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1534" name="onnx::MatMul_3191/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -22342,7 +22218,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -22353,13 +22229,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1536" name="/blocks.10/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.10/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -22371,7 +22247,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -22381,14 +22257,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1538" name="Constant_1255" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -22399,7 +22275,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -22412,7 +22288,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -22423,7 +22299,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -22432,7 +22308,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -22442,22 +22318,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -22467,16 +22343,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -22486,16 +22362,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -22505,21 +22381,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1545" name="onnx::Add_3193" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="73971933" size="4096" />
+			<data element_type="f32" shape="1024" offset="76019945" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3193">
 					<dim>1024</dim>
@@ -22527,7 +22403,7 @@
 			</output>
 		</layer>
 		<layer id="1546" name="onnx::MatMul_3194" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="73976029" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="76024041" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -22537,7 +22413,7 @@
 			</output>
 		</layer>
 		<layer id="1547" name="Convert_3058" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -22546,7 +22422,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22554,7 +22430,7 @@
 			</output>
 		</layer>
 		<layer id="1548" name="onnx::MatMul_3194/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -22562,32 +22438,32 @@
 			</output>
 		</layer>
 		<layer id="1549" name="Convert_3060" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1550" name="Subtract_3061" type="Subtract" version="opset1">
+		<layer id="1550" name="onnx::MatMul_3194/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22595,7 +22471,7 @@
 			</output>
 		</layer>
 		<layer id="1551" name="onnx::MatMul_3194/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="74500317" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="76548329" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -22604,57 +22480,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1552" name="onnx::MatMul_3194/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1553" name="onnx::MatMul_3194/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1552" name="onnx::MatMul_3194/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1554" name="Constant_3066" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1553" name="Constant_3065" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1555" name="Reshape_3067" type="Reshape" version="opset1">
+		<layer id="1554" name="Reshape_3066" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22664,7 +22523,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1555" name="onnx::MatMul_3194/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -22675,7 +22549,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -22686,7 +22560,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -22699,20 +22573,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1558" name="/blocks.10/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.10/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -22724,7 +22598,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -22734,14 +22608,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1560" name="Constant_1260" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -22752,7 +22626,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -22764,7 +22638,7 @@
 				<port id="2" precision="FP32" names="/blocks.10/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -22775,13 +22649,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -22789,13 +22663,13 @@
 				<port id="2" precision="FP32" names="/blocks.10/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1563" name="Constant_1267" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -22807,7 +22681,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -22817,7 +22691,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -22828,7 +22702,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -22840,7 +22714,7 @@
 			</output>
 		</layer>
 		<layer id="1566" name="/blocks.10/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.10/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -22893,7 +22767,7 @@
 			</output>
 		</layer>
 		<layer id="1571" name="/blocks.10/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.10/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -22925,7 +22799,7 @@
 			</output>
 		</layer>
 		<layer id="1573" name="/blocks.10/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.10/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -22953,7 +22827,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -22964,13 +22838,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1576" name="onnx::MatMul_3212" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="74516701" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="76564713" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -22980,7 +22854,7 @@
 			</output>
 		</layer>
 		<layer id="1577" name="Convert_3091" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -22989,7 +22863,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -22997,7 +22871,7 @@
 			</output>
 		</layer>
 		<layer id="1578" name="onnx::MatMul_3212/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -23005,32 +22879,32 @@
 			</output>
 		</layer>
 		<layer id="1579" name="Convert_3093" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1580" name="Subtract_3094" type="Subtract" version="opset1">
+		<layer id="1580" name="onnx::MatMul_3212/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -23038,7 +22912,7 @@
 			</output>
 		</layer>
 		<layer id="1581" name="onnx::MatMul_3212/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="75040989" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="77089001" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -23047,57 +22921,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1582" name="onnx::MatMul_3212/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1583" name="onnx::MatMul_3212/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1582" name="onnx::MatMul_3212/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1584" name="Constant_3099" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1583" name="Constant_3098" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1585" name="Reshape_3100" type="Reshape" version="opset1">
+		<layer id="1584" name="Reshape_3099" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -23107,7 +22964,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1585" name="onnx::MatMul_3212/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -23118,7 +22990,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -23129,7 +23001,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23142,14 +23014,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23159,19 +23031,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23181,20 +23053,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1590" name="onnx::Add_3217" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="75057373" size="4096" />
+			<data element_type="f32" shape="1024" offset="77105385" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3217">
 					<dim>1024</dim>
@@ -23202,7 +23074,7 @@
 			</output>
 		</layer>
 		<layer id="1591" name="onnx::Add_3214" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="75061469" size="16384" />
+			<data element_type="f32" shape="4096" offset="77109481" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3214">
 					<dim>4096</dim>
@@ -23210,13 +23082,13 @@
 			</output>
 		</layer>
 		<layer id="1592" name="Constant_1293" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1593" name="Constant_1292" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -23245,7 +23117,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -23255,13 +23127,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1597" name="blocks.10.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="75077853" size="4096" />
+			<data element_type="f32" shape="1024" offset="77125865" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.10.mlp_ln.weight">
 					<dim>1024</dim>
@@ -23273,7 +23145,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -23283,13 +23155,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1599" name="blocks.10.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="75081949" size="4096" />
+			<data element_type="f32" shape="1024" offset="77129961" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.10.mlp_ln.bias">
 					<dim>1024</dim>
@@ -23301,7 +23173,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -23311,7 +23183,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23321,20 +23193,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1602" name="onnx::MatMul_3215" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="75086045" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="77134057" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -23344,7 +23216,7 @@
 			</output>
 		</layer>
 		<layer id="1603" name="Convert_3102" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -23353,7 +23225,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -23361,7 +23233,7 @@
 			</output>
 		</layer>
 		<layer id="1604" name="onnx::MatMul_3215/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -23369,32 +23241,32 @@
 			</output>
 		</layer>
 		<layer id="1605" name="Convert_3104" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1606" name="Subtract_3105" type="Subtract" version="opset1">
+		<layer id="1606" name="onnx::MatMul_3215/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -23402,7 +23274,7 @@
 			</output>
 		</layer>
 		<layer id="1607" name="onnx::MatMul_3215/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="77183197" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="79231209" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -23411,57 +23283,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1608" name="onnx::MatMul_3215/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1609" name="onnx::MatMul_3215/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1608" name="onnx::MatMul_3215/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1610" name="Constant_3110" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="1609" name="Constant_3109" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1611" name="Reshape_3111" type="Reshape" version="opset1">
+		<layer id="1610" name="Reshape_3110" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -23471,7 +23326,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1611" name="onnx::MatMul_3215/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -23482,7 +23352,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -23493,7 +23363,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -23506,14 +23376,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -23529,7 +23399,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -23537,7 +23407,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -23546,14 +23416,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.10/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -23569,7 +23439,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -23577,7 +23447,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -23587,19 +23457,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -23615,7 +23485,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -23623,13 +23493,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1622" name="onnx::MatMul_3218" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="77248733" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="79296745" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -23639,7 +23509,7 @@
 			</output>
 		</layer>
 		<layer id="1623" name="Convert_3113" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -23648,7 +23518,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -23656,7 +23526,7 @@
 			</output>
 		</layer>
 		<layer id="1624" name="onnx::MatMul_3218/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -23664,32 +23534,32 @@
 			</output>
 		</layer>
 		<layer id="1625" name="Convert_3115" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1626" name="Subtract_3116" type="Subtract" version="opset1">
+		<layer id="1626" name="onnx::MatMul_3218/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -23697,7 +23567,7 @@
 			</output>
 		</layer>
 		<layer id="1627" name="onnx::MatMul_3218/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="79345885" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="81393897" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -23706,57 +23576,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1628" name="onnx::MatMul_3218/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1629" name="onnx::MatMul_3218/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1628" name="onnx::MatMul_3218/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1630" name="Constant_3121" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="1629" name="Constant_3120" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1631" name="Reshape_3122" type="Reshape" version="opset1">
+		<layer id="1630" name="Reshape_3121" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -23766,7 +23619,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1631" name="onnx::MatMul_3218/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -23777,7 +23645,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -23788,7 +23656,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23801,14 +23669,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23818,19 +23686,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.10/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23840,20 +23708,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1636" name="onnx::Add_3243" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="79411421" size="4096" />
+			<data element_type="f32" shape="1024" offset="81459433" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3243">
 					<dim>1024</dim>
@@ -23861,7 +23729,7 @@
 			</output>
 		</layer>
 		<layer id="1637" name="onnx::Add_3220" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="79415517" size="4096" />
+			<data element_type="f32" shape="1024" offset="81463529" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3220">
 					<dim>1024</dim>
@@ -23869,13 +23737,13 @@
 			</output>
 		</layer>
 		<layer id="1638" name="Constant_1315" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1639" name="Constant_1314" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -23904,7 +23772,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -23914,13 +23782,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1643" name="blocks.11.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="79419613" size="4096" />
+			<data element_type="f32" shape="1024" offset="81467625" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.11.attn_ln.weight">
 					<dim>1024</dim>
@@ -23932,7 +23800,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -23942,13 +23810,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1645" name="blocks.11.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="79423709" size="4096" />
+			<data element_type="f32" shape="1024" offset="81471721" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.11.attn_ln.bias">
 					<dim>1024</dim>
@@ -23960,7 +23828,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -23970,7 +23838,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -23980,20 +23848,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1648" name="onnx::MatMul_3221" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="79427805" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="81475817" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -24003,7 +23871,7 @@
 			</output>
 		</layer>
 		<layer id="1649" name="Convert_3146" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -24012,7 +23880,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24020,7 +23888,7 @@
 			</output>
 		</layer>
 		<layer id="1650" name="onnx::MatMul_3221/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="82000105" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -24028,32 +23896,32 @@
 			</output>
 		</layer>
 		<layer id="1651" name="Convert_3148" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1652" name="Subtract_3149" type="Subtract" version="opset1">
+		<layer id="1652" name="onnx::MatMul_3221/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24061,7 +23929,7 @@
 			</output>
 		</layer>
 		<layer id="1653" name="onnx::MatMul_3221/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="79952093" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="82000106" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -24070,57 +23938,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1654" name="onnx::MatMul_3221/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1655" name="onnx::MatMul_3221/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1654" name="onnx::MatMul_3221/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1656" name="Constant_3154" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1655" name="Constant_3153" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1657" name="Reshape_3155" type="Reshape" version="opset1">
+		<layer id="1656" name="Reshape_3154" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24130,7 +23981,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1657" name="onnx::MatMul_3221/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -24141,7 +24007,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -24152,7 +24018,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -24165,20 +24031,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1660" name="/blocks.11/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.11/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -24190,7 +24056,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -24200,14 +24066,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1662" name="Constant_1331" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -24218,7 +24084,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -24230,19 +24096,19 @@
 				<port id="2" precision="FP32" names="/blocks.11/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1664" name="/blocks.11/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.11/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="1665" name="/blocks.11/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.11/attn/Constant_1_output_0" />
 			</output>
@@ -24263,7 +24129,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -24272,13 +24138,13 @@
 				<port id="2" precision="FP32" names="/blocks.11/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1668" name="onnx::MatMul_3223" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="79968477" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="82016490" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -24288,7 +24154,7 @@
 			</output>
 		</layer>
 		<layer id="1669" name="Convert_3135" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -24297,7 +24163,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24305,7 +24171,7 @@
 			</output>
 		</layer>
 		<layer id="1670" name="onnx::MatMul_3223/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -24313,32 +24179,32 @@
 			</output>
 		</layer>
 		<layer id="1671" name="Convert_3137" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1672" name="Subtract_3138" type="Subtract" version="opset1">
+		<layer id="1672" name="onnx::MatMul_3223/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24346,7 +24212,7 @@
 			</output>
 		</layer>
 		<layer id="1673" name="onnx::MatMul_3223/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="80492765" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="82540778" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -24355,57 +24221,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1674" name="onnx::MatMul_3223/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1675" name="onnx::MatMul_3223/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1674" name="onnx::MatMul_3223/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1676" name="Constant_3143" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1675" name="Constant_3142" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1677" name="Reshape_3144" type="Reshape" version="opset1">
+		<layer id="1676" name="Reshape_3143" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24415,7 +24264,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1677" name="onnx::MatMul_3223/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -24426,7 +24290,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -24437,13 +24301,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1679" name="/blocks.11/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.11/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -24455,7 +24319,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -24465,14 +24329,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1681" name="Constant_1336" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -24483,7 +24347,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -24496,7 +24360,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -24507,7 +24371,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -24516,7 +24380,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -24526,22 +24390,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -24551,16 +24415,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -24570,16 +24434,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -24589,21 +24453,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1688" name="onnx::Add_3225" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="80509149" size="4096" />
+			<data element_type="f32" shape="1024" offset="82557162" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3225">
 					<dim>1024</dim>
@@ -24611,7 +24475,7 @@
 			</output>
 		</layer>
 		<layer id="1689" name="onnx::MatMul_3226" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="80513245" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="82561258" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -24621,7 +24485,7 @@
 			</output>
 		</layer>
 		<layer id="1690" name="Convert_3124" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -24630,7 +24494,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24638,7 +24502,7 @@
 			</output>
 		</layer>
 		<layer id="1691" name="onnx::MatMul_3226/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -24646,32 +24510,32 @@
 			</output>
 		</layer>
 		<layer id="1692" name="Convert_3126" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1693" name="Subtract_3127" type="Subtract" version="opset1">
+		<layer id="1693" name="onnx::MatMul_3226/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24679,7 +24543,7 @@
 			</output>
 		</layer>
 		<layer id="1694" name="onnx::MatMul_3226/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="81037533" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="83085546" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -24688,57 +24552,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1695" name="onnx::MatMul_3226/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1696" name="onnx::MatMul_3226/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1695" name="onnx::MatMul_3226/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1697" name="Constant_3132" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1696" name="Constant_3131" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1698" name="Reshape_3133" type="Reshape" version="opset1">
+		<layer id="1697" name="Reshape_3132" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -24748,7 +24595,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1698" name="onnx::MatMul_3226/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -24759,7 +24621,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -24770,7 +24632,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -24783,20 +24645,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1701" name="/blocks.11/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.11/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -24808,7 +24670,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -24818,14 +24680,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1703" name="Constant_1341" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -24836,7 +24698,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -24848,7 +24710,7 @@
 				<port id="2" precision="FP32" names="/blocks.11/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -24859,13 +24721,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -24873,13 +24735,13 @@
 				<port id="2" precision="FP32" names="/blocks.11/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1706" name="Constant_1348" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -24891,7 +24753,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -24901,7 +24763,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -24912,7 +24774,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -24924,7 +24786,7 @@
 			</output>
 		</layer>
 		<layer id="1709" name="/blocks.11/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.11/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -24977,7 +24839,7 @@
 			</output>
 		</layer>
 		<layer id="1714" name="/blocks.11/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.11/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -25009,7 +24871,7 @@
 			</output>
 		</layer>
 		<layer id="1716" name="/blocks.11/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.11/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -25037,7 +24899,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -25048,13 +24910,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1719" name="onnx::MatMul_3244" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="81053917" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="83101930" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -25064,7 +24926,7 @@
 			</output>
 		</layer>
 		<layer id="1720" name="Convert_3157" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -25073,7 +24935,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -25081,7 +24943,7 @@
 			</output>
 		</layer>
 		<layer id="1721" name="onnx::MatMul_3244/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -25089,32 +24951,32 @@
 			</output>
 		</layer>
 		<layer id="1722" name="Convert_3159" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1723" name="Subtract_3160" type="Subtract" version="opset1">
+		<layer id="1723" name="onnx::MatMul_3244/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -25122,7 +24984,7 @@
 			</output>
 		</layer>
 		<layer id="1724" name="onnx::MatMul_3244/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="81578205" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="83626218" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -25131,57 +24993,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1725" name="onnx::MatMul_3244/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1726" name="onnx::MatMul_3244/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1725" name="onnx::MatMul_3244/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1727" name="Constant_3165" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1726" name="Constant_3164" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1728" name="Reshape_3166" type="Reshape" version="opset1">
+		<layer id="1727" name="Reshape_3165" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -25191,7 +25036,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1728" name="onnx::MatMul_3244/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -25202,7 +25062,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -25213,7 +25073,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -25226,14 +25086,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -25243,19 +25103,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -25265,20 +25125,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1733" name="onnx::Add_3249" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="81594589" size="4096" />
+			<data element_type="f32" shape="1024" offset="83642602" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3249">
 					<dim>1024</dim>
@@ -25286,7 +25146,7 @@
 			</output>
 		</layer>
 		<layer id="1734" name="onnx::Add_3246" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="81598685" size="16384" />
+			<data element_type="f32" shape="4096" offset="83646698" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3246">
 					<dim>4096</dim>
@@ -25294,13 +25154,13 @@
 			</output>
 		</layer>
 		<layer id="1735" name="Constant_1374" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1736" name="Constant_1373" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -25329,7 +25189,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -25339,13 +25199,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1740" name="blocks.11.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="81615069" size="4096" />
+			<data element_type="f32" shape="1024" offset="83663082" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.11.mlp_ln.weight">
 					<dim>1024</dim>
@@ -25357,7 +25217,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -25367,13 +25227,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1742" name="blocks.11.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="81619165" size="4096" />
+			<data element_type="f32" shape="1024" offset="83667178" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.11.mlp_ln.bias">
 					<dim>1024</dim>
@@ -25385,7 +25245,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -25395,7 +25255,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -25405,20 +25265,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1745" name="onnx::MatMul_3247" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="81623261" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="83671274" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -25428,7 +25288,7 @@
 			</output>
 		</layer>
 		<layer id="1746" name="Convert_3168" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -25437,7 +25297,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -25445,7 +25305,7 @@
 			</output>
 		</layer>
 		<layer id="1747" name="onnx::MatMul_3247/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -25453,32 +25313,32 @@
 			</output>
 		</layer>
 		<layer id="1748" name="Convert_3170" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1749" name="Subtract_3171" type="Subtract" version="opset1">
+		<layer id="1749" name="onnx::MatMul_3247/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -25486,7 +25346,7 @@
 			</output>
 		</layer>
 		<layer id="1750" name="onnx::MatMul_3247/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="83720413" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="85768426" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -25495,57 +25355,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1751" name="onnx::MatMul_3247/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1752" name="onnx::MatMul_3247/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1751" name="onnx::MatMul_3247/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1753" name="Constant_3176" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="1752" name="Constant_3175" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1754" name="Reshape_3177" type="Reshape" version="opset1">
+		<layer id="1753" name="Reshape_3176" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -25555,7 +25398,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1754" name="onnx::MatMul_3247/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -25566,7 +25424,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -25577,7 +25435,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -25590,14 +25448,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -25613,7 +25471,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -25621,7 +25479,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -25630,14 +25488,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.11/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -25653,7 +25511,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -25661,7 +25519,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -25671,19 +25529,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -25699,7 +25557,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -25707,13 +25565,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1765" name="onnx::MatMul_3250" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="83785949" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="85833962" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -25723,7 +25581,7 @@
 			</output>
 		</layer>
 		<layer id="1766" name="Convert_3179" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -25732,7 +25590,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -25740,7 +25598,7 @@
 			</output>
 		</layer>
 		<layer id="1767" name="onnx::MatMul_3250/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -25748,32 +25606,32 @@
 			</output>
 		</layer>
 		<layer id="1768" name="Convert_3181" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1769" name="Subtract_3182" type="Subtract" version="opset1">
+		<layer id="1769" name="onnx::MatMul_3250/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -25781,7 +25639,7 @@
 			</output>
 		</layer>
 		<layer id="1770" name="onnx::MatMul_3250/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="85883101" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="87931114" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -25790,57 +25648,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1771" name="onnx::MatMul_3250/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1772" name="onnx::MatMul_3250/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1771" name="onnx::MatMul_3250/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1773" name="Constant_3187" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="1772" name="Constant_3186" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1774" name="Reshape_3188" type="Reshape" version="opset1">
+		<layer id="1773" name="Reshape_3187" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -25850,7 +25691,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1774" name="onnx::MatMul_3250/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -25861,7 +25717,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -25872,7 +25728,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -25885,14 +25741,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -25902,19 +25758,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.11/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -25924,20 +25780,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1779" name="onnx::Add_3275" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="85948637" size="4096" />
+			<data element_type="f32" shape="1024" offset="87996650" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3275">
 					<dim>1024</dim>
@@ -25945,7 +25801,7 @@
 			</output>
 		</layer>
 		<layer id="1780" name="onnx::Add_3252" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="85952733" size="4096" />
+			<data element_type="f32" shape="1024" offset="88000746" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3252">
 					<dim>1024</dim>
@@ -25953,13 +25809,13 @@
 			</output>
 		</layer>
 		<layer id="1781" name="Constant_1396" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1782" name="Constant_1395" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -25988,7 +25844,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -25998,13 +25854,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1786" name="blocks.12.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="85956829" size="4096" />
+			<data element_type="f32" shape="1024" offset="88004842" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.12.attn_ln.weight">
 					<dim>1024</dim>
@@ -26016,7 +25872,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -26026,13 +25882,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1788" name="blocks.12.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="85960925" size="4096" />
+			<data element_type="f32" shape="1024" offset="88008938" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.12.attn_ln.bias">
 					<dim>1024</dim>
@@ -26044,7 +25900,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -26054,7 +25910,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -26064,20 +25920,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1791" name="onnx::MatMul_3253" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="85965021" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="88013034" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -26087,7 +25943,7 @@
 			</output>
 		</layer>
 		<layer id="1792" name="Convert_3212" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -26096,7 +25952,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26104,7 +25960,7 @@
 			</output>
 		</layer>
 		<layer id="1793" name="onnx::MatMul_3253/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="50399462" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -26112,32 +25968,32 @@
 			</output>
 		</layer>
 		<layer id="1794" name="Convert_3214" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1795" name="Subtract_3215" type="Subtract" version="opset1">
+		<layer id="1795" name="onnx::MatMul_3253/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26145,7 +26001,7 @@
 			</output>
 		</layer>
 		<layer id="1796" name="onnx::MatMul_3253/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="86489309" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="88537322" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -26154,57 +26010,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1797" name="onnx::MatMul_3253/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1798" name="onnx::MatMul_3253/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1797" name="onnx::MatMul_3253/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1799" name="Constant_3220" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1798" name="Constant_3219" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1800" name="Reshape_3221" type="Reshape" version="opset1">
+		<layer id="1799" name="Reshape_3220" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26214,7 +26053,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1800" name="onnx::MatMul_3253/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -26225,7 +26079,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -26236,7 +26090,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -26249,20 +26103,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1803" name="/blocks.12/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.12/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -26274,7 +26128,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -26284,14 +26138,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1805" name="Constant_1412" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -26302,7 +26156,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -26314,19 +26168,19 @@
 				<port id="2" precision="FP32" names="/blocks.12/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1807" name="/blocks.12/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.12/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="1808" name="/blocks.12/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.12/attn/Constant_1_output_0" />
 			</output>
@@ -26347,7 +26201,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -26356,13 +26210,13 @@
 				<port id="2" precision="FP32" names="/blocks.12/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1811" name="onnx::MatMul_3255" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="86505693" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="88553706" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -26372,7 +26226,7 @@
 			</output>
 		</layer>
 		<layer id="1812" name="Convert_3201" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -26381,7 +26235,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26389,7 +26243,7 @@
 			</output>
 		</layer>
 		<layer id="1813" name="onnx::MatMul_3255/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -26397,32 +26251,32 @@
 			</output>
 		</layer>
 		<layer id="1814" name="Convert_3203" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1815" name="Subtract_3204" type="Subtract" version="opset1">
+		<layer id="1815" name="onnx::MatMul_3255/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26430,7 +26284,7 @@
 			</output>
 		</layer>
 		<layer id="1816" name="onnx::MatMul_3255/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="87029981" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="89077994" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -26439,57 +26293,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1817" name="onnx::MatMul_3255/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1818" name="onnx::MatMul_3255/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1817" name="onnx::MatMul_3255/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1819" name="Constant_3209" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1818" name="Constant_3208" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1820" name="Reshape_3210" type="Reshape" version="opset1">
+		<layer id="1819" name="Reshape_3209" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26499,7 +26336,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1820" name="onnx::MatMul_3255/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -26510,7 +26362,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -26521,13 +26373,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1822" name="/blocks.12/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.12/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -26539,7 +26391,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -26549,14 +26401,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1824" name="Constant_1417" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -26567,7 +26419,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -26580,7 +26432,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -26591,7 +26443,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -26600,7 +26452,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -26610,22 +26462,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -26635,16 +26487,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -26654,16 +26506,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -26673,21 +26525,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1831" name="onnx::Add_3257" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="87046365" size="4096" />
+			<data element_type="f32" shape="1024" offset="89094378" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3257">
 					<dim>1024</dim>
@@ -26695,7 +26547,7 @@
 			</output>
 		</layer>
 		<layer id="1832" name="onnx::MatMul_3258" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="87050461" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="89098474" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -26705,7 +26557,7 @@
 			</output>
 		</layer>
 		<layer id="1833" name="Convert_3190" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -26714,7 +26566,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26722,7 +26574,7 @@
 			</output>
 		</layer>
 		<layer id="1834" name="onnx::MatMul_3258/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -26730,32 +26582,32 @@
 			</output>
 		</layer>
 		<layer id="1835" name="Convert_3192" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1836" name="Subtract_3193" type="Subtract" version="opset1">
+		<layer id="1836" name="onnx::MatMul_3258/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26763,7 +26615,7 @@
 			</output>
 		</layer>
 		<layer id="1837" name="onnx::MatMul_3258/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="87574749" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="89622762" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -26772,57 +26624,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1838" name="onnx::MatMul_3258/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1839" name="onnx::MatMul_3258/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1838" name="onnx::MatMul_3258/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1840" name="Constant_3198" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1839" name="Constant_3197" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1841" name="Reshape_3199" type="Reshape" version="opset1">
+		<layer id="1840" name="Reshape_3198" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -26832,7 +26667,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1841" name="onnx::MatMul_3258/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -26843,7 +26693,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -26854,7 +26704,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -26867,20 +26717,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1844" name="/blocks.12/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.12/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -26892,7 +26742,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -26902,14 +26752,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1846" name="Constant_1422" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -26920,7 +26770,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -26932,7 +26782,7 @@
 				<port id="2" precision="FP32" names="/blocks.12/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -26943,13 +26793,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -26957,13 +26807,13 @@
 				<port id="2" precision="FP32" names="/blocks.12/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1849" name="Constant_1429" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -26975,7 +26825,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -26985,7 +26835,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -26996,7 +26846,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -27008,7 +26858,7 @@
 			</output>
 		</layer>
 		<layer id="1852" name="/blocks.12/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.12/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -27061,7 +26911,7 @@
 			</output>
 		</layer>
 		<layer id="1857" name="/blocks.12/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.12/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -27093,7 +26943,7 @@
 			</output>
 		</layer>
 		<layer id="1859" name="/blocks.12/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.12/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -27121,7 +26971,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -27132,13 +26982,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1862" name="onnx::MatMul_3276" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="87591133" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="89639146" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -27148,7 +26998,7 @@
 			</output>
 		</layer>
 		<layer id="1863" name="Convert_3223" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -27157,7 +27007,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -27165,7 +27015,7 @@
 			</output>
 		</layer>
 		<layer id="1864" name="onnx::MatMul_3276/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -27173,32 +27023,32 @@
 			</output>
 		</layer>
 		<layer id="1865" name="Convert_3225" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1866" name="Subtract_3226" type="Subtract" version="opset1">
+		<layer id="1866" name="onnx::MatMul_3276/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -27206,7 +27056,7 @@
 			</output>
 		</layer>
 		<layer id="1867" name="onnx::MatMul_3276/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="88115421" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="90163434" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -27215,57 +27065,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1868" name="onnx::MatMul_3276/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1869" name="onnx::MatMul_3276/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1868" name="onnx::MatMul_3276/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1870" name="Constant_3231" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1869" name="Constant_3230" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1871" name="Reshape_3232" type="Reshape" version="opset1">
+		<layer id="1870" name="Reshape_3231" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -27275,7 +27108,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1871" name="onnx::MatMul_3276/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -27286,7 +27134,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -27297,7 +27145,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -27310,14 +27158,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -27327,19 +27175,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -27349,20 +27197,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1876" name="onnx::Add_3281" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="88131805" size="4096" />
+			<data element_type="f32" shape="1024" offset="90179818" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3281">
 					<dim>1024</dim>
@@ -27370,7 +27218,7 @@
 			</output>
 		</layer>
 		<layer id="1877" name="onnx::Add_3278" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="88135901" size="16384" />
+			<data element_type="f32" shape="4096" offset="90183914" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3278">
 					<dim>4096</dim>
@@ -27378,13 +27226,13 @@
 			</output>
 		</layer>
 		<layer id="1878" name="Constant_1455" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1879" name="Constant_1454" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -27413,7 +27261,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -27423,13 +27271,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1883" name="blocks.12.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="88152285" size="4096" />
+			<data element_type="f32" shape="1024" offset="90200298" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.12.mlp_ln.weight">
 					<dim>1024</dim>
@@ -27441,7 +27289,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -27451,13 +27299,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1885" name="blocks.12.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="88156381" size="4096" />
+			<data element_type="f32" shape="1024" offset="90204394" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.12.mlp_ln.bias">
 					<dim>1024</dim>
@@ -27469,7 +27317,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -27479,7 +27327,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -27489,20 +27337,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1888" name="onnx::MatMul_3279" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="88160477" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="90208490" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -27512,7 +27360,7 @@
 			</output>
 		</layer>
 		<layer id="1889" name="Convert_3234" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -27521,7 +27369,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -27529,7 +27377,7 @@
 			</output>
 		</layer>
 		<layer id="1890" name="onnx::MatMul_3279/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -27537,32 +27385,32 @@
 			</output>
 		</layer>
 		<layer id="1891" name="Convert_3236" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1892" name="Subtract_3237" type="Subtract" version="opset1">
+		<layer id="1892" name="onnx::MatMul_3279/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -27570,7 +27418,7 @@
 			</output>
 		</layer>
 		<layer id="1893" name="onnx::MatMul_3279/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="90257629" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="92305642" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -27579,57 +27427,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1894" name="onnx::MatMul_3279/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1895" name="onnx::MatMul_3279/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1894" name="onnx::MatMul_3279/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1896" name="Constant_3242" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="1895" name="Constant_3241" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1897" name="Reshape_3243" type="Reshape" version="opset1">
+		<layer id="1896" name="Reshape_3242" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -27639,7 +27470,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1897" name="onnx::MatMul_3279/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -27650,7 +27496,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -27661,7 +27507,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -27674,14 +27520,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -27697,7 +27543,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -27705,7 +27551,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -27714,14 +27560,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.12/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -27737,7 +27583,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -27745,7 +27591,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -27755,19 +27601,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -27783,7 +27629,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -27791,13 +27637,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1908" name="onnx::MatMul_3282" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="90323165" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="92371178" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -27807,7 +27653,7 @@
 			</output>
 		</layer>
 		<layer id="1909" name="Convert_3245" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -27816,7 +27662,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -27824,7 +27670,7 @@
 			</output>
 		</layer>
 		<layer id="1910" name="onnx::MatMul_3282/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -27832,32 +27678,32 @@
 			</output>
 		</layer>
 		<layer id="1911" name="Convert_3247" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1912" name="Subtract_3248" type="Subtract" version="opset1">
+		<layer id="1912" name="onnx::MatMul_3282/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -27865,7 +27711,7 @@
 			</output>
 		</layer>
 		<layer id="1913" name="onnx::MatMul_3282/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="92420317" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="94468330" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -27874,57 +27720,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1914" name="onnx::MatMul_3282/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1915" name="onnx::MatMul_3282/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1914" name="onnx::MatMul_3282/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1916" name="Constant_3253" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="1915" name="Constant_3252" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1917" name="Reshape_3254" type="Reshape" version="opset1">
+		<layer id="1916" name="Reshape_3253" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -27934,7 +27763,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1917" name="onnx::MatMul_3282/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -27945,7 +27789,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -27956,7 +27800,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -27969,14 +27813,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -27986,19 +27830,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.12/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -28008,20 +27852,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1922" name="onnx::Add_3307" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="92485853" size="4096" />
+			<data element_type="f32" shape="1024" offset="94533866" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3307">
 					<dim>1024</dim>
@@ -28029,7 +27873,7 @@
 			</output>
 		</layer>
 		<layer id="1923" name="onnx::Add_3284" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="92489949" size="4096" />
+			<data element_type="f32" shape="1024" offset="94537962" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3284">
 					<dim>1024</dim>
@@ -28037,13 +27881,13 @@
 			</output>
 		</layer>
 		<layer id="1924" name="Constant_1477" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="1925" name="Constant_1476" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -28072,7 +27916,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -28082,13 +27926,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1929" name="blocks.13.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="92494045" size="4096" />
+			<data element_type="f32" shape="1024" offset="94542058" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.13.attn_ln.weight">
 					<dim>1024</dim>
@@ -28100,7 +27944,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -28110,13 +27954,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1931" name="blocks.13.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="92498141" size="4096" />
+			<data element_type="f32" shape="1024" offset="94546154" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.13.attn_ln.bias">
 					<dim>1024</dim>
@@ -28128,7 +27972,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -28138,7 +27982,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -28148,20 +27992,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1934" name="onnx::MatMul_3285" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="92502237" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="94550250" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -28171,7 +28015,7 @@
 			</output>
 		</layer>
 		<layer id="1935" name="Convert_3278" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -28180,7 +28024,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28188,7 +28032,7 @@
 			</output>
 		</layer>
 		<layer id="1936" name="onnx::MatMul_3285/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -28196,32 +28040,32 @@
 			</output>
 		</layer>
 		<layer id="1937" name="Convert_3280" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1938" name="Subtract_3281" type="Subtract" version="opset1">
+		<layer id="1938" name="onnx::MatMul_3285/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28229,7 +28073,7 @@
 			</output>
 		</layer>
 		<layer id="1939" name="onnx::MatMul_3285/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="93026525" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="95074538" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -28238,57 +28082,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1940" name="onnx::MatMul_3285/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1941" name="onnx::MatMul_3285/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1940" name="onnx::MatMul_3285/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1942" name="Constant_3286" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1941" name="Constant_3285" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1943" name="Reshape_3287" type="Reshape" version="opset1">
+		<layer id="1942" name="Reshape_3286" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28298,7 +28125,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1943" name="onnx::MatMul_3285/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -28309,7 +28151,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -28320,7 +28162,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -28333,20 +28175,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1946" name="/blocks.13/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.13/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -28358,7 +28200,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -28368,14 +28210,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1948" name="Constant_1493" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -28386,7 +28228,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -28398,19 +28240,19 @@
 				<port id="2" precision="FP32" names="/blocks.13/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1950" name="/blocks.13/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.13/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="1951" name="/blocks.13/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.13/attn/Constant_1_output_0" />
 			</output>
@@ -28431,7 +28273,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -28440,13 +28282,13 @@
 				<port id="2" precision="FP32" names="/blocks.13/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1954" name="onnx::MatMul_3287" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="93042909" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="95090922" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -28456,7 +28298,7 @@
 			</output>
 		</layer>
 		<layer id="1955" name="Convert_3267" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -28465,7 +28307,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28473,7 +28315,7 @@
 			</output>
 		</layer>
 		<layer id="1956" name="onnx::MatMul_3287/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10631325" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -28481,32 +28323,32 @@
 			</output>
 		</layer>
 		<layer id="1957" name="Convert_3269" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1958" name="Subtract_3270" type="Subtract" version="opset1">
+		<layer id="1958" name="onnx::MatMul_3287/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28514,7 +28356,7 @@
 			</output>
 		</layer>
 		<layer id="1959" name="onnx::MatMul_3287/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="93567197" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="95615210" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -28523,57 +28365,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1960" name="onnx::MatMul_3287/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1961" name="onnx::MatMul_3287/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1960" name="onnx::MatMul_3287/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1962" name="Constant_3275" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1961" name="Constant_3274" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1963" name="Reshape_3276" type="Reshape" version="opset1">
+		<layer id="1962" name="Reshape_3275" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28583,7 +28408,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1963" name="onnx::MatMul_3287/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -28594,7 +28434,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -28605,13 +28445,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1965" name="/blocks.13/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.13/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -28623,7 +28463,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -28633,14 +28473,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1967" name="Constant_1498" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -28651,7 +28491,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -28664,7 +28504,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -28675,7 +28515,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -28684,7 +28524,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -28694,22 +28534,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -28719,16 +28559,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -28738,16 +28578,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -28757,21 +28597,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1974" name="onnx::Add_3289" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="93583581" size="4096" />
+			<data element_type="f32" shape="1024" offset="95631594" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3289">
 					<dim>1024</dim>
@@ -28779,7 +28619,7 @@
 			</output>
 		</layer>
 		<layer id="1975" name="onnx::MatMul_3290" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="93587677" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="95635690" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -28789,7 +28629,7 @@
 			</output>
 		</layer>
 		<layer id="1976" name="Convert_3256" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -28798,7 +28638,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28806,7 +28646,7 @@
 			</output>
 		</layer>
 		<layer id="1977" name="onnx::MatMul_3290/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -28814,32 +28654,32 @@
 			</output>
 		</layer>
 		<layer id="1978" name="Convert_3258" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1979" name="Subtract_3259" type="Subtract" version="opset1">
+		<layer id="1979" name="onnx::MatMul_3290/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28847,7 +28687,7 @@
 			</output>
 		</layer>
 		<layer id="1980" name="onnx::MatMul_3290/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="94111965" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="96159978" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -28856,57 +28696,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="1981" name="onnx::MatMul_3290/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="1982" name="onnx::MatMul_3290/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="1981" name="onnx::MatMul_3290/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1983" name="Constant_3264" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="1982" name="Constant_3263" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1984" name="Reshape_3265" type="Reshape" version="opset1">
+		<layer id="1983" name="Reshape_3264" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -28916,7 +28739,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="1984" name="onnx::MatMul_3290/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -28927,7 +28765,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -28938,7 +28776,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -28951,20 +28789,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1987" name="/blocks.13/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.13/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -28976,7 +28814,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -28986,14 +28824,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1989" name="Constant_1503" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -29004,7 +28842,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -29016,7 +28854,7 @@
 				<port id="2" precision="FP32" names="/blocks.13/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -29027,13 +28865,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -29041,13 +28879,13 @@
 				<port id="2" precision="FP32" names="/blocks.13/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="1992" name="Constant_1510" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -29059,7 +28897,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -29069,7 +28907,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -29080,7 +28918,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -29092,7 +28930,7 @@
 			</output>
 		</layer>
 		<layer id="1995" name="/blocks.13/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.13/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -29145,7 +28983,7 @@
 			</output>
 		</layer>
 		<layer id="2000" name="/blocks.13/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.13/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -29177,7 +29015,7 @@
 			</output>
 		</layer>
 		<layer id="2002" name="/blocks.13/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.13/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -29205,7 +29043,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -29216,13 +29054,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2005" name="onnx::MatMul_3308" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="94128349" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="96176362" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -29232,7 +29070,7 @@
 			</output>
 		</layer>
 		<layer id="2006" name="Convert_3289" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -29241,7 +29079,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -29249,7 +29087,7 @@
 			</output>
 		</layer>
 		<layer id="2007" name="onnx::MatMul_3308/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="26933474" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -29257,32 +29095,32 @@
 			</output>
 		</layer>
 		<layer id="2008" name="Convert_3291" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2009" name="Subtract_3292" type="Subtract" version="opset1">
+		<layer id="2009" name="onnx::MatMul_3308/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -29290,7 +29128,7 @@
 			</output>
 		</layer>
 		<layer id="2010" name="onnx::MatMul_3308/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="94652637" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="96700650" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -29299,57 +29137,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2011" name="onnx::MatMul_3308/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2012" name="onnx::MatMul_3308/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2011" name="onnx::MatMul_3308/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2013" name="Constant_3297" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2012" name="Constant_3296" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2014" name="Reshape_3298" type="Reshape" version="opset1">
+		<layer id="2013" name="Reshape_3297" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -29359,7 +29180,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2014" name="onnx::MatMul_3308/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -29370,7 +29206,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -29381,7 +29217,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -29394,14 +29230,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -29411,19 +29247,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -29433,20 +29269,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2019" name="onnx::Add_3313" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="94669021" size="4096" />
+			<data element_type="f32" shape="1024" offset="96717034" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3313">
 					<dim>1024</dim>
@@ -29454,7 +29290,7 @@
 			</output>
 		</layer>
 		<layer id="2020" name="onnx::Add_3310" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="94673117" size="16384" />
+			<data element_type="f32" shape="4096" offset="96721130" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3310">
 					<dim>4096</dim>
@@ -29462,13 +29298,13 @@
 			</output>
 		</layer>
 		<layer id="2021" name="Constant_1536" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2022" name="Constant_1535" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -29497,7 +29333,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -29507,13 +29343,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2026" name="blocks.13.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="94689501" size="4096" />
+			<data element_type="f32" shape="1024" offset="96737514" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.13.mlp_ln.weight">
 					<dim>1024</dim>
@@ -29525,7 +29361,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -29535,13 +29371,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2028" name="blocks.13.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="94693597" size="4096" />
+			<data element_type="f32" shape="1024" offset="96741610" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.13.mlp_ln.bias">
 					<dim>1024</dim>
@@ -29553,7 +29389,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -29563,7 +29399,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -29573,20 +29409,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2031" name="onnx::MatMul_3311" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="94697693" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="96745706" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -29596,7 +29432,7 @@
 			</output>
 		</layer>
 		<layer id="2032" name="Convert_3300" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -29605,7 +29441,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -29613,7 +29449,7 @@
 			</output>
 		</layer>
 		<layer id="2033" name="onnx::MatMul_3311/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -29621,32 +29457,32 @@
 			</output>
 		</layer>
 		<layer id="2034" name="Convert_3302" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2035" name="Subtract_3303" type="Subtract" version="opset1">
+		<layer id="2035" name="onnx::MatMul_3311/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -29654,7 +29490,7 @@
 			</output>
 		</layer>
 		<layer id="2036" name="onnx::MatMul_3311/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="96794845" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="98842858" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -29663,57 +29499,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2037" name="onnx::MatMul_3311/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2038" name="onnx::MatMul_3311/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2037" name="onnx::MatMul_3311/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2039" name="Constant_3308" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="2038" name="Constant_3307" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2040" name="Reshape_3309" type="Reshape" version="opset1">
+		<layer id="2039" name="Reshape_3308" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -29723,7 +29542,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2040" name="onnx::MatMul_3311/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -29734,7 +29568,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -29745,7 +29579,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -29758,14 +29592,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -29781,7 +29615,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -29789,7 +29623,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -29798,14 +29632,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.13/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -29821,7 +29655,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -29829,7 +29663,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -29839,19 +29673,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -29867,7 +29701,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -29875,13 +29709,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2051" name="onnx::MatMul_3314" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="96860381" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="98908394" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -29891,7 +29725,7 @@
 			</output>
 		</layer>
 		<layer id="2052" name="Convert_3311" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -29900,7 +29734,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -29908,7 +29742,7 @@
 			</output>
 		</layer>
 		<layer id="2053" name="onnx::MatMul_3314/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -29916,32 +29750,32 @@
 			</output>
 		</layer>
 		<layer id="2054" name="Convert_3313" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2055" name="Subtract_3314" type="Subtract" version="opset1">
+		<layer id="2055" name="onnx::MatMul_3314/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -29949,7 +29783,7 @@
 			</output>
 		</layer>
 		<layer id="2056" name="onnx::MatMul_3314/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="98957533" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="101005546" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -29958,57 +29792,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2057" name="onnx::MatMul_3314/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2058" name="onnx::MatMul_3314/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2057" name="onnx::MatMul_3314/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2059" name="Constant_3319" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="2058" name="Constant_3318" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2060" name="Reshape_3320" type="Reshape" version="opset1">
+		<layer id="2059" name="Reshape_3319" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30018,7 +29835,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2060" name="onnx::MatMul_3314/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -30029,7 +29861,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -30040,7 +29872,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -30053,14 +29885,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -30070,19 +29902,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.13/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -30092,20 +29924,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2065" name="onnx::Add_3339" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="99023069" size="4096" />
+			<data element_type="f32" shape="1024" offset="101071082" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3339">
 					<dim>1024</dim>
@@ -30113,7 +29945,7 @@
 			</output>
 		</layer>
 		<layer id="2066" name="onnx::Add_3316" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="99027165" size="4096" />
+			<data element_type="f32" shape="1024" offset="101075178" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3316">
 					<dim>1024</dim>
@@ -30121,13 +29953,13 @@
 			</output>
 		</layer>
 		<layer id="2067" name="Constant_1558" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2068" name="Constant_1557" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -30156,7 +29988,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -30166,13 +29998,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2072" name="blocks.14.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="99031261" size="4096" />
+			<data element_type="f32" shape="1024" offset="101079274" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.14.attn_ln.weight">
 					<dim>1024</dim>
@@ -30184,7 +30016,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -30194,13 +30026,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2074" name="blocks.14.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="99035357" size="4096" />
+			<data element_type="f32" shape="1024" offset="101083370" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.14.attn_ln.bias">
 					<dim>1024</dim>
@@ -30212,7 +30044,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -30222,7 +30054,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -30232,20 +30064,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2077" name="onnx::MatMul_3317" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="99039453" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="101087466" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -30255,7 +30087,7 @@
 			</output>
 		</layer>
 		<layer id="2078" name="Convert_3344" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -30264,7 +30096,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30272,7 +30104,7 @@
 			</output>
 		</layer>
 		<layer id="2079" name="onnx::MatMul_3317/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="76003560" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -30280,32 +30112,32 @@
 			</output>
 		</layer>
 		<layer id="2080" name="Convert_3346" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2081" name="Subtract_3347" type="Subtract" version="opset1">
+		<layer id="2081" name="onnx::MatMul_3317/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30313,7 +30145,7 @@
 			</output>
 		</layer>
 		<layer id="2082" name="onnx::MatMul_3317/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="99563741" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="101611754" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -30322,57 +30154,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2083" name="onnx::MatMul_3317/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2084" name="onnx::MatMul_3317/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2083" name="onnx::MatMul_3317/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2085" name="Constant_3352" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2084" name="Constant_3351" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2086" name="Reshape_3353" type="Reshape" version="opset1">
+		<layer id="2085" name="Reshape_3352" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30382,7 +30197,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2086" name="onnx::MatMul_3317/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -30393,7 +30223,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -30404,7 +30234,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -30417,20 +30247,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2089" name="/blocks.14/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.14/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -30442,7 +30272,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -30452,14 +30282,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2091" name="Constant_1574" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -30470,7 +30300,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -30482,19 +30312,19 @@
 				<port id="2" precision="FP32" names="/blocks.14/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2093" name="/blocks.14/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.14/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="2094" name="/blocks.14/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.14/attn/Constant_1_output_0" />
 			</output>
@@ -30515,7 +30345,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -30524,13 +30354,13 @@
 				<port id="2" precision="FP32" names="/blocks.14/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2097" name="onnx::MatMul_3319" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="99580125" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="101628138" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -30540,7 +30370,7 @@
 			</output>
 		</layer>
 		<layer id="2098" name="Convert_3333" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -30549,7 +30379,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30557,7 +30387,7 @@
 			</output>
 		</layer>
 		<layer id="2099" name="onnx::MatMul_3319/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="40007909" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -30565,32 +30395,32 @@
 			</output>
 		</layer>
 		<layer id="2100" name="Convert_3335" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2101" name="Subtract_3336" type="Subtract" version="opset1">
+		<layer id="2101" name="onnx::MatMul_3319/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30598,7 +30428,7 @@
 			</output>
 		</layer>
 		<layer id="2102" name="onnx::MatMul_3319/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="100104413" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="102152426" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -30607,57 +30437,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2103" name="onnx::MatMul_3319/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2104" name="onnx::MatMul_3319/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2103" name="onnx::MatMul_3319/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2105" name="Constant_3341" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2104" name="Constant_3340" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2106" name="Reshape_3342" type="Reshape" version="opset1">
+		<layer id="2105" name="Reshape_3341" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30667,7 +30480,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2106" name="onnx::MatMul_3319/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -30678,7 +30506,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -30689,13 +30517,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2108" name="/blocks.14/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.14/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -30707,7 +30535,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -30717,14 +30545,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2110" name="Constant_1579" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -30735,7 +30563,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -30748,7 +30576,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -30759,7 +30587,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -30768,7 +30596,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -30778,22 +30606,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -30803,16 +30631,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -30822,16 +30650,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -30841,21 +30669,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2117" name="onnx::Add_3321" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="100120797" size="4096" />
+			<data element_type="f32" shape="1024" offset="102168810" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3321">
 					<dim>1024</dim>
@@ -30863,7 +30691,7 @@
 			</output>
 		</layer>
 		<layer id="2118" name="onnx::MatMul_3322" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="100124893" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="102172906" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -30873,7 +30701,7 @@
 			</output>
 		</layer>
 		<layer id="2119" name="Convert_3322" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -30882,7 +30710,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30890,7 +30718,7 @@
 			</output>
 		</layer>
 		<layer id="2120" name="onnx::MatMul_3322/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -30898,32 +30726,32 @@
 			</output>
 		</layer>
 		<layer id="2121" name="Convert_3324" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2122" name="Subtract_3325" type="Subtract" version="opset1">
+		<layer id="2122" name="onnx::MatMul_3322/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -30931,7 +30759,7 @@
 			</output>
 		</layer>
 		<layer id="2123" name="onnx::MatMul_3322/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="100649181" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="102697194" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -30940,57 +30768,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2124" name="onnx::MatMul_3322/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2125" name="onnx::MatMul_3322/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2124" name="onnx::MatMul_3322/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2126" name="Constant_3330" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2125" name="Constant_3329" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2127" name="Reshape_3331" type="Reshape" version="opset1">
+		<layer id="2126" name="Reshape_3330" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -31000,7 +30811,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2127" name="onnx::MatMul_3322/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -31011,7 +30837,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -31022,7 +30848,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -31035,20 +30861,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2130" name="/blocks.14/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.14/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -31060,7 +30886,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -31070,14 +30896,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2132" name="Constant_1584" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -31088,7 +30914,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -31100,7 +30926,7 @@
 				<port id="2" precision="FP32" names="/blocks.14/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -31111,13 +30937,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -31125,13 +30951,13 @@
 				<port id="2" precision="FP32" names="/blocks.14/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2135" name="Constant_1591" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -31143,7 +30969,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -31153,7 +30979,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -31164,7 +30990,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -31176,7 +31002,7 @@
 			</output>
 		</layer>
 		<layer id="2138" name="/blocks.14/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.14/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -31229,7 +31055,7 @@
 			</output>
 		</layer>
 		<layer id="2143" name="/blocks.14/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.14/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -31261,7 +31087,7 @@
 			</output>
 		</layer>
 		<layer id="2145" name="/blocks.14/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.14/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -31289,7 +31115,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -31300,13 +31126,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2148" name="onnx::MatMul_3340" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="100665565" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="102713578" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -31316,7 +31142,7 @@
 			</output>
 		</layer>
 		<layer id="2149" name="Convert_3355" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -31325,7 +31151,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -31333,7 +31159,7 @@
 			</output>
 		</layer>
 		<layer id="2150" name="onnx::MatMul_3340/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -31341,32 +31167,32 @@
 			</output>
 		</layer>
 		<layer id="2151" name="Convert_3357" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2152" name="Subtract_3358" type="Subtract" version="opset1">
+		<layer id="2152" name="onnx::MatMul_3340/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -31374,7 +31200,7 @@
 			</output>
 		</layer>
 		<layer id="2153" name="onnx::MatMul_3340/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="101189853" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="103237866" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -31383,57 +31209,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2154" name="onnx::MatMul_3340/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2155" name="onnx::MatMul_3340/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2154" name="onnx::MatMul_3340/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2156" name="Constant_3363" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2155" name="Constant_3362" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2157" name="Reshape_3364" type="Reshape" version="opset1">
+		<layer id="2156" name="Reshape_3363" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -31443,7 +31252,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2157" name="onnx::MatMul_3340/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -31454,7 +31278,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -31465,7 +31289,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -31478,14 +31302,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -31495,19 +31319,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -31517,20 +31341,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2162" name="onnx::Add_3345" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="101206237" size="4096" />
+			<data element_type="f32" shape="1024" offset="103254250" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3345">
 					<dim>1024</dim>
@@ -31538,7 +31362,7 @@
 			</output>
 		</layer>
 		<layer id="2163" name="onnx::Add_3342" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="101210333" size="16384" />
+			<data element_type="f32" shape="4096" offset="103258346" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3342">
 					<dim>4096</dim>
@@ -31546,13 +31370,13 @@
 			</output>
 		</layer>
 		<layer id="2164" name="Constant_1617" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2165" name="Constant_1616" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -31581,7 +31405,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -31591,13 +31415,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2169" name="blocks.14.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="101226717" size="4096" />
+			<data element_type="f32" shape="1024" offset="103274730" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.14.mlp_ln.weight">
 					<dim>1024</dim>
@@ -31609,7 +31433,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -31619,13 +31443,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2171" name="blocks.14.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="101230813" size="4096" />
+			<data element_type="f32" shape="1024" offset="103278826" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.14.mlp_ln.bias">
 					<dim>1024</dim>
@@ -31637,7 +31461,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -31647,7 +31471,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -31657,20 +31481,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2174" name="onnx::MatMul_3343" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="101234909" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="103282922" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -31680,7 +31504,7 @@
 			</output>
 		</layer>
 		<layer id="2175" name="Convert_3366" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -31689,7 +31513,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -31697,7 +31521,7 @@
 			</output>
 		</layer>
 		<layer id="2176" name="onnx::MatMul_3343/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -31705,32 +31529,32 @@
 			</output>
 		</layer>
 		<layer id="2177" name="Convert_3368" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2178" name="Subtract_3369" type="Subtract" version="opset1">
+		<layer id="2178" name="onnx::MatMul_3343/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -31738,7 +31562,7 @@
 			</output>
 		</layer>
 		<layer id="2179" name="onnx::MatMul_3343/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="103332061" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="105380074" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -31747,57 +31571,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2180" name="onnx::MatMul_3343/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2181" name="onnx::MatMul_3343/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2180" name="onnx::MatMul_3343/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2182" name="Constant_3374" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="2181" name="Constant_3373" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2183" name="Reshape_3375" type="Reshape" version="opset1">
+		<layer id="2182" name="Reshape_3374" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -31807,7 +31614,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2183" name="onnx::MatMul_3343/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -31818,7 +31640,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -31829,7 +31651,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -31842,14 +31664,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -31865,7 +31687,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -31873,7 +31695,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -31882,14 +31704,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.14/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -31905,7 +31727,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -31913,7 +31735,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -31923,19 +31745,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -31951,7 +31773,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -31959,13 +31781,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2194" name="onnx::MatMul_3346" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="103397597" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="105445610" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -31975,7 +31797,7 @@
 			</output>
 		</layer>
 		<layer id="2195" name="Convert_3377" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -31984,7 +31806,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -31992,7 +31814,7 @@
 			</output>
 		</layer>
 		<layer id="2196" name="onnx::MatMul_3346/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="17168607" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -32000,32 +31822,32 @@
 			</output>
 		</layer>
 		<layer id="2197" name="Convert_3379" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2198" name="Subtract_3380" type="Subtract" version="opset1">
+		<layer id="2198" name="onnx::MatMul_3346/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32033,7 +31855,7 @@
 			</output>
 		</layer>
 		<layer id="2199" name="onnx::MatMul_3346/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="105494749" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="107542762" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -32042,57 +31864,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2200" name="onnx::MatMul_3346/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2201" name="onnx::MatMul_3346/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2200" name="onnx::MatMul_3346/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2202" name="Constant_3385" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="2201" name="Constant_3384" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2203" name="Reshape_3386" type="Reshape" version="opset1">
+		<layer id="2202" name="Reshape_3385" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32102,7 +31907,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2203" name="onnx::MatMul_3346/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -32113,7 +31933,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -32124,7 +31944,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -32137,14 +31957,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -32154,19 +31974,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.14/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -32176,20 +31996,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2208" name="onnx::Add_3371" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="105560285" size="4096" />
+			<data element_type="f32" shape="1024" offset="107608298" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3371">
 					<dim>1024</dim>
@@ -32197,7 +32017,7 @@
 			</output>
 		</layer>
 		<layer id="2209" name="onnx::Add_3348" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="105564381" size="4096" />
+			<data element_type="f32" shape="1024" offset="107612394" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3348">
 					<dim>1024</dim>
@@ -32205,13 +32025,13 @@
 			</output>
 		</layer>
 		<layer id="2210" name="Constant_1639" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2211" name="Constant_1638" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -32240,7 +32060,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -32250,13 +32070,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2215" name="blocks.15.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="105568477" size="4096" />
+			<data element_type="f32" shape="1024" offset="107616490" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.15.attn_ln.weight">
 					<dim>1024</dim>
@@ -32268,7 +32088,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -32278,13 +32098,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2217" name="blocks.15.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="105572573" size="4096" />
+			<data element_type="f32" shape="1024" offset="107620586" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.15.attn_ln.bias">
 					<dim>1024</dim>
@@ -32296,7 +32116,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -32306,7 +32126,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -32316,20 +32136,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2220" name="onnx::MatMul_3349" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="105576669" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="107624682" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -32339,7 +32159,7 @@
 			</output>
 		</layer>
 		<layer id="2221" name="Convert_3410" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -32348,7 +32168,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32356,7 +32176,7 @@
 			</output>
 		</layer>
 		<layer id="2222" name="onnx::MatMul_3349/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -32364,32 +32184,32 @@
 			</output>
 		</layer>
 		<layer id="2223" name="Convert_3412" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2224" name="Subtract_3413" type="Subtract" version="opset1">
+		<layer id="2224" name="onnx::MatMul_3349/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32397,7 +32217,7 @@
 			</output>
 		</layer>
 		<layer id="2225" name="onnx::MatMul_3349/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="106100957" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="108148970" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -32406,57 +32226,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2226" name="onnx::MatMul_3349/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2227" name="onnx::MatMul_3349/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2226" name="onnx::MatMul_3349/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2228" name="Constant_3418" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2227" name="Constant_3417" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2229" name="Reshape_3419" type="Reshape" version="opset1">
+		<layer id="2228" name="Reshape_3418" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32466,7 +32269,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2229" name="onnx::MatMul_3349/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -32477,7 +32295,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -32488,7 +32306,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -32501,20 +32319,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2232" name="/blocks.15/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.15/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -32526,7 +32344,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -32536,14 +32354,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2234" name="Constant_1655" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -32554,7 +32372,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -32566,19 +32384,19 @@
 				<port id="2" precision="FP32" names="/blocks.15/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2236" name="/blocks.15/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.15/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="2237" name="/blocks.15/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.15/attn/Constant_1_output_0" />
 			</output>
@@ -32599,7 +32417,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -32608,13 +32426,13 @@
 				<port id="2" precision="FP32" names="/blocks.15/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2240" name="onnx::MatMul_3351" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="106117341" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="108165354" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -32624,7 +32442,7 @@
 			</output>
 		</layer>
 		<layer id="2241" name="Convert_3399" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -32633,7 +32451,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32641,7 +32459,7 @@
 			</output>
 		</layer>
 		<layer id="2242" name="onnx::MatMul_3351/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -32649,32 +32467,32 @@
 			</output>
 		</layer>
 		<layer id="2243" name="Convert_3401" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2244" name="Subtract_3402" type="Subtract" version="opset1">
+		<layer id="2244" name="onnx::MatMul_3351/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32682,7 +32500,7 @@
 			</output>
 		</layer>
 		<layer id="2245" name="onnx::MatMul_3351/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="106641629" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="108689642" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -32691,57 +32509,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2246" name="onnx::MatMul_3351/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2247" name="onnx::MatMul_3351/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2246" name="onnx::MatMul_3351/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2248" name="Constant_3407" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2247" name="Constant_3406" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2249" name="Reshape_3408" type="Reshape" version="opset1">
+		<layer id="2248" name="Reshape_3407" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32751,7 +32552,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2249" name="onnx::MatMul_3351/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -32762,7 +32578,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -32773,13 +32589,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2251" name="/blocks.15/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.15/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -32791,7 +32607,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -32801,14 +32617,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2253" name="Constant_1660" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -32819,7 +32635,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -32832,7 +32648,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -32843,7 +32659,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -32852,7 +32668,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -32862,22 +32678,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -32887,16 +32703,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -32906,16 +32722,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -32925,21 +32741,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2260" name="onnx::Add_3353" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="106658013" size="4096" />
+			<data element_type="f32" shape="1024" offset="108706026" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3353">
 					<dim>1024</dim>
@@ -32947,7 +32763,7 @@
 			</output>
 		</layer>
 		<layer id="2261" name="onnx::MatMul_3354" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="106662109" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="108710122" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -32957,7 +32773,7 @@
 			</output>
 		</layer>
 		<layer id="2262" name="Convert_3388" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -32966,7 +32782,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -32974,7 +32790,7 @@
 			</output>
 		</layer>
 		<layer id="2263" name="onnx::MatMul_3354/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="76003560" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -32982,32 +32798,32 @@
 			</output>
 		</layer>
 		<layer id="2264" name="Convert_3390" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2265" name="Subtract_3391" type="Subtract" version="opset1">
+		<layer id="2265" name="onnx::MatMul_3354/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -33015,7 +32831,7 @@
 			</output>
 		</layer>
 		<layer id="2266" name="onnx::MatMul_3354/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="107186397" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="109234410" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -33024,57 +32840,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2267" name="onnx::MatMul_3354/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2268" name="onnx::MatMul_3354/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2267" name="onnx::MatMul_3354/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2269" name="Constant_3396" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2268" name="Constant_3395" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2270" name="Reshape_3397" type="Reshape" version="opset1">
+		<layer id="2269" name="Reshape_3396" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -33084,7 +32883,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2270" name="onnx::MatMul_3354/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -33095,7 +32909,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -33106,7 +32920,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -33119,20 +32933,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2273" name="/blocks.15/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.15/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -33144,7 +32958,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -33154,14 +32968,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2275" name="Constant_1665" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -33172,7 +32986,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -33184,7 +32998,7 @@
 				<port id="2" precision="FP32" names="/blocks.15/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -33195,13 +33009,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -33209,13 +33023,13 @@
 				<port id="2" precision="FP32" names="/blocks.15/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2278" name="Constant_1672" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -33227,7 +33041,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -33237,7 +33051,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -33248,7 +33062,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -33260,7 +33074,7 @@
 			</output>
 		</layer>
 		<layer id="2281" name="/blocks.15/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.15/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -33313,7 +33127,7 @@
 			</output>
 		</layer>
 		<layer id="2286" name="/blocks.15/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.15/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -33345,7 +33159,7 @@
 			</output>
 		</layer>
 		<layer id="2288" name="/blocks.15/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.15/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -33373,7 +33187,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -33384,13 +33198,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2291" name="onnx::MatMul_3372" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="107202781" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="109250794" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -33400,7 +33214,7 @@
 			</output>
 		</layer>
 		<layer id="2292" name="Convert_3421" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -33409,7 +33223,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -33417,7 +33231,7 @@
 			</output>
 		</layer>
 		<layer id="2293" name="onnx::MatMul_3372/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -33425,32 +33239,32 @@
 			</output>
 		</layer>
 		<layer id="2294" name="Convert_3423" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2295" name="Subtract_3424" type="Subtract" version="opset1">
+		<layer id="2295" name="onnx::MatMul_3372/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -33458,7 +33272,7 @@
 			</output>
 		</layer>
 		<layer id="2296" name="onnx::MatMul_3372/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="107727069" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="109775082" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -33467,57 +33281,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2297" name="onnx::MatMul_3372/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2298" name="onnx::MatMul_3372/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2297" name="onnx::MatMul_3372/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2299" name="Constant_3429" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2298" name="Constant_3428" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2300" name="Reshape_3430" type="Reshape" version="opset1">
+		<layer id="2299" name="Reshape_3429" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -33527,7 +33324,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2300" name="onnx::MatMul_3372/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -33538,7 +33350,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -33549,7 +33361,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -33562,14 +33374,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -33579,19 +33391,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -33601,20 +33413,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2305" name="onnx::Add_3377" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="107743453" size="4096" />
+			<data element_type="f32" shape="1024" offset="109791466" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3377">
 					<dim>1024</dim>
@@ -33622,7 +33434,7 @@
 			</output>
 		</layer>
 		<layer id="2306" name="onnx::Add_3374" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="107747549" size="16384" />
+			<data element_type="f32" shape="4096" offset="109795562" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3374">
 					<dim>4096</dim>
@@ -33630,13 +33442,13 @@
 			</output>
 		</layer>
 		<layer id="2307" name="Constant_1698" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2308" name="Constant_1697" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -33665,7 +33477,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -33675,13 +33487,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2312" name="blocks.15.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="107763933" size="4096" />
+			<data element_type="f32" shape="1024" offset="109811946" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.15.mlp_ln.weight">
 					<dim>1024</dim>
@@ -33693,7 +33505,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -33703,13 +33515,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2314" name="blocks.15.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="107768029" size="4096" />
+			<data element_type="f32" shape="1024" offset="109816042" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.15.mlp_ln.bias">
 					<dim>1024</dim>
@@ -33721,7 +33533,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -33731,7 +33543,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -33741,20 +33553,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2317" name="onnx::MatMul_3375" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="107772125" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="109820138" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -33764,7 +33576,7 @@
 			</output>
 		</layer>
 		<layer id="2318" name="Convert_3432" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -33773,7 +33585,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -33781,7 +33593,7 @@
 			</output>
 		</layer>
 		<layer id="2319" name="onnx::MatMul_3375/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -33789,32 +33601,32 @@
 			</output>
 		</layer>
 		<layer id="2320" name="Convert_3434" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2321" name="Subtract_3435" type="Subtract" version="opset1">
+		<layer id="2321" name="onnx::MatMul_3375/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -33822,7 +33634,7 @@
 			</output>
 		</layer>
 		<layer id="2322" name="onnx::MatMul_3375/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="109869277" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="111917290" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -33831,57 +33643,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2323" name="onnx::MatMul_3375/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2324" name="onnx::MatMul_3375/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2323" name="onnx::MatMul_3375/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2325" name="Constant_3440" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="2324" name="Constant_3439" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2326" name="Reshape_3441" type="Reshape" version="opset1">
+		<layer id="2325" name="Reshape_3440" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -33891,7 +33686,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2326" name="onnx::MatMul_3375/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -33902,7 +33712,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -33913,7 +33723,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -33926,14 +33736,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -33949,7 +33759,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -33957,7 +33767,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -33966,14 +33776,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.15/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -33989,7 +33799,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -33997,7 +33807,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -34007,19 +33817,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -34035,7 +33845,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -34043,13 +33853,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2337" name="onnx::MatMul_3378" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="109934813" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="111982826" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -34059,7 +33869,7 @@
 			</output>
 		</layer>
 		<layer id="2338" name="Convert_3443" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -34068,7 +33878,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34076,7 +33886,7 @@
 			</output>
 		</layer>
 		<layer id="2339" name="onnx::MatMul_3378/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -34084,32 +33894,32 @@
 			</output>
 		</layer>
 		<layer id="2340" name="Convert_3445" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2341" name="Subtract_3446" type="Subtract" version="opset1">
+		<layer id="2341" name="onnx::MatMul_3378/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34117,7 +33927,7 @@
 			</output>
 		</layer>
 		<layer id="2342" name="onnx::MatMul_3378/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="112031965" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="114079978" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -34126,57 +33936,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2343" name="onnx::MatMul_3378/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2344" name="onnx::MatMul_3378/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2343" name="onnx::MatMul_3378/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2345" name="Constant_3451" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="2344" name="Constant_3450" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2346" name="Reshape_3452" type="Reshape" version="opset1">
+		<layer id="2345" name="Reshape_3451" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34186,7 +33979,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2346" name="onnx::MatMul_3378/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -34197,7 +34005,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -34208,7 +34016,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -34221,14 +34029,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -34238,19 +34046,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.15/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -34260,20 +34068,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2351" name="onnx::Add_3403" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="112097501" size="4096" />
+			<data element_type="f32" shape="1024" offset="114145514" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3403">
 					<dim>1024</dim>
@@ -34281,7 +34089,7 @@
 			</output>
 		</layer>
 		<layer id="2352" name="onnx::Add_3380" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="112101597" size="4096" />
+			<data element_type="f32" shape="1024" offset="114149610" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3380">
 					<dim>1024</dim>
@@ -34289,13 +34097,13 @@
 			</output>
 		</layer>
 		<layer id="2353" name="Constant_1720" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2354" name="Constant_1719" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -34324,7 +34132,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -34334,13 +34142,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2358" name="blocks.16.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="112105693" size="4096" />
+			<data element_type="f32" shape="1024" offset="114153706" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.16.attn_ln.weight">
 					<dim>1024</dim>
@@ -34352,7 +34160,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -34362,13 +34170,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2360" name="blocks.16.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="112109789" size="4096" />
+			<data element_type="f32" shape="1024" offset="114157802" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.16.attn_ln.bias">
 					<dim>1024</dim>
@@ -34380,7 +34188,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -34390,7 +34198,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -34400,20 +34208,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2363" name="onnx::MatMul_3381" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="112113885" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="114161898" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -34423,7 +34231,7 @@
 			</output>
 		</layer>
 		<layer id="2364" name="Convert_3476" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -34432,7 +34240,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34440,7 +34248,7 @@
 			</output>
 		</layer>
 		<layer id="2365" name="onnx::MatMul_3381/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -34448,32 +34256,32 @@
 			</output>
 		</layer>
 		<layer id="2366" name="Convert_3478" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2367" name="Subtract_3479" type="Subtract" version="opset1">
+		<layer id="2367" name="onnx::MatMul_3381/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34481,7 +34289,7 @@
 			</output>
 		</layer>
 		<layer id="2368" name="onnx::MatMul_3381/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="112638173" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="114686186" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -34490,57 +34298,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2369" name="onnx::MatMul_3381/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2370" name="onnx::MatMul_3381/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2369" name="onnx::MatMul_3381/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2371" name="Constant_3484" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2370" name="Constant_3483" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2372" name="Reshape_3485" type="Reshape" version="opset1">
+		<layer id="2371" name="Reshape_3484" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34550,7 +34341,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2372" name="onnx::MatMul_3381/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -34561,7 +34367,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -34572,7 +34378,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -34585,20 +34391,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2375" name="/blocks.16/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.16/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -34610,7 +34416,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -34620,14 +34426,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2377" name="Constant_1736" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -34638,7 +34444,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -34650,19 +34456,19 @@
 				<port id="2" precision="FP32" names="/blocks.16/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2379" name="/blocks.16/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.16/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="2380" name="/blocks.16/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.16/attn/Constant_1_output_0" />
 			</output>
@@ -34683,7 +34489,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -34692,13 +34498,13 @@
 				<port id="2" precision="FP32" names="/blocks.16/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2383" name="onnx::MatMul_3383" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="112654557" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="114702570" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -34708,7 +34514,7 @@
 			</output>
 		</layer>
 		<layer id="2384" name="Convert_3465" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -34717,7 +34523,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34725,7 +34531,7 @@
 			</output>
 		</layer>
 		<layer id="2385" name="onnx::MatMul_3383/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="17168607" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -34733,32 +34539,32 @@
 			</output>
 		</layer>
 		<layer id="2386" name="Convert_3467" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2387" name="Subtract_3468" type="Subtract" version="opset1">
+		<layer id="2387" name="onnx::MatMul_3383/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34766,7 +34572,7 @@
 			</output>
 		</layer>
 		<layer id="2388" name="onnx::MatMul_3383/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="113178845" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="115226858" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -34775,57 +34581,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2389" name="onnx::MatMul_3383/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2390" name="onnx::MatMul_3383/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2389" name="onnx::MatMul_3383/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2391" name="Constant_3473" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2390" name="Constant_3472" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2392" name="Reshape_3474" type="Reshape" version="opset1">
+		<layer id="2391" name="Reshape_3473" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -34835,7 +34624,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2392" name="onnx::MatMul_3383/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -34846,7 +34650,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -34857,13 +34661,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2394" name="/blocks.16/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.16/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -34875,7 +34679,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -34885,14 +34689,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2396" name="Constant_1741" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -34903,7 +34707,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -34916,7 +34720,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -34927,7 +34731,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -34936,7 +34740,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -34946,22 +34750,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -34971,16 +34775,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -34990,16 +34794,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -35009,21 +34813,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2403" name="onnx::Add_3385" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="113195229" size="4096" />
+			<data element_type="f32" shape="1024" offset="115243242" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3385">
 					<dim>1024</dim>
@@ -35031,7 +34835,7 @@
 			</output>
 		</layer>
 		<layer id="2404" name="onnx::MatMul_3386" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="113199325" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="115247338" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -35041,7 +34845,7 @@
 			</output>
 		</layer>
 		<layer id="2405" name="Convert_3454" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -35050,7 +34854,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -35058,7 +34862,7 @@
 			</output>
 		</layer>
 		<layer id="2406" name="onnx::MatMul_3386/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="74856679" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -35066,32 +34870,32 @@
 			</output>
 		</layer>
 		<layer id="2407" name="Convert_3456" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2408" name="Subtract_3457" type="Subtract" version="opset1">
+		<layer id="2408" name="onnx::MatMul_3386/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -35099,7 +34903,7 @@
 			</output>
 		</layer>
 		<layer id="2409" name="onnx::MatMul_3386/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="113723613" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="115771626" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -35108,57 +34912,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2410" name="onnx::MatMul_3386/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2411" name="onnx::MatMul_3386/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2410" name="onnx::MatMul_3386/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2412" name="Constant_3462" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2411" name="Constant_3461" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2413" name="Reshape_3463" type="Reshape" version="opset1">
+		<layer id="2412" name="Reshape_3462" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -35168,7 +34955,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2413" name="onnx::MatMul_3386/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -35179,7 +34981,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -35190,7 +34992,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -35203,20 +35005,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2416" name="/blocks.16/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.16/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -35228,7 +35030,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -35238,14 +35040,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2418" name="Constant_1746" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -35256,7 +35058,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -35268,7 +35070,7 @@
 				<port id="2" precision="FP32" names="/blocks.16/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -35279,13 +35081,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -35293,13 +35095,13 @@
 				<port id="2" precision="FP32" names="/blocks.16/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2421" name="Constant_1753" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -35311,7 +35113,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -35321,7 +35123,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -35332,7 +35134,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -35344,7 +35146,7 @@
 			</output>
 		</layer>
 		<layer id="2424" name="/blocks.16/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.16/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -35397,7 +35199,7 @@
 			</output>
 		</layer>
 		<layer id="2429" name="/blocks.16/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.16/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -35429,7 +35231,7 @@
 			</output>
 		</layer>
 		<layer id="2431" name="/blocks.16/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.16/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -35457,7 +35259,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -35468,13 +35270,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2434" name="onnx::MatMul_3404" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="113739997" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="115788010" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -35484,7 +35286,7 @@
 			</output>
 		</layer>
 		<layer id="2435" name="Convert_3487" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -35493,7 +35295,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -35501,7 +35303,7 @@
 			</output>
 		</layer>
 		<layer id="2436" name="onnx::MatMul_3404/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -35509,32 +35311,32 @@
 			</output>
 		</layer>
 		<layer id="2437" name="Convert_3489" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2438" name="Subtract_3490" type="Subtract" version="opset1">
+		<layer id="2438" name="onnx::MatMul_3404/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -35542,7 +35344,7 @@
 			</output>
 		</layer>
 		<layer id="2439" name="onnx::MatMul_3404/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="114264285" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="116312298" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -35551,57 +35353,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2440" name="onnx::MatMul_3404/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2441" name="onnx::MatMul_3404/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2440" name="onnx::MatMul_3404/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2442" name="Constant_3495" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2441" name="Constant_3494" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2443" name="Reshape_3496" type="Reshape" version="opset1">
+		<layer id="2442" name="Reshape_3495" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -35611,7 +35396,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2443" name="onnx::MatMul_3404/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -35622,7 +35422,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -35633,7 +35433,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -35646,14 +35446,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -35663,19 +35463,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -35685,20 +35485,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2448" name="onnx::Add_3409" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="114280669" size="4096" />
+			<data element_type="f32" shape="1024" offset="116328682" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3409">
 					<dim>1024</dim>
@@ -35706,7 +35506,7 @@
 			</output>
 		</layer>
 		<layer id="2449" name="onnx::Add_3406" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="114284765" size="16384" />
+			<data element_type="f32" shape="4096" offset="116332778" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3406">
 					<dim>4096</dim>
@@ -35714,13 +35514,13 @@
 			</output>
 		</layer>
 		<layer id="2450" name="Constant_1779" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2451" name="Constant_1778" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -35749,7 +35549,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -35759,13 +35559,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2455" name="blocks.16.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="114301149" size="4096" />
+			<data element_type="f32" shape="1024" offset="116349162" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.16.mlp_ln.weight">
 					<dim>1024</dim>
@@ -35777,7 +35577,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -35787,13 +35587,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2457" name="blocks.16.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="114305245" size="4096" />
+			<data element_type="f32" shape="1024" offset="116353258" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.16.mlp_ln.bias">
 					<dim>1024</dim>
@@ -35805,7 +35605,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -35815,7 +35615,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -35825,20 +35625,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2460" name="onnx::MatMul_3407" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="114309341" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="116357354" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -35848,7 +35648,7 @@
 			</output>
 		</layer>
 		<layer id="2461" name="Convert_3498" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -35857,7 +35657,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -35865,7 +35665,7 @@
 			</output>
 		</layer>
 		<layer id="2462" name="onnx::MatMul_3407/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -35873,32 +35673,32 @@
 			</output>
 		</layer>
 		<layer id="2463" name="Convert_3500" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2464" name="Subtract_3501" type="Subtract" version="opset1">
+		<layer id="2464" name="onnx::MatMul_3407/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -35906,7 +35706,7 @@
 			</output>
 		</layer>
 		<layer id="2465" name="onnx::MatMul_3407/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="116406493" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="118454506" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -35915,57 +35715,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2466" name="onnx::MatMul_3407/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2467" name="onnx::MatMul_3407/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2466" name="onnx::MatMul_3407/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2468" name="Constant_3506" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="2467" name="Constant_3505" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2469" name="Reshape_3507" type="Reshape" version="opset1">
+		<layer id="2468" name="Reshape_3506" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -35975,7 +35758,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2469" name="onnx::MatMul_3407/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -35986,7 +35784,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -35997,7 +35795,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -36010,14 +35808,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -36033,7 +35831,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -36041,7 +35839,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -36050,14 +35848,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.16/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -36073,7 +35871,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -36081,7 +35879,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -36091,19 +35889,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -36119,7 +35917,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -36127,13 +35925,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2480" name="onnx::MatMul_3410" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="116472029" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="118520042" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -36143,7 +35941,7 @@
 			</output>
 		</layer>
 		<layer id="2481" name="Convert_3509" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -36152,7 +35950,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36160,7 +35958,7 @@
 			</output>
 		</layer>
 		<layer id="2482" name="onnx::MatMul_3410/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="31328483" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -36168,32 +35966,32 @@
 			</output>
 		</layer>
 		<layer id="2483" name="Convert_3511" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2484" name="Subtract_3512" type="Subtract" version="opset1">
+		<layer id="2484" name="onnx::MatMul_3410/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36201,7 +35999,7 @@
 			</output>
 		</layer>
 		<layer id="2485" name="onnx::MatMul_3410/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="118569181" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="120617194" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -36210,57 +36008,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2486" name="onnx::MatMul_3410/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2487" name="onnx::MatMul_3410/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2486" name="onnx::MatMul_3410/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2488" name="Constant_3517" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="2487" name="Constant_3516" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2489" name="Reshape_3518" type="Reshape" version="opset1">
+		<layer id="2488" name="Reshape_3517" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36270,7 +36051,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2489" name="onnx::MatMul_3410/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -36281,7 +36077,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -36292,7 +36088,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -36305,14 +36101,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -36322,19 +36118,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.16/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -36344,20 +36140,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2494" name="onnx::Add_3435" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="118634717" size="4096" />
+			<data element_type="f32" shape="1024" offset="120682730" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3435">
 					<dim>1024</dim>
@@ -36365,7 +36161,7 @@
 			</output>
 		</layer>
 		<layer id="2495" name="onnx::Add_3412" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="118638813" size="4096" />
+			<data element_type="f32" shape="1024" offset="120686826" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3412">
 					<dim>1024</dim>
@@ -36373,13 +36169,13 @@
 			</output>
 		</layer>
 		<layer id="2496" name="Constant_1801" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2497" name="Constant_1800" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -36408,7 +36204,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -36418,13 +36214,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2501" name="blocks.17.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="118642909" size="4096" />
+			<data element_type="f32" shape="1024" offset="120690922" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.17.attn_ln.weight">
 					<dim>1024</dim>
@@ -36436,7 +36232,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -36446,13 +36242,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2503" name="blocks.17.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="118647005" size="4096" />
+			<data element_type="f32" shape="1024" offset="120695018" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.17.attn_ln.bias">
 					<dim>1024</dim>
@@ -36464,7 +36260,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -36474,7 +36270,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -36484,20 +36280,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2506" name="onnx::MatMul_3413" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="118651101" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="120699114" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -36507,7 +36303,7 @@
 			</output>
 		</layer>
 		<layer id="2507" name="Convert_3542" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -36516,7 +36312,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36524,7 +36320,7 @@
 			</output>
 		</layer>
 		<layer id="2508" name="onnx::MatMul_3413/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -36532,32 +36328,32 @@
 			</output>
 		</layer>
 		<layer id="2509" name="Convert_3544" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2510" name="Subtract_3545" type="Subtract" version="opset1">
+		<layer id="2510" name="onnx::MatMul_3413/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36565,7 +36361,7 @@
 			</output>
 		</layer>
 		<layer id="2511" name="onnx::MatMul_3413/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="119175389" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="121223402" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -36574,57 +36370,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2512" name="onnx::MatMul_3413/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2513" name="onnx::MatMul_3413/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2512" name="onnx::MatMul_3413/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2514" name="Constant_3550" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2513" name="Constant_3549" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2515" name="Reshape_3551" type="Reshape" version="opset1">
+		<layer id="2514" name="Reshape_3550" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36634,7 +36413,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2515" name="onnx::MatMul_3413/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -36645,7 +36439,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -36656,7 +36450,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -36669,20 +36463,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2518" name="/blocks.17/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.17/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -36694,7 +36488,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -36704,14 +36498,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2520" name="Constant_1817" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -36722,7 +36516,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -36734,19 +36528,19 @@
 				<port id="2" precision="FP32" names="/blocks.17/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2522" name="/blocks.17/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.17/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="2523" name="/blocks.17/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.17/attn/Constant_1_output_0" />
 			</output>
@@ -36767,7 +36561,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -36776,13 +36570,13 @@
 				<port id="2" precision="FP32" names="/blocks.17/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2526" name="onnx::MatMul_3415" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="119191773" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="121239786" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -36792,7 +36586,7 @@
 			</output>
 		</layer>
 		<layer id="2527" name="Convert_3531" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -36801,7 +36595,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36809,7 +36603,7 @@
 			</output>
 		</layer>
 		<layer id="2528" name="onnx::MatMul_3415/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -36817,32 +36611,32 @@
 			</output>
 		</layer>
 		<layer id="2529" name="Convert_3533" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2530" name="Subtract_3534" type="Subtract" version="opset1">
+		<layer id="2530" name="onnx::MatMul_3415/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36850,7 +36644,7 @@
 			</output>
 		</layer>
 		<layer id="2531" name="onnx::MatMul_3415/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="119716061" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="121764074" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -36859,57 +36653,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2532" name="onnx::MatMul_3415/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2533" name="onnx::MatMul_3415/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2532" name="onnx::MatMul_3415/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2534" name="Constant_3539" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2533" name="Constant_3538" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2535" name="Reshape_3540" type="Reshape" version="opset1">
+		<layer id="2534" name="Reshape_3539" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -36919,7 +36696,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2535" name="onnx::MatMul_3415/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -36930,7 +36722,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -36941,13 +36733,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2537" name="/blocks.17/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.17/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -36959,7 +36751,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -36969,14 +36761,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2539" name="Constant_1822" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -36987,7 +36779,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -37000,7 +36792,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -37011,7 +36803,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -37020,7 +36812,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -37030,22 +36822,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -37055,16 +36847,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -37074,16 +36866,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -37093,21 +36885,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2546" name="onnx::Add_3417" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="119732445" size="4096" />
+			<data element_type="f32" shape="1024" offset="121780458" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3417">
 					<dim>1024</dim>
@@ -37115,7 +36907,7 @@
 			</output>
 		</layer>
 		<layer id="2547" name="onnx::MatMul_3418" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="119736541" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="121784554" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -37125,7 +36917,7 @@
 			</output>
 		</layer>
 		<layer id="2548" name="Convert_3520" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -37134,7 +36926,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -37142,7 +36934,7 @@
 			</output>
 		</layer>
 		<layer id="2549" name="onnx::MatMul_3418/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -37150,32 +36942,32 @@
 			</output>
 		</layer>
 		<layer id="2550" name="Convert_3522" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2551" name="Subtract_3523" type="Subtract" version="opset1">
+		<layer id="2551" name="onnx::MatMul_3418/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -37183,7 +36975,7 @@
 			</output>
 		</layer>
 		<layer id="2552" name="onnx::MatMul_3418/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="120260829" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="122308842" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -37192,57 +36984,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2553" name="onnx::MatMul_3418/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2554" name="onnx::MatMul_3418/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2553" name="onnx::MatMul_3418/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2555" name="Constant_3528" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2554" name="Constant_3527" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2556" name="Reshape_3529" type="Reshape" version="opset1">
+		<layer id="2555" name="Reshape_3528" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -37252,7 +37027,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2556" name="onnx::MatMul_3418/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -37263,7 +37053,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -37274,7 +37064,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -37287,20 +37077,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2559" name="/blocks.17/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.17/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -37312,7 +37102,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -37322,14 +37112,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2561" name="Constant_1827" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -37340,7 +37130,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -37352,7 +37142,7 @@
 				<port id="2" precision="FP32" names="/blocks.17/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -37363,13 +37153,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -37377,13 +37167,13 @@
 				<port id="2" precision="FP32" names="/blocks.17/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2564" name="Constant_1834" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -37395,7 +37185,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -37405,7 +37195,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -37416,7 +37206,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -37428,7 +37218,7 @@
 			</output>
 		</layer>
 		<layer id="2567" name="/blocks.17/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.17/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -37481,7 +37271,7 @@
 			</output>
 		</layer>
 		<layer id="2572" name="/blocks.17/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.17/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -37513,7 +37303,7 @@
 			</output>
 		</layer>
 		<layer id="2574" name="/blocks.17/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.17/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -37541,7 +37331,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -37552,13 +37342,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2577" name="onnx::MatMul_3436" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="120277213" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="122325226" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -37568,7 +37358,7 @@
 			</output>
 		</layer>
 		<layer id="2578" name="Convert_3553" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -37577,7 +37367,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -37585,7 +37375,7 @@
 			</output>
 		</layer>
 		<layer id="2579" name="onnx::MatMul_3436/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -37593,32 +37383,32 @@
 			</output>
 		</layer>
 		<layer id="2580" name="Convert_3555" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2581" name="Subtract_3556" type="Subtract" version="opset1">
+		<layer id="2581" name="onnx::MatMul_3436/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -37626,7 +37416,7 @@
 			</output>
 		</layer>
 		<layer id="2582" name="onnx::MatMul_3436/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="120801501" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="122849514" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -37635,57 +37425,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2583" name="onnx::MatMul_3436/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2584" name="onnx::MatMul_3436/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2583" name="onnx::MatMul_3436/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2585" name="Constant_3561" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2584" name="Constant_3560" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2586" name="Reshape_3562" type="Reshape" version="opset1">
+		<layer id="2585" name="Reshape_3561" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -37695,7 +37468,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2586" name="onnx::MatMul_3436/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -37706,7 +37494,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -37717,7 +37505,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -37730,14 +37518,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -37747,19 +37535,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -37769,20 +37557,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2591" name="onnx::Add_3441" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="120817885" size="4096" />
+			<data element_type="f32" shape="1024" offset="122865898" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3441">
 					<dim>1024</dim>
@@ -37790,7 +37578,7 @@
 			</output>
 		</layer>
 		<layer id="2592" name="onnx::Add_3438" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="120821981" size="16384" />
+			<data element_type="f32" shape="4096" offset="122869994" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3438">
 					<dim>4096</dim>
@@ -37798,13 +37586,13 @@
 			</output>
 		</layer>
 		<layer id="2593" name="Constant_1860" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2594" name="Constant_1859" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -37833,7 +37621,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -37843,13 +37631,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2598" name="blocks.17.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="120838365" size="4096" />
+			<data element_type="f32" shape="1024" offset="122886378" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.17.mlp_ln.weight">
 					<dim>1024</dim>
@@ -37861,7 +37649,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -37871,13 +37659,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2600" name="blocks.17.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="120842461" size="4096" />
+			<data element_type="f32" shape="1024" offset="122890474" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.17.mlp_ln.bias">
 					<dim>1024</dim>
@@ -37889,7 +37677,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -37899,7 +37687,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -37909,20 +37697,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2603" name="onnx::MatMul_3439" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="120846557" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="122894570" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -37932,7 +37720,7 @@
 			</output>
 		</layer>
 		<layer id="2604" name="Convert_3564" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -37941,7 +37729,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -37949,7 +37737,7 @@
 			</output>
 		</layer>
 		<layer id="2605" name="onnx::MatMul_3439/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -37957,32 +37745,32 @@
 			</output>
 		</layer>
 		<layer id="2606" name="Convert_3566" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2607" name="Subtract_3567" type="Subtract" version="opset1">
+		<layer id="2607" name="onnx::MatMul_3439/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -37990,7 +37778,7 @@
 			</output>
 		</layer>
 		<layer id="2608" name="onnx::MatMul_3439/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="122943709" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="124991722" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -37999,57 +37787,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2609" name="onnx::MatMul_3439/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2610" name="onnx::MatMul_3439/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2609" name="onnx::MatMul_3439/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2611" name="Constant_3572" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="2610" name="Constant_3571" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2612" name="Reshape_3573" type="Reshape" version="opset1">
+		<layer id="2611" name="Reshape_3572" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -38059,7 +37830,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2612" name="onnx::MatMul_3439/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -38070,7 +37856,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -38081,7 +37867,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -38094,14 +37880,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -38117,7 +37903,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -38125,7 +37911,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -38134,14 +37920,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.17/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -38157,7 +37943,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -38165,7 +37951,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -38175,19 +37961,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -38203,7 +37989,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -38211,13 +37997,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2623" name="onnx::MatMul_3442" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="123009245" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="125057258" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -38227,7 +38013,7 @@
 			</output>
 		</layer>
 		<layer id="2624" name="Convert_3575" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -38236,7 +38022,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38244,7 +38030,7 @@
 			</output>
 		</layer>
 		<layer id="2625" name="onnx::MatMul_3442/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -38252,32 +38038,32 @@
 			</output>
 		</layer>
 		<layer id="2626" name="Convert_3577" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2627" name="Subtract_3578" type="Subtract" version="opset1">
+		<layer id="2627" name="onnx::MatMul_3442/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38285,7 +38071,7 @@
 			</output>
 		</layer>
 		<layer id="2628" name="onnx::MatMul_3442/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="125106397" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="127154410" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -38294,57 +38080,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2629" name="onnx::MatMul_3442/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2630" name="onnx::MatMul_3442/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2629" name="onnx::MatMul_3442/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2631" name="Constant_3583" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="2630" name="Constant_3582" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2632" name="Reshape_3584" type="Reshape" version="opset1">
+		<layer id="2631" name="Reshape_3583" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38354,7 +38123,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2632" name="onnx::MatMul_3442/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -38365,7 +38149,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -38376,7 +38160,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -38389,14 +38173,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -38406,19 +38190,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.17/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -38428,20 +38212,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2637" name="onnx::Add_3467" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="125171933" size="4096" />
+			<data element_type="f32" shape="1024" offset="127219946" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3467">
 					<dim>1024</dim>
@@ -38449,7 +38233,7 @@
 			</output>
 		</layer>
 		<layer id="2638" name="onnx::Add_3444" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="125176029" size="4096" />
+			<data element_type="f32" shape="1024" offset="127224042" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3444">
 					<dim>1024</dim>
@@ -38457,13 +38241,13 @@
 			</output>
 		</layer>
 		<layer id="2639" name="Constant_1882" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2640" name="Constant_1881" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -38492,7 +38276,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -38502,13 +38286,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2644" name="blocks.18.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="125180125" size="4096" />
+			<data element_type="f32" shape="1024" offset="127228138" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.18.attn_ln.weight">
 					<dim>1024</dim>
@@ -38520,7 +38304,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -38530,13 +38314,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2646" name="blocks.18.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="125184221" size="4096" />
+			<data element_type="f32" shape="1024" offset="127232234" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.18.attn_ln.bias">
 					<dim>1024</dim>
@@ -38548,7 +38332,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -38558,7 +38342,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -38568,20 +38352,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2649" name="onnx::MatMul_3445" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="125188317" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="127236330" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -38591,7 +38375,7 @@
 			</output>
 		</layer>
 		<layer id="2650" name="Convert_3608" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -38600,7 +38384,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38608,7 +38392,7 @@
 			</output>
 		</layer>
 		<layer id="2651" name="onnx::MatMul_3445/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -38616,32 +38400,32 @@
 			</output>
 		</layer>
 		<layer id="2652" name="Convert_3610" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2653" name="Subtract_3611" type="Subtract" version="opset1">
+		<layer id="2653" name="onnx::MatMul_3445/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38649,7 +38433,7 @@
 			</output>
 		</layer>
 		<layer id="2654" name="onnx::MatMul_3445/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="125712605" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="127760618" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -38658,57 +38442,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2655" name="onnx::MatMul_3445/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2656" name="onnx::MatMul_3445/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2655" name="onnx::MatMul_3445/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2657" name="Constant_3616" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2656" name="Constant_3615" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2658" name="Reshape_3617" type="Reshape" version="opset1">
+		<layer id="2657" name="Reshape_3616" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38718,7 +38485,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2658" name="onnx::MatMul_3445/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -38729,7 +38511,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -38740,7 +38522,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -38753,20 +38535,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2661" name="/blocks.18/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.18/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -38778,7 +38560,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -38788,14 +38570,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2663" name="Constant_1898" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -38806,7 +38588,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -38818,19 +38600,19 @@
 				<port id="2" precision="FP32" names="/blocks.18/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2665" name="/blocks.18/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.18/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="2666" name="/blocks.18/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.18/attn/Constant_1_output_0" />
 			</output>
@@ -38851,7 +38633,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -38860,13 +38642,13 @@
 				<port id="2" precision="FP32" names="/blocks.18/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2669" name="onnx::MatMul_3447" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="125728989" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="127777002" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -38876,7 +38658,7 @@
 			</output>
 		</layer>
 		<layer id="2670" name="Convert_3597" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -38885,7 +38667,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38893,7 +38675,7 @@
 			</output>
 		</layer>
 		<layer id="2671" name="onnx::MatMul_3447/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -38901,32 +38683,32 @@
 			</output>
 		</layer>
 		<layer id="2672" name="Convert_3599" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2673" name="Subtract_3600" type="Subtract" version="opset1">
+		<layer id="2673" name="onnx::MatMul_3447/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -38934,7 +38716,7 @@
 			</output>
 		</layer>
 		<layer id="2674" name="onnx::MatMul_3447/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="126253277" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="128301290" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -38943,57 +38725,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2675" name="onnx::MatMul_3447/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2676" name="onnx::MatMul_3447/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2675" name="onnx::MatMul_3447/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2677" name="Constant_3605" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2676" name="Constant_3604" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2678" name="Reshape_3606" type="Reshape" version="opset1">
+		<layer id="2677" name="Reshape_3605" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -39003,7 +38768,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2678" name="onnx::MatMul_3447/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -39014,7 +38794,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -39025,13 +38805,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2680" name="/blocks.18/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.18/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -39043,7 +38823,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -39053,14 +38833,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2682" name="Constant_1903" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -39071,7 +38851,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -39084,7 +38864,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -39095,7 +38875,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -39104,7 +38884,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -39114,22 +38894,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -39139,16 +38919,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -39158,16 +38938,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -39177,21 +38957,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2689" name="onnx::Add_3449" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="126269661" size="4096" />
+			<data element_type="f32" shape="1024" offset="128317674" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3449">
 					<dim>1024</dim>
@@ -39199,7 +38979,7 @@
 			</output>
 		</layer>
 		<layer id="2690" name="onnx::MatMul_3450" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="126273757" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="128321770" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -39209,7 +38989,7 @@
 			</output>
 		</layer>
 		<layer id="2691" name="Convert_3586" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -39218,7 +38998,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -39226,7 +39006,7 @@
 			</output>
 		</layer>
 		<layer id="2692" name="onnx::MatMul_3450/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -39234,32 +39014,32 @@
 			</output>
 		</layer>
 		<layer id="2693" name="Convert_3588" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2694" name="Subtract_3589" type="Subtract" version="opset1">
+		<layer id="2694" name="onnx::MatMul_3450/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -39267,7 +39047,7 @@
 			</output>
 		</layer>
 		<layer id="2695" name="onnx::MatMul_3450/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="126798045" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="128846058" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -39276,57 +39056,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2696" name="onnx::MatMul_3450/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2697" name="onnx::MatMul_3450/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2696" name="onnx::MatMul_3450/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2698" name="Constant_3594" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2697" name="Constant_3593" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2699" name="Reshape_3595" type="Reshape" version="opset1">
+		<layer id="2698" name="Reshape_3594" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -39336,7 +39099,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2699" name="onnx::MatMul_3450/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -39347,7 +39125,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -39358,7 +39136,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -39371,20 +39149,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2702" name="/blocks.18/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.18/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -39396,7 +39174,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -39406,14 +39184,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2704" name="Constant_1908" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -39424,7 +39202,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -39436,7 +39214,7 @@
 				<port id="2" precision="FP32" names="/blocks.18/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -39447,13 +39225,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -39461,13 +39239,13 @@
 				<port id="2" precision="FP32" names="/blocks.18/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2707" name="Constant_1915" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -39479,7 +39257,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -39489,7 +39267,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -39500,7 +39278,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -39512,7 +39290,7 @@
 			</output>
 		</layer>
 		<layer id="2710" name="/blocks.18/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.18/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -39565,7 +39343,7 @@
 			</output>
 		</layer>
 		<layer id="2715" name="/blocks.18/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.18/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -39597,7 +39375,7 @@
 			</output>
 		</layer>
 		<layer id="2717" name="/blocks.18/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.18/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -39625,7 +39403,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -39636,13 +39414,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2720" name="onnx::MatMul_3468" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="126814429" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="128862442" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -39652,7 +39430,7 @@
 			</output>
 		</layer>
 		<layer id="2721" name="Convert_3619" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -39661,7 +39439,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -39669,7 +39447,7 @@
 			</output>
 		</layer>
 		<layer id="2722" name="onnx::MatMul_3468/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -39677,32 +39455,32 @@
 			</output>
 		</layer>
 		<layer id="2723" name="Convert_3621" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2724" name="Subtract_3622" type="Subtract" version="opset1">
+		<layer id="2724" name="onnx::MatMul_3468/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -39710,7 +39488,7 @@
 			</output>
 		</layer>
 		<layer id="2725" name="onnx::MatMul_3468/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="127338717" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="129386730" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -39719,57 +39497,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2726" name="onnx::MatMul_3468/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2727" name="onnx::MatMul_3468/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2726" name="onnx::MatMul_3468/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2728" name="Constant_3627" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2727" name="Constant_3626" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2729" name="Reshape_3628" type="Reshape" version="opset1">
+		<layer id="2728" name="Reshape_3627" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -39779,7 +39540,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2729" name="onnx::MatMul_3468/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -39790,7 +39566,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -39801,7 +39577,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -39814,14 +39590,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -39831,19 +39607,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -39853,20 +39629,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2734" name="onnx::Add_3473" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="127355101" size="4096" />
+			<data element_type="f32" shape="1024" offset="129403114" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3473">
 					<dim>1024</dim>
@@ -39874,7 +39650,7 @@
 			</output>
 		</layer>
 		<layer id="2735" name="onnx::Add_3470" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="127359197" size="16384" />
+			<data element_type="f32" shape="4096" offset="129407210" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3470">
 					<dim>4096</dim>
@@ -39882,13 +39658,13 @@
 			</output>
 		</layer>
 		<layer id="2736" name="Constant_1941" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2737" name="Constant_1940" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -39917,7 +39693,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -39927,13 +39703,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2741" name="blocks.18.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="127375581" size="4096" />
+			<data element_type="f32" shape="1024" offset="129423594" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.18.mlp_ln.weight">
 					<dim>1024</dim>
@@ -39945,7 +39721,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -39955,13 +39731,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2743" name="blocks.18.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="127379677" size="4096" />
+			<data element_type="f32" shape="1024" offset="129427690" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.18.mlp_ln.bias">
 					<dim>1024</dim>
@@ -39973,7 +39749,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -39983,7 +39759,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -39993,20 +39769,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2746" name="onnx::MatMul_3471" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="127383773" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="129431786" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -40016,7 +39792,7 @@
 			</output>
 		</layer>
 		<layer id="2747" name="Convert_3630" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -40025,7 +39801,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -40033,7 +39809,7 @@
 			</output>
 		</layer>
 		<layer id="2748" name="onnx::MatMul_3471/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -40041,32 +39817,32 @@
 			</output>
 		</layer>
 		<layer id="2749" name="Convert_3632" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2750" name="Subtract_3633" type="Subtract" version="opset1">
+		<layer id="2750" name="onnx::MatMul_3471/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -40074,7 +39850,7 @@
 			</output>
 		</layer>
 		<layer id="2751" name="onnx::MatMul_3471/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="129480925" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="131528938" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -40083,57 +39859,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2752" name="onnx::MatMul_3471/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2753" name="onnx::MatMul_3471/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2752" name="onnx::MatMul_3471/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2754" name="Constant_3638" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="2753" name="Constant_3637" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2755" name="Reshape_3639" type="Reshape" version="opset1">
+		<layer id="2754" name="Reshape_3638" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -40143,7 +39902,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2755" name="onnx::MatMul_3471/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -40154,7 +39928,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -40165,7 +39939,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -40178,14 +39952,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -40201,7 +39975,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -40209,7 +39983,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -40218,14 +39992,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.18/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -40241,7 +40015,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -40249,7 +40023,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -40259,19 +40033,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -40287,7 +40061,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -40295,13 +40069,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2766" name="onnx::MatMul_3474" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="129546461" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="131594474" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -40311,7 +40085,7 @@
 			</output>
 		</layer>
 		<layer id="2767" name="Convert_3641" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -40320,7 +40094,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -40328,7 +40102,7 @@
 			</output>
 		</layer>
 		<layer id="2768" name="onnx::MatMul_3474/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="23165152" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -40336,32 +40110,32 @@
 			</output>
 		</layer>
 		<layer id="2769" name="Convert_3643" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2770" name="Subtract_3644" type="Subtract" version="opset1">
+		<layer id="2770" name="onnx::MatMul_3474/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -40369,7 +40143,7 @@
 			</output>
 		</layer>
 		<layer id="2771" name="onnx::MatMul_3474/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="131643613" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="133691626" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -40378,57 +40152,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2772" name="onnx::MatMul_3474/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2773" name="onnx::MatMul_3474/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2772" name="onnx::MatMul_3474/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2774" name="Constant_3649" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="2773" name="Constant_3648" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2775" name="Reshape_3650" type="Reshape" version="opset1">
+		<layer id="2774" name="Reshape_3649" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -40438,7 +40195,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2775" name="onnx::MatMul_3474/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -40449,7 +40221,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -40460,7 +40232,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -40473,14 +40245,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -40490,19 +40262,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.18/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -40512,20 +40284,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2780" name="onnx::Add_3499" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="131709149" size="4096" />
+			<data element_type="f32" shape="1024" offset="133757162" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3499">
 					<dim>1024</dim>
@@ -40533,7 +40305,7 @@
 			</output>
 		</layer>
 		<layer id="2781" name="onnx::Add_3476" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="131713245" size="4096" />
+			<data element_type="f32" shape="1024" offset="133761258" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3476">
 					<dim>1024</dim>
@@ -40541,13 +40313,13 @@
 			</output>
 		</layer>
 		<layer id="2782" name="Constant_1963" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2783" name="Constant_1962" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -40576,7 +40348,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -40586,13 +40358,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2787" name="blocks.19.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="131717341" size="4096" />
+			<data element_type="f32" shape="1024" offset="133765354" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.19.attn_ln.weight">
 					<dim>1024</dim>
@@ -40604,7 +40376,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -40614,13 +40386,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2789" name="blocks.19.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="131721437" size="4096" />
+			<data element_type="f32" shape="1024" offset="133769450" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.19.attn_ln.bias">
 					<dim>1024</dim>
@@ -40632,7 +40404,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -40642,7 +40414,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -40652,20 +40424,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2792" name="onnx::MatMul_3477" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="131725533" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="133773546" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -40675,7 +40447,7 @@
 			</output>
 		</layer>
 		<layer id="2793" name="Convert_3674" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -40684,7 +40456,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -40692,7 +40464,7 @@
 			</output>
 		</layer>
 		<layer id="2794" name="onnx::MatMul_3477/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -40700,32 +40472,32 @@
 			</output>
 		</layer>
 		<layer id="2795" name="Convert_3676" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2796" name="Subtract_3677" type="Subtract" version="opset1">
+		<layer id="2796" name="onnx::MatMul_3477/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -40733,7 +40505,7 @@
 			</output>
 		</layer>
 		<layer id="2797" name="onnx::MatMul_3477/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="132249821" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="134297834" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -40742,57 +40514,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2798" name="onnx::MatMul_3477/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2799" name="onnx::MatMul_3477/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2798" name="onnx::MatMul_3477/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2800" name="Constant_3682" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2799" name="Constant_3681" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2801" name="Reshape_3683" type="Reshape" version="opset1">
+		<layer id="2800" name="Reshape_3682" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -40802,7 +40557,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2801" name="onnx::MatMul_3477/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -40813,7 +40583,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -40824,7 +40594,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -40837,20 +40607,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2804" name="/blocks.19/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.19/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -40862,7 +40632,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -40872,14 +40642,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2806" name="Constant_1979" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -40890,7 +40660,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -40902,19 +40672,19 @@
 				<port id="2" precision="FP32" names="/blocks.19/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2808" name="/blocks.19/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.19/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="2809" name="/blocks.19/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.19/attn/Constant_1_output_0" />
 			</output>
@@ -40935,7 +40705,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -40944,13 +40714,13 @@
 				<port id="2" precision="FP32" names="/blocks.19/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2812" name="onnx::MatMul_3479" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="132266205" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="134314218" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -40960,7 +40730,7 @@
 			</output>
 		</layer>
 		<layer id="2813" name="Convert_3663" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -40969,7 +40739,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -40977,7 +40747,7 @@
 			</output>
 		</layer>
 		<layer id="2814" name="onnx::MatMul_3479/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -40985,32 +40755,32 @@
 			</output>
 		</layer>
 		<layer id="2815" name="Convert_3665" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2816" name="Subtract_3666" type="Subtract" version="opset1">
+		<layer id="2816" name="onnx::MatMul_3479/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41018,7 +40788,7 @@
 			</output>
 		</layer>
 		<layer id="2817" name="onnx::MatMul_3479/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="132790493" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="134838506" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -41027,57 +40797,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2818" name="onnx::MatMul_3479/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2819" name="onnx::MatMul_3479/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2818" name="onnx::MatMul_3479/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2820" name="Constant_3671" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2819" name="Constant_3670" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2821" name="Reshape_3672" type="Reshape" version="opset1">
+		<layer id="2820" name="Reshape_3671" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41087,7 +40840,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2821" name="onnx::MatMul_3479/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -41098,7 +40866,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -41109,13 +40877,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2823" name="/blocks.19/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.19/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -41127,7 +40895,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -41137,14 +40905,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2825" name="Constant_1984" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -41155,7 +40923,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -41168,7 +40936,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -41179,7 +40947,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -41188,7 +40956,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -41198,22 +40966,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -41223,16 +40991,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -41242,16 +41010,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -41261,21 +41029,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2832" name="onnx::Add_3481" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="132806877" size="4096" />
+			<data element_type="f32" shape="1024" offset="134854890" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3481">
 					<dim>1024</dim>
@@ -41283,7 +41051,7 @@
 			</output>
 		</layer>
 		<layer id="2833" name="onnx::MatMul_3482" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="132810973" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="134858986" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -41293,7 +41061,7 @@
 			</output>
 		</layer>
 		<layer id="2834" name="Convert_3652" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -41302,7 +41070,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41310,7 +41078,7 @@
 			</output>
 		</layer>
 		<layer id="2835" name="onnx::MatMul_3482/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="40007909" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -41318,32 +41086,32 @@
 			</output>
 		</layer>
 		<layer id="2836" name="Convert_3654" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2837" name="Subtract_3655" type="Subtract" version="opset1">
+		<layer id="2837" name="onnx::MatMul_3482/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41351,7 +41119,7 @@
 			</output>
 		</layer>
 		<layer id="2838" name="onnx::MatMul_3482/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="133335261" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="135383274" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -41360,57 +41128,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2839" name="onnx::MatMul_3482/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2840" name="onnx::MatMul_3482/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2839" name="onnx::MatMul_3482/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2841" name="Constant_3660" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2840" name="Constant_3659" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2842" name="Reshape_3661" type="Reshape" version="opset1">
+		<layer id="2841" name="Reshape_3660" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41420,7 +41171,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2842" name="onnx::MatMul_3482/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -41431,7 +41197,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -41442,7 +41208,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -41455,20 +41221,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2845" name="/blocks.19/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.19/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -41480,7 +41246,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -41490,14 +41256,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2847" name="Constant_1989" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -41508,7 +41274,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -41520,7 +41286,7 @@
 				<port id="2" precision="FP32" names="/blocks.19/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -41531,13 +41297,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -41545,13 +41311,13 @@
 				<port id="2" precision="FP32" names="/blocks.19/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2850" name="Constant_1996" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -41563,7 +41329,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -41573,7 +41339,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -41584,7 +41350,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -41596,7 +41362,7 @@
 			</output>
 		</layer>
 		<layer id="2853" name="/blocks.19/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.19/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -41649,7 +41415,7 @@
 			</output>
 		</layer>
 		<layer id="2858" name="/blocks.19/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.19/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -41681,7 +41447,7 @@
 			</output>
 		</layer>
 		<layer id="2860" name="/blocks.19/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.19/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -41709,7 +41475,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -41720,13 +41486,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2863" name="onnx::MatMul_3500" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="133351645" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="135399658" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -41736,7 +41502,7 @@
 			</output>
 		</layer>
 		<layer id="2864" name="Convert_3685" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -41745,7 +41511,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41753,7 +41519,7 @@
 			</output>
 		</layer>
 		<layer id="2865" name="onnx::MatMul_3500/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -41761,32 +41527,32 @@
 			</output>
 		</layer>
 		<layer id="2866" name="Convert_3687" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2867" name="Subtract_3688" type="Subtract" version="opset1">
+		<layer id="2867" name="onnx::MatMul_3500/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41794,7 +41560,7 @@
 			</output>
 		</layer>
 		<layer id="2868" name="onnx::MatMul_3500/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="133875933" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="135923946" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -41803,57 +41569,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2869" name="onnx::MatMul_3500/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2870" name="onnx::MatMul_3500/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2869" name="onnx::MatMul_3500/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2871" name="Constant_3693" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2870" name="Constant_3692" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2872" name="Reshape_3694" type="Reshape" version="opset1">
+		<layer id="2871" name="Reshape_3693" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -41863,7 +41612,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2872" name="onnx::MatMul_3500/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -41874,7 +41638,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -41885,7 +41649,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -41898,14 +41662,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -41915,19 +41679,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -41937,20 +41701,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2877" name="onnx::Add_3505" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="133892317" size="4096" />
+			<data element_type="f32" shape="1024" offset="135940330" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3505">
 					<dim>1024</dim>
@@ -41958,7 +41722,7 @@
 			</output>
 		</layer>
 		<layer id="2878" name="onnx::Add_3502" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="133896413" size="16384" />
+			<data element_type="f32" shape="4096" offset="135944426" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3502">
 					<dim>4096</dim>
@@ -41966,13 +41730,13 @@
 			</output>
 		</layer>
 		<layer id="2879" name="Constant_2022" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2880" name="Constant_2021" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -42001,7 +41765,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -42011,13 +41775,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2884" name="blocks.19.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="133912797" size="4096" />
+			<data element_type="f32" shape="1024" offset="135960810" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.19.mlp_ln.weight">
 					<dim>1024</dim>
@@ -42029,7 +41793,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -42039,13 +41803,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2886" name="blocks.19.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="133916893" size="4096" />
+			<data element_type="f32" shape="1024" offset="135964906" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.19.mlp_ln.bias">
 					<dim>1024</dim>
@@ -42057,7 +41821,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -42067,7 +41831,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -42077,20 +41841,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2889" name="onnx::MatMul_3503" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="133920989" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="135969002" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -42100,7 +41864,7 @@
 			</output>
 		</layer>
 		<layer id="2890" name="Convert_3696" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -42109,7 +41873,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -42117,7 +41881,7 @@
 			</output>
 		</layer>
 		<layer id="2891" name="onnx::MatMul_3503/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -42125,32 +41889,32 @@
 			</output>
 		</layer>
 		<layer id="2892" name="Convert_3698" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2893" name="Subtract_3699" type="Subtract" version="opset1">
+		<layer id="2893" name="onnx::MatMul_3503/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -42158,7 +41922,7 @@
 			</output>
 		</layer>
 		<layer id="2894" name="onnx::MatMul_3503/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="136018141" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="138066154" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -42167,57 +41931,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2895" name="onnx::MatMul_3503/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2896" name="onnx::MatMul_3503/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2895" name="onnx::MatMul_3503/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2897" name="Constant_3704" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="2896" name="Constant_3703" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2898" name="Reshape_3705" type="Reshape" version="opset1">
+		<layer id="2897" name="Reshape_3704" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -42227,7 +41974,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2898" name="onnx::MatMul_3503/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -42238,7 +42000,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -42249,7 +42011,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -42262,14 +42024,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -42285,7 +42047,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -42293,7 +42055,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -42302,14 +42064,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.19/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -42325,7 +42087,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -42333,7 +42095,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -42343,19 +42105,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -42371,7 +42133,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -42379,13 +42141,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2909" name="onnx::MatMul_3506" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="136083677" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="138131690" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -42395,7 +42157,7 @@
 			</output>
 		</layer>
 		<layer id="2910" name="Convert_3707" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -42404,7 +42166,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -42412,7 +42174,7 @@
 			</output>
 		</layer>
 		<layer id="2911" name="onnx::MatMul_3506/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -42420,32 +42182,32 @@
 			</output>
 		</layer>
 		<layer id="2912" name="Convert_3709" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2913" name="Subtract_3710" type="Subtract" version="opset1">
+		<layer id="2913" name="onnx::MatMul_3506/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -42453,7 +42215,7 @@
 			</output>
 		</layer>
 		<layer id="2914" name="onnx::MatMul_3506/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="138180829" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="140228842" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -42462,57 +42224,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2915" name="onnx::MatMul_3506/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2916" name="onnx::MatMul_3506/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2915" name="onnx::MatMul_3506/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2917" name="Constant_3715" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="2916" name="Constant_3714" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2918" name="Reshape_3716" type="Reshape" version="opset1">
+		<layer id="2917" name="Reshape_3715" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -42522,7 +42267,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2918" name="onnx::MatMul_3506/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -42533,7 +42293,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -42544,7 +42304,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -42557,14 +42317,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -42574,19 +42334,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.19/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -42596,20 +42356,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2923" name="onnx::Add_3531" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="138246365" size="4096" />
+			<data element_type="f32" shape="1024" offset="140294378" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3531">
 					<dim>1024</dim>
@@ -42617,7 +42377,7 @@
 			</output>
 		</layer>
 		<layer id="2924" name="onnx::Add_3508" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="138250461" size="4096" />
+			<data element_type="f32" shape="1024" offset="140298474" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3508">
 					<dim>1024</dim>
@@ -42625,13 +42385,13 @@
 			</output>
 		</layer>
 		<layer id="2925" name="Constant_2044" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="2926" name="Constant_2043" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -42660,7 +42420,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -42670,13 +42430,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2930" name="blocks.20.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="138254557" size="4096" />
+			<data element_type="f32" shape="1024" offset="140302570" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.20.attn_ln.weight">
 					<dim>1024</dim>
@@ -42688,7 +42448,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -42698,13 +42458,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2932" name="blocks.20.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="138258653" size="4096" />
+			<data element_type="f32" shape="1024" offset="140306666" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.20.attn_ln.bias">
 					<dim>1024</dim>
@@ -42716,7 +42476,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -42726,7 +42486,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -42736,20 +42496,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2935" name="onnx::MatMul_3509" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="138262749" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="140310762" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -42759,7 +42519,7 @@
 			</output>
 		</layer>
 		<layer id="2936" name="Convert_3740" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -42768,7 +42528,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -42776,7 +42536,7 @@
 			</output>
 		</layer>
 		<layer id="2937" name="onnx::MatMul_3509/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -42784,32 +42544,32 @@
 			</output>
 		</layer>
 		<layer id="2938" name="Convert_3742" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2939" name="Subtract_3743" type="Subtract" version="opset1">
+		<layer id="2939" name="onnx::MatMul_3509/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -42817,7 +42577,7 @@
 			</output>
 		</layer>
 		<layer id="2940" name="onnx::MatMul_3509/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="138787037" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="140835050" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -42826,57 +42586,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2941" name="onnx::MatMul_3509/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2942" name="onnx::MatMul_3509/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2941" name="onnx::MatMul_3509/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2943" name="Constant_3748" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2942" name="Constant_3747" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2944" name="Reshape_3749" type="Reshape" version="opset1">
+		<layer id="2943" name="Reshape_3748" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -42886,7 +42629,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2944" name="onnx::MatMul_3509/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -42897,7 +42655,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -42908,7 +42666,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -42921,20 +42679,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2947" name="/blocks.20/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.20/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -42946,7 +42704,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -42956,14 +42714,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2949" name="Constant_2060" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -42974,7 +42732,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -42986,19 +42744,19 @@
 				<port id="2" precision="FP32" names="/blocks.20/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2951" name="/blocks.20/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.20/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="2952" name="/blocks.20/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.20/attn/Constant_1_output_0" />
 			</output>
@@ -43019,7 +42777,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -43028,13 +42786,13 @@
 				<port id="2" precision="FP32" names="/blocks.20/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2955" name="onnx::MatMul_3511" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="138803421" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="140851434" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -43044,7 +42802,7 @@
 			</output>
 		</layer>
 		<layer id="2956" name="Convert_3729" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -43053,7 +42811,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43061,7 +42819,7 @@
 			</output>
 		</layer>
 		<layer id="2957" name="onnx::MatMul_3511/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -43069,32 +42827,32 @@
 			</output>
 		</layer>
 		<layer id="2958" name="Convert_3731" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2959" name="Subtract_3732" type="Subtract" version="opset1">
+		<layer id="2959" name="onnx::MatMul_3511/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43102,7 +42860,7 @@
 			</output>
 		</layer>
 		<layer id="2960" name="onnx::MatMul_3511/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="139327709" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="141375722" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -43111,57 +42869,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2961" name="onnx::MatMul_3511/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2962" name="onnx::MatMul_3511/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2961" name="onnx::MatMul_3511/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2963" name="Constant_3737" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2962" name="Constant_3736" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2964" name="Reshape_3738" type="Reshape" version="opset1">
+		<layer id="2963" name="Reshape_3737" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43171,7 +42912,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2964" name="onnx::MatMul_3511/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -43182,7 +42938,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -43193,13 +42949,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2966" name="/blocks.20/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.20/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -43211,7 +42967,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -43221,14 +42977,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2968" name="Constant_2065" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -43239,7 +42995,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -43252,7 +43008,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -43263,7 +43019,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -43272,7 +43028,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -43282,22 +43038,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -43307,16 +43063,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -43326,16 +43082,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -43345,21 +43101,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2975" name="onnx::Add_3513" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="139344093" size="4096" />
+			<data element_type="f32" shape="1024" offset="141392106" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3513">
 					<dim>1024</dim>
@@ -43367,7 +43123,7 @@
 			</output>
 		</layer>
 		<layer id="2976" name="onnx::MatMul_3514" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="139348189" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="141396202" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -43377,7 +43133,7 @@
 			</output>
 		</layer>
 		<layer id="2977" name="Convert_3718" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -43386,7 +43142,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43394,7 +43150,7 @@
 			</output>
 		</layer>
 		<layer id="2978" name="onnx::MatMul_3514/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -43402,32 +43158,32 @@
 			</output>
 		</layer>
 		<layer id="2979" name="Convert_3720" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2980" name="Subtract_3721" type="Subtract" version="opset1">
+		<layer id="2980" name="onnx::MatMul_3514/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43435,7 +43191,7 @@
 			</output>
 		</layer>
 		<layer id="2981" name="onnx::MatMul_3514/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="139872477" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="141920490" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -43444,57 +43200,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2982" name="onnx::MatMul_3514/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="2983" name="onnx::MatMul_3514/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="2982" name="onnx::MatMul_3514/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2984" name="Constant_3726" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="2983" name="Constant_3725" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="2985" name="Reshape_3727" type="Reshape" version="opset1">
+		<layer id="2984" name="Reshape_3726" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43504,7 +43243,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="2985" name="onnx::MatMul_3514/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -43515,7 +43269,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -43526,7 +43280,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -43539,20 +43293,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2988" name="/blocks.20/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.20/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -43564,7 +43318,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -43574,14 +43328,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2990" name="Constant_2070" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -43592,7 +43346,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -43604,7 +43358,7 @@
 				<port id="2" precision="FP32" names="/blocks.20/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -43615,13 +43369,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -43629,13 +43383,13 @@
 				<port id="2" precision="FP32" names="/blocks.20/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="2993" name="Constant_2077" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -43647,7 +43401,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -43657,7 +43411,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -43668,7 +43422,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -43680,7 +43434,7 @@
 			</output>
 		</layer>
 		<layer id="2996" name="/blocks.20/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.20/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -43733,7 +43487,7 @@
 			</output>
 		</layer>
 		<layer id="3001" name="/blocks.20/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.20/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -43765,7 +43519,7 @@
 			</output>
 		</layer>
 		<layer id="3003" name="/blocks.20/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.20/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -43793,7 +43547,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -43804,13 +43558,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3006" name="onnx::MatMul_3532" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="139888861" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="141936874" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -43820,7 +43574,7 @@
 			</output>
 		</layer>
 		<layer id="3007" name="Convert_3751" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -43829,7 +43583,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43837,7 +43591,7 @@
 			</output>
 		</layer>
 		<layer id="3008" name="onnx::MatMul_3532/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -43845,32 +43599,32 @@
 			</output>
 		</layer>
 		<layer id="3009" name="Convert_3753" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3010" name="Subtract_3754" type="Subtract" version="opset1">
+		<layer id="3010" name="onnx::MatMul_3532/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43878,7 +43632,7 @@
 			</output>
 		</layer>
 		<layer id="3011" name="onnx::MatMul_3532/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="140413149" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="142461162" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -43887,57 +43641,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3012" name="onnx::MatMul_3532/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3013" name="onnx::MatMul_3532/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3012" name="onnx::MatMul_3532/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3014" name="Constant_3759" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3013" name="Constant_3758" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3015" name="Reshape_3760" type="Reshape" version="opset1">
+		<layer id="3014" name="Reshape_3759" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -43947,7 +43684,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3015" name="onnx::MatMul_3532/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -43958,7 +43710,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -43969,7 +43721,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -43982,14 +43734,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -43999,19 +43751,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -44021,20 +43773,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3020" name="onnx::Add_3537" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="140429533" size="4096" />
+			<data element_type="f32" shape="1024" offset="142477546" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3537">
 					<dim>1024</dim>
@@ -44042,7 +43794,7 @@
 			</output>
 		</layer>
 		<layer id="3021" name="onnx::Add_3534" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="140433629" size="16384" />
+			<data element_type="f32" shape="4096" offset="142481642" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3534">
 					<dim>4096</dim>
@@ -44050,13 +43802,13 @@
 			</output>
 		</layer>
 		<layer id="3022" name="Constant_2103" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3023" name="Constant_2102" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -44085,7 +43837,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -44095,13 +43847,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3027" name="blocks.20.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="140450013" size="4096" />
+			<data element_type="f32" shape="1024" offset="142498026" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.20.mlp_ln.weight">
 					<dim>1024</dim>
@@ -44113,7 +43865,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -44123,13 +43875,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3029" name="blocks.20.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="140454109" size="4096" />
+			<data element_type="f32" shape="1024" offset="142502122" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.20.mlp_ln.bias">
 					<dim>1024</dim>
@@ -44141,7 +43893,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -44151,7 +43903,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -44161,20 +43913,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3032" name="onnx::MatMul_3535" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="140458205" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="142506218" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -44184,7 +43936,7 @@
 			</output>
 		</layer>
 		<layer id="3033" name="Convert_3762" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -44193,7 +43945,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -44201,7 +43953,7 @@
 			</output>
 		</layer>
 		<layer id="3034" name="onnx::MatMul_3535/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -44209,32 +43961,32 @@
 			</output>
 		</layer>
 		<layer id="3035" name="Convert_3764" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3036" name="Subtract_3765" type="Subtract" version="opset1">
+		<layer id="3036" name="onnx::MatMul_3535/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -44242,7 +43994,7 @@
 			</output>
 		</layer>
 		<layer id="3037" name="onnx::MatMul_3535/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="142555357" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="144603370" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -44251,57 +44003,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3038" name="onnx::MatMul_3535/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3039" name="onnx::MatMul_3535/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3038" name="onnx::MatMul_3535/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3040" name="Constant_3770" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="3039" name="Constant_3769" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3041" name="Reshape_3771" type="Reshape" version="opset1">
+		<layer id="3040" name="Reshape_3770" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -44311,7 +44046,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3041" name="onnx::MatMul_3535/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -44322,7 +44072,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -44333,7 +44083,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -44346,14 +44096,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -44369,7 +44119,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -44377,7 +44127,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -44386,14 +44136,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.20/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -44409,7 +44159,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -44417,7 +44167,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -44427,19 +44177,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -44455,7 +44205,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -44463,13 +44213,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3052" name="onnx::MatMul_3538" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="142620893" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="144668906" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -44479,7 +44229,7 @@
 			</output>
 		</layer>
 		<layer id="3053" name="Convert_3773" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -44488,7 +44238,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -44496,7 +44246,7 @@
 			</output>
 		</layer>
 		<layer id="3054" name="onnx::MatMul_3538/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -44504,32 +44254,32 @@
 			</output>
 		</layer>
 		<layer id="3055" name="Convert_3775" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3056" name="Subtract_3776" type="Subtract" version="opset1">
+		<layer id="3056" name="onnx::MatMul_3538/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -44537,7 +44287,7 @@
 			</output>
 		</layer>
 		<layer id="3057" name="onnx::MatMul_3538/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="144718045" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="146766058" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -44546,57 +44296,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3058" name="onnx::MatMul_3538/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3059" name="onnx::MatMul_3538/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3058" name="onnx::MatMul_3538/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3060" name="Constant_3781" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="3059" name="Constant_3780" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3061" name="Reshape_3782" type="Reshape" version="opset1">
+		<layer id="3060" name="Reshape_3781" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -44606,7 +44339,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3061" name="onnx::MatMul_3538/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -44617,7 +44365,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -44628,7 +44376,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -44641,14 +44389,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -44658,19 +44406,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.20/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -44680,20 +44428,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3066" name="onnx::Add_3563" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="144783581" size="4096" />
+			<data element_type="f32" shape="1024" offset="146831594" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3563">
 					<dim>1024</dim>
@@ -44701,7 +44449,7 @@
 			</output>
 		</layer>
 		<layer id="3067" name="onnx::Add_3540" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="144787677" size="4096" />
+			<data element_type="f32" shape="1024" offset="146835690" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3540">
 					<dim>1024</dim>
@@ -44709,13 +44457,13 @@
 			</output>
 		</layer>
 		<layer id="3068" name="Constant_2125" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3069" name="Constant_2124" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -44744,7 +44492,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -44754,13 +44502,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3073" name="blocks.21.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="144791773" size="4096" />
+			<data element_type="f32" shape="1024" offset="146839786" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.21.attn_ln.weight">
 					<dim>1024</dim>
@@ -44772,7 +44520,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -44782,13 +44530,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3075" name="blocks.21.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="144795869" size="4096" />
+			<data element_type="f32" shape="1024" offset="146843882" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.21.attn_ln.bias">
 					<dim>1024</dim>
@@ -44800,7 +44548,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -44810,7 +44558,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -44820,20 +44568,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3078" name="onnx::MatMul_3541" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="144799965" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="146847978" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -44843,7 +44591,7 @@
 			</output>
 		</layer>
 		<layer id="3079" name="Convert_3806" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -44852,7 +44600,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -44860,7 +44608,7 @@
 			</output>
 		</layer>
 		<layer id="3080" name="onnx::MatMul_3541/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -44868,32 +44616,32 @@
 			</output>
 		</layer>
 		<layer id="3081" name="Convert_3808" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3082" name="Subtract_3809" type="Subtract" version="opset1">
+		<layer id="3082" name="onnx::MatMul_3541/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -44901,7 +44649,7 @@
 			</output>
 		</layer>
 		<layer id="3083" name="onnx::MatMul_3541/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="145324253" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="147372266" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -44910,57 +44658,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3084" name="onnx::MatMul_3541/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3085" name="onnx::MatMul_3541/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3084" name="onnx::MatMul_3541/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3086" name="Constant_3814" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3085" name="Constant_3813" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3087" name="Reshape_3815" type="Reshape" version="opset1">
+		<layer id="3086" name="Reshape_3814" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -44970,7 +44701,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3087" name="onnx::MatMul_3541/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -44981,7 +44727,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -44992,7 +44738,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -45005,20 +44751,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3090" name="/blocks.21/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.21/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -45030,7 +44776,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -45040,14 +44786,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3092" name="Constant_2141" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -45058,7 +44804,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -45070,19 +44816,19 @@
 				<port id="2" precision="FP32" names="/blocks.21/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3094" name="/blocks.21/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.21/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="3095" name="/blocks.21/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.21/attn/Constant_1_output_0" />
 			</output>
@@ -45103,7 +44849,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -45112,13 +44858,13 @@
 				<port id="2" precision="FP32" names="/blocks.21/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3098" name="onnx::MatMul_3543" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="145340637" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="147388650" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -45128,7 +44874,7 @@
 			</output>
 		</layer>
 		<layer id="3099" name="Convert_3795" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -45137,7 +44883,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45145,7 +44891,7 @@
 			</output>
 		</layer>
 		<layer id="3100" name="onnx::MatMul_3543/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -45153,32 +44899,32 @@
 			</output>
 		</layer>
 		<layer id="3101" name="Convert_3797" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3102" name="Subtract_3798" type="Subtract" version="opset1">
+		<layer id="3102" name="onnx::MatMul_3543/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45186,7 +44932,7 @@
 			</output>
 		</layer>
 		<layer id="3103" name="onnx::MatMul_3543/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="145864925" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="147912938" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -45195,57 +44941,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3104" name="onnx::MatMul_3543/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3105" name="onnx::MatMul_3543/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3104" name="onnx::MatMul_3543/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3106" name="Constant_3803" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3105" name="Constant_3802" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3107" name="Reshape_3804" type="Reshape" version="opset1">
+		<layer id="3106" name="Reshape_3803" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45255,7 +44984,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3107" name="onnx::MatMul_3543/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -45266,7 +45010,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -45277,13 +45021,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3109" name="/blocks.21/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.21/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -45295,7 +45039,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -45305,14 +45049,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3111" name="Constant_2146" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -45323,7 +45067,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -45336,7 +45080,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -45347,7 +45091,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -45356,7 +45100,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -45366,22 +45110,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -45391,16 +45135,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -45410,16 +45154,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -45429,21 +45173,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3118" name="onnx::Add_3545" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="145881309" size="4096" />
+			<data element_type="f32" shape="1024" offset="147929322" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3545">
 					<dim>1024</dim>
@@ -45451,7 +45195,7 @@
 			</output>
 		</layer>
 		<layer id="3119" name="onnx::MatMul_3546" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="145885405" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="147933418" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -45461,7 +45205,7 @@
 			</output>
 		</layer>
 		<layer id="3120" name="Convert_3784" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -45470,7 +45214,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45478,7 +45222,7 @@
 			</output>
 		</layer>
 		<layer id="3121" name="onnx::MatMul_3546/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -45486,32 +45230,32 @@
 			</output>
 		</layer>
 		<layer id="3122" name="Convert_3786" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3123" name="Subtract_3787" type="Subtract" version="opset1">
+		<layer id="3123" name="onnx::MatMul_3546/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45519,7 +45263,7 @@
 			</output>
 		</layer>
 		<layer id="3124" name="onnx::MatMul_3546/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="146409693" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="148457706" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -45528,57 +45272,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3125" name="onnx::MatMul_3546/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3126" name="onnx::MatMul_3546/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3125" name="onnx::MatMul_3546/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3127" name="Constant_3792" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3126" name="Constant_3791" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3128" name="Reshape_3793" type="Reshape" version="opset1">
+		<layer id="3127" name="Reshape_3792" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45588,7 +45315,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3128" name="onnx::MatMul_3546/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -45599,7 +45341,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -45610,7 +45352,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -45623,20 +45365,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3131" name="/blocks.21/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.21/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -45648,7 +45390,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -45658,14 +45400,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3133" name="Constant_2151" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -45676,7 +45418,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -45688,7 +45430,7 @@
 				<port id="2" precision="FP32" names="/blocks.21/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -45699,13 +45441,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -45713,13 +45455,13 @@
 				<port id="2" precision="FP32" names="/blocks.21/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3136" name="Constant_2158" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -45731,7 +45473,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -45741,7 +45483,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -45752,7 +45494,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -45764,7 +45506,7 @@
 			</output>
 		</layer>
 		<layer id="3139" name="/blocks.21/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.21/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -45817,7 +45559,7 @@
 			</output>
 		</layer>
 		<layer id="3144" name="/blocks.21/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.21/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -45849,7 +45591,7 @@
 			</output>
 		</layer>
 		<layer id="3146" name="/blocks.21/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.21/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -45877,7 +45619,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -45888,13 +45630,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3149" name="onnx::MatMul_3564" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="146426077" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="148474090" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -45904,7 +45646,7 @@
 			</output>
 		</layer>
 		<layer id="3150" name="Convert_3817" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -45913,7 +45655,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45921,7 +45663,7 @@
 			</output>
 		</layer>
 		<layer id="3151" name="onnx::MatMul_3564/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="33470692" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -45929,32 +45671,32 @@
 			</output>
 		</layer>
 		<layer id="3152" name="Convert_3819" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3153" name="Subtract_3820" type="Subtract" version="opset1">
+		<layer id="3153" name="onnx::MatMul_3564/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -45962,7 +45704,7 @@
 			</output>
 		</layer>
 		<layer id="3154" name="onnx::MatMul_3564/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="146950365" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="148998378" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -45971,57 +45713,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3155" name="onnx::MatMul_3564/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3156" name="onnx::MatMul_3564/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3155" name="onnx::MatMul_3564/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3157" name="Constant_3825" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3156" name="Constant_3824" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3158" name="Reshape_3826" type="Reshape" version="opset1">
+		<layer id="3157" name="Reshape_3825" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -46031,7 +45756,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3158" name="onnx::MatMul_3564/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -46042,7 +45782,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -46053,7 +45793,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46066,14 +45806,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46083,19 +45823,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46105,20 +45845,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3163" name="onnx::Add_3569" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="146966749" size="4096" />
+			<data element_type="f32" shape="1024" offset="149014762" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3569">
 					<dim>1024</dim>
@@ -46126,7 +45866,7 @@
 			</output>
 		</layer>
 		<layer id="3164" name="onnx::Add_3566" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="146970845" size="16384" />
+			<data element_type="f32" shape="4096" offset="149018858" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3566">
 					<dim>4096</dim>
@@ -46134,13 +45874,13 @@
 			</output>
 		</layer>
 		<layer id="3165" name="Constant_2184" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3166" name="Constant_2183" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -46169,7 +45909,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -46179,13 +45919,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3170" name="blocks.21.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="146987229" size="4096" />
+			<data element_type="f32" shape="1024" offset="149035242" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.21.mlp_ln.weight">
 					<dim>1024</dim>
@@ -46197,7 +45937,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -46207,13 +45947,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3172" name="blocks.21.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="146991325" size="4096" />
+			<data element_type="f32" shape="1024" offset="149039338" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.21.mlp_ln.bias">
 					<dim>1024</dim>
@@ -46225,7 +45965,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -46235,7 +45975,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46245,20 +45985,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3175" name="onnx::MatMul_3567" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="146995421" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="149043434" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -46268,7 +46008,7 @@
 			</output>
 		</layer>
 		<layer id="3176" name="Convert_3828" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -46277,7 +46017,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -46285,7 +46025,7 @@
 			</output>
 		</layer>
 		<layer id="3177" name="onnx::MatMul_3567/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -46293,32 +46033,32 @@
 			</output>
 		</layer>
 		<layer id="3178" name="Convert_3830" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3179" name="Subtract_3831" type="Subtract" version="opset1">
+		<layer id="3179" name="onnx::MatMul_3567/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -46326,7 +46066,7 @@
 			</output>
 		</layer>
 		<layer id="3180" name="onnx::MatMul_3567/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="149092573" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="151140586" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -46335,57 +46075,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3181" name="onnx::MatMul_3567/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3182" name="onnx::MatMul_3567/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3181" name="onnx::MatMul_3567/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3183" name="Constant_3836" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="3182" name="Constant_3835" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3184" name="Reshape_3837" type="Reshape" version="opset1">
+		<layer id="3183" name="Reshape_3836" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -46395,7 +46118,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3184" name="onnx::MatMul_3567/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -46406,7 +46144,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -46417,7 +46155,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -46430,14 +46168,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -46453,7 +46191,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -46461,7 +46199,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -46470,14 +46208,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.21/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -46493,7 +46231,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -46501,7 +46239,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -46511,19 +46249,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -46539,7 +46277,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -46547,13 +46285,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3195" name="onnx::MatMul_3570" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="149158109" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="151206122" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -46563,7 +46301,7 @@
 			</output>
 		</layer>
 		<layer id="3196" name="Convert_3839" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -46572,7 +46310,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -46580,7 +46318,7 @@
 			</output>
 		</layer>
 		<layer id="3197" name="onnx::MatMul_3570/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -46588,32 +46326,32 @@
 			</output>
 		</layer>
 		<layer id="3198" name="Convert_3841" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3199" name="Subtract_3842" type="Subtract" version="opset1">
+		<layer id="3199" name="onnx::MatMul_3570/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -46621,7 +46359,7 @@
 			</output>
 		</layer>
 		<layer id="3200" name="onnx::MatMul_3570/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="151255261" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="153303274" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -46630,57 +46368,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3201" name="onnx::MatMul_3570/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3202" name="onnx::MatMul_3570/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3201" name="onnx::MatMul_3570/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3203" name="Constant_3847" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="3202" name="Constant_3846" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3204" name="Reshape_3848" type="Reshape" version="opset1">
+		<layer id="3203" name="Reshape_3847" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -46690,7 +46411,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3204" name="onnx::MatMul_3570/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -46701,7 +46437,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -46712,7 +46448,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46725,14 +46461,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46742,19 +46478,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.21/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46764,20 +46500,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3209" name="onnx::Add_3595" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="151320797" size="4096" />
+			<data element_type="f32" shape="1024" offset="153368810" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3595">
 					<dim>1024</dim>
@@ -46785,7 +46521,7 @@
 			</output>
 		</layer>
 		<layer id="3210" name="onnx::Add_3572" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="151324893" size="4096" />
+			<data element_type="f32" shape="1024" offset="153372906" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3572">
 					<dim>1024</dim>
@@ -46793,13 +46529,13 @@
 			</output>
 		</layer>
 		<layer id="3211" name="Constant_2206" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3212" name="Constant_2205" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -46828,7 +46564,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -46838,13 +46574,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3216" name="blocks.22.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="151328989" size="4096" />
+			<data element_type="f32" shape="1024" offset="153377002" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.22.attn_ln.weight">
 					<dim>1024</dim>
@@ -46856,7 +46592,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -46866,13 +46602,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3218" name="blocks.22.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="151333085" size="4096" />
+			<data element_type="f32" shape="1024" offset="153381098" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.22.attn_ln.bias">
 					<dim>1024</dim>
@@ -46884,7 +46620,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -46894,7 +46630,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -46904,20 +46640,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3221" name="onnx::MatMul_3573" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="151337181" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="153385194" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -46927,7 +46663,7 @@
 			</output>
 		</layer>
 		<layer id="3222" name="Convert_3872" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -46936,7 +46672,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -46944,7 +46680,7 @@
 			</output>
 		</layer>
 		<layer id="3223" name="onnx::MatMul_3573/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -46952,32 +46688,32 @@
 			</output>
 		</layer>
 		<layer id="3224" name="Convert_3874" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3225" name="Subtract_3875" type="Subtract" version="opset1">
+		<layer id="3225" name="onnx::MatMul_3573/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -46985,7 +46721,7 @@
 			</output>
 		</layer>
 		<layer id="3226" name="onnx::MatMul_3573/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="151861469" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="153909482" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -46994,57 +46730,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3227" name="onnx::MatMul_3573/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3228" name="onnx::MatMul_3573/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3227" name="onnx::MatMul_3573/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3229" name="Constant_3880" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3228" name="Constant_3879" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3230" name="Reshape_3881" type="Reshape" version="opset1">
+		<layer id="3229" name="Reshape_3880" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -47054,7 +46773,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3230" name="onnx::MatMul_3573/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -47065,7 +46799,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -47076,7 +46810,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -47089,20 +46823,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3233" name="/blocks.22/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.22/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -47114,7 +46848,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -47124,14 +46858,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3235" name="Constant_2222" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -47142,7 +46876,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -47154,19 +46888,19 @@
 				<port id="2" precision="FP32" names="/blocks.22/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3237" name="/blocks.22/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.22/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="3238" name="/blocks.22/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.22/attn/Constant_1_output_0" />
 			</output>
@@ -47187,7 +46921,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -47196,13 +46930,13 @@
 				<port id="2" precision="FP32" names="/blocks.22/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3241" name="onnx::MatMul_3575" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="151877853" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="153925866" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -47212,7 +46946,7 @@
 			</output>
 		</layer>
 		<layer id="3242" name="Convert_3861" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -47221,7 +46955,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -47229,7 +46963,7 @@
 			</output>
 		</layer>
 		<layer id="3243" name="onnx::MatMul_3575/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -47237,32 +46971,32 @@
 			</output>
 		</layer>
 		<layer id="3244" name="Convert_3863" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3245" name="Subtract_3864" type="Subtract" version="opset1">
+		<layer id="3245" name="onnx::MatMul_3575/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -47270,7 +47004,7 @@
 			</output>
 		</layer>
 		<layer id="3246" name="onnx::MatMul_3575/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="152402141" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="154450154" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -47279,57 +47013,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3247" name="onnx::MatMul_3575/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3248" name="onnx::MatMul_3575/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3247" name="onnx::MatMul_3575/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3249" name="Constant_3869" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3248" name="Constant_3868" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3250" name="Reshape_3870" type="Reshape" version="opset1">
+		<layer id="3249" name="Reshape_3869" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -47339,7 +47056,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3250" name="onnx::MatMul_3575/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -47350,7 +47082,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -47361,13 +47093,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3252" name="/blocks.22/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.22/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -47379,7 +47111,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -47389,14 +47121,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3254" name="Constant_2227" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -47407,7 +47139,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -47420,7 +47152,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -47431,7 +47163,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -47440,7 +47172,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -47450,22 +47182,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -47475,16 +47207,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -47494,16 +47226,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -47513,21 +47245,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3261" name="onnx::Add_3577" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="152418525" size="4096" />
+			<data element_type="f32" shape="1024" offset="154466538" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3577">
 					<dim>1024</dim>
@@ -47535,7 +47267,7 @@
 			</output>
 		</layer>
 		<layer id="3262" name="onnx::MatMul_3578" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="152422621" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="154470634" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -47545,7 +47277,7 @@
 			</output>
 		</layer>
 		<layer id="3263" name="Convert_3850" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -47554,7 +47286,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -47562,7 +47294,7 @@
 			</output>
 		</layer>
 		<layer id="3264" name="onnx::MatMul_3578/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -47570,32 +47302,32 @@
 			</output>
 		</layer>
 		<layer id="3265" name="Convert_3852" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3266" name="Subtract_3853" type="Subtract" version="opset1">
+		<layer id="3266" name="onnx::MatMul_3578/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -47603,7 +47335,7 @@
 			</output>
 		</layer>
 		<layer id="3267" name="onnx::MatMul_3578/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="152946909" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="154994922" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -47612,57 +47344,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3268" name="onnx::MatMul_3578/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3269" name="onnx::MatMul_3578/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3268" name="onnx::MatMul_3578/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3270" name="Constant_3858" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3269" name="Constant_3857" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3271" name="Reshape_3859" type="Reshape" version="opset1">
+		<layer id="3270" name="Reshape_3858" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -47672,7 +47387,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3271" name="onnx::MatMul_3578/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -47683,7 +47413,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -47694,7 +47424,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -47707,20 +47437,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3274" name="/blocks.22/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.22/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -47732,7 +47462,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -47742,14 +47472,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3276" name="Constant_2232" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -47760,7 +47490,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -47772,7 +47502,7 @@
 				<port id="2" precision="FP32" names="/blocks.22/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -47783,13 +47513,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -47797,13 +47527,13 @@
 				<port id="2" precision="FP32" names="/blocks.22/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3279" name="Constant_2239" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -47815,7 +47545,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -47825,7 +47555,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -47836,7 +47566,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -47848,7 +47578,7 @@
 			</output>
 		</layer>
 		<layer id="3282" name="/blocks.22/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.22/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -47901,7 +47631,7 @@
 			</output>
 		</layer>
 		<layer id="3287" name="/blocks.22/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.22/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -47933,7 +47663,7 @@
 			</output>
 		</layer>
 		<layer id="3289" name="/blocks.22/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.22/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -47961,7 +47691,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -47972,13 +47702,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3292" name="onnx::MatMul_3596" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="152963293" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="155011306" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -47988,7 +47718,7 @@
 			</output>
 		</layer>
 		<layer id="3293" name="Convert_3883" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -47997,7 +47727,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -48005,7 +47735,7 @@
 			</output>
 		</layer>
 		<layer id="3294" name="onnx::MatMul_3596/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="74856679" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -48013,32 +47743,32 @@
 			</output>
 		</layer>
 		<layer id="3295" name="Convert_3885" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3296" name="Subtract_3886" type="Subtract" version="opset1">
+		<layer id="3296" name="onnx::MatMul_3596/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -48046,7 +47776,7 @@
 			</output>
 		</layer>
 		<layer id="3297" name="onnx::MatMul_3596/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="153487581" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="155535594" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -48055,57 +47785,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3298" name="onnx::MatMul_3596/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3299" name="onnx::MatMul_3596/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3298" name="onnx::MatMul_3596/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3300" name="Constant_3891" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3299" name="Constant_3890" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3301" name="Reshape_3892" type="Reshape" version="opset1">
+		<layer id="3300" name="Reshape_3891" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -48115,7 +47828,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3301" name="onnx::MatMul_3596/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -48126,7 +47854,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -48137,7 +47865,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48150,14 +47878,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48167,19 +47895,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48189,20 +47917,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3306" name="onnx::Add_3601" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="153503965" size="4096" />
+			<data element_type="f32" shape="1024" offset="155551978" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3601">
 					<dim>1024</dim>
@@ -48210,7 +47938,7 @@
 			</output>
 		</layer>
 		<layer id="3307" name="onnx::Add_3598" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="153508061" size="16384" />
+			<data element_type="f32" shape="4096" offset="155556074" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3598">
 					<dim>4096</dim>
@@ -48218,13 +47946,13 @@
 			</output>
 		</layer>
 		<layer id="3308" name="Constant_2265" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3309" name="Constant_2264" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -48253,7 +47981,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -48263,13 +47991,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3313" name="blocks.22.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="153524445" size="4096" />
+			<data element_type="f32" shape="1024" offset="155572458" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.22.mlp_ln.weight">
 					<dim>1024</dim>
@@ -48281,7 +48009,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -48291,13 +48019,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3315" name="blocks.22.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="153528541" size="4096" />
+			<data element_type="f32" shape="1024" offset="155576554" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.22.mlp_ln.bias">
 					<dim>1024</dim>
@@ -48309,7 +48037,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -48319,7 +48047,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48329,20 +48057,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3318" name="onnx::MatMul_3599" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="153532637" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="155580650" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -48352,7 +48080,7 @@
 			</output>
 		</layer>
 		<layer id="3319" name="Convert_3894" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -48361,7 +48089,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -48369,7 +48097,7 @@
 			</output>
 		</layer>
 		<layer id="3320" name="onnx::MatMul_3599/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="17168607" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -48377,32 +48105,32 @@
 			</output>
 		</layer>
 		<layer id="3321" name="Convert_3896" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3322" name="Subtract_3897" type="Subtract" version="opset1">
+		<layer id="3322" name="onnx::MatMul_3599/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -48410,7 +48138,7 @@
 			</output>
 		</layer>
 		<layer id="3323" name="onnx::MatMul_3599/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="155629789" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="157677802" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -48419,57 +48147,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3324" name="onnx::MatMul_3599/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3325" name="onnx::MatMul_3599/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3324" name="onnx::MatMul_3599/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3326" name="Constant_3902" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="3325" name="Constant_3901" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3327" name="Reshape_3903" type="Reshape" version="opset1">
+		<layer id="3326" name="Reshape_3902" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -48479,7 +48190,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3327" name="onnx::MatMul_3599/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -48490,7 +48216,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -48501,7 +48227,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -48514,14 +48240,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -48537,7 +48263,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -48545,7 +48271,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -48554,14 +48280,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.22/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -48577,7 +48303,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -48585,7 +48311,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -48595,19 +48321,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -48623,7 +48349,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -48631,13 +48357,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3338" name="onnx::MatMul_3602" type="Const" version="opset1">
-			<data element_type="u4" shape="32, 128, 1024" offset="155695325" size="2097152" />
+			<data element_type="u4" shape="32, 128, 1024" offset="157743338" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -48647,7 +48373,7 @@
 			</output>
 		</layer>
 		<layer id="3339" name="Convert_3905" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>32</dim>
@@ -48656,7 +48382,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -48664,7 +48390,7 @@
 			</output>
 		</layer>
 		<layer id="3340" name="onnx::MatMul_3602/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -48672,32 +48398,32 @@
 			</output>
 		</layer>
 		<layer id="3341" name="Convert_3907" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3342" name="Subtract_3908" type="Subtract" version="opset1">
+		<layer id="3342" name="onnx::MatMul_3602/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -48705,7 +48431,7 @@
 			</output>
 		</layer>
 		<layer id="3343" name="onnx::MatMul_3602/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="32, 1, 1024" offset="157792477" size="65536" />
+			<data element_type="f16" shape="32, 1, 1024" offset="159840490" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>32</dim>
@@ -48714,57 +48440,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3344" name="onnx::MatMul_3602/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>32</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3345" name="onnx::MatMul_3602/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3344" name="onnx::MatMul_3602/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>32</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3346" name="Constant_3913" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="14039245" size="16" />
+		<layer id="3345" name="Constant_3912" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="16087247" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3347" name="Reshape_3914" type="Reshape" version="opset1">
+		<layer id="3346" name="Reshape_3913" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>32</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -48774,7 +48483,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3347" name="onnx::MatMul_3602/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -48785,7 +48509,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -48796,7 +48520,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48809,14 +48533,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48826,19 +48550,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.22/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48848,20 +48572,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/attn_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3352" name="onnx::Add_3627" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="157858013" size="4096" />
+			<data element_type="f32" shape="1024" offset="159906026" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3627">
 					<dim>1024</dim>
@@ -48869,7 +48593,7 @@
 			</output>
 		</layer>
 		<layer id="3353" name="onnx::Add_3604" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="157862109" size="4096" />
+			<data element_type="f32" shape="1024" offset="159910122" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3604">
 					<dim>1024</dim>
@@ -48877,13 +48601,13 @@
 			</output>
 		</layer>
 		<layer id="3354" name="Constant_2287" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3355" name="Constant_2286" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -48912,7 +48636,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -48922,13 +48646,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3359" name="blocks.23.attn_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="157866205" size="4096" />
+			<data element_type="f32" shape="1024" offset="159914218" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.23.attn_ln.weight">
 					<dim>1024</dim>
@@ -48940,7 +48664,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -48950,13 +48674,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3361" name="blocks.23.attn_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="157870301" size="4096" />
+			<data element_type="f32" shape="1024" offset="159918314" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.23.attn_ln.bias">
 					<dim>1024</dim>
@@ -48968,7 +48692,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -48978,7 +48702,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -48988,20 +48712,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/attn_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3364" name="onnx::MatMul_3605" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="157874397" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="159922410" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -49011,7 +48735,7 @@
 			</output>
 		</layer>
 		<layer id="3365" name="Convert_3938" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -49020,7 +48744,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49028,7 +48752,7 @@
 			</output>
 		</layer>
 		<layer id="3366" name="onnx::MatMul_3605/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="74856679" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -49036,32 +48760,32 @@
 			</output>
 		</layer>
 		<layer id="3367" name="Convert_3940" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3368" name="Subtract_3941" type="Subtract" version="opset1">
+		<layer id="3368" name="onnx::MatMul_3605/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49069,7 +48793,7 @@
 			</output>
 		</layer>
 		<layer id="3369" name="onnx::MatMul_3605/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="158398685" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="160446698" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -49078,57 +48802,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3370" name="onnx::MatMul_3605/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3371" name="onnx::MatMul_3605/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3370" name="onnx::MatMul_3605/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3372" name="Constant_3946" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3371" name="Constant_3945" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3373" name="Reshape_3947" type="Reshape" version="opset1">
+		<layer id="3372" name="Reshape_3946" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49138,7 +48845,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3373" name="onnx::MatMul_3605/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -49149,7 +48871,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -49160,7 +48882,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/query/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -49173,20 +48895,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/query/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3376" name="/blocks.23/attn/Constant_2" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.23/attn/Constant_2_output_0">
 					<dim>4</dim>
@@ -49198,7 +48920,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -49208,14 +48930,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/Reshape_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3378" name="Constant_2303" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -49226,7 +48948,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -49238,19 +48960,19 @@
 				<port id="2" precision="FP32" names="/blocks.23/attn/Transpose_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3380" name="/blocks.23/attn/Constant" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059029" size="4" />
+			<data element_type="f32" shape="" offset="10107029" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.23/attn/Constant_output_0" />
 			</output>
 		</layer>
 		<layer id="3381" name="/blocks.23/attn/Constant_1" type="Const" version="opset1">
-			<data element_type="f32" shape="" offset="8059033" size="4" />
+			<data element_type="f32" shape="" offset="10107033" size="4" />
 			<output>
 				<port id="0" precision="FP32" names="/blocks.23/attn/Constant_1_output_0" />
 			</output>
@@ -49271,7 +48993,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -49280,13 +49002,13 @@
 				<port id="2" precision="FP32" names="/blocks.23/attn/Mul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3384" name="onnx::MatMul_3607" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="158415069" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="160463082" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -49296,7 +49018,7 @@
 			</output>
 		</layer>
 		<layer id="3385" name="Convert_3927" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -49305,7 +49027,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49313,7 +49035,7 @@
 			</output>
 		</layer>
 		<layer id="3386" name="onnx::MatMul_3607/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="24791265" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -49321,32 +49043,32 @@
 			</output>
 		</layer>
 		<layer id="3387" name="Convert_3929" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3388" name="Subtract_3930" type="Subtract" version="opset1">
+		<layer id="3388" name="onnx::MatMul_3607/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49354,7 +49076,7 @@
 			</output>
 		</layer>
 		<layer id="3389" name="onnx::MatMul_3607/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="158939357" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="160987370" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -49363,57 +49085,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3390" name="onnx::MatMul_3607/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3391" name="onnx::MatMul_3607/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3390" name="onnx::MatMul_3607/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3392" name="Constant_3935" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3391" name="Constant_3934" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3393" name="Reshape_3936" type="Reshape" version="opset1">
+		<layer id="3392" name="Reshape_3935" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49423,7 +49128,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3393" name="onnx::MatMul_3607/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -49434,7 +49154,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -49445,13 +49165,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/key/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3395" name="/blocks.23/attn/Constant_3" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.23/attn/Constant_3_output_0">
 					<dim>4</dim>
@@ -49463,7 +49183,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -49473,14 +49193,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/Reshape_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3397" name="Constant_2308" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8599709" size="32" />
+			<data element_type="i64" shape="4" offset="10647710" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -49491,7 +49211,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -49504,7 +49224,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -49515,7 +49235,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32" />
 			</input>
@@ -49524,7 +49244,7 @@
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -49534,22 +49254,22 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
 					<dim>64</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/MatMul_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -49559,16 +49279,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/attn/Cast_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -49578,16 +49298,16 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/attn/Softmax_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
@@ -49597,21 +49317,21 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/attn/Cast_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3404" name="onnx::Add_3609" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="158955741" size="4096" />
+			<data element_type="f32" shape="1024" offset="161003754" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3609">
 					<dim>1024</dim>
@@ -49619,7 +49339,7 @@
 			</output>
 		</layer>
 		<layer id="3405" name="onnx::MatMul_3610" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="158959837" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="161007850" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -49629,7 +49349,7 @@
 			</output>
 		</layer>
 		<layer id="3406" name="Convert_3916" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -49638,7 +49358,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49646,7 +49366,7 @@
 			</output>
 		</layer>
 		<layer id="3407" name="onnx::MatMul_3610/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="10090564" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -49654,32 +49374,32 @@
 			</output>
 		</layer>
 		<layer id="3408" name="Convert_3918" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3409" name="Subtract_3919" type="Subtract" version="opset1">
+		<layer id="3409" name="onnx::MatMul_3610/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49687,7 +49407,7 @@
 			</output>
 		</layer>
 		<layer id="3410" name="onnx::MatMul_3610/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="159484125" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="161532138" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -49696,57 +49416,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3411" name="onnx::MatMul_3610/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3412" name="onnx::MatMul_3610/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3411" name="onnx::MatMul_3610/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3413" name="Constant_3924" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3412" name="Constant_3923" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3414" name="Reshape_3925" type="Reshape" version="opset1">
+		<layer id="3413" name="Reshape_3924" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -49756,7 +49459,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3414" name="onnx::MatMul_3610/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -49767,7 +49485,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -49778,7 +49496,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/value/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -49791,20 +49509,20 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/value/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3417" name="/blocks.23/attn/Constant_4" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058965" size="32" />
+			<data element_type="i64" shape="4" offset="10106965" size="32" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.23/attn/Constant_4_output_0">
 					<dim>4</dim>
@@ -49816,7 +49534,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -49826,14 +49544,14 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/Reshape_2_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3419" name="Constant_2313" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -49844,7 +49562,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -49856,7 +49574,7 @@
 				<port id="2" precision="FP32" names="/blocks.23/attn/Transpose_2_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
@@ -49867,13 +49585,13 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
+					<dim>1500</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</input>
@@ -49881,13 +49599,13 @@
 				<port id="2" precision="FP32" names="/blocks.23/attn/MatMul_1_output_0">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3422" name="Constant_2320" type="Const" version="opset1">
-			<data element_type="i64" shape="4" offset="8058997" size="32" />
+			<data element_type="i64" shape="4" offset="10106997" size="32" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>4</dim>
@@ -49899,7 +49617,7 @@
 				<port id="0" precision="FP32">
 					<dim>1</dim>
 					<dim>16</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>64</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -49909,7 +49627,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/Transpose_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -49920,7 +49638,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -49932,7 +49650,7 @@
 			</output>
 		</layer>
 		<layer id="3425" name="/blocks.23/attn/Constant_6" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.23/attn/Constant_6_output_0">
 					<dim>1</dim>
@@ -49985,7 +49703,7 @@
 			</output>
 		</layer>
 		<layer id="3430" name="/blocks.23/attn/Constant_5" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510076" size="8" />
+			<data element_type="i64" shape="1" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.23/attn/Constant_5_output_0">
 					<dim>1</dim>
@@ -50017,7 +49735,7 @@
 			</output>
 		</layer>
 		<layer id="3432" name="/blocks.23/attn/Constant_8" type="Const" version="opset1">
-			<data element_type="i64" shape="1" offset="7510068" size="8" />
+			<data element_type="i64" shape="1" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" names="/blocks.23/attn/Constant_8_output_0">
 					<dim>1</dim>
@@ -50045,7 +49763,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>16</dim>
 					<dim>64</dim>
 				</port>
@@ -50056,13 +49774,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/Reshape_3_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3435" name="onnx::MatMul_3628" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 1024" offset="159500509" size="524288" />
+			<data element_type="u4" shape="8, 128, 1024" offset="161548522" size="524288" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -50072,7 +49790,7 @@
 			</output>
 		</layer>
 		<layer id="3436" name="Convert_3949" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -50081,7 +49799,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -50089,7 +49807,7 @@
 			</output>
 		</layer>
 		<layer id="3437" name="onnx::MatMul_3628/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="11716798" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -50097,32 +49815,32 @@
 			</output>
 		</layer>
 		<layer id="3438" name="Convert_3951" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3439" name="Subtract_3952" type="Subtract" version="opset1">
+		<layer id="3439" name="onnx::MatMul_3628/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -50130,7 +49848,7 @@
 			</output>
 		</layer>
 		<layer id="3440" name="onnx::MatMul_3628/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 1024" offset="160024797" size="16384" />
+			<data element_type="f16" shape="8, 1, 1024" offset="162072810" size="16384" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -50139,57 +49857,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3441" name="onnx::MatMul_3628/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3442" name="onnx::MatMul_3628/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3441" name="onnx::MatMul_3628/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3443" name="Constant_3957" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="8058949" size="16" />
+		<layer id="3442" name="Constant_3956" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="10106949" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3444" name="Reshape_3958" type="Reshape" version="opset1">
+		<layer id="3443" name="Reshape_3957" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>1024</dim>
@@ -50199,7 +49900,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3444" name="onnx::MatMul_3628/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>1024</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>1024</dim>
 				</port>
@@ -50210,7 +49926,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -50221,7 +49937,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/out/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -50234,14 +49950,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/attn/out/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -50251,19 +49967,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -50273,20 +49989,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/mlp_ln/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3449" name="onnx::Add_3633" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="160041181" size="4096" />
+			<data element_type="f32" shape="1024" offset="162089194" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3633">
 					<dim>1024</dim>
@@ -50294,7 +50010,7 @@
 			</output>
 		</layer>
 		<layer id="3450" name="onnx::Add_3630" type="Const" version="opset1">
-			<data element_type="f32" shape="4096" offset="160045277" size="16384" />
+			<data element_type="f32" shape="4096" offset="162093290" size="16384" />
 			<output>
 				<port id="0" precision="FP32" names="onnx::Add_3630">
 					<dim>4096</dim>
@@ -50302,13 +50018,13 @@
 			</output>
 		</layer>
 		<layer id="3451" name="Constant_2346" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3452" name="Constant_2345" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -50337,7 +50053,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -50347,13 +50063,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3456" name="blocks.23.mlp_ln.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="160061661" size="4096" />
+			<data element_type="f32" shape="1024" offset="162109674" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.23.mlp_ln.weight">
 					<dim>1024</dim>
@@ -50365,7 +50081,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -50375,13 +50091,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3458" name="blocks.23.mlp_ln.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="160065757" size="4096" />
+			<data element_type="f32" shape="1024" offset="162113770" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="blocks.23.mlp_ln.bias">
 					<dim>1024</dim>
@@ -50393,7 +50109,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -50403,7 +50119,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp_ln/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -50413,20 +50129,20 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/mlp_ln/Cast_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3461" name="onnx::MatMul_3631" type="Const" version="opset1">
-			<data element_type="u4" shape="8, 128, 4096" offset="160069853" size="2097152" />
+			<data element_type="u4" shape="8, 128, 4096" offset="162117866" size="2097152" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -50436,7 +50152,7 @@
 			</output>
 		</layer>
 		<layer id="3462" name="Convert_3960" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>8</dim>
@@ -50445,7 +50161,7 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -50453,7 +50169,7 @@
 			</output>
 		</layer>
 		<layer id="3463" name="onnx::MatMul_3631/zero_point" type="Const" version="opset1">
-			<data element_type="u4" shape="1" offset="8042564" size="1" />
+			<data element_type="u4" shape="1" offset="31328483" size="1" />
 			<output>
 				<port id="0" precision="U4">
 					<dim>1</dim>
@@ -50461,32 +50177,32 @@
 			</output>
 		</layer>
 		<layer id="3464" name="Convert_3962" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U4">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3465" name="Subtract_3963" type="Subtract" version="opset1">
+		<layer id="3465" name="onnx::MatMul_3631/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -50494,7 +50210,7 @@
 			</output>
 		</layer>
 		<layer id="3466" name="onnx::MatMul_3631/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="8, 1, 4096" offset="162167005" size="65536" />
+			<data element_type="f16" shape="8, 1, 4096" offset="164215018" size="65536" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>8</dim>
@@ -50503,57 +50219,40 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3467" name="onnx::MatMul_3631/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
-			<input>
-				<port id="0" precision="FP16">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</input>
-			<output>
-				<port id="1" precision="FP32">
-					<dim>8</dim>
-					<dim>1</dim>
-					<dim>4096</dim>
-				</port>
-			</output>
-		</layer>
-		<layer id="3468" name="onnx::MatMul_3631/fq_weights_1" type="Multiply" version="opset1">
+		<layer id="3467" name="onnx::MatMul_3631/fq_weights_1" type="Multiply" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>8</dim>
 					<dim>1</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3469" name="Constant_3968" type="Const" version="opset1">
-			<data element_type="i64" shape="2" offset="11876541" size="16" />
+		<layer id="3468" name="Constant_3967" type="Const" version="opset1">
+			<data element_type="i64" shape="2" offset="13924543" size="16" />
 			<output>
 				<port id="0" precision="I64">
 					<dim>2</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3470" name="Reshape_3969" type="Reshape" version="opset1">
+		<layer id="3469" name="Reshape_3968" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>8</dim>
 					<dim>128</dim>
 					<dim>4096</dim>
@@ -50563,7 +50262,22 @@
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</output>
+		</layer>
+		<layer id="3470" name="onnx::MatMul_3631/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
+			<input>
+				<port id="0" precision="FP16">
+					<dim>1024</dim>
+					<dim>4096</dim>
+				</port>
+			</input>
+			<output>
+				<port id="1" precision="FP32">
 					<dim>1024</dim>
 					<dim>4096</dim>
 				</port>
@@ -50574,7 +50288,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -50585,7 +50299,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.0/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -50598,14 +50312,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.0/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -50621,7 +50335,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -50629,7 +50343,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.1/Div_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -50638,14 +50352,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/blocks.23/mlp/mlp.1/Erf_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -50661,7 +50375,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -50669,7 +50383,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.1/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -50679,19 +50393,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.1/Mul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
@@ -50707,7 +50421,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32" />
@@ -50715,13 +50429,13 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.1/Mul_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3481" name="onnx::MatMul_3634" type="Const" version="opset1">
-			<data element_type="u8" shape="4096, 1024" offset="162232541" size="4194304" />
+			<data element_type="u8" shape="4096, 1024" offset="164280554" size="4194304" />
 			<output>
 				<port id="0" precision="U8">
 					<dim>4096</dim>
@@ -50730,7 +50444,7 @@
 			</output>
 		</layer>
 		<layer id="3482" name="Convert_3972" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U8">
 					<dim>4096</dim>
@@ -50738,14 +50452,14 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3483" name="onnx::MatMul_3634/zero_point" type="Const" version="opset1">
-			<data element_type="u8" shape="1, 1024" offset="166426845" size="1024" />
+			<data element_type="u8" shape="1, 1024" offset="168474858" size="1024" />
 			<output>
 				<port id="0" precision="U8">
 					<dim>1</dim>
@@ -50754,7 +50468,7 @@
 			</output>
 		</layer>
 		<layer id="3484" name="Convert_3975" type="Convert" version="opset1">
-			<data destination_type="f32" />
+			<data destination_type="f16" />
 			<input>
 				<port id="0" precision="U8">
 					<dim>1</dim>
@@ -50762,33 +50476,33 @@
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3485" name="Subtract_3976" type="Subtract" version="opset1">
+		<layer id="3485" name="onnx::MatMul_3634/zero_point/subtract" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="2" precision="FP16">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3486" name="onnx::MatMul_3634/scale" type="Const" version="opset1">
-			<data element_type="f16" shape="1, 1024" offset="166427869" size="2048" />
+			<data element_type="f16" shape="1, 1024" offset="168475882" size="2048" />
 			<output>
 				<port id="0" precision="FP16">
 					<dim>1</dim>
@@ -50796,35 +50510,35 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3487" name="onnx::MatMul_3634/scale_convert" type="Convert" version="opset1">
-			<data destination_type="f32" />
+		<layer id="3487" name="onnx::MatMul_3634/fq_weights_1" type="Multiply" version="opset1">
+			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="FP16">
+					<dim>4096</dim>
+					<dim>1024</dim>
+				</port>
+				<port id="1" precision="FP16">
 					<dim>1</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
-				<port id="1" precision="FP32">
-					<dim>1</dim>
+				<port id="2" precision="FP16">
+					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="3488" name="onnx::MatMul_3634/fq_weights_1" type="Multiply" version="opset1">
-			<data auto_broadcast="numpy" />
+		<layer id="3488" name="onnx::MatMul_3634/fq_weights_1/convert" type="Convert" version="opset1">
+			<data destination_type="f32" />
 			<input>
-				<port id="0" precision="FP32">
+				<port id="0" precision="FP16">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
-				<port id="1" precision="FP32">
-					<dim>1</dim>
-					<dim>1024</dim>
-				</port>
 			</input>
 			<output>
-				<port id="2" precision="FP32">
+				<port id="1" precision="FP32">
 					<dim>4096</dim>
 					<dim>1024</dim>
 				</port>
@@ -50835,7 +50549,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>4096</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -50846,7 +50560,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.2/MatMul_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -50859,14 +50573,14 @@
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/mlp/mlp.2/Add_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -50876,19 +50590,19 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="2" precision="FP32" names="/blocks.23/Add_1_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -50898,26 +50612,26 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="/ln_post/Cast_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3493" name="Constant_2368" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510068" size="8" />
+			<data element_type="i64" shape="" offset="9558068" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
 		<layer id="3494" name="Constant_2367" type="Const" version="opset1">
-			<data element_type="i64" shape="" offset="7510076" size="8" />
+			<data element_type="i64" shape="" offset="9558076" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
@@ -50946,7 +50660,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="I64">
@@ -50956,13 +50670,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3498" name="ln_post.weight" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="166429917" size="4096" />
+			<data element_type="f32" shape="1024" offset="168477930" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="ln_post.weight">
 					<dim>1024</dim>
@@ -50974,7 +50688,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -50984,13 +50698,13 @@
 			<output>
 				<port id="2" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
 		</layer>
 		<layer id="3500" name="ln_post.bias" type="Const" version="opset1">
-			<data element_type="f32" shape="1024" offset="166434013" size="4096" />
+			<data element_type="f32" shape="1024" offset="168482026" size="4096" />
 			<output>
 				<port id="0" precision="FP32" names="ln_post.bias">
 					<dim>1024</dim>
@@ -51002,7 +50716,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 				<port id="1" precision="FP32">
@@ -51012,7 +50726,7 @@
 			<output>
 				<port id="2" precision="FP32" names="/ln_post/LayerNormalization_output_0">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -51022,14 +50736,14 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
 			<output>
 				<port id="1" precision="FP32" names="output_features">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</output>
@@ -51038,7 +50752,7 @@
 			<input>
 				<port id="0" precision="FP32">
 					<dim>1</dim>
-					<dim>1000</dim>
+					<dim>1500</dim>
 					<dim>1024</dim>
 				</port>
 			</input>
@@ -51050,10 +50764,10 @@
 		<edge from-layer="2" from-port="1" to-layer="5" to-port="0" />
 		<edge from-layer="3" from-port="0" to-layer="4" to-port="0" />
 		<edge from-layer="4" from-port="1" to-layer="5" to-port="1" />
-		<edge from-layer="5" from-port="2" to-layer="8" to-port="0" />
-		<edge from-layer="6" from-port="0" to-layer="7" to-port="0" />
-		<edge from-layer="7" from-port="1" to-layer="8" to-port="1" />
-		<edge from-layer="8" from-port="2" to-layer="9" to-port="1" />
+		<edge from-layer="5" from-port="2" to-layer="7" to-port="0" />
+		<edge from-layer="6" from-port="0" to-layer="7" to-port="1" />
+		<edge from-layer="7" from-port="2" to-layer="8" to-port="0" />
+		<edge from-layer="8" from-port="1" to-layer="9" to-port="1" />
 		<edge from-layer="9" from-port="2" to-layer="20" to-port="0" />
 		<edge from-layer="9" from-port="2" to-layer="13" to-port="0" />
 		<edge from-layer="10" from-port="0" to-layer="19" to-port="0" />
@@ -51082,10 +50796,10 @@
 		<edge from-layer="30" from-port="1" to-layer="33" to-port="0" />
 		<edge from-layer="31" from-port="0" to-layer="32" to-port="0" />
 		<edge from-layer="32" from-port="1" to-layer="33" to-port="1" />
-		<edge from-layer="33" from-port="2" to-layer="36" to-port="0" />
-		<edge from-layer="34" from-port="0" to-layer="35" to-port="0" />
-		<edge from-layer="35" from-port="1" to-layer="36" to-port="1" />
-		<edge from-layer="36" from-port="2" to-layer="37" to-port="1" />
+		<edge from-layer="33" from-port="2" to-layer="35" to-port="0" />
+		<edge from-layer="34" from-port="0" to-layer="35" to-port="1" />
+		<edge from-layer="35" from-port="2" to-layer="36" to-port="0" />
+		<edge from-layer="36" from-port="1" to-layer="37" to-port="1" />
 		<edge from-layer="37" from-port="2" to-layer="41" to-port="0" />
 		<edge from-layer="37" from-port="2" to-layer="48" to-port="0" />
 		<edge from-layer="38" from-port="0" to-layer="40" to-port="0" />
@@ -51135,12 +50849,12 @@
 		<edge from-layer="76" from-port="1" to-layer="79" to-port="0" />
 		<edge from-layer="77" from-port="0" to-layer="78" to-port="0" />
 		<edge from-layer="78" from-port="1" to-layer="79" to-port="1" />
-		<edge from-layer="79" from-port="2" to-layer="82" to-port="0" />
-		<edge from-layer="80" from-port="0" to-layer="81" to-port="0" />
-		<edge from-layer="81" from-port="1" to-layer="82" to-port="1" />
-		<edge from-layer="82" from-port="2" to-layer="84" to-port="0" />
-		<edge from-layer="83" from-port="0" to-layer="84" to-port="1" />
-		<edge from-layer="84" from-port="2" to-layer="85" to-port="1" />
+		<edge from-layer="79" from-port="2" to-layer="81" to-port="0" />
+		<edge from-layer="80" from-port="0" to-layer="81" to-port="1" />
+		<edge from-layer="81" from-port="2" to-layer="83" to-port="0" />
+		<edge from-layer="82" from-port="0" to-layer="83" to-port="1" />
+		<edge from-layer="83" from-port="2" to-layer="84" to-port="0" />
+		<edge from-layer="84" from-port="1" to-layer="85" to-port="1" />
 		<edge from-layer="85" from-port="2" to-layer="86" to-port="1" />
 		<edge from-layer="86" from-port="2" to-layer="88" to-port="0" />
 		<edge from-layer="87" from-port="0" to-layer="88" to-port="1" />
@@ -51156,12 +50870,12 @@
 		<edge from-layer="96" from-port="1" to-layer="99" to-port="0" />
 		<edge from-layer="97" from-port="0" to-layer="98" to-port="0" />
 		<edge from-layer="98" from-port="1" to-layer="99" to-port="1" />
-		<edge from-layer="99" from-port="2" to-layer="102" to-port="0" />
-		<edge from-layer="100" from-port="0" to-layer="101" to-port="0" />
-		<edge from-layer="101" from-port="1" to-layer="102" to-port="1" />
-		<edge from-layer="102" from-port="2" to-layer="104" to-port="0" />
-		<edge from-layer="103" from-port="0" to-layer="104" to-port="1" />
-		<edge from-layer="104" from-port="2" to-layer="105" to-port="1" />
+		<edge from-layer="99" from-port="2" to-layer="101" to-port="0" />
+		<edge from-layer="100" from-port="0" to-layer="101" to-port="1" />
+		<edge from-layer="101" from-port="2" to-layer="103" to-port="0" />
+		<edge from-layer="102" from-port="0" to-layer="103" to-port="1" />
+		<edge from-layer="103" from-port="2" to-layer="104" to-port="0" />
+		<edge from-layer="104" from-port="1" to-layer="105" to-port="1" />
 		<edge from-layer="105" from-port="2" to-layer="107" to-port="0" />
 		<edge from-layer="106" from-port="0" to-layer="107" to-port="1" />
 		<edge from-layer="107" from-port="2" to-layer="109" to-port="0" />
@@ -51177,12 +50891,12 @@
 		<edge from-layer="117" from-port="1" to-layer="120" to-port="0" />
 		<edge from-layer="118" from-port="0" to-layer="119" to-port="0" />
 		<edge from-layer="119" from-port="1" to-layer="120" to-port="1" />
-		<edge from-layer="120" from-port="2" to-layer="123" to-port="0" />
-		<edge from-layer="121" from-port="0" to-layer="122" to-port="0" />
-		<edge from-layer="122" from-port="1" to-layer="123" to-port="1" />
-		<edge from-layer="123" from-port="2" to-layer="125" to-port="0" />
-		<edge from-layer="124" from-port="0" to-layer="125" to-port="1" />
-		<edge from-layer="125" from-port="2" to-layer="126" to-port="1" />
+		<edge from-layer="120" from-port="2" to-layer="122" to-port="0" />
+		<edge from-layer="121" from-port="0" to-layer="122" to-port="1" />
+		<edge from-layer="122" from-port="2" to-layer="124" to-port="0" />
+		<edge from-layer="123" from-port="0" to-layer="124" to-port="1" />
+		<edge from-layer="124" from-port="2" to-layer="125" to-port="0" />
+		<edge from-layer="125" from-port="1" to-layer="126" to-port="1" />
 		<edge from-layer="126" from-port="2" to-layer="127" to-port="1" />
 		<edge from-layer="127" from-port="2" to-layer="129" to-port="0" />
 		<edge from-layer="128" from-port="0" to-layer="129" to-port="1" />
@@ -51191,11 +50905,11 @@
 		<edge from-layer="131" from-port="2" to-layer="132" to-port="1" />
 		<edge from-layer="132" from-port="2" to-layer="134" to-port="0" />
 		<edge from-layer="133" from-port="0" to-layer="134" to-port="1" />
-		<edge from-layer="134" from-port="2" to-layer="145" to-port="0" />
 		<edge from-layer="134" from-port="2" to-layer="135" to-port="0" />
+		<edge from-layer="134" from-port="2" to-layer="145" to-port="0" />
 		<edge from-layer="135" from-port="1" to-layer="142" to-port="0" />
-		<edge from-layer="136" from-port="0" to-layer="142" to-port="1" />
 		<edge from-layer="136" from-port="0" to-layer="139" to-port="0" />
+		<edge from-layer="136" from-port="0" to-layer="142" to-port="1" />
 		<edge from-layer="137" from-port="0" to-layer="142" to-port="2" />
 		<edge from-layer="138" from-port="0" to-layer="140" to-port="0" />
 		<edge from-layer="139" from-port="1" to-layer="140" to-port="1" />
@@ -51209,12 +50923,12 @@
 		<edge from-layer="147" from-port="1" to-layer="150" to-port="0" />
 		<edge from-layer="148" from-port="0" to-layer="149" to-port="0" />
 		<edge from-layer="149" from-port="1" to-layer="150" to-port="1" />
-		<edge from-layer="150" from-port="2" to-layer="153" to-port="0" />
-		<edge from-layer="151" from-port="0" to-layer="152" to-port="0" />
-		<edge from-layer="152" from-port="1" to-layer="153" to-port="1" />
-		<edge from-layer="153" from-port="2" to-layer="155" to-port="0" />
-		<edge from-layer="154" from-port="0" to-layer="155" to-port="1" />
-		<edge from-layer="155" from-port="2" to-layer="156" to-port="1" />
+		<edge from-layer="150" from-port="2" to-layer="152" to-port="0" />
+		<edge from-layer="151" from-port="0" to-layer="152" to-port="1" />
+		<edge from-layer="152" from-port="2" to-layer="154" to-port="0" />
+		<edge from-layer="153" from-port="0" to-layer="154" to-port="1" />
+		<edge from-layer="154" from-port="2" to-layer="155" to-port="0" />
+		<edge from-layer="155" from-port="1" to-layer="156" to-port="1" />
 		<edge from-layer="156" from-port="2" to-layer="157" to-port="1" />
 		<edge from-layer="157" from-port="2" to-layer="158" to-port="1" />
 		<edge from-layer="158" from-port="2" to-layer="159" to-port="0" />
@@ -51236,12 +50950,12 @@
 		<edge from-layer="173" from-port="1" to-layer="176" to-port="0" />
 		<edge from-layer="174" from-port="0" to-layer="175" to-port="0" />
 		<edge from-layer="175" from-port="1" to-layer="176" to-port="1" />
-		<edge from-layer="176" from-port="2" to-layer="179" to-port="0" />
-		<edge from-layer="177" from-port="0" to-layer="178" to-port="0" />
-		<edge from-layer="178" from-port="1" to-layer="179" to-port="1" />
-		<edge from-layer="179" from-port="2" to-layer="181" to-port="0" />
-		<edge from-layer="180" from-port="0" to-layer="181" to-port="1" />
-		<edge from-layer="181" from-port="2" to-layer="182" to-port="1" />
+		<edge from-layer="176" from-port="2" to-layer="178" to-port="0" />
+		<edge from-layer="177" from-port="0" to-layer="178" to-port="1" />
+		<edge from-layer="178" from-port="2" to-layer="180" to-port="0" />
+		<edge from-layer="179" from-port="0" to-layer="180" to-port="1" />
+		<edge from-layer="180" from-port="2" to-layer="181" to-port="0" />
+		<edge from-layer="181" from-port="1" to-layer="182" to-port="1" />
 		<edge from-layer="182" from-port="2" to-layer="183" to-port="1" />
 		<edge from-layer="183" from-port="2" to-layer="185" to-port="0" />
 		<edge from-layer="183" from-port="2" to-layer="189" to-port="0" />
@@ -51257,17 +50971,17 @@
 		<edge from-layer="193" from-port="1" to-layer="196" to-port="0" />
 		<edge from-layer="194" from-port="0" to-layer="195" to-port="0" />
 		<edge from-layer="195" from-port="1" to-layer="196" to-port="1" />
-		<edge from-layer="196" from-port="2" to-layer="199" to-port="0" />
-		<edge from-layer="197" from-port="0" to-layer="198" to-port="0" />
-		<edge from-layer="198" from-port="1" to-layer="199" to-port="1" />
-		<edge from-layer="199" from-port="2" to-layer="201" to-port="0" />
-		<edge from-layer="200" from-port="0" to-layer="201" to-port="1" />
-		<edge from-layer="201" from-port="2" to-layer="202" to-port="1" />
+		<edge from-layer="196" from-port="2" to-layer="198" to-port="0" />
+		<edge from-layer="197" from-port="0" to-layer="198" to-port="1" />
+		<edge from-layer="198" from-port="2" to-layer="200" to-port="0" />
+		<edge from-layer="199" from-port="0" to-layer="200" to-port="1" />
+		<edge from-layer="200" from-port="2" to-layer="201" to-port="0" />
+		<edge from-layer="201" from-port="1" to-layer="202" to-port="1" />
 		<edge from-layer="202" from-port="2" to-layer="203" to-port="1" />
 		<edge from-layer="203" from-port="2" to-layer="204" to-port="1" />
 		<edge from-layer="204" from-port="2" to-layer="205" to-port="0" />
-		<edge from-layer="205" from-port="1" to-layer="212" to-port="0" />
 		<edge from-layer="205" from-port="1" to-layer="301" to-port="0" />
+		<edge from-layer="205" from-port="1" to-layer="212" to-port="0" />
 		<edge from-layer="206" from-port="0" to-layer="300" to-port="0" />
 		<edge from-layer="207" from-port="0" to-layer="229" to-port="0" />
 		<edge from-layer="208" from-port="0" to-layer="211" to-port="0" />
@@ -51280,18 +50994,18 @@
 		<edge from-layer="215" from-port="0" to-layer="216" to-port="1" />
 		<edge from-layer="216" from-port="2" to-layer="217" to-port="0" />
 		<edge from-layer="217" from-port="1" to-layer="269" to-port="0" />
-		<edge from-layer="217" from-port="1" to-layer="228" to-port="0" />
 		<edge from-layer="217" from-port="1" to-layer="248" to-port="0" />
+		<edge from-layer="217" from-port="1" to-layer="228" to-port="0" />
 		<edge from-layer="218" from-port="0" to-layer="219" to-port="0" />
 		<edge from-layer="219" from-port="1" to-layer="222" to-port="0" />
 		<edge from-layer="220" from-port="0" to-layer="221" to-port="0" />
 		<edge from-layer="221" from-port="1" to-layer="222" to-port="1" />
-		<edge from-layer="222" from-port="2" to-layer="225" to-port="0" />
-		<edge from-layer="223" from-port="0" to-layer="224" to-port="0" />
-		<edge from-layer="224" from-port="1" to-layer="225" to-port="1" />
-		<edge from-layer="225" from-port="2" to-layer="227" to-port="0" />
-		<edge from-layer="226" from-port="0" to-layer="227" to-port="1" />
-		<edge from-layer="227" from-port="2" to-layer="228" to-port="1" />
+		<edge from-layer="222" from-port="2" to-layer="224" to-port="0" />
+		<edge from-layer="223" from-port="0" to-layer="224" to-port="1" />
+		<edge from-layer="224" from-port="2" to-layer="226" to-port="0" />
+		<edge from-layer="225" from-port="0" to-layer="226" to-port="1" />
+		<edge from-layer="226" from-port="2" to-layer="227" to-port="0" />
+		<edge from-layer="227" from-port="1" to-layer="228" to-port="1" />
 		<edge from-layer="228" from-port="2" to-layer="229" to-port="1" />
 		<edge from-layer="229" from-port="2" to-layer="231" to-port="0" />
 		<edge from-layer="230" from-port="0" to-layer="231" to-port="1" />
@@ -51307,12 +51021,12 @@
 		<edge from-layer="239" from-port="1" to-layer="242" to-port="0" />
 		<edge from-layer="240" from-port="0" to-layer="241" to-port="0" />
 		<edge from-layer="241" from-port="1" to-layer="242" to-port="1" />
-		<edge from-layer="242" from-port="2" to-layer="245" to-port="0" />
-		<edge from-layer="243" from-port="0" to-layer="244" to-port="0" />
-		<edge from-layer="244" from-port="1" to-layer="245" to-port="1" />
-		<edge from-layer="245" from-port="2" to-layer="247" to-port="0" />
-		<edge from-layer="246" from-port="0" to-layer="247" to-port="1" />
-		<edge from-layer="247" from-port="2" to-layer="248" to-port="1" />
+		<edge from-layer="242" from-port="2" to-layer="244" to-port="0" />
+		<edge from-layer="243" from-port="0" to-layer="244" to-port="1" />
+		<edge from-layer="244" from-port="2" to-layer="246" to-port="0" />
+		<edge from-layer="245" from-port="0" to-layer="246" to-port="1" />
+		<edge from-layer="246" from-port="2" to-layer="247" to-port="0" />
+		<edge from-layer="247" from-port="1" to-layer="248" to-port="1" />
 		<edge from-layer="248" from-port="2" to-layer="250" to-port="0" />
 		<edge from-layer="249" from-port="0" to-layer="250" to-port="1" />
 		<edge from-layer="250" from-port="2" to-layer="252" to-port="0" />
@@ -51328,12 +51042,12 @@
 		<edge from-layer="260" from-port="1" to-layer="263" to-port="0" />
 		<edge from-layer="261" from-port="0" to-layer="262" to-port="0" />
 		<edge from-layer="262" from-port="1" to-layer="263" to-port="1" />
-		<edge from-layer="263" from-port="2" to-layer="266" to-port="0" />
-		<edge from-layer="264" from-port="0" to-layer="265" to-port="0" />
-		<edge from-layer="265" from-port="1" to-layer="266" to-port="1" />
-		<edge from-layer="266" from-port="2" to-layer="268" to-port="0" />
-		<edge from-layer="267" from-port="0" to-layer="268" to-port="1" />
-		<edge from-layer="268" from-port="2" to-layer="269" to-port="1" />
+		<edge from-layer="263" from-port="2" to-layer="265" to-port="0" />
+		<edge from-layer="264" from-port="0" to-layer="265" to-port="1" />
+		<edge from-layer="265" from-port="2" to-layer="267" to-port="0" />
+		<edge from-layer="266" from-port="0" to-layer="267" to-port="1" />
+		<edge from-layer="267" from-port="2" to-layer="268" to-port="0" />
+		<edge from-layer="268" from-port="1" to-layer="269" to-port="1" />
 		<edge from-layer="269" from-port="2" to-layer="270" to-port="1" />
 		<edge from-layer="270" from-port="2" to-layer="272" to-port="0" />
 		<edge from-layer="271" from-port="0" to-layer="272" to-port="1" />
@@ -51360,12 +51074,12 @@
 		<edge from-layer="290" from-port="1" to-layer="293" to-port="0" />
 		<edge from-layer="291" from-port="0" to-layer="292" to-port="0" />
 		<edge from-layer="292" from-port="1" to-layer="293" to-port="1" />
-		<edge from-layer="293" from-port="2" to-layer="296" to-port="0" />
-		<edge from-layer="294" from-port="0" to-layer="295" to-port="0" />
-		<edge from-layer="295" from-port="1" to-layer="296" to-port="1" />
-		<edge from-layer="296" from-port="2" to-layer="298" to-port="0" />
-		<edge from-layer="297" from-port="0" to-layer="298" to-port="1" />
-		<edge from-layer="298" from-port="2" to-layer="299" to-port="1" />
+		<edge from-layer="293" from-port="2" to-layer="295" to-port="0" />
+		<edge from-layer="294" from-port="0" to-layer="295" to-port="1" />
+		<edge from-layer="295" from-port="2" to-layer="297" to-port="0" />
+		<edge from-layer="296" from-port="0" to-layer="297" to-port="1" />
+		<edge from-layer="297" from-port="2" to-layer="298" to-port="0" />
+		<edge from-layer="298" from-port="1" to-layer="299" to-port="1" />
 		<edge from-layer="299" from-port="2" to-layer="300" to-port="1" />
 		<edge from-layer="300" from-port="2" to-layer="301" to-port="1" />
 		<edge from-layer="301" from-port="2" to-layer="302" to-port="0" />
@@ -51387,12 +51101,12 @@
 		<edge from-layer="316" from-port="1" to-layer="319" to-port="0" />
 		<edge from-layer="317" from-port="0" to-layer="318" to-port="0" />
 		<edge from-layer="318" from-port="1" to-layer="319" to-port="1" />
-		<edge from-layer="319" from-port="2" to-layer="322" to-port="0" />
-		<edge from-layer="320" from-port="0" to-layer="321" to-port="0" />
-		<edge from-layer="321" from-port="1" to-layer="322" to-port="1" />
-		<edge from-layer="322" from-port="2" to-layer="324" to-port="0" />
-		<edge from-layer="323" from-port="0" to-layer="324" to-port="1" />
-		<edge from-layer="324" from-port="2" to-layer="325" to-port="1" />
+		<edge from-layer="319" from-port="2" to-layer="321" to-port="0" />
+		<edge from-layer="320" from-port="0" to-layer="321" to-port="1" />
+		<edge from-layer="321" from-port="2" to-layer="323" to-port="0" />
+		<edge from-layer="322" from-port="0" to-layer="323" to-port="1" />
+		<edge from-layer="323" from-port="2" to-layer="324" to-port="0" />
+		<edge from-layer="324" from-port="1" to-layer="325" to-port="1" />
 		<edge from-layer="325" from-port="2" to-layer="326" to-port="1" />
 		<edge from-layer="326" from-port="2" to-layer="328" to-port="0" />
 		<edge from-layer="326" from-port="2" to-layer="332" to-port="0" />
@@ -51408,12 +51122,12 @@
 		<edge from-layer="336" from-port="1" to-layer="339" to-port="0" />
 		<edge from-layer="337" from-port="0" to-layer="338" to-port="0" />
 		<edge from-layer="338" from-port="1" to-layer="339" to-port="1" />
-		<edge from-layer="339" from-port="2" to-layer="342" to-port="0" />
-		<edge from-layer="340" from-port="0" to-layer="341" to-port="0" />
-		<edge from-layer="341" from-port="1" to-layer="342" to-port="1" />
-		<edge from-layer="342" from-port="2" to-layer="344" to-port="0" />
-		<edge from-layer="343" from-port="0" to-layer="344" to-port="1" />
-		<edge from-layer="344" from-port="2" to-layer="345" to-port="1" />
+		<edge from-layer="339" from-port="2" to-layer="341" to-port="0" />
+		<edge from-layer="340" from-port="0" to-layer="341" to-port="1" />
+		<edge from-layer="341" from-port="2" to-layer="343" to-port="0" />
+		<edge from-layer="342" from-port="0" to-layer="343" to-port="1" />
+		<edge from-layer="343" from-port="2" to-layer="344" to-port="0" />
+		<edge from-layer="344" from-port="1" to-layer="345" to-port="1" />
 		<edge from-layer="345" from-port="2" to-layer="346" to-port="1" />
 		<edge from-layer="346" from-port="2" to-layer="347" to-port="1" />
 		<edge from-layer="347" from-port="2" to-layer="348" to-port="0" />
@@ -51431,18 +51145,18 @@
 		<edge from-layer="358" from-port="0" to-layer="359" to-port="1" />
 		<edge from-layer="359" from-port="2" to-layer="360" to-port="0" />
 		<edge from-layer="360" from-port="1" to-layer="412" to-port="0" />
-		<edge from-layer="360" from-port="1" to-layer="371" to-port="0" />
 		<edge from-layer="360" from-port="1" to-layer="391" to-port="0" />
+		<edge from-layer="360" from-port="1" to-layer="371" to-port="0" />
 		<edge from-layer="361" from-port="0" to-layer="362" to-port="0" />
 		<edge from-layer="362" from-port="1" to-layer="365" to-port="0" />
 		<edge from-layer="363" from-port="0" to-layer="364" to-port="0" />
 		<edge from-layer="364" from-port="1" to-layer="365" to-port="1" />
-		<edge from-layer="365" from-port="2" to-layer="368" to-port="0" />
-		<edge from-layer="366" from-port="0" to-layer="367" to-port="0" />
-		<edge from-layer="367" from-port="1" to-layer="368" to-port="1" />
-		<edge from-layer="368" from-port="2" to-layer="370" to-port="0" />
-		<edge from-layer="369" from-port="0" to-layer="370" to-port="1" />
-		<edge from-layer="370" from-port="2" to-layer="371" to-port="1" />
+		<edge from-layer="365" from-port="2" to-layer="367" to-port="0" />
+		<edge from-layer="366" from-port="0" to-layer="367" to-port="1" />
+		<edge from-layer="367" from-port="2" to-layer="369" to-port="0" />
+		<edge from-layer="368" from-port="0" to-layer="369" to-port="1" />
+		<edge from-layer="369" from-port="2" to-layer="370" to-port="0" />
+		<edge from-layer="370" from-port="1" to-layer="371" to-port="1" />
 		<edge from-layer="371" from-port="2" to-layer="372" to-port="1" />
 		<edge from-layer="372" from-port="2" to-layer="374" to-port="0" />
 		<edge from-layer="373" from-port="0" to-layer="374" to-port="1" />
@@ -51458,12 +51172,12 @@
 		<edge from-layer="382" from-port="1" to-layer="385" to-port="0" />
 		<edge from-layer="383" from-port="0" to-layer="384" to-port="0" />
 		<edge from-layer="384" from-port="1" to-layer="385" to-port="1" />
-		<edge from-layer="385" from-port="2" to-layer="388" to-port="0" />
-		<edge from-layer="386" from-port="0" to-layer="387" to-port="0" />
-		<edge from-layer="387" from-port="1" to-layer="388" to-port="1" />
-		<edge from-layer="388" from-port="2" to-layer="390" to-port="0" />
-		<edge from-layer="389" from-port="0" to-layer="390" to-port="1" />
-		<edge from-layer="390" from-port="2" to-layer="391" to-port="1" />
+		<edge from-layer="385" from-port="2" to-layer="387" to-port="0" />
+		<edge from-layer="386" from-port="0" to-layer="387" to-port="1" />
+		<edge from-layer="387" from-port="2" to-layer="389" to-port="0" />
+		<edge from-layer="388" from-port="0" to-layer="389" to-port="1" />
+		<edge from-layer="389" from-port="2" to-layer="390" to-port="0" />
+		<edge from-layer="390" from-port="1" to-layer="391" to-port="1" />
 		<edge from-layer="391" from-port="2" to-layer="393" to-port="0" />
 		<edge from-layer="392" from-port="0" to-layer="393" to-port="1" />
 		<edge from-layer="393" from-port="2" to-layer="395" to-port="0" />
@@ -51479,12 +51193,12 @@
 		<edge from-layer="403" from-port="1" to-layer="406" to-port="0" />
 		<edge from-layer="404" from-port="0" to-layer="405" to-port="0" />
 		<edge from-layer="405" from-port="1" to-layer="406" to-port="1" />
-		<edge from-layer="406" from-port="2" to-layer="409" to-port="0" />
-		<edge from-layer="407" from-port="0" to-layer="408" to-port="0" />
-		<edge from-layer="408" from-port="1" to-layer="409" to-port="1" />
-		<edge from-layer="409" from-port="2" to-layer="411" to-port="0" />
-		<edge from-layer="410" from-port="0" to-layer="411" to-port="1" />
-		<edge from-layer="411" from-port="2" to-layer="412" to-port="1" />
+		<edge from-layer="406" from-port="2" to-layer="408" to-port="0" />
+		<edge from-layer="407" from-port="0" to-layer="408" to-port="1" />
+		<edge from-layer="408" from-port="2" to-layer="410" to-port="0" />
+		<edge from-layer="409" from-port="0" to-layer="410" to-port="1" />
+		<edge from-layer="410" from-port="2" to-layer="411" to-port="0" />
+		<edge from-layer="411" from-port="1" to-layer="412" to-port="1" />
 		<edge from-layer="412" from-port="2" to-layer="413" to-port="1" />
 		<edge from-layer="413" from-port="2" to-layer="415" to-port="0" />
 		<edge from-layer="414" from-port="0" to-layer="415" to-port="1" />
@@ -51496,8 +51210,8 @@
 		<edge from-layer="420" from-port="2" to-layer="421" to-port="0" />
 		<edge from-layer="420" from-port="2" to-layer="431" to-port="0" />
 		<edge from-layer="421" from-port="1" to-layer="428" to-port="0" />
-		<edge from-layer="422" from-port="0" to-layer="425" to-port="0" />
 		<edge from-layer="422" from-port="0" to-layer="428" to-port="1" />
+		<edge from-layer="422" from-port="0" to-layer="425" to-port="0" />
 		<edge from-layer="423" from-port="0" to-layer="428" to-port="2" />
 		<edge from-layer="424" from-port="0" to-layer="426" to-port="0" />
 		<edge from-layer="425" from-port="1" to-layer="426" to-port="1" />
@@ -51511,12 +51225,12 @@
 		<edge from-layer="433" from-port="1" to-layer="436" to-port="0" />
 		<edge from-layer="434" from-port="0" to-layer="435" to-port="0" />
 		<edge from-layer="435" from-port="1" to-layer="436" to-port="1" />
-		<edge from-layer="436" from-port="2" to-layer="439" to-port="0" />
-		<edge from-layer="437" from-port="0" to-layer="438" to-port="0" />
-		<edge from-layer="438" from-port="1" to-layer="439" to-port="1" />
-		<edge from-layer="439" from-port="2" to-layer="441" to-port="0" />
-		<edge from-layer="440" from-port="0" to-layer="441" to-port="1" />
-		<edge from-layer="441" from-port="2" to-layer="442" to-port="1" />
+		<edge from-layer="436" from-port="2" to-layer="438" to-port="0" />
+		<edge from-layer="437" from-port="0" to-layer="438" to-port="1" />
+		<edge from-layer="438" from-port="2" to-layer="440" to-port="0" />
+		<edge from-layer="439" from-port="0" to-layer="440" to-port="1" />
+		<edge from-layer="440" from-port="2" to-layer="441" to-port="0" />
+		<edge from-layer="441" from-port="1" to-layer="442" to-port="1" />
 		<edge from-layer="442" from-port="2" to-layer="443" to-port="1" />
 		<edge from-layer="443" from-port="2" to-layer="444" to-port="1" />
 		<edge from-layer="444" from-port="2" to-layer="445" to-port="0" />
@@ -51538,12 +51252,12 @@
 		<edge from-layer="459" from-port="1" to-layer="462" to-port="0" />
 		<edge from-layer="460" from-port="0" to-layer="461" to-port="0" />
 		<edge from-layer="461" from-port="1" to-layer="462" to-port="1" />
-		<edge from-layer="462" from-port="2" to-layer="465" to-port="0" />
-		<edge from-layer="463" from-port="0" to-layer="464" to-port="0" />
-		<edge from-layer="464" from-port="1" to-layer="465" to-port="1" />
-		<edge from-layer="465" from-port="2" to-layer="467" to-port="0" />
-		<edge from-layer="466" from-port="0" to-layer="467" to-port="1" />
-		<edge from-layer="467" from-port="2" to-layer="468" to-port="1" />
+		<edge from-layer="462" from-port="2" to-layer="464" to-port="0" />
+		<edge from-layer="463" from-port="0" to-layer="464" to-port="1" />
+		<edge from-layer="464" from-port="2" to-layer="466" to-port="0" />
+		<edge from-layer="465" from-port="0" to-layer="466" to-port="1" />
+		<edge from-layer="466" from-port="2" to-layer="467" to-port="0" />
+		<edge from-layer="467" from-port="1" to-layer="468" to-port="1" />
 		<edge from-layer="468" from-port="2" to-layer="469" to-port="1" />
 		<edge from-layer="469" from-port="2" to-layer="471" to-port="0" />
 		<edge from-layer="469" from-port="2" to-layer="475" to-port="0" />
@@ -51559,12 +51273,12 @@
 		<edge from-layer="479" from-port="1" to-layer="482" to-port="0" />
 		<edge from-layer="480" from-port="0" to-layer="481" to-port="0" />
 		<edge from-layer="481" from-port="1" to-layer="482" to-port="1" />
-		<edge from-layer="482" from-port="2" to-layer="485" to-port="0" />
-		<edge from-layer="483" from-port="0" to-layer="484" to-port="0" />
-		<edge from-layer="484" from-port="1" to-layer="485" to-port="1" />
-		<edge from-layer="485" from-port="2" to-layer="487" to-port="0" />
-		<edge from-layer="486" from-port="0" to-layer="487" to-port="1" />
-		<edge from-layer="487" from-port="2" to-layer="488" to-port="1" />
+		<edge from-layer="482" from-port="2" to-layer="484" to-port="0" />
+		<edge from-layer="483" from-port="0" to-layer="484" to-port="1" />
+		<edge from-layer="484" from-port="2" to-layer="486" to-port="0" />
+		<edge from-layer="485" from-port="0" to-layer="486" to-port="1" />
+		<edge from-layer="486" from-port="2" to-layer="487" to-port="0" />
+		<edge from-layer="487" from-port="1" to-layer="488" to-port="1" />
 		<edge from-layer="488" from-port="2" to-layer="489" to-port="1" />
 		<edge from-layer="489" from-port="2" to-layer="490" to-port="1" />
 		<edge from-layer="490" from-port="2" to-layer="491" to-port="0" />
@@ -51588,12 +51302,12 @@
 		<edge from-layer="505" from-port="1" to-layer="508" to-port="0" />
 		<edge from-layer="506" from-port="0" to-layer="507" to-port="0" />
 		<edge from-layer="507" from-port="1" to-layer="508" to-port="1" />
-		<edge from-layer="508" from-port="2" to-layer="511" to-port="0" />
-		<edge from-layer="509" from-port="0" to-layer="510" to-port="0" />
-		<edge from-layer="510" from-port="1" to-layer="511" to-port="1" />
-		<edge from-layer="511" from-port="2" to-layer="513" to-port="0" />
-		<edge from-layer="512" from-port="0" to-layer="513" to-port="1" />
-		<edge from-layer="513" from-port="2" to-layer="514" to-port="1" />
+		<edge from-layer="508" from-port="2" to-layer="510" to-port="0" />
+		<edge from-layer="509" from-port="0" to-layer="510" to-port="1" />
+		<edge from-layer="510" from-port="2" to-layer="512" to-port="0" />
+		<edge from-layer="511" from-port="0" to-layer="512" to-port="1" />
+		<edge from-layer="512" from-port="2" to-layer="513" to-port="0" />
+		<edge from-layer="513" from-port="1" to-layer="514" to-port="1" />
 		<edge from-layer="514" from-port="2" to-layer="515" to-port="1" />
 		<edge from-layer="515" from-port="2" to-layer="517" to-port="0" />
 		<edge from-layer="516" from-port="0" to-layer="517" to-port="1" />
@@ -51609,12 +51323,12 @@
 		<edge from-layer="525" from-port="1" to-layer="528" to-port="0" />
 		<edge from-layer="526" from-port="0" to-layer="527" to-port="0" />
 		<edge from-layer="527" from-port="1" to-layer="528" to-port="1" />
-		<edge from-layer="528" from-port="2" to-layer="531" to-port="0" />
-		<edge from-layer="529" from-port="0" to-layer="530" to-port="0" />
-		<edge from-layer="530" from-port="1" to-layer="531" to-port="1" />
-		<edge from-layer="531" from-port="2" to-layer="533" to-port="0" />
-		<edge from-layer="532" from-port="0" to-layer="533" to-port="1" />
-		<edge from-layer="533" from-port="2" to-layer="534" to-port="1" />
+		<edge from-layer="528" from-port="2" to-layer="530" to-port="0" />
+		<edge from-layer="529" from-port="0" to-layer="530" to-port="1" />
+		<edge from-layer="530" from-port="2" to-layer="532" to-port="0" />
+		<edge from-layer="531" from-port="0" to-layer="532" to-port="1" />
+		<edge from-layer="532" from-port="2" to-layer="533" to-port="0" />
+		<edge from-layer="533" from-port="1" to-layer="534" to-port="1" />
 		<edge from-layer="534" from-port="2" to-layer="536" to-port="0" />
 		<edge from-layer="535" from-port="0" to-layer="536" to-port="1" />
 		<edge from-layer="536" from-port="2" to-layer="538" to-port="0" />
@@ -51630,12 +51344,12 @@
 		<edge from-layer="546" from-port="1" to-layer="549" to-port="0" />
 		<edge from-layer="547" from-port="0" to-layer="548" to-port="0" />
 		<edge from-layer="548" from-port="1" to-layer="549" to-port="1" />
-		<edge from-layer="549" from-port="2" to-layer="552" to-port="0" />
-		<edge from-layer="550" from-port="0" to-layer="551" to-port="0" />
-		<edge from-layer="551" from-port="1" to-layer="552" to-port="1" />
-		<edge from-layer="552" from-port="2" to-layer="554" to-port="0" />
-		<edge from-layer="553" from-port="0" to-layer="554" to-port="1" />
-		<edge from-layer="554" from-port="2" to-layer="555" to-port="1" />
+		<edge from-layer="549" from-port="2" to-layer="551" to-port="0" />
+		<edge from-layer="550" from-port="0" to-layer="551" to-port="1" />
+		<edge from-layer="551" from-port="2" to-layer="553" to-port="0" />
+		<edge from-layer="552" from-port="0" to-layer="553" to-port="1" />
+		<edge from-layer="553" from-port="2" to-layer="554" to-port="0" />
+		<edge from-layer="554" from-port="1" to-layer="555" to-port="1" />
 		<edge from-layer="555" from-port="2" to-layer="556" to-port="1" />
 		<edge from-layer="556" from-port="2" to-layer="558" to-port="0" />
 		<edge from-layer="557" from-port="0" to-layer="558" to-port="1" />
@@ -51662,12 +51376,12 @@
 		<edge from-layer="576" from-port="1" to-layer="579" to-port="0" />
 		<edge from-layer="577" from-port="0" to-layer="578" to-port="0" />
 		<edge from-layer="578" from-port="1" to-layer="579" to-port="1" />
-		<edge from-layer="579" from-port="2" to-layer="582" to-port="0" />
-		<edge from-layer="580" from-port="0" to-layer="581" to-port="0" />
-		<edge from-layer="581" from-port="1" to-layer="582" to-port="1" />
-		<edge from-layer="582" from-port="2" to-layer="584" to-port="0" />
-		<edge from-layer="583" from-port="0" to-layer="584" to-port="1" />
-		<edge from-layer="584" from-port="2" to-layer="585" to-port="1" />
+		<edge from-layer="579" from-port="2" to-layer="581" to-port="0" />
+		<edge from-layer="580" from-port="0" to-layer="581" to-port="1" />
+		<edge from-layer="581" from-port="2" to-layer="583" to-port="0" />
+		<edge from-layer="582" from-port="0" to-layer="583" to-port="1" />
+		<edge from-layer="583" from-port="2" to-layer="584" to-port="0" />
+		<edge from-layer="584" from-port="1" to-layer="585" to-port="1" />
 		<edge from-layer="585" from-port="2" to-layer="586" to-port="1" />
 		<edge from-layer="586" from-port="2" to-layer="587" to-port="1" />
 		<edge from-layer="587" from-port="2" to-layer="588" to-port="0" />
@@ -51689,12 +51403,12 @@
 		<edge from-layer="602" from-port="1" to-layer="605" to-port="0" />
 		<edge from-layer="603" from-port="0" to-layer="604" to-port="0" />
 		<edge from-layer="604" from-port="1" to-layer="605" to-port="1" />
-		<edge from-layer="605" from-port="2" to-layer="608" to-port="0" />
-		<edge from-layer="606" from-port="0" to-layer="607" to-port="0" />
-		<edge from-layer="607" from-port="1" to-layer="608" to-port="1" />
-		<edge from-layer="608" from-port="2" to-layer="610" to-port="0" />
-		<edge from-layer="609" from-port="0" to-layer="610" to-port="1" />
-		<edge from-layer="610" from-port="2" to-layer="611" to-port="1" />
+		<edge from-layer="605" from-port="2" to-layer="607" to-port="0" />
+		<edge from-layer="606" from-port="0" to-layer="607" to-port="1" />
+		<edge from-layer="607" from-port="2" to-layer="609" to-port="0" />
+		<edge from-layer="608" from-port="0" to-layer="609" to-port="1" />
+		<edge from-layer="609" from-port="2" to-layer="610" to-port="0" />
+		<edge from-layer="610" from-port="1" to-layer="611" to-port="1" />
 		<edge from-layer="611" from-port="2" to-layer="612" to-port="1" />
 		<edge from-layer="612" from-port="2" to-layer="614" to-port="0" />
 		<edge from-layer="612" from-port="2" to-layer="618" to-port="0" />
@@ -51710,12 +51424,12 @@
 		<edge from-layer="622" from-port="1" to-layer="625" to-port="0" />
 		<edge from-layer="623" from-port="0" to-layer="624" to-port="0" />
 		<edge from-layer="624" from-port="1" to-layer="625" to-port="1" />
-		<edge from-layer="625" from-port="2" to-layer="628" to-port="0" />
-		<edge from-layer="626" from-port="0" to-layer="627" to-port="0" />
-		<edge from-layer="627" from-port="1" to-layer="628" to-port="1" />
-		<edge from-layer="628" from-port="2" to-layer="630" to-port="0" />
-		<edge from-layer="629" from-port="0" to-layer="630" to-port="1" />
-		<edge from-layer="630" from-port="2" to-layer="631" to-port="1" />
+		<edge from-layer="625" from-port="2" to-layer="627" to-port="0" />
+		<edge from-layer="626" from-port="0" to-layer="627" to-port="1" />
+		<edge from-layer="627" from-port="2" to-layer="629" to-port="0" />
+		<edge from-layer="628" from-port="0" to-layer="629" to-port="1" />
+		<edge from-layer="629" from-port="2" to-layer="630" to-port="0" />
+		<edge from-layer="630" from-port="1" to-layer="631" to-port="1" />
 		<edge from-layer="631" from-port="2" to-layer="632" to-port="1" />
 		<edge from-layer="632" from-port="2" to-layer="633" to-port="1" />
 		<edge from-layer="633" from-port="2" to-layer="634" to-port="0" />
@@ -51739,12 +51453,12 @@
 		<edge from-layer="648" from-port="1" to-layer="651" to-port="0" />
 		<edge from-layer="649" from-port="0" to-layer="650" to-port="0" />
 		<edge from-layer="650" from-port="1" to-layer="651" to-port="1" />
-		<edge from-layer="651" from-port="2" to-layer="654" to-port="0" />
-		<edge from-layer="652" from-port="0" to-layer="653" to-port="0" />
-		<edge from-layer="653" from-port="1" to-layer="654" to-port="1" />
-		<edge from-layer="654" from-port="2" to-layer="656" to-port="0" />
-		<edge from-layer="655" from-port="0" to-layer="656" to-port="1" />
-		<edge from-layer="656" from-port="2" to-layer="657" to-port="1" />
+		<edge from-layer="651" from-port="2" to-layer="653" to-port="0" />
+		<edge from-layer="652" from-port="0" to-layer="653" to-port="1" />
+		<edge from-layer="653" from-port="2" to-layer="655" to-port="0" />
+		<edge from-layer="654" from-port="0" to-layer="655" to-port="1" />
+		<edge from-layer="655" from-port="2" to-layer="656" to-port="0" />
+		<edge from-layer="656" from-port="1" to-layer="657" to-port="1" />
 		<edge from-layer="657" from-port="2" to-layer="658" to-port="1" />
 		<edge from-layer="658" from-port="2" to-layer="660" to-port="0" />
 		<edge from-layer="659" from-port="0" to-layer="660" to-port="1" />
@@ -51760,12 +51474,12 @@
 		<edge from-layer="668" from-port="1" to-layer="671" to-port="0" />
 		<edge from-layer="669" from-port="0" to-layer="670" to-port="0" />
 		<edge from-layer="670" from-port="1" to-layer="671" to-port="1" />
-		<edge from-layer="671" from-port="2" to-layer="674" to-port="0" />
-		<edge from-layer="672" from-port="0" to-layer="673" to-port="0" />
-		<edge from-layer="673" from-port="1" to-layer="674" to-port="1" />
-		<edge from-layer="674" from-port="2" to-layer="676" to-port="0" />
-		<edge from-layer="675" from-port="0" to-layer="676" to-port="1" />
-		<edge from-layer="676" from-port="2" to-layer="677" to-port="1" />
+		<edge from-layer="671" from-port="2" to-layer="673" to-port="0" />
+		<edge from-layer="672" from-port="0" to-layer="673" to-port="1" />
+		<edge from-layer="673" from-port="2" to-layer="675" to-port="0" />
+		<edge from-layer="674" from-port="0" to-layer="675" to-port="1" />
+		<edge from-layer="675" from-port="2" to-layer="676" to-port="0" />
+		<edge from-layer="676" from-port="1" to-layer="677" to-port="1" />
 		<edge from-layer="677" from-port="2" to-layer="679" to-port="0" />
 		<edge from-layer="678" from-port="0" to-layer="679" to-port="1" />
 		<edge from-layer="679" from-port="2" to-layer="681" to-port="0" />
@@ -51781,12 +51495,12 @@
 		<edge from-layer="689" from-port="1" to-layer="692" to-port="0" />
 		<edge from-layer="690" from-port="0" to-layer="691" to-port="0" />
 		<edge from-layer="691" from-port="1" to-layer="692" to-port="1" />
-		<edge from-layer="692" from-port="2" to-layer="695" to-port="0" />
-		<edge from-layer="693" from-port="0" to-layer="694" to-port="0" />
-		<edge from-layer="694" from-port="1" to-layer="695" to-port="1" />
-		<edge from-layer="695" from-port="2" to-layer="697" to-port="0" />
-		<edge from-layer="696" from-port="0" to-layer="697" to-port="1" />
-		<edge from-layer="697" from-port="2" to-layer="698" to-port="1" />
+		<edge from-layer="692" from-port="2" to-layer="694" to-port="0" />
+		<edge from-layer="693" from-port="0" to-layer="694" to-port="1" />
+		<edge from-layer="694" from-port="2" to-layer="696" to-port="0" />
+		<edge from-layer="695" from-port="0" to-layer="696" to-port="1" />
+		<edge from-layer="696" from-port="2" to-layer="697" to-port="0" />
+		<edge from-layer="697" from-port="1" to-layer="698" to-port="1" />
 		<edge from-layer="698" from-port="2" to-layer="699" to-port="1" />
 		<edge from-layer="699" from-port="2" to-layer="701" to-port="0" />
 		<edge from-layer="700" from-port="0" to-layer="701" to-port="1" />
@@ -51813,12 +51527,12 @@
 		<edge from-layer="719" from-port="1" to-layer="722" to-port="0" />
 		<edge from-layer="720" from-port="0" to-layer="721" to-port="0" />
 		<edge from-layer="721" from-port="1" to-layer="722" to-port="1" />
-		<edge from-layer="722" from-port="2" to-layer="725" to-port="0" />
-		<edge from-layer="723" from-port="0" to-layer="724" to-port="0" />
-		<edge from-layer="724" from-port="1" to-layer="725" to-port="1" />
-		<edge from-layer="725" from-port="2" to-layer="727" to-port="0" />
-		<edge from-layer="726" from-port="0" to-layer="727" to-port="1" />
-		<edge from-layer="727" from-port="2" to-layer="728" to-port="1" />
+		<edge from-layer="722" from-port="2" to-layer="724" to-port="0" />
+		<edge from-layer="723" from-port="0" to-layer="724" to-port="1" />
+		<edge from-layer="724" from-port="2" to-layer="726" to-port="0" />
+		<edge from-layer="725" from-port="0" to-layer="726" to-port="1" />
+		<edge from-layer="726" from-port="2" to-layer="727" to-port="0" />
+		<edge from-layer="727" from-port="1" to-layer="728" to-port="1" />
 		<edge from-layer="728" from-port="2" to-layer="729" to-port="1" />
 		<edge from-layer="729" from-port="2" to-layer="730" to-port="1" />
 		<edge from-layer="730" from-port="2" to-layer="731" to-port="0" />
@@ -51840,12 +51554,12 @@
 		<edge from-layer="745" from-port="1" to-layer="748" to-port="0" />
 		<edge from-layer="746" from-port="0" to-layer="747" to-port="0" />
 		<edge from-layer="747" from-port="1" to-layer="748" to-port="1" />
-		<edge from-layer="748" from-port="2" to-layer="751" to-port="0" />
-		<edge from-layer="749" from-port="0" to-layer="750" to-port="0" />
-		<edge from-layer="750" from-port="1" to-layer="751" to-port="1" />
-		<edge from-layer="751" from-port="2" to-layer="753" to-port="0" />
-		<edge from-layer="752" from-port="0" to-layer="753" to-port="1" />
-		<edge from-layer="753" from-port="2" to-layer="754" to-port="1" />
+		<edge from-layer="748" from-port="2" to-layer="750" to-port="0" />
+		<edge from-layer="749" from-port="0" to-layer="750" to-port="1" />
+		<edge from-layer="750" from-port="2" to-layer="752" to-port="0" />
+		<edge from-layer="751" from-port="0" to-layer="752" to-port="1" />
+		<edge from-layer="752" from-port="2" to-layer="753" to-port="0" />
+		<edge from-layer="753" from-port="1" to-layer="754" to-port="1" />
 		<edge from-layer="754" from-port="2" to-layer="755" to-port="1" />
 		<edge from-layer="755" from-port="2" to-layer="761" to-port="0" />
 		<edge from-layer="755" from-port="2" to-layer="757" to-port="0" />
@@ -51861,12 +51575,12 @@
 		<edge from-layer="765" from-port="1" to-layer="768" to-port="0" />
 		<edge from-layer="766" from-port="0" to-layer="767" to-port="0" />
 		<edge from-layer="767" from-port="1" to-layer="768" to-port="1" />
-		<edge from-layer="768" from-port="2" to-layer="771" to-port="0" />
-		<edge from-layer="769" from-port="0" to-layer="770" to-port="0" />
-		<edge from-layer="770" from-port="1" to-layer="771" to-port="1" />
-		<edge from-layer="771" from-port="2" to-layer="773" to-port="0" />
-		<edge from-layer="772" from-port="0" to-layer="773" to-port="1" />
-		<edge from-layer="773" from-port="2" to-layer="774" to-port="1" />
+		<edge from-layer="768" from-port="2" to-layer="770" to-port="0" />
+		<edge from-layer="769" from-port="0" to-layer="770" to-port="1" />
+		<edge from-layer="770" from-port="2" to-layer="772" to-port="0" />
+		<edge from-layer="771" from-port="0" to-layer="772" to-port="1" />
+		<edge from-layer="772" from-port="2" to-layer="773" to-port="0" />
+		<edge from-layer="773" from-port="1" to-layer="774" to-port="1" />
 		<edge from-layer="774" from-port="2" to-layer="775" to-port="1" />
 		<edge from-layer="775" from-port="2" to-layer="776" to-port="1" />
 		<edge from-layer="776" from-port="2" to-layer="777" to-port="0" />
@@ -51890,12 +51604,12 @@
 		<edge from-layer="791" from-port="1" to-layer="794" to-port="0" />
 		<edge from-layer="792" from-port="0" to-layer="793" to-port="0" />
 		<edge from-layer="793" from-port="1" to-layer="794" to-port="1" />
-		<edge from-layer="794" from-port="2" to-layer="797" to-port="0" />
-		<edge from-layer="795" from-port="0" to-layer="796" to-port="0" />
-		<edge from-layer="796" from-port="1" to-layer="797" to-port="1" />
-		<edge from-layer="797" from-port="2" to-layer="799" to-port="0" />
-		<edge from-layer="798" from-port="0" to-layer="799" to-port="1" />
-		<edge from-layer="799" from-port="2" to-layer="800" to-port="1" />
+		<edge from-layer="794" from-port="2" to-layer="796" to-port="0" />
+		<edge from-layer="795" from-port="0" to-layer="796" to-port="1" />
+		<edge from-layer="796" from-port="2" to-layer="798" to-port="0" />
+		<edge from-layer="797" from-port="0" to-layer="798" to-port="1" />
+		<edge from-layer="798" from-port="2" to-layer="799" to-port="0" />
+		<edge from-layer="799" from-port="1" to-layer="800" to-port="1" />
 		<edge from-layer="800" from-port="2" to-layer="801" to-port="1" />
 		<edge from-layer="801" from-port="2" to-layer="803" to-port="0" />
 		<edge from-layer="802" from-port="0" to-layer="803" to-port="1" />
@@ -51911,12 +51625,12 @@
 		<edge from-layer="811" from-port="1" to-layer="814" to-port="0" />
 		<edge from-layer="812" from-port="0" to-layer="813" to-port="0" />
 		<edge from-layer="813" from-port="1" to-layer="814" to-port="1" />
-		<edge from-layer="814" from-port="2" to-layer="817" to-port="0" />
-		<edge from-layer="815" from-port="0" to-layer="816" to-port="0" />
-		<edge from-layer="816" from-port="1" to-layer="817" to-port="1" />
-		<edge from-layer="817" from-port="2" to-layer="819" to-port="0" />
-		<edge from-layer="818" from-port="0" to-layer="819" to-port="1" />
-		<edge from-layer="819" from-port="2" to-layer="820" to-port="1" />
+		<edge from-layer="814" from-port="2" to-layer="816" to-port="0" />
+		<edge from-layer="815" from-port="0" to-layer="816" to-port="1" />
+		<edge from-layer="816" from-port="2" to-layer="818" to-port="0" />
+		<edge from-layer="817" from-port="0" to-layer="818" to-port="1" />
+		<edge from-layer="818" from-port="2" to-layer="819" to-port="0" />
+		<edge from-layer="819" from-port="1" to-layer="820" to-port="1" />
 		<edge from-layer="820" from-port="2" to-layer="822" to-port="0" />
 		<edge from-layer="821" from-port="0" to-layer="822" to-port="1" />
 		<edge from-layer="822" from-port="2" to-layer="824" to-port="0" />
@@ -51932,12 +51646,12 @@
 		<edge from-layer="832" from-port="1" to-layer="835" to-port="0" />
 		<edge from-layer="833" from-port="0" to-layer="834" to-port="0" />
 		<edge from-layer="834" from-port="1" to-layer="835" to-port="1" />
-		<edge from-layer="835" from-port="2" to-layer="838" to-port="0" />
-		<edge from-layer="836" from-port="0" to-layer="837" to-port="0" />
-		<edge from-layer="837" from-port="1" to-layer="838" to-port="1" />
-		<edge from-layer="838" from-port="2" to-layer="840" to-port="0" />
-		<edge from-layer="839" from-port="0" to-layer="840" to-port="1" />
-		<edge from-layer="840" from-port="2" to-layer="841" to-port="1" />
+		<edge from-layer="835" from-port="2" to-layer="837" to-port="0" />
+		<edge from-layer="836" from-port="0" to-layer="837" to-port="1" />
+		<edge from-layer="837" from-port="2" to-layer="839" to-port="0" />
+		<edge from-layer="838" from-port="0" to-layer="839" to-port="1" />
+		<edge from-layer="839" from-port="2" to-layer="840" to-port="0" />
+		<edge from-layer="840" from-port="1" to-layer="841" to-port="1" />
 		<edge from-layer="841" from-port="2" to-layer="842" to-port="1" />
 		<edge from-layer="842" from-port="2" to-layer="844" to-port="0" />
 		<edge from-layer="843" from-port="0" to-layer="844" to-port="1" />
@@ -51946,11 +51660,11 @@
 		<edge from-layer="846" from-port="2" to-layer="847" to-port="1" />
 		<edge from-layer="847" from-port="2" to-layer="849" to-port="0" />
 		<edge from-layer="848" from-port="0" to-layer="849" to-port="1" />
-		<edge from-layer="849" from-port="2" to-layer="850" to-port="0" />
 		<edge from-layer="849" from-port="2" to-layer="860" to-port="0" />
+		<edge from-layer="849" from-port="2" to-layer="850" to-port="0" />
 		<edge from-layer="850" from-port="1" to-layer="857" to-port="0" />
-		<edge from-layer="851" from-port="0" to-layer="854" to-port="0" />
 		<edge from-layer="851" from-port="0" to-layer="857" to-port="1" />
+		<edge from-layer="851" from-port="0" to-layer="854" to-port="0" />
 		<edge from-layer="852" from-port="0" to-layer="857" to-port="2" />
 		<edge from-layer="853" from-port="0" to-layer="855" to-port="0" />
 		<edge from-layer="854" from-port="1" to-layer="855" to-port="1" />
@@ -51964,12 +51678,12 @@
 		<edge from-layer="862" from-port="1" to-layer="865" to-port="0" />
 		<edge from-layer="863" from-port="0" to-layer="864" to-port="0" />
 		<edge from-layer="864" from-port="1" to-layer="865" to-port="1" />
-		<edge from-layer="865" from-port="2" to-layer="868" to-port="0" />
-		<edge from-layer="866" from-port="0" to-layer="867" to-port="0" />
-		<edge from-layer="867" from-port="1" to-layer="868" to-port="1" />
-		<edge from-layer="868" from-port="2" to-layer="870" to-port="0" />
-		<edge from-layer="869" from-port="0" to-layer="870" to-port="1" />
-		<edge from-layer="870" from-port="2" to-layer="871" to-port="1" />
+		<edge from-layer="865" from-port="2" to-layer="867" to-port="0" />
+		<edge from-layer="866" from-port="0" to-layer="867" to-port="1" />
+		<edge from-layer="867" from-port="2" to-layer="869" to-port="0" />
+		<edge from-layer="868" from-port="0" to-layer="869" to-port="1" />
+		<edge from-layer="869" from-port="2" to-layer="870" to-port="0" />
+		<edge from-layer="870" from-port="1" to-layer="871" to-port="1" />
 		<edge from-layer="871" from-port="2" to-layer="872" to-port="1" />
 		<edge from-layer="872" from-port="2" to-layer="873" to-port="1" />
 		<edge from-layer="873" from-port="2" to-layer="874" to-port="0" />
@@ -51991,12 +51705,12 @@
 		<edge from-layer="888" from-port="1" to-layer="891" to-port="0" />
 		<edge from-layer="889" from-port="0" to-layer="890" to-port="0" />
 		<edge from-layer="890" from-port="1" to-layer="891" to-port="1" />
-		<edge from-layer="891" from-port="2" to-layer="894" to-port="0" />
-		<edge from-layer="892" from-port="0" to-layer="893" to-port="0" />
-		<edge from-layer="893" from-port="1" to-layer="894" to-port="1" />
-		<edge from-layer="894" from-port="2" to-layer="896" to-port="0" />
-		<edge from-layer="895" from-port="0" to-layer="896" to-port="1" />
-		<edge from-layer="896" from-port="2" to-layer="897" to-port="1" />
+		<edge from-layer="891" from-port="2" to-layer="893" to-port="0" />
+		<edge from-layer="892" from-port="0" to-layer="893" to-port="1" />
+		<edge from-layer="893" from-port="2" to-layer="895" to-port="0" />
+		<edge from-layer="894" from-port="0" to-layer="895" to-port="1" />
+		<edge from-layer="895" from-port="2" to-layer="896" to-port="0" />
+		<edge from-layer="896" from-port="1" to-layer="897" to-port="1" />
 		<edge from-layer="897" from-port="2" to-layer="898" to-port="1" />
 		<edge from-layer="898" from-port="2" to-layer="900" to-port="0" />
 		<edge from-layer="898" from-port="2" to-layer="904" to-port="0" />
@@ -52012,12 +51726,12 @@
 		<edge from-layer="908" from-port="1" to-layer="911" to-port="0" />
 		<edge from-layer="909" from-port="0" to-layer="910" to-port="0" />
 		<edge from-layer="910" from-port="1" to-layer="911" to-port="1" />
-		<edge from-layer="911" from-port="2" to-layer="914" to-port="0" />
-		<edge from-layer="912" from-port="0" to-layer="913" to-port="0" />
-		<edge from-layer="913" from-port="1" to-layer="914" to-port="1" />
-		<edge from-layer="914" from-port="2" to-layer="916" to-port="0" />
-		<edge from-layer="915" from-port="0" to-layer="916" to-port="1" />
-		<edge from-layer="916" from-port="2" to-layer="917" to-port="1" />
+		<edge from-layer="911" from-port="2" to-layer="913" to-port="0" />
+		<edge from-layer="912" from-port="0" to-layer="913" to-port="1" />
+		<edge from-layer="913" from-port="2" to-layer="915" to-port="0" />
+		<edge from-layer="914" from-port="0" to-layer="915" to-port="1" />
+		<edge from-layer="915" from-port="2" to-layer="916" to-port="0" />
+		<edge from-layer="916" from-port="1" to-layer="917" to-port="1" />
 		<edge from-layer="917" from-port="2" to-layer="918" to-port="1" />
 		<edge from-layer="918" from-port="2" to-layer="919" to-port="1" />
 		<edge from-layer="919" from-port="2" to-layer="920" to-port="0" />
@@ -52041,12 +51755,12 @@
 		<edge from-layer="934" from-port="1" to-layer="937" to-port="0" />
 		<edge from-layer="935" from-port="0" to-layer="936" to-port="0" />
 		<edge from-layer="936" from-port="1" to-layer="937" to-port="1" />
-		<edge from-layer="937" from-port="2" to-layer="940" to-port="0" />
-		<edge from-layer="938" from-port="0" to-layer="939" to-port="0" />
-		<edge from-layer="939" from-port="1" to-layer="940" to-port="1" />
-		<edge from-layer="940" from-port="2" to-layer="942" to-port="0" />
-		<edge from-layer="941" from-port="0" to-layer="942" to-port="1" />
-		<edge from-layer="942" from-port="2" to-layer="943" to-port="1" />
+		<edge from-layer="937" from-port="2" to-layer="939" to-port="0" />
+		<edge from-layer="938" from-port="0" to-layer="939" to-port="1" />
+		<edge from-layer="939" from-port="2" to-layer="941" to-port="0" />
+		<edge from-layer="940" from-port="0" to-layer="941" to-port="1" />
+		<edge from-layer="941" from-port="2" to-layer="942" to-port="0" />
+		<edge from-layer="942" from-port="1" to-layer="943" to-port="1" />
 		<edge from-layer="943" from-port="2" to-layer="944" to-port="1" />
 		<edge from-layer="944" from-port="2" to-layer="946" to-port="0" />
 		<edge from-layer="945" from-port="0" to-layer="946" to-port="1" />
@@ -52062,12 +51776,12 @@
 		<edge from-layer="954" from-port="1" to-layer="957" to-port="0" />
 		<edge from-layer="955" from-port="0" to-layer="956" to-port="0" />
 		<edge from-layer="956" from-port="1" to-layer="957" to-port="1" />
-		<edge from-layer="957" from-port="2" to-layer="960" to-port="0" />
-		<edge from-layer="958" from-port="0" to-layer="959" to-port="0" />
-		<edge from-layer="959" from-port="1" to-layer="960" to-port="1" />
-		<edge from-layer="960" from-port="2" to-layer="962" to-port="0" />
-		<edge from-layer="961" from-port="0" to-layer="962" to-port="1" />
-		<edge from-layer="962" from-port="2" to-layer="963" to-port="1" />
+		<edge from-layer="957" from-port="2" to-layer="959" to-port="0" />
+		<edge from-layer="958" from-port="0" to-layer="959" to-port="1" />
+		<edge from-layer="959" from-port="2" to-layer="961" to-port="0" />
+		<edge from-layer="960" from-port="0" to-layer="961" to-port="1" />
+		<edge from-layer="961" from-port="2" to-layer="962" to-port="0" />
+		<edge from-layer="962" from-port="1" to-layer="963" to-port="1" />
 		<edge from-layer="963" from-port="2" to-layer="965" to-port="0" />
 		<edge from-layer="964" from-port="0" to-layer="965" to-port="1" />
 		<edge from-layer="965" from-port="2" to-layer="967" to-port="0" />
@@ -52083,12 +51797,12 @@
 		<edge from-layer="975" from-port="1" to-layer="978" to-port="0" />
 		<edge from-layer="976" from-port="0" to-layer="977" to-port="0" />
 		<edge from-layer="977" from-port="1" to-layer="978" to-port="1" />
-		<edge from-layer="978" from-port="2" to-layer="981" to-port="0" />
-		<edge from-layer="979" from-port="0" to-layer="980" to-port="0" />
-		<edge from-layer="980" from-port="1" to-layer="981" to-port="1" />
-		<edge from-layer="981" from-port="2" to-layer="983" to-port="0" />
-		<edge from-layer="982" from-port="0" to-layer="983" to-port="1" />
-		<edge from-layer="983" from-port="2" to-layer="984" to-port="1" />
+		<edge from-layer="978" from-port="2" to-layer="980" to-port="0" />
+		<edge from-layer="979" from-port="0" to-layer="980" to-port="1" />
+		<edge from-layer="980" from-port="2" to-layer="982" to-port="0" />
+		<edge from-layer="981" from-port="0" to-layer="982" to-port="1" />
+		<edge from-layer="982" from-port="2" to-layer="983" to-port="0" />
+		<edge from-layer="983" from-port="1" to-layer="984" to-port="1" />
 		<edge from-layer="984" from-port="2" to-layer="985" to-port="1" />
 		<edge from-layer="985" from-port="2" to-layer="987" to-port="0" />
 		<edge from-layer="986" from-port="0" to-layer="987" to-port="1" />
@@ -52115,12 +51829,12 @@
 		<edge from-layer="1005" from-port="1" to-layer="1008" to-port="0" />
 		<edge from-layer="1006" from-port="0" to-layer="1007" to-port="0" />
 		<edge from-layer="1007" from-port="1" to-layer="1008" to-port="1" />
-		<edge from-layer="1008" from-port="2" to-layer="1011" to-port="0" />
-		<edge from-layer="1009" from-port="0" to-layer="1010" to-port="0" />
-		<edge from-layer="1010" from-port="1" to-layer="1011" to-port="1" />
-		<edge from-layer="1011" from-port="2" to-layer="1013" to-port="0" />
-		<edge from-layer="1012" from-port="0" to-layer="1013" to-port="1" />
-		<edge from-layer="1013" from-port="2" to-layer="1014" to-port="1" />
+		<edge from-layer="1008" from-port="2" to-layer="1010" to-port="0" />
+		<edge from-layer="1009" from-port="0" to-layer="1010" to-port="1" />
+		<edge from-layer="1010" from-port="2" to-layer="1012" to-port="0" />
+		<edge from-layer="1011" from-port="0" to-layer="1012" to-port="1" />
+		<edge from-layer="1012" from-port="2" to-layer="1013" to-port="0" />
+		<edge from-layer="1013" from-port="1" to-layer="1014" to-port="1" />
 		<edge from-layer="1014" from-port="2" to-layer="1015" to-port="1" />
 		<edge from-layer="1015" from-port="2" to-layer="1016" to-port="1" />
 		<edge from-layer="1016" from-port="2" to-layer="1017" to-port="0" />
@@ -52142,12 +51856,12 @@
 		<edge from-layer="1031" from-port="1" to-layer="1034" to-port="0" />
 		<edge from-layer="1032" from-port="0" to-layer="1033" to-port="0" />
 		<edge from-layer="1033" from-port="1" to-layer="1034" to-port="1" />
-		<edge from-layer="1034" from-port="2" to-layer="1037" to-port="0" />
-		<edge from-layer="1035" from-port="0" to-layer="1036" to-port="0" />
-		<edge from-layer="1036" from-port="1" to-layer="1037" to-port="1" />
-		<edge from-layer="1037" from-port="2" to-layer="1039" to-port="0" />
-		<edge from-layer="1038" from-port="0" to-layer="1039" to-port="1" />
-		<edge from-layer="1039" from-port="2" to-layer="1040" to-port="1" />
+		<edge from-layer="1034" from-port="2" to-layer="1036" to-port="0" />
+		<edge from-layer="1035" from-port="0" to-layer="1036" to-port="1" />
+		<edge from-layer="1036" from-port="2" to-layer="1038" to-port="0" />
+		<edge from-layer="1037" from-port="0" to-layer="1038" to-port="1" />
+		<edge from-layer="1038" from-port="2" to-layer="1039" to-port="0" />
+		<edge from-layer="1039" from-port="1" to-layer="1040" to-port="1" />
 		<edge from-layer="1040" from-port="2" to-layer="1041" to-port="1" />
 		<edge from-layer="1041" from-port="2" to-layer="1043" to-port="0" />
 		<edge from-layer="1041" from-port="2" to-layer="1047" to-port="0" />
@@ -52163,12 +51877,12 @@
 		<edge from-layer="1051" from-port="1" to-layer="1054" to-port="0" />
 		<edge from-layer="1052" from-port="0" to-layer="1053" to-port="0" />
 		<edge from-layer="1053" from-port="1" to-layer="1054" to-port="1" />
-		<edge from-layer="1054" from-port="2" to-layer="1057" to-port="0" />
-		<edge from-layer="1055" from-port="0" to-layer="1056" to-port="0" />
-		<edge from-layer="1056" from-port="1" to-layer="1057" to-port="1" />
-		<edge from-layer="1057" from-port="2" to-layer="1059" to-port="0" />
-		<edge from-layer="1058" from-port="0" to-layer="1059" to-port="1" />
-		<edge from-layer="1059" from-port="2" to-layer="1060" to-port="1" />
+		<edge from-layer="1054" from-port="2" to-layer="1056" to-port="0" />
+		<edge from-layer="1055" from-port="0" to-layer="1056" to-port="1" />
+		<edge from-layer="1056" from-port="2" to-layer="1058" to-port="0" />
+		<edge from-layer="1057" from-port="0" to-layer="1058" to-port="1" />
+		<edge from-layer="1058" from-port="2" to-layer="1059" to-port="0" />
+		<edge from-layer="1059" from-port="1" to-layer="1060" to-port="1" />
 		<edge from-layer="1060" from-port="2" to-layer="1061" to-port="1" />
 		<edge from-layer="1061" from-port="2" to-layer="1062" to-port="1" />
 		<edge from-layer="1062" from-port="2" to-layer="1063" to-port="0" />
@@ -52192,12 +51906,12 @@
 		<edge from-layer="1077" from-port="1" to-layer="1080" to-port="0" />
 		<edge from-layer="1078" from-port="0" to-layer="1079" to-port="0" />
 		<edge from-layer="1079" from-port="1" to-layer="1080" to-port="1" />
-		<edge from-layer="1080" from-port="2" to-layer="1083" to-port="0" />
-		<edge from-layer="1081" from-port="0" to-layer="1082" to-port="0" />
-		<edge from-layer="1082" from-port="1" to-layer="1083" to-port="1" />
-		<edge from-layer="1083" from-port="2" to-layer="1085" to-port="0" />
-		<edge from-layer="1084" from-port="0" to-layer="1085" to-port="1" />
-		<edge from-layer="1085" from-port="2" to-layer="1086" to-port="1" />
+		<edge from-layer="1080" from-port="2" to-layer="1082" to-port="0" />
+		<edge from-layer="1081" from-port="0" to-layer="1082" to-port="1" />
+		<edge from-layer="1082" from-port="2" to-layer="1084" to-port="0" />
+		<edge from-layer="1083" from-port="0" to-layer="1084" to-port="1" />
+		<edge from-layer="1084" from-port="2" to-layer="1085" to-port="0" />
+		<edge from-layer="1085" from-port="1" to-layer="1086" to-port="1" />
 		<edge from-layer="1086" from-port="2" to-layer="1087" to-port="1" />
 		<edge from-layer="1087" from-port="2" to-layer="1089" to-port="0" />
 		<edge from-layer="1088" from-port="0" to-layer="1089" to-port="1" />
@@ -52213,12 +51927,12 @@
 		<edge from-layer="1097" from-port="1" to-layer="1100" to-port="0" />
 		<edge from-layer="1098" from-port="0" to-layer="1099" to-port="0" />
 		<edge from-layer="1099" from-port="1" to-layer="1100" to-port="1" />
-		<edge from-layer="1100" from-port="2" to-layer="1103" to-port="0" />
-		<edge from-layer="1101" from-port="0" to-layer="1102" to-port="0" />
-		<edge from-layer="1102" from-port="1" to-layer="1103" to-port="1" />
-		<edge from-layer="1103" from-port="2" to-layer="1105" to-port="0" />
-		<edge from-layer="1104" from-port="0" to-layer="1105" to-port="1" />
-		<edge from-layer="1105" from-port="2" to-layer="1106" to-port="1" />
+		<edge from-layer="1100" from-port="2" to-layer="1102" to-port="0" />
+		<edge from-layer="1101" from-port="0" to-layer="1102" to-port="1" />
+		<edge from-layer="1102" from-port="2" to-layer="1104" to-port="0" />
+		<edge from-layer="1103" from-port="0" to-layer="1104" to-port="1" />
+		<edge from-layer="1104" from-port="2" to-layer="1105" to-port="0" />
+		<edge from-layer="1105" from-port="1" to-layer="1106" to-port="1" />
 		<edge from-layer="1106" from-port="2" to-layer="1108" to-port="0" />
 		<edge from-layer="1107" from-port="0" to-layer="1108" to-port="1" />
 		<edge from-layer="1108" from-port="2" to-layer="1110" to-port="0" />
@@ -52234,12 +51948,12 @@
 		<edge from-layer="1118" from-port="1" to-layer="1121" to-port="0" />
 		<edge from-layer="1119" from-port="0" to-layer="1120" to-port="0" />
 		<edge from-layer="1120" from-port="1" to-layer="1121" to-port="1" />
-		<edge from-layer="1121" from-port="2" to-layer="1124" to-port="0" />
-		<edge from-layer="1122" from-port="0" to-layer="1123" to-port="0" />
-		<edge from-layer="1123" from-port="1" to-layer="1124" to-port="1" />
-		<edge from-layer="1124" from-port="2" to-layer="1126" to-port="0" />
-		<edge from-layer="1125" from-port="0" to-layer="1126" to-port="1" />
-		<edge from-layer="1126" from-port="2" to-layer="1127" to-port="1" />
+		<edge from-layer="1121" from-port="2" to-layer="1123" to-port="0" />
+		<edge from-layer="1122" from-port="0" to-layer="1123" to-port="1" />
+		<edge from-layer="1123" from-port="2" to-layer="1125" to-port="0" />
+		<edge from-layer="1124" from-port="0" to-layer="1125" to-port="1" />
+		<edge from-layer="1125" from-port="2" to-layer="1126" to-port="0" />
+		<edge from-layer="1126" from-port="1" to-layer="1127" to-port="1" />
 		<edge from-layer="1127" from-port="2" to-layer="1128" to-port="1" />
 		<edge from-layer="1128" from-port="2" to-layer="1130" to-port="0" />
 		<edge from-layer="1129" from-port="0" to-layer="1130" to-port="1" />
@@ -52266,12 +51980,12 @@
 		<edge from-layer="1148" from-port="1" to-layer="1151" to-port="0" />
 		<edge from-layer="1149" from-port="0" to-layer="1150" to-port="0" />
 		<edge from-layer="1150" from-port="1" to-layer="1151" to-port="1" />
-		<edge from-layer="1151" from-port="2" to-layer="1154" to-port="0" />
-		<edge from-layer="1152" from-port="0" to-layer="1153" to-port="0" />
-		<edge from-layer="1153" from-port="1" to-layer="1154" to-port="1" />
-		<edge from-layer="1154" from-port="2" to-layer="1156" to-port="0" />
-		<edge from-layer="1155" from-port="0" to-layer="1156" to-port="1" />
-		<edge from-layer="1156" from-port="2" to-layer="1157" to-port="1" />
+		<edge from-layer="1151" from-port="2" to-layer="1153" to-port="0" />
+		<edge from-layer="1152" from-port="0" to-layer="1153" to-port="1" />
+		<edge from-layer="1153" from-port="2" to-layer="1155" to-port="0" />
+		<edge from-layer="1154" from-port="0" to-layer="1155" to-port="1" />
+		<edge from-layer="1155" from-port="2" to-layer="1156" to-port="0" />
+		<edge from-layer="1156" from-port="1" to-layer="1157" to-port="1" />
 		<edge from-layer="1157" from-port="2" to-layer="1158" to-port="1" />
 		<edge from-layer="1158" from-port="2" to-layer="1159" to-port="1" />
 		<edge from-layer="1159" from-port="2" to-layer="1160" to-port="0" />
@@ -52293,12 +52007,12 @@
 		<edge from-layer="1174" from-port="1" to-layer="1177" to-port="0" />
 		<edge from-layer="1175" from-port="0" to-layer="1176" to-port="0" />
 		<edge from-layer="1176" from-port="1" to-layer="1177" to-port="1" />
-		<edge from-layer="1177" from-port="2" to-layer="1180" to-port="0" />
-		<edge from-layer="1178" from-port="0" to-layer="1179" to-port="0" />
-		<edge from-layer="1179" from-port="1" to-layer="1180" to-port="1" />
-		<edge from-layer="1180" from-port="2" to-layer="1182" to-port="0" />
-		<edge from-layer="1181" from-port="0" to-layer="1182" to-port="1" />
-		<edge from-layer="1182" from-port="2" to-layer="1183" to-port="1" />
+		<edge from-layer="1177" from-port="2" to-layer="1179" to-port="0" />
+		<edge from-layer="1178" from-port="0" to-layer="1179" to-port="1" />
+		<edge from-layer="1179" from-port="2" to-layer="1181" to-port="0" />
+		<edge from-layer="1180" from-port="0" to-layer="1181" to-port="1" />
+		<edge from-layer="1181" from-port="2" to-layer="1182" to-port="0" />
+		<edge from-layer="1182" from-port="1" to-layer="1183" to-port="1" />
 		<edge from-layer="1183" from-port="2" to-layer="1184" to-port="1" />
 		<edge from-layer="1184" from-port="2" to-layer="1190" to-port="0" />
 		<edge from-layer="1184" from-port="2" to-layer="1186" to-port="0" />
@@ -52314,12 +52028,12 @@
 		<edge from-layer="1194" from-port="1" to-layer="1197" to-port="0" />
 		<edge from-layer="1195" from-port="0" to-layer="1196" to-port="0" />
 		<edge from-layer="1196" from-port="1" to-layer="1197" to-port="1" />
-		<edge from-layer="1197" from-port="2" to-layer="1200" to-port="0" />
-		<edge from-layer="1198" from-port="0" to-layer="1199" to-port="0" />
-		<edge from-layer="1199" from-port="1" to-layer="1200" to-port="1" />
-		<edge from-layer="1200" from-port="2" to-layer="1202" to-port="0" />
-		<edge from-layer="1201" from-port="0" to-layer="1202" to-port="1" />
-		<edge from-layer="1202" from-port="2" to-layer="1203" to-port="1" />
+		<edge from-layer="1197" from-port="2" to-layer="1199" to-port="0" />
+		<edge from-layer="1198" from-port="0" to-layer="1199" to-port="1" />
+		<edge from-layer="1199" from-port="2" to-layer="1201" to-port="0" />
+		<edge from-layer="1200" from-port="0" to-layer="1201" to-port="1" />
+		<edge from-layer="1201" from-port="2" to-layer="1202" to-port="0" />
+		<edge from-layer="1202" from-port="1" to-layer="1203" to-port="1" />
 		<edge from-layer="1203" from-port="2" to-layer="1204" to-port="1" />
 		<edge from-layer="1204" from-port="2" to-layer="1205" to-port="1" />
 		<edge from-layer="1205" from-port="2" to-layer="1206" to-port="0" />
@@ -52343,12 +52057,12 @@
 		<edge from-layer="1220" from-port="1" to-layer="1223" to-port="0" />
 		<edge from-layer="1221" from-port="0" to-layer="1222" to-port="0" />
 		<edge from-layer="1222" from-port="1" to-layer="1223" to-port="1" />
-		<edge from-layer="1223" from-port="2" to-layer="1226" to-port="0" />
-		<edge from-layer="1224" from-port="0" to-layer="1225" to-port="0" />
-		<edge from-layer="1225" from-port="1" to-layer="1226" to-port="1" />
-		<edge from-layer="1226" from-port="2" to-layer="1228" to-port="0" />
-		<edge from-layer="1227" from-port="0" to-layer="1228" to-port="1" />
-		<edge from-layer="1228" from-port="2" to-layer="1229" to-port="1" />
+		<edge from-layer="1223" from-port="2" to-layer="1225" to-port="0" />
+		<edge from-layer="1224" from-port="0" to-layer="1225" to-port="1" />
+		<edge from-layer="1225" from-port="2" to-layer="1227" to-port="0" />
+		<edge from-layer="1226" from-port="0" to-layer="1227" to-port="1" />
+		<edge from-layer="1227" from-port="2" to-layer="1228" to-port="0" />
+		<edge from-layer="1228" from-port="1" to-layer="1229" to-port="1" />
 		<edge from-layer="1229" from-port="2" to-layer="1230" to-port="1" />
 		<edge from-layer="1230" from-port="2" to-layer="1232" to-port="0" />
 		<edge from-layer="1231" from-port="0" to-layer="1232" to-port="1" />
@@ -52364,12 +52078,12 @@
 		<edge from-layer="1240" from-port="1" to-layer="1243" to-port="0" />
 		<edge from-layer="1241" from-port="0" to-layer="1242" to-port="0" />
 		<edge from-layer="1242" from-port="1" to-layer="1243" to-port="1" />
-		<edge from-layer="1243" from-port="2" to-layer="1246" to-port="0" />
-		<edge from-layer="1244" from-port="0" to-layer="1245" to-port="0" />
-		<edge from-layer="1245" from-port="1" to-layer="1246" to-port="1" />
-		<edge from-layer="1246" from-port="2" to-layer="1248" to-port="0" />
-		<edge from-layer="1247" from-port="0" to-layer="1248" to-port="1" />
-		<edge from-layer="1248" from-port="2" to-layer="1249" to-port="1" />
+		<edge from-layer="1243" from-port="2" to-layer="1245" to-port="0" />
+		<edge from-layer="1244" from-port="0" to-layer="1245" to-port="1" />
+		<edge from-layer="1245" from-port="2" to-layer="1247" to-port="0" />
+		<edge from-layer="1246" from-port="0" to-layer="1247" to-port="1" />
+		<edge from-layer="1247" from-port="2" to-layer="1248" to-port="0" />
+		<edge from-layer="1248" from-port="1" to-layer="1249" to-port="1" />
 		<edge from-layer="1249" from-port="2" to-layer="1251" to-port="0" />
 		<edge from-layer="1250" from-port="0" to-layer="1251" to-port="1" />
 		<edge from-layer="1251" from-port="2" to-layer="1253" to-port="0" />
@@ -52385,12 +52099,12 @@
 		<edge from-layer="1261" from-port="1" to-layer="1264" to-port="0" />
 		<edge from-layer="1262" from-port="0" to-layer="1263" to-port="0" />
 		<edge from-layer="1263" from-port="1" to-layer="1264" to-port="1" />
-		<edge from-layer="1264" from-port="2" to-layer="1267" to-port="0" />
-		<edge from-layer="1265" from-port="0" to-layer="1266" to-port="0" />
-		<edge from-layer="1266" from-port="1" to-layer="1267" to-port="1" />
-		<edge from-layer="1267" from-port="2" to-layer="1269" to-port="0" />
-		<edge from-layer="1268" from-port="0" to-layer="1269" to-port="1" />
-		<edge from-layer="1269" from-port="2" to-layer="1270" to-port="1" />
+		<edge from-layer="1264" from-port="2" to-layer="1266" to-port="0" />
+		<edge from-layer="1265" from-port="0" to-layer="1266" to-port="1" />
+		<edge from-layer="1266" from-port="2" to-layer="1268" to-port="0" />
+		<edge from-layer="1267" from-port="0" to-layer="1268" to-port="1" />
+		<edge from-layer="1268" from-port="2" to-layer="1269" to-port="0" />
+		<edge from-layer="1269" from-port="1" to-layer="1270" to-port="1" />
 		<edge from-layer="1270" from-port="2" to-layer="1271" to-port="1" />
 		<edge from-layer="1271" from-port="2" to-layer="1273" to-port="0" />
 		<edge from-layer="1272" from-port="0" to-layer="1273" to-port="1" />
@@ -52417,12 +52131,12 @@
 		<edge from-layer="1291" from-port="1" to-layer="1294" to-port="0" />
 		<edge from-layer="1292" from-port="0" to-layer="1293" to-port="0" />
 		<edge from-layer="1293" from-port="1" to-layer="1294" to-port="1" />
-		<edge from-layer="1294" from-port="2" to-layer="1297" to-port="0" />
-		<edge from-layer="1295" from-port="0" to-layer="1296" to-port="0" />
-		<edge from-layer="1296" from-port="1" to-layer="1297" to-port="1" />
-		<edge from-layer="1297" from-port="2" to-layer="1299" to-port="0" />
-		<edge from-layer="1298" from-port="0" to-layer="1299" to-port="1" />
-		<edge from-layer="1299" from-port="2" to-layer="1300" to-port="1" />
+		<edge from-layer="1294" from-port="2" to-layer="1296" to-port="0" />
+		<edge from-layer="1295" from-port="0" to-layer="1296" to-port="1" />
+		<edge from-layer="1296" from-port="2" to-layer="1298" to-port="0" />
+		<edge from-layer="1297" from-port="0" to-layer="1298" to-port="1" />
+		<edge from-layer="1298" from-port="2" to-layer="1299" to-port="0" />
+		<edge from-layer="1299" from-port="1" to-layer="1300" to-port="1" />
 		<edge from-layer="1300" from-port="2" to-layer="1301" to-port="1" />
 		<edge from-layer="1301" from-port="2" to-layer="1302" to-port="1" />
 		<edge from-layer="1302" from-port="2" to-layer="1303" to-port="0" />
@@ -52444,12 +52158,12 @@
 		<edge from-layer="1317" from-port="1" to-layer="1320" to-port="0" />
 		<edge from-layer="1318" from-port="0" to-layer="1319" to-port="0" />
 		<edge from-layer="1319" from-port="1" to-layer="1320" to-port="1" />
-		<edge from-layer="1320" from-port="2" to-layer="1323" to-port="0" />
-		<edge from-layer="1321" from-port="0" to-layer="1322" to-port="0" />
-		<edge from-layer="1322" from-port="1" to-layer="1323" to-port="1" />
-		<edge from-layer="1323" from-port="2" to-layer="1325" to-port="0" />
-		<edge from-layer="1324" from-port="0" to-layer="1325" to-port="1" />
-		<edge from-layer="1325" from-port="2" to-layer="1326" to-port="1" />
+		<edge from-layer="1320" from-port="2" to-layer="1322" to-port="0" />
+		<edge from-layer="1321" from-port="0" to-layer="1322" to-port="1" />
+		<edge from-layer="1322" from-port="2" to-layer="1324" to-port="0" />
+		<edge from-layer="1323" from-port="0" to-layer="1324" to-port="1" />
+		<edge from-layer="1324" from-port="2" to-layer="1325" to-port="0" />
+		<edge from-layer="1325" from-port="1" to-layer="1326" to-port="1" />
 		<edge from-layer="1326" from-port="2" to-layer="1327" to-port="1" />
 		<edge from-layer="1327" from-port="2" to-layer="1333" to-port="0" />
 		<edge from-layer="1327" from-port="2" to-layer="1329" to-port="0" />
@@ -52465,12 +52179,12 @@
 		<edge from-layer="1337" from-port="1" to-layer="1340" to-port="0" />
 		<edge from-layer="1338" from-port="0" to-layer="1339" to-port="0" />
 		<edge from-layer="1339" from-port="1" to-layer="1340" to-port="1" />
-		<edge from-layer="1340" from-port="2" to-layer="1343" to-port="0" />
-		<edge from-layer="1341" from-port="0" to-layer="1342" to-port="0" />
-		<edge from-layer="1342" from-port="1" to-layer="1343" to-port="1" />
-		<edge from-layer="1343" from-port="2" to-layer="1345" to-port="0" />
-		<edge from-layer="1344" from-port="0" to-layer="1345" to-port="1" />
-		<edge from-layer="1345" from-port="2" to-layer="1346" to-port="1" />
+		<edge from-layer="1340" from-port="2" to-layer="1342" to-port="0" />
+		<edge from-layer="1341" from-port="0" to-layer="1342" to-port="1" />
+		<edge from-layer="1342" from-port="2" to-layer="1344" to-port="0" />
+		<edge from-layer="1343" from-port="0" to-layer="1344" to-port="1" />
+		<edge from-layer="1344" from-port="2" to-layer="1345" to-port="0" />
+		<edge from-layer="1345" from-port="1" to-layer="1346" to-port="1" />
 		<edge from-layer="1346" from-port="2" to-layer="1347" to-port="1" />
 		<edge from-layer="1347" from-port="2" to-layer="1348" to-port="1" />
 		<edge from-layer="1348" from-port="2" to-layer="1349" to-port="0" />
@@ -52494,12 +52208,12 @@
 		<edge from-layer="1363" from-port="1" to-layer="1366" to-port="0" />
 		<edge from-layer="1364" from-port="0" to-layer="1365" to-port="0" />
 		<edge from-layer="1365" from-port="1" to-layer="1366" to-port="1" />
-		<edge from-layer="1366" from-port="2" to-layer="1369" to-port="0" />
-		<edge from-layer="1367" from-port="0" to-layer="1368" to-port="0" />
-		<edge from-layer="1368" from-port="1" to-layer="1369" to-port="1" />
-		<edge from-layer="1369" from-port="2" to-layer="1371" to-port="0" />
-		<edge from-layer="1370" from-port="0" to-layer="1371" to-port="1" />
-		<edge from-layer="1371" from-port="2" to-layer="1372" to-port="1" />
+		<edge from-layer="1366" from-port="2" to-layer="1368" to-port="0" />
+		<edge from-layer="1367" from-port="0" to-layer="1368" to-port="1" />
+		<edge from-layer="1368" from-port="2" to-layer="1370" to-port="0" />
+		<edge from-layer="1369" from-port="0" to-layer="1370" to-port="1" />
+		<edge from-layer="1370" from-port="2" to-layer="1371" to-port="0" />
+		<edge from-layer="1371" from-port="1" to-layer="1372" to-port="1" />
 		<edge from-layer="1372" from-port="2" to-layer="1373" to-port="1" />
 		<edge from-layer="1373" from-port="2" to-layer="1375" to-port="0" />
 		<edge from-layer="1374" from-port="0" to-layer="1375" to-port="1" />
@@ -52515,12 +52229,12 @@
 		<edge from-layer="1383" from-port="1" to-layer="1386" to-port="0" />
 		<edge from-layer="1384" from-port="0" to-layer="1385" to-port="0" />
 		<edge from-layer="1385" from-port="1" to-layer="1386" to-port="1" />
-		<edge from-layer="1386" from-port="2" to-layer="1389" to-port="0" />
-		<edge from-layer="1387" from-port="0" to-layer="1388" to-port="0" />
-		<edge from-layer="1388" from-port="1" to-layer="1389" to-port="1" />
-		<edge from-layer="1389" from-port="2" to-layer="1391" to-port="0" />
-		<edge from-layer="1390" from-port="0" to-layer="1391" to-port="1" />
-		<edge from-layer="1391" from-port="2" to-layer="1392" to-port="1" />
+		<edge from-layer="1386" from-port="2" to-layer="1388" to-port="0" />
+		<edge from-layer="1387" from-port="0" to-layer="1388" to-port="1" />
+		<edge from-layer="1388" from-port="2" to-layer="1390" to-port="0" />
+		<edge from-layer="1389" from-port="0" to-layer="1390" to-port="1" />
+		<edge from-layer="1390" from-port="2" to-layer="1391" to-port="0" />
+		<edge from-layer="1391" from-port="1" to-layer="1392" to-port="1" />
 		<edge from-layer="1392" from-port="2" to-layer="1394" to-port="0" />
 		<edge from-layer="1393" from-port="0" to-layer="1394" to-port="1" />
 		<edge from-layer="1394" from-port="2" to-layer="1396" to-port="0" />
@@ -52536,12 +52250,12 @@
 		<edge from-layer="1404" from-port="1" to-layer="1407" to-port="0" />
 		<edge from-layer="1405" from-port="0" to-layer="1406" to-port="0" />
 		<edge from-layer="1406" from-port="1" to-layer="1407" to-port="1" />
-		<edge from-layer="1407" from-port="2" to-layer="1410" to-port="0" />
-		<edge from-layer="1408" from-port="0" to-layer="1409" to-port="0" />
-		<edge from-layer="1409" from-port="1" to-layer="1410" to-port="1" />
-		<edge from-layer="1410" from-port="2" to-layer="1412" to-port="0" />
-		<edge from-layer="1411" from-port="0" to-layer="1412" to-port="1" />
-		<edge from-layer="1412" from-port="2" to-layer="1413" to-port="1" />
+		<edge from-layer="1407" from-port="2" to-layer="1409" to-port="0" />
+		<edge from-layer="1408" from-port="0" to-layer="1409" to-port="1" />
+		<edge from-layer="1409" from-port="2" to-layer="1411" to-port="0" />
+		<edge from-layer="1410" from-port="0" to-layer="1411" to-port="1" />
+		<edge from-layer="1411" from-port="2" to-layer="1412" to-port="0" />
+		<edge from-layer="1412" from-port="1" to-layer="1413" to-port="1" />
 		<edge from-layer="1413" from-port="2" to-layer="1414" to-port="1" />
 		<edge from-layer="1414" from-port="2" to-layer="1416" to-port="0" />
 		<edge from-layer="1415" from-port="0" to-layer="1416" to-port="1" />
@@ -52568,12 +52282,12 @@
 		<edge from-layer="1434" from-port="1" to-layer="1437" to-port="0" />
 		<edge from-layer="1435" from-port="0" to-layer="1436" to-port="0" />
 		<edge from-layer="1436" from-port="1" to-layer="1437" to-port="1" />
-		<edge from-layer="1437" from-port="2" to-layer="1440" to-port="0" />
-		<edge from-layer="1438" from-port="0" to-layer="1439" to-port="0" />
-		<edge from-layer="1439" from-port="1" to-layer="1440" to-port="1" />
-		<edge from-layer="1440" from-port="2" to-layer="1442" to-port="0" />
-		<edge from-layer="1441" from-port="0" to-layer="1442" to-port="1" />
-		<edge from-layer="1442" from-port="2" to-layer="1443" to-port="1" />
+		<edge from-layer="1437" from-port="2" to-layer="1439" to-port="0" />
+		<edge from-layer="1438" from-port="0" to-layer="1439" to-port="1" />
+		<edge from-layer="1439" from-port="2" to-layer="1441" to-port="0" />
+		<edge from-layer="1440" from-port="0" to-layer="1441" to-port="1" />
+		<edge from-layer="1441" from-port="2" to-layer="1442" to-port="0" />
+		<edge from-layer="1442" from-port="1" to-layer="1443" to-port="1" />
 		<edge from-layer="1443" from-port="2" to-layer="1444" to-port="1" />
 		<edge from-layer="1444" from-port="2" to-layer="1445" to-port="1" />
 		<edge from-layer="1445" from-port="2" to-layer="1446" to-port="0" />
@@ -52595,12 +52309,12 @@
 		<edge from-layer="1460" from-port="1" to-layer="1463" to-port="0" />
 		<edge from-layer="1461" from-port="0" to-layer="1462" to-port="0" />
 		<edge from-layer="1462" from-port="1" to-layer="1463" to-port="1" />
-		<edge from-layer="1463" from-port="2" to-layer="1466" to-port="0" />
-		<edge from-layer="1464" from-port="0" to-layer="1465" to-port="0" />
-		<edge from-layer="1465" from-port="1" to-layer="1466" to-port="1" />
-		<edge from-layer="1466" from-port="2" to-layer="1468" to-port="0" />
-		<edge from-layer="1467" from-port="0" to-layer="1468" to-port="1" />
-		<edge from-layer="1468" from-port="2" to-layer="1469" to-port="1" />
+		<edge from-layer="1463" from-port="2" to-layer="1465" to-port="0" />
+		<edge from-layer="1464" from-port="0" to-layer="1465" to-port="1" />
+		<edge from-layer="1465" from-port="2" to-layer="1467" to-port="0" />
+		<edge from-layer="1466" from-port="0" to-layer="1467" to-port="1" />
+		<edge from-layer="1467" from-port="2" to-layer="1468" to-port="0" />
+		<edge from-layer="1468" from-port="1" to-layer="1469" to-port="1" />
 		<edge from-layer="1469" from-port="2" to-layer="1470" to-port="1" />
 		<edge from-layer="1470" from-port="2" to-layer="1472" to-port="0" />
 		<edge from-layer="1470" from-port="2" to-layer="1476" to-port="0" />
@@ -52616,12 +52330,12 @@
 		<edge from-layer="1480" from-port="1" to-layer="1483" to-port="0" />
 		<edge from-layer="1481" from-port="0" to-layer="1482" to-port="0" />
 		<edge from-layer="1482" from-port="1" to-layer="1483" to-port="1" />
-		<edge from-layer="1483" from-port="2" to-layer="1486" to-port="0" />
-		<edge from-layer="1484" from-port="0" to-layer="1485" to-port="0" />
-		<edge from-layer="1485" from-port="1" to-layer="1486" to-port="1" />
-		<edge from-layer="1486" from-port="2" to-layer="1488" to-port="0" />
-		<edge from-layer="1487" from-port="0" to-layer="1488" to-port="1" />
-		<edge from-layer="1488" from-port="2" to-layer="1489" to-port="1" />
+		<edge from-layer="1483" from-port="2" to-layer="1485" to-port="0" />
+		<edge from-layer="1484" from-port="0" to-layer="1485" to-port="1" />
+		<edge from-layer="1485" from-port="2" to-layer="1487" to-port="0" />
+		<edge from-layer="1486" from-port="0" to-layer="1487" to-port="1" />
+		<edge from-layer="1487" from-port="2" to-layer="1488" to-port="0" />
+		<edge from-layer="1488" from-port="1" to-layer="1489" to-port="1" />
 		<edge from-layer="1489" from-port="2" to-layer="1490" to-port="1" />
 		<edge from-layer="1490" from-port="2" to-layer="1491" to-port="1" />
 		<edge from-layer="1491" from-port="2" to-layer="1492" to-port="0" />
@@ -52645,12 +52359,12 @@
 		<edge from-layer="1506" from-port="1" to-layer="1509" to-port="0" />
 		<edge from-layer="1507" from-port="0" to-layer="1508" to-port="0" />
 		<edge from-layer="1508" from-port="1" to-layer="1509" to-port="1" />
-		<edge from-layer="1509" from-port="2" to-layer="1512" to-port="0" />
-		<edge from-layer="1510" from-port="0" to-layer="1511" to-port="0" />
-		<edge from-layer="1511" from-port="1" to-layer="1512" to-port="1" />
-		<edge from-layer="1512" from-port="2" to-layer="1514" to-port="0" />
-		<edge from-layer="1513" from-port="0" to-layer="1514" to-port="1" />
-		<edge from-layer="1514" from-port="2" to-layer="1515" to-port="1" />
+		<edge from-layer="1509" from-port="2" to-layer="1511" to-port="0" />
+		<edge from-layer="1510" from-port="0" to-layer="1511" to-port="1" />
+		<edge from-layer="1511" from-port="2" to-layer="1513" to-port="0" />
+		<edge from-layer="1512" from-port="0" to-layer="1513" to-port="1" />
+		<edge from-layer="1513" from-port="2" to-layer="1514" to-port="0" />
+		<edge from-layer="1514" from-port="1" to-layer="1515" to-port="1" />
 		<edge from-layer="1515" from-port="2" to-layer="1516" to-port="1" />
 		<edge from-layer="1516" from-port="2" to-layer="1518" to-port="0" />
 		<edge from-layer="1517" from-port="0" to-layer="1518" to-port="1" />
@@ -52666,12 +52380,12 @@
 		<edge from-layer="1526" from-port="1" to-layer="1529" to-port="0" />
 		<edge from-layer="1527" from-port="0" to-layer="1528" to-port="0" />
 		<edge from-layer="1528" from-port="1" to-layer="1529" to-port="1" />
-		<edge from-layer="1529" from-port="2" to-layer="1532" to-port="0" />
-		<edge from-layer="1530" from-port="0" to-layer="1531" to-port="0" />
-		<edge from-layer="1531" from-port="1" to-layer="1532" to-port="1" />
-		<edge from-layer="1532" from-port="2" to-layer="1534" to-port="0" />
-		<edge from-layer="1533" from-port="0" to-layer="1534" to-port="1" />
-		<edge from-layer="1534" from-port="2" to-layer="1535" to-port="1" />
+		<edge from-layer="1529" from-port="2" to-layer="1531" to-port="0" />
+		<edge from-layer="1530" from-port="0" to-layer="1531" to-port="1" />
+		<edge from-layer="1531" from-port="2" to-layer="1533" to-port="0" />
+		<edge from-layer="1532" from-port="0" to-layer="1533" to-port="1" />
+		<edge from-layer="1533" from-port="2" to-layer="1534" to-port="0" />
+		<edge from-layer="1534" from-port="1" to-layer="1535" to-port="1" />
 		<edge from-layer="1535" from-port="2" to-layer="1537" to-port="0" />
 		<edge from-layer="1536" from-port="0" to-layer="1537" to-port="1" />
 		<edge from-layer="1537" from-port="2" to-layer="1539" to-port="0" />
@@ -52687,12 +52401,12 @@
 		<edge from-layer="1547" from-port="1" to-layer="1550" to-port="0" />
 		<edge from-layer="1548" from-port="0" to-layer="1549" to-port="0" />
 		<edge from-layer="1549" from-port="1" to-layer="1550" to-port="1" />
-		<edge from-layer="1550" from-port="2" to-layer="1553" to-port="0" />
-		<edge from-layer="1551" from-port="0" to-layer="1552" to-port="0" />
-		<edge from-layer="1552" from-port="1" to-layer="1553" to-port="1" />
-		<edge from-layer="1553" from-port="2" to-layer="1555" to-port="0" />
-		<edge from-layer="1554" from-port="0" to-layer="1555" to-port="1" />
-		<edge from-layer="1555" from-port="2" to-layer="1556" to-port="1" />
+		<edge from-layer="1550" from-port="2" to-layer="1552" to-port="0" />
+		<edge from-layer="1551" from-port="0" to-layer="1552" to-port="1" />
+		<edge from-layer="1552" from-port="2" to-layer="1554" to-port="0" />
+		<edge from-layer="1553" from-port="0" to-layer="1554" to-port="1" />
+		<edge from-layer="1554" from-port="2" to-layer="1555" to-port="0" />
+		<edge from-layer="1555" from-port="1" to-layer="1556" to-port="1" />
 		<edge from-layer="1556" from-port="2" to-layer="1557" to-port="1" />
 		<edge from-layer="1557" from-port="2" to-layer="1559" to-port="0" />
 		<edge from-layer="1558" from-port="0" to-layer="1559" to-port="1" />
@@ -52719,12 +52433,12 @@
 		<edge from-layer="1577" from-port="1" to-layer="1580" to-port="0" />
 		<edge from-layer="1578" from-port="0" to-layer="1579" to-port="0" />
 		<edge from-layer="1579" from-port="1" to-layer="1580" to-port="1" />
-		<edge from-layer="1580" from-port="2" to-layer="1583" to-port="0" />
-		<edge from-layer="1581" from-port="0" to-layer="1582" to-port="0" />
-		<edge from-layer="1582" from-port="1" to-layer="1583" to-port="1" />
-		<edge from-layer="1583" from-port="2" to-layer="1585" to-port="0" />
-		<edge from-layer="1584" from-port="0" to-layer="1585" to-port="1" />
-		<edge from-layer="1585" from-port="2" to-layer="1586" to-port="1" />
+		<edge from-layer="1580" from-port="2" to-layer="1582" to-port="0" />
+		<edge from-layer="1581" from-port="0" to-layer="1582" to-port="1" />
+		<edge from-layer="1582" from-port="2" to-layer="1584" to-port="0" />
+		<edge from-layer="1583" from-port="0" to-layer="1584" to-port="1" />
+		<edge from-layer="1584" from-port="2" to-layer="1585" to-port="0" />
+		<edge from-layer="1585" from-port="1" to-layer="1586" to-port="1" />
 		<edge from-layer="1586" from-port="2" to-layer="1587" to-port="1" />
 		<edge from-layer="1587" from-port="2" to-layer="1588" to-port="1" />
 		<edge from-layer="1588" from-port="2" to-layer="1589" to-port="0" />
@@ -52746,12 +52460,12 @@
 		<edge from-layer="1603" from-port="1" to-layer="1606" to-port="0" />
 		<edge from-layer="1604" from-port="0" to-layer="1605" to-port="0" />
 		<edge from-layer="1605" from-port="1" to-layer="1606" to-port="1" />
-		<edge from-layer="1606" from-port="2" to-layer="1609" to-port="0" />
-		<edge from-layer="1607" from-port="0" to-layer="1608" to-port="0" />
-		<edge from-layer="1608" from-port="1" to-layer="1609" to-port="1" />
-		<edge from-layer="1609" from-port="2" to-layer="1611" to-port="0" />
-		<edge from-layer="1610" from-port="0" to-layer="1611" to-port="1" />
-		<edge from-layer="1611" from-port="2" to-layer="1612" to-port="1" />
+		<edge from-layer="1606" from-port="2" to-layer="1608" to-port="0" />
+		<edge from-layer="1607" from-port="0" to-layer="1608" to-port="1" />
+		<edge from-layer="1608" from-port="2" to-layer="1610" to-port="0" />
+		<edge from-layer="1609" from-port="0" to-layer="1610" to-port="1" />
+		<edge from-layer="1610" from-port="2" to-layer="1611" to-port="0" />
+		<edge from-layer="1611" from-port="1" to-layer="1612" to-port="1" />
 		<edge from-layer="1612" from-port="2" to-layer="1613" to-port="1" />
 		<edge from-layer="1613" from-port="2" to-layer="1615" to-port="0" />
 		<edge from-layer="1613" from-port="2" to-layer="1619" to-port="0" />
@@ -52767,12 +52481,12 @@
 		<edge from-layer="1623" from-port="1" to-layer="1626" to-port="0" />
 		<edge from-layer="1624" from-port="0" to-layer="1625" to-port="0" />
 		<edge from-layer="1625" from-port="1" to-layer="1626" to-port="1" />
-		<edge from-layer="1626" from-port="2" to-layer="1629" to-port="0" />
-		<edge from-layer="1627" from-port="0" to-layer="1628" to-port="0" />
-		<edge from-layer="1628" from-port="1" to-layer="1629" to-port="1" />
-		<edge from-layer="1629" from-port="2" to-layer="1631" to-port="0" />
-		<edge from-layer="1630" from-port="0" to-layer="1631" to-port="1" />
-		<edge from-layer="1631" from-port="2" to-layer="1632" to-port="1" />
+		<edge from-layer="1626" from-port="2" to-layer="1628" to-port="0" />
+		<edge from-layer="1627" from-port="0" to-layer="1628" to-port="1" />
+		<edge from-layer="1628" from-port="2" to-layer="1630" to-port="0" />
+		<edge from-layer="1629" from-port="0" to-layer="1630" to-port="1" />
+		<edge from-layer="1630" from-port="2" to-layer="1631" to-port="0" />
+		<edge from-layer="1631" from-port="1" to-layer="1632" to-port="1" />
 		<edge from-layer="1632" from-port="2" to-layer="1633" to-port="1" />
 		<edge from-layer="1633" from-port="2" to-layer="1634" to-port="1" />
 		<edge from-layer="1634" from-port="2" to-layer="1635" to-port="0" />
@@ -52796,12 +52510,12 @@
 		<edge from-layer="1649" from-port="1" to-layer="1652" to-port="0" />
 		<edge from-layer="1650" from-port="0" to-layer="1651" to-port="0" />
 		<edge from-layer="1651" from-port="1" to-layer="1652" to-port="1" />
-		<edge from-layer="1652" from-port="2" to-layer="1655" to-port="0" />
-		<edge from-layer="1653" from-port="0" to-layer="1654" to-port="0" />
-		<edge from-layer="1654" from-port="1" to-layer="1655" to-port="1" />
-		<edge from-layer="1655" from-port="2" to-layer="1657" to-port="0" />
-		<edge from-layer="1656" from-port="0" to-layer="1657" to-port="1" />
-		<edge from-layer="1657" from-port="2" to-layer="1658" to-port="1" />
+		<edge from-layer="1652" from-port="2" to-layer="1654" to-port="0" />
+		<edge from-layer="1653" from-port="0" to-layer="1654" to-port="1" />
+		<edge from-layer="1654" from-port="2" to-layer="1656" to-port="0" />
+		<edge from-layer="1655" from-port="0" to-layer="1656" to-port="1" />
+		<edge from-layer="1656" from-port="2" to-layer="1657" to-port="0" />
+		<edge from-layer="1657" from-port="1" to-layer="1658" to-port="1" />
 		<edge from-layer="1658" from-port="2" to-layer="1659" to-port="1" />
 		<edge from-layer="1659" from-port="2" to-layer="1661" to-port="0" />
 		<edge from-layer="1660" from-port="0" to-layer="1661" to-port="1" />
@@ -52817,12 +52531,12 @@
 		<edge from-layer="1669" from-port="1" to-layer="1672" to-port="0" />
 		<edge from-layer="1670" from-port="0" to-layer="1671" to-port="0" />
 		<edge from-layer="1671" from-port="1" to-layer="1672" to-port="1" />
-		<edge from-layer="1672" from-port="2" to-layer="1675" to-port="0" />
-		<edge from-layer="1673" from-port="0" to-layer="1674" to-port="0" />
-		<edge from-layer="1674" from-port="1" to-layer="1675" to-port="1" />
-		<edge from-layer="1675" from-port="2" to-layer="1677" to-port="0" />
-		<edge from-layer="1676" from-port="0" to-layer="1677" to-port="1" />
-		<edge from-layer="1677" from-port="2" to-layer="1678" to-port="1" />
+		<edge from-layer="1672" from-port="2" to-layer="1674" to-port="0" />
+		<edge from-layer="1673" from-port="0" to-layer="1674" to-port="1" />
+		<edge from-layer="1674" from-port="2" to-layer="1676" to-port="0" />
+		<edge from-layer="1675" from-port="0" to-layer="1676" to-port="1" />
+		<edge from-layer="1676" from-port="2" to-layer="1677" to-port="0" />
+		<edge from-layer="1677" from-port="1" to-layer="1678" to-port="1" />
 		<edge from-layer="1678" from-port="2" to-layer="1680" to-port="0" />
 		<edge from-layer="1679" from-port="0" to-layer="1680" to-port="1" />
 		<edge from-layer="1680" from-port="2" to-layer="1682" to-port="0" />
@@ -52838,12 +52552,12 @@
 		<edge from-layer="1690" from-port="1" to-layer="1693" to-port="0" />
 		<edge from-layer="1691" from-port="0" to-layer="1692" to-port="0" />
 		<edge from-layer="1692" from-port="1" to-layer="1693" to-port="1" />
-		<edge from-layer="1693" from-port="2" to-layer="1696" to-port="0" />
-		<edge from-layer="1694" from-port="0" to-layer="1695" to-port="0" />
-		<edge from-layer="1695" from-port="1" to-layer="1696" to-port="1" />
-		<edge from-layer="1696" from-port="2" to-layer="1698" to-port="0" />
-		<edge from-layer="1697" from-port="0" to-layer="1698" to-port="1" />
-		<edge from-layer="1698" from-port="2" to-layer="1699" to-port="1" />
+		<edge from-layer="1693" from-port="2" to-layer="1695" to-port="0" />
+		<edge from-layer="1694" from-port="0" to-layer="1695" to-port="1" />
+		<edge from-layer="1695" from-port="2" to-layer="1697" to-port="0" />
+		<edge from-layer="1696" from-port="0" to-layer="1697" to-port="1" />
+		<edge from-layer="1697" from-port="2" to-layer="1698" to-port="0" />
+		<edge from-layer="1698" from-port="1" to-layer="1699" to-port="1" />
 		<edge from-layer="1699" from-port="2" to-layer="1700" to-port="1" />
 		<edge from-layer="1700" from-port="2" to-layer="1702" to-port="0" />
 		<edge from-layer="1701" from-port="0" to-layer="1702" to-port="1" />
@@ -52870,12 +52584,12 @@
 		<edge from-layer="1720" from-port="1" to-layer="1723" to-port="0" />
 		<edge from-layer="1721" from-port="0" to-layer="1722" to-port="0" />
 		<edge from-layer="1722" from-port="1" to-layer="1723" to-port="1" />
-		<edge from-layer="1723" from-port="2" to-layer="1726" to-port="0" />
-		<edge from-layer="1724" from-port="0" to-layer="1725" to-port="0" />
-		<edge from-layer="1725" from-port="1" to-layer="1726" to-port="1" />
-		<edge from-layer="1726" from-port="2" to-layer="1728" to-port="0" />
-		<edge from-layer="1727" from-port="0" to-layer="1728" to-port="1" />
-		<edge from-layer="1728" from-port="2" to-layer="1729" to-port="1" />
+		<edge from-layer="1723" from-port="2" to-layer="1725" to-port="0" />
+		<edge from-layer="1724" from-port="0" to-layer="1725" to-port="1" />
+		<edge from-layer="1725" from-port="2" to-layer="1727" to-port="0" />
+		<edge from-layer="1726" from-port="0" to-layer="1727" to-port="1" />
+		<edge from-layer="1727" from-port="2" to-layer="1728" to-port="0" />
+		<edge from-layer="1728" from-port="1" to-layer="1729" to-port="1" />
 		<edge from-layer="1729" from-port="2" to-layer="1730" to-port="1" />
 		<edge from-layer="1730" from-port="2" to-layer="1731" to-port="1" />
 		<edge from-layer="1731" from-port="2" to-layer="1732" to-port="0" />
@@ -52897,12 +52611,12 @@
 		<edge from-layer="1746" from-port="1" to-layer="1749" to-port="0" />
 		<edge from-layer="1747" from-port="0" to-layer="1748" to-port="0" />
 		<edge from-layer="1748" from-port="1" to-layer="1749" to-port="1" />
-		<edge from-layer="1749" from-port="2" to-layer="1752" to-port="0" />
-		<edge from-layer="1750" from-port="0" to-layer="1751" to-port="0" />
-		<edge from-layer="1751" from-port="1" to-layer="1752" to-port="1" />
-		<edge from-layer="1752" from-port="2" to-layer="1754" to-port="0" />
-		<edge from-layer="1753" from-port="0" to-layer="1754" to-port="1" />
-		<edge from-layer="1754" from-port="2" to-layer="1755" to-port="1" />
+		<edge from-layer="1749" from-port="2" to-layer="1751" to-port="0" />
+		<edge from-layer="1750" from-port="0" to-layer="1751" to-port="1" />
+		<edge from-layer="1751" from-port="2" to-layer="1753" to-port="0" />
+		<edge from-layer="1752" from-port="0" to-layer="1753" to-port="1" />
+		<edge from-layer="1753" from-port="2" to-layer="1754" to-port="0" />
+		<edge from-layer="1754" from-port="1" to-layer="1755" to-port="1" />
 		<edge from-layer="1755" from-port="2" to-layer="1756" to-port="1" />
 		<edge from-layer="1756" from-port="2" to-layer="1758" to-port="0" />
 		<edge from-layer="1756" from-port="2" to-layer="1762" to-port="0" />
@@ -52918,12 +52632,12 @@
 		<edge from-layer="1766" from-port="1" to-layer="1769" to-port="0" />
 		<edge from-layer="1767" from-port="0" to-layer="1768" to-port="0" />
 		<edge from-layer="1768" from-port="1" to-layer="1769" to-port="1" />
-		<edge from-layer="1769" from-port="2" to-layer="1772" to-port="0" />
-		<edge from-layer="1770" from-port="0" to-layer="1771" to-port="0" />
-		<edge from-layer="1771" from-port="1" to-layer="1772" to-port="1" />
-		<edge from-layer="1772" from-port="2" to-layer="1774" to-port="0" />
-		<edge from-layer="1773" from-port="0" to-layer="1774" to-port="1" />
-		<edge from-layer="1774" from-port="2" to-layer="1775" to-port="1" />
+		<edge from-layer="1769" from-port="2" to-layer="1771" to-port="0" />
+		<edge from-layer="1770" from-port="0" to-layer="1771" to-port="1" />
+		<edge from-layer="1771" from-port="2" to-layer="1773" to-port="0" />
+		<edge from-layer="1772" from-port="0" to-layer="1773" to-port="1" />
+		<edge from-layer="1773" from-port="2" to-layer="1774" to-port="0" />
+		<edge from-layer="1774" from-port="1" to-layer="1775" to-port="1" />
 		<edge from-layer="1775" from-port="2" to-layer="1776" to-port="1" />
 		<edge from-layer="1776" from-port="2" to-layer="1777" to-port="1" />
 		<edge from-layer="1777" from-port="2" to-layer="1778" to-port="0" />
@@ -52947,12 +52661,12 @@
 		<edge from-layer="1792" from-port="1" to-layer="1795" to-port="0" />
 		<edge from-layer="1793" from-port="0" to-layer="1794" to-port="0" />
 		<edge from-layer="1794" from-port="1" to-layer="1795" to-port="1" />
-		<edge from-layer="1795" from-port="2" to-layer="1798" to-port="0" />
-		<edge from-layer="1796" from-port="0" to-layer="1797" to-port="0" />
-		<edge from-layer="1797" from-port="1" to-layer="1798" to-port="1" />
-		<edge from-layer="1798" from-port="2" to-layer="1800" to-port="0" />
-		<edge from-layer="1799" from-port="0" to-layer="1800" to-port="1" />
-		<edge from-layer="1800" from-port="2" to-layer="1801" to-port="1" />
+		<edge from-layer="1795" from-port="2" to-layer="1797" to-port="0" />
+		<edge from-layer="1796" from-port="0" to-layer="1797" to-port="1" />
+		<edge from-layer="1797" from-port="2" to-layer="1799" to-port="0" />
+		<edge from-layer="1798" from-port="0" to-layer="1799" to-port="1" />
+		<edge from-layer="1799" from-port="2" to-layer="1800" to-port="0" />
+		<edge from-layer="1800" from-port="1" to-layer="1801" to-port="1" />
 		<edge from-layer="1801" from-port="2" to-layer="1802" to-port="1" />
 		<edge from-layer="1802" from-port="2" to-layer="1804" to-port="0" />
 		<edge from-layer="1803" from-port="0" to-layer="1804" to-port="1" />
@@ -52968,12 +52682,12 @@
 		<edge from-layer="1812" from-port="1" to-layer="1815" to-port="0" />
 		<edge from-layer="1813" from-port="0" to-layer="1814" to-port="0" />
 		<edge from-layer="1814" from-port="1" to-layer="1815" to-port="1" />
-		<edge from-layer="1815" from-port="2" to-layer="1818" to-port="0" />
-		<edge from-layer="1816" from-port="0" to-layer="1817" to-port="0" />
-		<edge from-layer="1817" from-port="1" to-layer="1818" to-port="1" />
-		<edge from-layer="1818" from-port="2" to-layer="1820" to-port="0" />
-		<edge from-layer="1819" from-port="0" to-layer="1820" to-port="1" />
-		<edge from-layer="1820" from-port="2" to-layer="1821" to-port="1" />
+		<edge from-layer="1815" from-port="2" to-layer="1817" to-port="0" />
+		<edge from-layer="1816" from-port="0" to-layer="1817" to-port="1" />
+		<edge from-layer="1817" from-port="2" to-layer="1819" to-port="0" />
+		<edge from-layer="1818" from-port="0" to-layer="1819" to-port="1" />
+		<edge from-layer="1819" from-port="2" to-layer="1820" to-port="0" />
+		<edge from-layer="1820" from-port="1" to-layer="1821" to-port="1" />
 		<edge from-layer="1821" from-port="2" to-layer="1823" to-port="0" />
 		<edge from-layer="1822" from-port="0" to-layer="1823" to-port="1" />
 		<edge from-layer="1823" from-port="2" to-layer="1825" to-port="0" />
@@ -52989,12 +52703,12 @@
 		<edge from-layer="1833" from-port="1" to-layer="1836" to-port="0" />
 		<edge from-layer="1834" from-port="0" to-layer="1835" to-port="0" />
 		<edge from-layer="1835" from-port="1" to-layer="1836" to-port="1" />
-		<edge from-layer="1836" from-port="2" to-layer="1839" to-port="0" />
-		<edge from-layer="1837" from-port="0" to-layer="1838" to-port="0" />
-		<edge from-layer="1838" from-port="1" to-layer="1839" to-port="1" />
-		<edge from-layer="1839" from-port="2" to-layer="1841" to-port="0" />
-		<edge from-layer="1840" from-port="0" to-layer="1841" to-port="1" />
-		<edge from-layer="1841" from-port="2" to-layer="1842" to-port="1" />
+		<edge from-layer="1836" from-port="2" to-layer="1838" to-port="0" />
+		<edge from-layer="1837" from-port="0" to-layer="1838" to-port="1" />
+		<edge from-layer="1838" from-port="2" to-layer="1840" to-port="0" />
+		<edge from-layer="1839" from-port="0" to-layer="1840" to-port="1" />
+		<edge from-layer="1840" from-port="2" to-layer="1841" to-port="0" />
+		<edge from-layer="1841" from-port="1" to-layer="1842" to-port="1" />
 		<edge from-layer="1842" from-port="2" to-layer="1843" to-port="1" />
 		<edge from-layer="1843" from-port="2" to-layer="1845" to-port="0" />
 		<edge from-layer="1844" from-port="0" to-layer="1845" to-port="1" />
@@ -53006,8 +52720,8 @@
 		<edge from-layer="1850" from-port="2" to-layer="1851" to-port="0" />
 		<edge from-layer="1850" from-port="2" to-layer="1861" to-port="0" />
 		<edge from-layer="1851" from-port="1" to-layer="1858" to-port="0" />
-		<edge from-layer="1852" from-port="0" to-layer="1855" to-port="0" />
 		<edge from-layer="1852" from-port="0" to-layer="1858" to-port="1" />
+		<edge from-layer="1852" from-port="0" to-layer="1855" to-port="0" />
 		<edge from-layer="1853" from-port="0" to-layer="1858" to-port="2" />
 		<edge from-layer="1854" from-port="0" to-layer="1856" to-port="0" />
 		<edge from-layer="1855" from-port="1" to-layer="1856" to-port="1" />
@@ -53021,12 +52735,12 @@
 		<edge from-layer="1863" from-port="1" to-layer="1866" to-port="0" />
 		<edge from-layer="1864" from-port="0" to-layer="1865" to-port="0" />
 		<edge from-layer="1865" from-port="1" to-layer="1866" to-port="1" />
-		<edge from-layer="1866" from-port="2" to-layer="1869" to-port="0" />
-		<edge from-layer="1867" from-port="0" to-layer="1868" to-port="0" />
-		<edge from-layer="1868" from-port="1" to-layer="1869" to-port="1" />
-		<edge from-layer="1869" from-port="2" to-layer="1871" to-port="0" />
-		<edge from-layer="1870" from-port="0" to-layer="1871" to-port="1" />
-		<edge from-layer="1871" from-port="2" to-layer="1872" to-port="1" />
+		<edge from-layer="1866" from-port="2" to-layer="1868" to-port="0" />
+		<edge from-layer="1867" from-port="0" to-layer="1868" to-port="1" />
+		<edge from-layer="1868" from-port="2" to-layer="1870" to-port="0" />
+		<edge from-layer="1869" from-port="0" to-layer="1870" to-port="1" />
+		<edge from-layer="1870" from-port="2" to-layer="1871" to-port="0" />
+		<edge from-layer="1871" from-port="1" to-layer="1872" to-port="1" />
 		<edge from-layer="1872" from-port="2" to-layer="1873" to-port="1" />
 		<edge from-layer="1873" from-port="2" to-layer="1874" to-port="1" />
 		<edge from-layer="1874" from-port="2" to-layer="1875" to-port="0" />
@@ -53048,12 +52762,12 @@
 		<edge from-layer="1889" from-port="1" to-layer="1892" to-port="0" />
 		<edge from-layer="1890" from-port="0" to-layer="1891" to-port="0" />
 		<edge from-layer="1891" from-port="1" to-layer="1892" to-port="1" />
-		<edge from-layer="1892" from-port="2" to-layer="1895" to-port="0" />
-		<edge from-layer="1893" from-port="0" to-layer="1894" to-port="0" />
-		<edge from-layer="1894" from-port="1" to-layer="1895" to-port="1" />
-		<edge from-layer="1895" from-port="2" to-layer="1897" to-port="0" />
-		<edge from-layer="1896" from-port="0" to-layer="1897" to-port="1" />
-		<edge from-layer="1897" from-port="2" to-layer="1898" to-port="1" />
+		<edge from-layer="1892" from-port="2" to-layer="1894" to-port="0" />
+		<edge from-layer="1893" from-port="0" to-layer="1894" to-port="1" />
+		<edge from-layer="1894" from-port="2" to-layer="1896" to-port="0" />
+		<edge from-layer="1895" from-port="0" to-layer="1896" to-port="1" />
+		<edge from-layer="1896" from-port="2" to-layer="1897" to-port="0" />
+		<edge from-layer="1897" from-port="1" to-layer="1898" to-port="1" />
 		<edge from-layer="1898" from-port="2" to-layer="1899" to-port="1" />
 		<edge from-layer="1899" from-port="2" to-layer="1901" to-port="0" />
 		<edge from-layer="1899" from-port="2" to-layer="1905" to-port="0" />
@@ -53069,12 +52783,12 @@
 		<edge from-layer="1909" from-port="1" to-layer="1912" to-port="0" />
 		<edge from-layer="1910" from-port="0" to-layer="1911" to-port="0" />
 		<edge from-layer="1911" from-port="1" to-layer="1912" to-port="1" />
-		<edge from-layer="1912" from-port="2" to-layer="1915" to-port="0" />
-		<edge from-layer="1913" from-port="0" to-layer="1914" to-port="0" />
-		<edge from-layer="1914" from-port="1" to-layer="1915" to-port="1" />
-		<edge from-layer="1915" from-port="2" to-layer="1917" to-port="0" />
-		<edge from-layer="1916" from-port="0" to-layer="1917" to-port="1" />
-		<edge from-layer="1917" from-port="2" to-layer="1918" to-port="1" />
+		<edge from-layer="1912" from-port="2" to-layer="1914" to-port="0" />
+		<edge from-layer="1913" from-port="0" to-layer="1914" to-port="1" />
+		<edge from-layer="1914" from-port="2" to-layer="1916" to-port="0" />
+		<edge from-layer="1915" from-port="0" to-layer="1916" to-port="1" />
+		<edge from-layer="1916" from-port="2" to-layer="1917" to-port="0" />
+		<edge from-layer="1917" from-port="1" to-layer="1918" to-port="1" />
 		<edge from-layer="1918" from-port="2" to-layer="1919" to-port="1" />
 		<edge from-layer="1919" from-port="2" to-layer="1920" to-port="1" />
 		<edge from-layer="1920" from-port="2" to-layer="1921" to-port="0" />
@@ -53098,12 +52812,12 @@
 		<edge from-layer="1935" from-port="1" to-layer="1938" to-port="0" />
 		<edge from-layer="1936" from-port="0" to-layer="1937" to-port="0" />
 		<edge from-layer="1937" from-port="1" to-layer="1938" to-port="1" />
-		<edge from-layer="1938" from-port="2" to-layer="1941" to-port="0" />
-		<edge from-layer="1939" from-port="0" to-layer="1940" to-port="0" />
-		<edge from-layer="1940" from-port="1" to-layer="1941" to-port="1" />
-		<edge from-layer="1941" from-port="2" to-layer="1943" to-port="0" />
-		<edge from-layer="1942" from-port="0" to-layer="1943" to-port="1" />
-		<edge from-layer="1943" from-port="2" to-layer="1944" to-port="1" />
+		<edge from-layer="1938" from-port="2" to-layer="1940" to-port="0" />
+		<edge from-layer="1939" from-port="0" to-layer="1940" to-port="1" />
+		<edge from-layer="1940" from-port="2" to-layer="1942" to-port="0" />
+		<edge from-layer="1941" from-port="0" to-layer="1942" to-port="1" />
+		<edge from-layer="1942" from-port="2" to-layer="1943" to-port="0" />
+		<edge from-layer="1943" from-port="1" to-layer="1944" to-port="1" />
 		<edge from-layer="1944" from-port="2" to-layer="1945" to-port="1" />
 		<edge from-layer="1945" from-port="2" to-layer="1947" to-port="0" />
 		<edge from-layer="1946" from-port="0" to-layer="1947" to-port="1" />
@@ -53119,12 +52833,12 @@
 		<edge from-layer="1955" from-port="1" to-layer="1958" to-port="0" />
 		<edge from-layer="1956" from-port="0" to-layer="1957" to-port="0" />
 		<edge from-layer="1957" from-port="1" to-layer="1958" to-port="1" />
-		<edge from-layer="1958" from-port="2" to-layer="1961" to-port="0" />
-		<edge from-layer="1959" from-port="0" to-layer="1960" to-port="0" />
-		<edge from-layer="1960" from-port="1" to-layer="1961" to-port="1" />
-		<edge from-layer="1961" from-port="2" to-layer="1963" to-port="0" />
-		<edge from-layer="1962" from-port="0" to-layer="1963" to-port="1" />
-		<edge from-layer="1963" from-port="2" to-layer="1964" to-port="1" />
+		<edge from-layer="1958" from-port="2" to-layer="1960" to-port="0" />
+		<edge from-layer="1959" from-port="0" to-layer="1960" to-port="1" />
+		<edge from-layer="1960" from-port="2" to-layer="1962" to-port="0" />
+		<edge from-layer="1961" from-port="0" to-layer="1962" to-port="1" />
+		<edge from-layer="1962" from-port="2" to-layer="1963" to-port="0" />
+		<edge from-layer="1963" from-port="1" to-layer="1964" to-port="1" />
 		<edge from-layer="1964" from-port="2" to-layer="1966" to-port="0" />
 		<edge from-layer="1965" from-port="0" to-layer="1966" to-port="1" />
 		<edge from-layer="1966" from-port="2" to-layer="1968" to-port="0" />
@@ -53140,12 +52854,12 @@
 		<edge from-layer="1976" from-port="1" to-layer="1979" to-port="0" />
 		<edge from-layer="1977" from-port="0" to-layer="1978" to-port="0" />
 		<edge from-layer="1978" from-port="1" to-layer="1979" to-port="1" />
-		<edge from-layer="1979" from-port="2" to-layer="1982" to-port="0" />
-		<edge from-layer="1980" from-port="0" to-layer="1981" to-port="0" />
-		<edge from-layer="1981" from-port="1" to-layer="1982" to-port="1" />
-		<edge from-layer="1982" from-port="2" to-layer="1984" to-port="0" />
-		<edge from-layer="1983" from-port="0" to-layer="1984" to-port="1" />
-		<edge from-layer="1984" from-port="2" to-layer="1985" to-port="1" />
+		<edge from-layer="1979" from-port="2" to-layer="1981" to-port="0" />
+		<edge from-layer="1980" from-port="0" to-layer="1981" to-port="1" />
+		<edge from-layer="1981" from-port="2" to-layer="1983" to-port="0" />
+		<edge from-layer="1982" from-port="0" to-layer="1983" to-port="1" />
+		<edge from-layer="1983" from-port="2" to-layer="1984" to-port="0" />
+		<edge from-layer="1984" from-port="1" to-layer="1985" to-port="1" />
 		<edge from-layer="1985" from-port="2" to-layer="1986" to-port="1" />
 		<edge from-layer="1986" from-port="2" to-layer="1988" to-port="0" />
 		<edge from-layer="1987" from-port="0" to-layer="1988" to-port="1" />
@@ -53172,12 +52886,12 @@
 		<edge from-layer="2006" from-port="1" to-layer="2009" to-port="0" />
 		<edge from-layer="2007" from-port="0" to-layer="2008" to-port="0" />
 		<edge from-layer="2008" from-port="1" to-layer="2009" to-port="1" />
-		<edge from-layer="2009" from-port="2" to-layer="2012" to-port="0" />
-		<edge from-layer="2010" from-port="0" to-layer="2011" to-port="0" />
-		<edge from-layer="2011" from-port="1" to-layer="2012" to-port="1" />
-		<edge from-layer="2012" from-port="2" to-layer="2014" to-port="0" />
-		<edge from-layer="2013" from-port="0" to-layer="2014" to-port="1" />
-		<edge from-layer="2014" from-port="2" to-layer="2015" to-port="1" />
+		<edge from-layer="2009" from-port="2" to-layer="2011" to-port="0" />
+		<edge from-layer="2010" from-port="0" to-layer="2011" to-port="1" />
+		<edge from-layer="2011" from-port="2" to-layer="2013" to-port="0" />
+		<edge from-layer="2012" from-port="0" to-layer="2013" to-port="1" />
+		<edge from-layer="2013" from-port="2" to-layer="2014" to-port="0" />
+		<edge from-layer="2014" from-port="1" to-layer="2015" to-port="1" />
 		<edge from-layer="2015" from-port="2" to-layer="2016" to-port="1" />
 		<edge from-layer="2016" from-port="2" to-layer="2017" to-port="1" />
 		<edge from-layer="2017" from-port="2" to-layer="2018" to-port="0" />
@@ -53199,12 +52913,12 @@
 		<edge from-layer="2032" from-port="1" to-layer="2035" to-port="0" />
 		<edge from-layer="2033" from-port="0" to-layer="2034" to-port="0" />
 		<edge from-layer="2034" from-port="1" to-layer="2035" to-port="1" />
-		<edge from-layer="2035" from-port="2" to-layer="2038" to-port="0" />
-		<edge from-layer="2036" from-port="0" to-layer="2037" to-port="0" />
-		<edge from-layer="2037" from-port="1" to-layer="2038" to-port="1" />
-		<edge from-layer="2038" from-port="2" to-layer="2040" to-port="0" />
-		<edge from-layer="2039" from-port="0" to-layer="2040" to-port="1" />
-		<edge from-layer="2040" from-port="2" to-layer="2041" to-port="1" />
+		<edge from-layer="2035" from-port="2" to-layer="2037" to-port="0" />
+		<edge from-layer="2036" from-port="0" to-layer="2037" to-port="1" />
+		<edge from-layer="2037" from-port="2" to-layer="2039" to-port="0" />
+		<edge from-layer="2038" from-port="0" to-layer="2039" to-port="1" />
+		<edge from-layer="2039" from-port="2" to-layer="2040" to-port="0" />
+		<edge from-layer="2040" from-port="1" to-layer="2041" to-port="1" />
 		<edge from-layer="2041" from-port="2" to-layer="2042" to-port="1" />
 		<edge from-layer="2042" from-port="2" to-layer="2044" to-port="0" />
 		<edge from-layer="2042" from-port="2" to-layer="2048" to-port="0" />
@@ -53220,12 +52934,12 @@
 		<edge from-layer="2052" from-port="1" to-layer="2055" to-port="0" />
 		<edge from-layer="2053" from-port="0" to-layer="2054" to-port="0" />
 		<edge from-layer="2054" from-port="1" to-layer="2055" to-port="1" />
-		<edge from-layer="2055" from-port="2" to-layer="2058" to-port="0" />
-		<edge from-layer="2056" from-port="0" to-layer="2057" to-port="0" />
-		<edge from-layer="2057" from-port="1" to-layer="2058" to-port="1" />
-		<edge from-layer="2058" from-port="2" to-layer="2060" to-port="0" />
-		<edge from-layer="2059" from-port="0" to-layer="2060" to-port="1" />
-		<edge from-layer="2060" from-port="2" to-layer="2061" to-port="1" />
+		<edge from-layer="2055" from-port="2" to-layer="2057" to-port="0" />
+		<edge from-layer="2056" from-port="0" to-layer="2057" to-port="1" />
+		<edge from-layer="2057" from-port="2" to-layer="2059" to-port="0" />
+		<edge from-layer="2058" from-port="0" to-layer="2059" to-port="1" />
+		<edge from-layer="2059" from-port="2" to-layer="2060" to-port="0" />
+		<edge from-layer="2060" from-port="1" to-layer="2061" to-port="1" />
 		<edge from-layer="2061" from-port="2" to-layer="2062" to-port="1" />
 		<edge from-layer="2062" from-port="2" to-layer="2063" to-port="1" />
 		<edge from-layer="2063" from-port="2" to-layer="2064" to-port="0" />
@@ -53249,12 +52963,12 @@
 		<edge from-layer="2078" from-port="1" to-layer="2081" to-port="0" />
 		<edge from-layer="2079" from-port="0" to-layer="2080" to-port="0" />
 		<edge from-layer="2080" from-port="1" to-layer="2081" to-port="1" />
-		<edge from-layer="2081" from-port="2" to-layer="2084" to-port="0" />
-		<edge from-layer="2082" from-port="0" to-layer="2083" to-port="0" />
-		<edge from-layer="2083" from-port="1" to-layer="2084" to-port="1" />
-		<edge from-layer="2084" from-port="2" to-layer="2086" to-port="0" />
-		<edge from-layer="2085" from-port="0" to-layer="2086" to-port="1" />
-		<edge from-layer="2086" from-port="2" to-layer="2087" to-port="1" />
+		<edge from-layer="2081" from-port="2" to-layer="2083" to-port="0" />
+		<edge from-layer="2082" from-port="0" to-layer="2083" to-port="1" />
+		<edge from-layer="2083" from-port="2" to-layer="2085" to-port="0" />
+		<edge from-layer="2084" from-port="0" to-layer="2085" to-port="1" />
+		<edge from-layer="2085" from-port="2" to-layer="2086" to-port="0" />
+		<edge from-layer="2086" from-port="1" to-layer="2087" to-port="1" />
 		<edge from-layer="2087" from-port="2" to-layer="2088" to-port="1" />
 		<edge from-layer="2088" from-port="2" to-layer="2090" to-port="0" />
 		<edge from-layer="2089" from-port="0" to-layer="2090" to-port="1" />
@@ -53270,12 +52984,12 @@
 		<edge from-layer="2098" from-port="1" to-layer="2101" to-port="0" />
 		<edge from-layer="2099" from-port="0" to-layer="2100" to-port="0" />
 		<edge from-layer="2100" from-port="1" to-layer="2101" to-port="1" />
-		<edge from-layer="2101" from-port="2" to-layer="2104" to-port="0" />
-		<edge from-layer="2102" from-port="0" to-layer="2103" to-port="0" />
-		<edge from-layer="2103" from-port="1" to-layer="2104" to-port="1" />
-		<edge from-layer="2104" from-port="2" to-layer="2106" to-port="0" />
-		<edge from-layer="2105" from-port="0" to-layer="2106" to-port="1" />
-		<edge from-layer="2106" from-port="2" to-layer="2107" to-port="1" />
+		<edge from-layer="2101" from-port="2" to-layer="2103" to-port="0" />
+		<edge from-layer="2102" from-port="0" to-layer="2103" to-port="1" />
+		<edge from-layer="2103" from-port="2" to-layer="2105" to-port="0" />
+		<edge from-layer="2104" from-port="0" to-layer="2105" to-port="1" />
+		<edge from-layer="2105" from-port="2" to-layer="2106" to-port="0" />
+		<edge from-layer="2106" from-port="1" to-layer="2107" to-port="1" />
 		<edge from-layer="2107" from-port="2" to-layer="2109" to-port="0" />
 		<edge from-layer="2108" from-port="0" to-layer="2109" to-port="1" />
 		<edge from-layer="2109" from-port="2" to-layer="2111" to-port="0" />
@@ -53291,12 +53005,12 @@
 		<edge from-layer="2119" from-port="1" to-layer="2122" to-port="0" />
 		<edge from-layer="2120" from-port="0" to-layer="2121" to-port="0" />
 		<edge from-layer="2121" from-port="1" to-layer="2122" to-port="1" />
-		<edge from-layer="2122" from-port="2" to-layer="2125" to-port="0" />
-		<edge from-layer="2123" from-port="0" to-layer="2124" to-port="0" />
-		<edge from-layer="2124" from-port="1" to-layer="2125" to-port="1" />
-		<edge from-layer="2125" from-port="2" to-layer="2127" to-port="0" />
-		<edge from-layer="2126" from-port="0" to-layer="2127" to-port="1" />
-		<edge from-layer="2127" from-port="2" to-layer="2128" to-port="1" />
+		<edge from-layer="2122" from-port="2" to-layer="2124" to-port="0" />
+		<edge from-layer="2123" from-port="0" to-layer="2124" to-port="1" />
+		<edge from-layer="2124" from-port="2" to-layer="2126" to-port="0" />
+		<edge from-layer="2125" from-port="0" to-layer="2126" to-port="1" />
+		<edge from-layer="2126" from-port="2" to-layer="2127" to-port="0" />
+		<edge from-layer="2127" from-port="1" to-layer="2128" to-port="1" />
 		<edge from-layer="2128" from-port="2" to-layer="2129" to-port="1" />
 		<edge from-layer="2129" from-port="2" to-layer="2131" to-port="0" />
 		<edge from-layer="2130" from-port="0" to-layer="2131" to-port="1" />
@@ -53323,12 +53037,12 @@
 		<edge from-layer="2149" from-port="1" to-layer="2152" to-port="0" />
 		<edge from-layer="2150" from-port="0" to-layer="2151" to-port="0" />
 		<edge from-layer="2151" from-port="1" to-layer="2152" to-port="1" />
-		<edge from-layer="2152" from-port="2" to-layer="2155" to-port="0" />
-		<edge from-layer="2153" from-port="0" to-layer="2154" to-port="0" />
-		<edge from-layer="2154" from-port="1" to-layer="2155" to-port="1" />
-		<edge from-layer="2155" from-port="2" to-layer="2157" to-port="0" />
-		<edge from-layer="2156" from-port="0" to-layer="2157" to-port="1" />
-		<edge from-layer="2157" from-port="2" to-layer="2158" to-port="1" />
+		<edge from-layer="2152" from-port="2" to-layer="2154" to-port="0" />
+		<edge from-layer="2153" from-port="0" to-layer="2154" to-port="1" />
+		<edge from-layer="2154" from-port="2" to-layer="2156" to-port="0" />
+		<edge from-layer="2155" from-port="0" to-layer="2156" to-port="1" />
+		<edge from-layer="2156" from-port="2" to-layer="2157" to-port="0" />
+		<edge from-layer="2157" from-port="1" to-layer="2158" to-port="1" />
 		<edge from-layer="2158" from-port="2" to-layer="2159" to-port="1" />
 		<edge from-layer="2159" from-port="2" to-layer="2160" to-port="1" />
 		<edge from-layer="2160" from-port="2" to-layer="2161" to-port="0" />
@@ -53350,12 +53064,12 @@
 		<edge from-layer="2175" from-port="1" to-layer="2178" to-port="0" />
 		<edge from-layer="2176" from-port="0" to-layer="2177" to-port="0" />
 		<edge from-layer="2177" from-port="1" to-layer="2178" to-port="1" />
-		<edge from-layer="2178" from-port="2" to-layer="2181" to-port="0" />
-		<edge from-layer="2179" from-port="0" to-layer="2180" to-port="0" />
-		<edge from-layer="2180" from-port="1" to-layer="2181" to-port="1" />
-		<edge from-layer="2181" from-port="2" to-layer="2183" to-port="0" />
-		<edge from-layer="2182" from-port="0" to-layer="2183" to-port="1" />
-		<edge from-layer="2183" from-port="2" to-layer="2184" to-port="1" />
+		<edge from-layer="2178" from-port="2" to-layer="2180" to-port="0" />
+		<edge from-layer="2179" from-port="0" to-layer="2180" to-port="1" />
+		<edge from-layer="2180" from-port="2" to-layer="2182" to-port="0" />
+		<edge from-layer="2181" from-port="0" to-layer="2182" to-port="1" />
+		<edge from-layer="2182" from-port="2" to-layer="2183" to-port="0" />
+		<edge from-layer="2183" from-port="1" to-layer="2184" to-port="1" />
 		<edge from-layer="2184" from-port="2" to-layer="2185" to-port="1" />
 		<edge from-layer="2185" from-port="2" to-layer="2187" to-port="0" />
 		<edge from-layer="2185" from-port="2" to-layer="2191" to-port="0" />
@@ -53371,12 +53085,12 @@
 		<edge from-layer="2195" from-port="1" to-layer="2198" to-port="0" />
 		<edge from-layer="2196" from-port="0" to-layer="2197" to-port="0" />
 		<edge from-layer="2197" from-port="1" to-layer="2198" to-port="1" />
-		<edge from-layer="2198" from-port="2" to-layer="2201" to-port="0" />
-		<edge from-layer="2199" from-port="0" to-layer="2200" to-port="0" />
-		<edge from-layer="2200" from-port="1" to-layer="2201" to-port="1" />
-		<edge from-layer="2201" from-port="2" to-layer="2203" to-port="0" />
-		<edge from-layer="2202" from-port="0" to-layer="2203" to-port="1" />
-		<edge from-layer="2203" from-port="2" to-layer="2204" to-port="1" />
+		<edge from-layer="2198" from-port="2" to-layer="2200" to-port="0" />
+		<edge from-layer="2199" from-port="0" to-layer="2200" to-port="1" />
+		<edge from-layer="2200" from-port="2" to-layer="2202" to-port="0" />
+		<edge from-layer="2201" from-port="0" to-layer="2202" to-port="1" />
+		<edge from-layer="2202" from-port="2" to-layer="2203" to-port="0" />
+		<edge from-layer="2203" from-port="1" to-layer="2204" to-port="1" />
 		<edge from-layer="2204" from-port="2" to-layer="2205" to-port="1" />
 		<edge from-layer="2205" from-port="2" to-layer="2206" to-port="1" />
 		<edge from-layer="2206" from-port="2" to-layer="2207" to-port="0" />
@@ -53400,12 +53114,12 @@
 		<edge from-layer="2221" from-port="1" to-layer="2224" to-port="0" />
 		<edge from-layer="2222" from-port="0" to-layer="2223" to-port="0" />
 		<edge from-layer="2223" from-port="1" to-layer="2224" to-port="1" />
-		<edge from-layer="2224" from-port="2" to-layer="2227" to-port="0" />
-		<edge from-layer="2225" from-port="0" to-layer="2226" to-port="0" />
-		<edge from-layer="2226" from-port="1" to-layer="2227" to-port="1" />
-		<edge from-layer="2227" from-port="2" to-layer="2229" to-port="0" />
-		<edge from-layer="2228" from-port="0" to-layer="2229" to-port="1" />
-		<edge from-layer="2229" from-port="2" to-layer="2230" to-port="1" />
+		<edge from-layer="2224" from-port="2" to-layer="2226" to-port="0" />
+		<edge from-layer="2225" from-port="0" to-layer="2226" to-port="1" />
+		<edge from-layer="2226" from-port="2" to-layer="2228" to-port="0" />
+		<edge from-layer="2227" from-port="0" to-layer="2228" to-port="1" />
+		<edge from-layer="2228" from-port="2" to-layer="2229" to-port="0" />
+		<edge from-layer="2229" from-port="1" to-layer="2230" to-port="1" />
 		<edge from-layer="2230" from-port="2" to-layer="2231" to-port="1" />
 		<edge from-layer="2231" from-port="2" to-layer="2233" to-port="0" />
 		<edge from-layer="2232" from-port="0" to-layer="2233" to-port="1" />
@@ -53421,12 +53135,12 @@
 		<edge from-layer="2241" from-port="1" to-layer="2244" to-port="0" />
 		<edge from-layer="2242" from-port="0" to-layer="2243" to-port="0" />
 		<edge from-layer="2243" from-port="1" to-layer="2244" to-port="1" />
-		<edge from-layer="2244" from-port="2" to-layer="2247" to-port="0" />
-		<edge from-layer="2245" from-port="0" to-layer="2246" to-port="0" />
-		<edge from-layer="2246" from-port="1" to-layer="2247" to-port="1" />
-		<edge from-layer="2247" from-port="2" to-layer="2249" to-port="0" />
-		<edge from-layer="2248" from-port="0" to-layer="2249" to-port="1" />
-		<edge from-layer="2249" from-port="2" to-layer="2250" to-port="1" />
+		<edge from-layer="2244" from-port="2" to-layer="2246" to-port="0" />
+		<edge from-layer="2245" from-port="0" to-layer="2246" to-port="1" />
+		<edge from-layer="2246" from-port="2" to-layer="2248" to-port="0" />
+		<edge from-layer="2247" from-port="0" to-layer="2248" to-port="1" />
+		<edge from-layer="2248" from-port="2" to-layer="2249" to-port="0" />
+		<edge from-layer="2249" from-port="1" to-layer="2250" to-port="1" />
 		<edge from-layer="2250" from-port="2" to-layer="2252" to-port="0" />
 		<edge from-layer="2251" from-port="0" to-layer="2252" to-port="1" />
 		<edge from-layer="2252" from-port="2" to-layer="2254" to-port="0" />
@@ -53442,12 +53156,12 @@
 		<edge from-layer="2262" from-port="1" to-layer="2265" to-port="0" />
 		<edge from-layer="2263" from-port="0" to-layer="2264" to-port="0" />
 		<edge from-layer="2264" from-port="1" to-layer="2265" to-port="1" />
-		<edge from-layer="2265" from-port="2" to-layer="2268" to-port="0" />
-		<edge from-layer="2266" from-port="0" to-layer="2267" to-port="0" />
-		<edge from-layer="2267" from-port="1" to-layer="2268" to-port="1" />
-		<edge from-layer="2268" from-port="2" to-layer="2270" to-port="0" />
-		<edge from-layer="2269" from-port="0" to-layer="2270" to-port="1" />
-		<edge from-layer="2270" from-port="2" to-layer="2271" to-port="1" />
+		<edge from-layer="2265" from-port="2" to-layer="2267" to-port="0" />
+		<edge from-layer="2266" from-port="0" to-layer="2267" to-port="1" />
+		<edge from-layer="2267" from-port="2" to-layer="2269" to-port="0" />
+		<edge from-layer="2268" from-port="0" to-layer="2269" to-port="1" />
+		<edge from-layer="2269" from-port="2" to-layer="2270" to-port="0" />
+		<edge from-layer="2270" from-port="1" to-layer="2271" to-port="1" />
 		<edge from-layer="2271" from-port="2" to-layer="2272" to-port="1" />
 		<edge from-layer="2272" from-port="2" to-layer="2274" to-port="0" />
 		<edge from-layer="2273" from-port="0" to-layer="2274" to-port="1" />
@@ -53474,12 +53188,12 @@
 		<edge from-layer="2292" from-port="1" to-layer="2295" to-port="0" />
 		<edge from-layer="2293" from-port="0" to-layer="2294" to-port="0" />
 		<edge from-layer="2294" from-port="1" to-layer="2295" to-port="1" />
-		<edge from-layer="2295" from-port="2" to-layer="2298" to-port="0" />
-		<edge from-layer="2296" from-port="0" to-layer="2297" to-port="0" />
-		<edge from-layer="2297" from-port="1" to-layer="2298" to-port="1" />
-		<edge from-layer="2298" from-port="2" to-layer="2300" to-port="0" />
-		<edge from-layer="2299" from-port="0" to-layer="2300" to-port="1" />
-		<edge from-layer="2300" from-port="2" to-layer="2301" to-port="1" />
+		<edge from-layer="2295" from-port="2" to-layer="2297" to-port="0" />
+		<edge from-layer="2296" from-port="0" to-layer="2297" to-port="1" />
+		<edge from-layer="2297" from-port="2" to-layer="2299" to-port="0" />
+		<edge from-layer="2298" from-port="0" to-layer="2299" to-port="1" />
+		<edge from-layer="2299" from-port="2" to-layer="2300" to-port="0" />
+		<edge from-layer="2300" from-port="1" to-layer="2301" to-port="1" />
 		<edge from-layer="2301" from-port="2" to-layer="2302" to-port="1" />
 		<edge from-layer="2302" from-port="2" to-layer="2303" to-port="1" />
 		<edge from-layer="2303" from-port="2" to-layer="2304" to-port="0" />
@@ -53501,12 +53215,12 @@
 		<edge from-layer="2318" from-port="1" to-layer="2321" to-port="0" />
 		<edge from-layer="2319" from-port="0" to-layer="2320" to-port="0" />
 		<edge from-layer="2320" from-port="1" to-layer="2321" to-port="1" />
-		<edge from-layer="2321" from-port="2" to-layer="2324" to-port="0" />
-		<edge from-layer="2322" from-port="0" to-layer="2323" to-port="0" />
-		<edge from-layer="2323" from-port="1" to-layer="2324" to-port="1" />
-		<edge from-layer="2324" from-port="2" to-layer="2326" to-port="0" />
-		<edge from-layer="2325" from-port="0" to-layer="2326" to-port="1" />
-		<edge from-layer="2326" from-port="2" to-layer="2327" to-port="1" />
+		<edge from-layer="2321" from-port="2" to-layer="2323" to-port="0" />
+		<edge from-layer="2322" from-port="0" to-layer="2323" to-port="1" />
+		<edge from-layer="2323" from-port="2" to-layer="2325" to-port="0" />
+		<edge from-layer="2324" from-port="0" to-layer="2325" to-port="1" />
+		<edge from-layer="2325" from-port="2" to-layer="2326" to-port="0" />
+		<edge from-layer="2326" from-port="1" to-layer="2327" to-port="1" />
 		<edge from-layer="2327" from-port="2" to-layer="2328" to-port="1" />
 		<edge from-layer="2328" from-port="2" to-layer="2330" to-port="0" />
 		<edge from-layer="2328" from-port="2" to-layer="2334" to-port="0" />
@@ -53522,12 +53236,12 @@
 		<edge from-layer="2338" from-port="1" to-layer="2341" to-port="0" />
 		<edge from-layer="2339" from-port="0" to-layer="2340" to-port="0" />
 		<edge from-layer="2340" from-port="1" to-layer="2341" to-port="1" />
-		<edge from-layer="2341" from-port="2" to-layer="2344" to-port="0" />
-		<edge from-layer="2342" from-port="0" to-layer="2343" to-port="0" />
-		<edge from-layer="2343" from-port="1" to-layer="2344" to-port="1" />
-		<edge from-layer="2344" from-port="2" to-layer="2346" to-port="0" />
-		<edge from-layer="2345" from-port="0" to-layer="2346" to-port="1" />
-		<edge from-layer="2346" from-port="2" to-layer="2347" to-port="1" />
+		<edge from-layer="2341" from-port="2" to-layer="2343" to-port="0" />
+		<edge from-layer="2342" from-port="0" to-layer="2343" to-port="1" />
+		<edge from-layer="2343" from-port="2" to-layer="2345" to-port="0" />
+		<edge from-layer="2344" from-port="0" to-layer="2345" to-port="1" />
+		<edge from-layer="2345" from-port="2" to-layer="2346" to-port="0" />
+		<edge from-layer="2346" from-port="1" to-layer="2347" to-port="1" />
 		<edge from-layer="2347" from-port="2" to-layer="2348" to-port="1" />
 		<edge from-layer="2348" from-port="2" to-layer="2349" to-port="1" />
 		<edge from-layer="2349" from-port="2" to-layer="2350" to-port="0" />
@@ -53545,18 +53259,18 @@
 		<edge from-layer="2360" from-port="0" to-layer="2361" to-port="1" />
 		<edge from-layer="2361" from-port="2" to-layer="2362" to-port="0" />
 		<edge from-layer="2362" from-port="1" to-layer="2414" to-port="0" />
-		<edge from-layer="2362" from-port="1" to-layer="2393" to-port="0" />
 		<edge from-layer="2362" from-port="1" to-layer="2373" to-port="0" />
+		<edge from-layer="2362" from-port="1" to-layer="2393" to-port="0" />
 		<edge from-layer="2363" from-port="0" to-layer="2364" to-port="0" />
 		<edge from-layer="2364" from-port="1" to-layer="2367" to-port="0" />
 		<edge from-layer="2365" from-port="0" to-layer="2366" to-port="0" />
 		<edge from-layer="2366" from-port="1" to-layer="2367" to-port="1" />
-		<edge from-layer="2367" from-port="2" to-layer="2370" to-port="0" />
-		<edge from-layer="2368" from-port="0" to-layer="2369" to-port="0" />
-		<edge from-layer="2369" from-port="1" to-layer="2370" to-port="1" />
-		<edge from-layer="2370" from-port="2" to-layer="2372" to-port="0" />
-		<edge from-layer="2371" from-port="0" to-layer="2372" to-port="1" />
-		<edge from-layer="2372" from-port="2" to-layer="2373" to-port="1" />
+		<edge from-layer="2367" from-port="2" to-layer="2369" to-port="0" />
+		<edge from-layer="2368" from-port="0" to-layer="2369" to-port="1" />
+		<edge from-layer="2369" from-port="2" to-layer="2371" to-port="0" />
+		<edge from-layer="2370" from-port="0" to-layer="2371" to-port="1" />
+		<edge from-layer="2371" from-port="2" to-layer="2372" to-port="0" />
+		<edge from-layer="2372" from-port="1" to-layer="2373" to-port="1" />
 		<edge from-layer="2373" from-port="2" to-layer="2374" to-port="1" />
 		<edge from-layer="2374" from-port="2" to-layer="2376" to-port="0" />
 		<edge from-layer="2375" from-port="0" to-layer="2376" to-port="1" />
@@ -53572,12 +53286,12 @@
 		<edge from-layer="2384" from-port="1" to-layer="2387" to-port="0" />
 		<edge from-layer="2385" from-port="0" to-layer="2386" to-port="0" />
 		<edge from-layer="2386" from-port="1" to-layer="2387" to-port="1" />
-		<edge from-layer="2387" from-port="2" to-layer="2390" to-port="0" />
-		<edge from-layer="2388" from-port="0" to-layer="2389" to-port="0" />
-		<edge from-layer="2389" from-port="1" to-layer="2390" to-port="1" />
-		<edge from-layer="2390" from-port="2" to-layer="2392" to-port="0" />
-		<edge from-layer="2391" from-port="0" to-layer="2392" to-port="1" />
-		<edge from-layer="2392" from-port="2" to-layer="2393" to-port="1" />
+		<edge from-layer="2387" from-port="2" to-layer="2389" to-port="0" />
+		<edge from-layer="2388" from-port="0" to-layer="2389" to-port="1" />
+		<edge from-layer="2389" from-port="2" to-layer="2391" to-port="0" />
+		<edge from-layer="2390" from-port="0" to-layer="2391" to-port="1" />
+		<edge from-layer="2391" from-port="2" to-layer="2392" to-port="0" />
+		<edge from-layer="2392" from-port="1" to-layer="2393" to-port="1" />
 		<edge from-layer="2393" from-port="2" to-layer="2395" to-port="0" />
 		<edge from-layer="2394" from-port="0" to-layer="2395" to-port="1" />
 		<edge from-layer="2395" from-port="2" to-layer="2397" to-port="0" />
@@ -53593,12 +53307,12 @@
 		<edge from-layer="2405" from-port="1" to-layer="2408" to-port="0" />
 		<edge from-layer="2406" from-port="0" to-layer="2407" to-port="0" />
 		<edge from-layer="2407" from-port="1" to-layer="2408" to-port="1" />
-		<edge from-layer="2408" from-port="2" to-layer="2411" to-port="0" />
-		<edge from-layer="2409" from-port="0" to-layer="2410" to-port="0" />
-		<edge from-layer="2410" from-port="1" to-layer="2411" to-port="1" />
-		<edge from-layer="2411" from-port="2" to-layer="2413" to-port="0" />
-		<edge from-layer="2412" from-port="0" to-layer="2413" to-port="1" />
-		<edge from-layer="2413" from-port="2" to-layer="2414" to-port="1" />
+		<edge from-layer="2408" from-port="2" to-layer="2410" to-port="0" />
+		<edge from-layer="2409" from-port="0" to-layer="2410" to-port="1" />
+		<edge from-layer="2410" from-port="2" to-layer="2412" to-port="0" />
+		<edge from-layer="2411" from-port="0" to-layer="2412" to-port="1" />
+		<edge from-layer="2412" from-port="2" to-layer="2413" to-port="0" />
+		<edge from-layer="2413" from-port="1" to-layer="2414" to-port="1" />
 		<edge from-layer="2414" from-port="2" to-layer="2415" to-port="1" />
 		<edge from-layer="2415" from-port="2" to-layer="2417" to-port="0" />
 		<edge from-layer="2416" from-port="0" to-layer="2417" to-port="1" />
@@ -53625,12 +53339,12 @@
 		<edge from-layer="2435" from-port="1" to-layer="2438" to-port="0" />
 		<edge from-layer="2436" from-port="0" to-layer="2437" to-port="0" />
 		<edge from-layer="2437" from-port="1" to-layer="2438" to-port="1" />
-		<edge from-layer="2438" from-port="2" to-layer="2441" to-port="0" />
-		<edge from-layer="2439" from-port="0" to-layer="2440" to-port="0" />
-		<edge from-layer="2440" from-port="1" to-layer="2441" to-port="1" />
-		<edge from-layer="2441" from-port="2" to-layer="2443" to-port="0" />
-		<edge from-layer="2442" from-port="0" to-layer="2443" to-port="1" />
-		<edge from-layer="2443" from-port="2" to-layer="2444" to-port="1" />
+		<edge from-layer="2438" from-port="2" to-layer="2440" to-port="0" />
+		<edge from-layer="2439" from-port="0" to-layer="2440" to-port="1" />
+		<edge from-layer="2440" from-port="2" to-layer="2442" to-port="0" />
+		<edge from-layer="2441" from-port="0" to-layer="2442" to-port="1" />
+		<edge from-layer="2442" from-port="2" to-layer="2443" to-port="0" />
+		<edge from-layer="2443" from-port="1" to-layer="2444" to-port="1" />
 		<edge from-layer="2444" from-port="2" to-layer="2445" to-port="1" />
 		<edge from-layer="2445" from-port="2" to-layer="2446" to-port="1" />
 		<edge from-layer="2446" from-port="2" to-layer="2447" to-port="0" />
@@ -53652,12 +53366,12 @@
 		<edge from-layer="2461" from-port="1" to-layer="2464" to-port="0" />
 		<edge from-layer="2462" from-port="0" to-layer="2463" to-port="0" />
 		<edge from-layer="2463" from-port="1" to-layer="2464" to-port="1" />
-		<edge from-layer="2464" from-port="2" to-layer="2467" to-port="0" />
-		<edge from-layer="2465" from-port="0" to-layer="2466" to-port="0" />
-		<edge from-layer="2466" from-port="1" to-layer="2467" to-port="1" />
-		<edge from-layer="2467" from-port="2" to-layer="2469" to-port="0" />
-		<edge from-layer="2468" from-port="0" to-layer="2469" to-port="1" />
-		<edge from-layer="2469" from-port="2" to-layer="2470" to-port="1" />
+		<edge from-layer="2464" from-port="2" to-layer="2466" to-port="0" />
+		<edge from-layer="2465" from-port="0" to-layer="2466" to-port="1" />
+		<edge from-layer="2466" from-port="2" to-layer="2468" to-port="0" />
+		<edge from-layer="2467" from-port="0" to-layer="2468" to-port="1" />
+		<edge from-layer="2468" from-port="2" to-layer="2469" to-port="0" />
+		<edge from-layer="2469" from-port="1" to-layer="2470" to-port="1" />
 		<edge from-layer="2470" from-port="2" to-layer="2471" to-port="1" />
 		<edge from-layer="2471" from-port="2" to-layer="2473" to-port="0" />
 		<edge from-layer="2471" from-port="2" to-layer="2477" to-port="0" />
@@ -53673,12 +53387,12 @@
 		<edge from-layer="2481" from-port="1" to-layer="2484" to-port="0" />
 		<edge from-layer="2482" from-port="0" to-layer="2483" to-port="0" />
 		<edge from-layer="2483" from-port="1" to-layer="2484" to-port="1" />
-		<edge from-layer="2484" from-port="2" to-layer="2487" to-port="0" />
-		<edge from-layer="2485" from-port="0" to-layer="2486" to-port="0" />
-		<edge from-layer="2486" from-port="1" to-layer="2487" to-port="1" />
-		<edge from-layer="2487" from-port="2" to-layer="2489" to-port="0" />
-		<edge from-layer="2488" from-port="0" to-layer="2489" to-port="1" />
-		<edge from-layer="2489" from-port="2" to-layer="2490" to-port="1" />
+		<edge from-layer="2484" from-port="2" to-layer="2486" to-port="0" />
+		<edge from-layer="2485" from-port="0" to-layer="2486" to-port="1" />
+		<edge from-layer="2486" from-port="2" to-layer="2488" to-port="0" />
+		<edge from-layer="2487" from-port="0" to-layer="2488" to-port="1" />
+		<edge from-layer="2488" from-port="2" to-layer="2489" to-port="0" />
+		<edge from-layer="2489" from-port="1" to-layer="2490" to-port="1" />
 		<edge from-layer="2490" from-port="2" to-layer="2491" to-port="1" />
 		<edge from-layer="2491" from-port="2" to-layer="2492" to-port="1" />
 		<edge from-layer="2492" from-port="2" to-layer="2493" to-port="0" />
@@ -53702,12 +53416,12 @@
 		<edge from-layer="2507" from-port="1" to-layer="2510" to-port="0" />
 		<edge from-layer="2508" from-port="0" to-layer="2509" to-port="0" />
 		<edge from-layer="2509" from-port="1" to-layer="2510" to-port="1" />
-		<edge from-layer="2510" from-port="2" to-layer="2513" to-port="0" />
-		<edge from-layer="2511" from-port="0" to-layer="2512" to-port="0" />
-		<edge from-layer="2512" from-port="1" to-layer="2513" to-port="1" />
-		<edge from-layer="2513" from-port="2" to-layer="2515" to-port="0" />
-		<edge from-layer="2514" from-port="0" to-layer="2515" to-port="1" />
-		<edge from-layer="2515" from-port="2" to-layer="2516" to-port="1" />
+		<edge from-layer="2510" from-port="2" to-layer="2512" to-port="0" />
+		<edge from-layer="2511" from-port="0" to-layer="2512" to-port="1" />
+		<edge from-layer="2512" from-port="2" to-layer="2514" to-port="0" />
+		<edge from-layer="2513" from-port="0" to-layer="2514" to-port="1" />
+		<edge from-layer="2514" from-port="2" to-layer="2515" to-port="0" />
+		<edge from-layer="2515" from-port="1" to-layer="2516" to-port="1" />
 		<edge from-layer="2516" from-port="2" to-layer="2517" to-port="1" />
 		<edge from-layer="2517" from-port="2" to-layer="2519" to-port="0" />
 		<edge from-layer="2518" from-port="0" to-layer="2519" to-port="1" />
@@ -53723,12 +53437,12 @@
 		<edge from-layer="2527" from-port="1" to-layer="2530" to-port="0" />
 		<edge from-layer="2528" from-port="0" to-layer="2529" to-port="0" />
 		<edge from-layer="2529" from-port="1" to-layer="2530" to-port="1" />
-		<edge from-layer="2530" from-port="2" to-layer="2533" to-port="0" />
-		<edge from-layer="2531" from-port="0" to-layer="2532" to-port="0" />
-		<edge from-layer="2532" from-port="1" to-layer="2533" to-port="1" />
-		<edge from-layer="2533" from-port="2" to-layer="2535" to-port="0" />
-		<edge from-layer="2534" from-port="0" to-layer="2535" to-port="1" />
-		<edge from-layer="2535" from-port="2" to-layer="2536" to-port="1" />
+		<edge from-layer="2530" from-port="2" to-layer="2532" to-port="0" />
+		<edge from-layer="2531" from-port="0" to-layer="2532" to-port="1" />
+		<edge from-layer="2532" from-port="2" to-layer="2534" to-port="0" />
+		<edge from-layer="2533" from-port="0" to-layer="2534" to-port="1" />
+		<edge from-layer="2534" from-port="2" to-layer="2535" to-port="0" />
+		<edge from-layer="2535" from-port="1" to-layer="2536" to-port="1" />
 		<edge from-layer="2536" from-port="2" to-layer="2538" to-port="0" />
 		<edge from-layer="2537" from-port="0" to-layer="2538" to-port="1" />
 		<edge from-layer="2538" from-port="2" to-layer="2540" to-port="0" />
@@ -53744,12 +53458,12 @@
 		<edge from-layer="2548" from-port="1" to-layer="2551" to-port="0" />
 		<edge from-layer="2549" from-port="0" to-layer="2550" to-port="0" />
 		<edge from-layer="2550" from-port="1" to-layer="2551" to-port="1" />
-		<edge from-layer="2551" from-port="2" to-layer="2554" to-port="0" />
-		<edge from-layer="2552" from-port="0" to-layer="2553" to-port="0" />
-		<edge from-layer="2553" from-port="1" to-layer="2554" to-port="1" />
-		<edge from-layer="2554" from-port="2" to-layer="2556" to-port="0" />
-		<edge from-layer="2555" from-port="0" to-layer="2556" to-port="1" />
-		<edge from-layer="2556" from-port="2" to-layer="2557" to-port="1" />
+		<edge from-layer="2551" from-port="2" to-layer="2553" to-port="0" />
+		<edge from-layer="2552" from-port="0" to-layer="2553" to-port="1" />
+		<edge from-layer="2553" from-port="2" to-layer="2555" to-port="0" />
+		<edge from-layer="2554" from-port="0" to-layer="2555" to-port="1" />
+		<edge from-layer="2555" from-port="2" to-layer="2556" to-port="0" />
+		<edge from-layer="2556" from-port="1" to-layer="2557" to-port="1" />
 		<edge from-layer="2557" from-port="2" to-layer="2558" to-port="1" />
 		<edge from-layer="2558" from-port="2" to-layer="2560" to-port="0" />
 		<edge from-layer="2559" from-port="0" to-layer="2560" to-port="1" />
@@ -53776,12 +53490,12 @@
 		<edge from-layer="2578" from-port="1" to-layer="2581" to-port="0" />
 		<edge from-layer="2579" from-port="0" to-layer="2580" to-port="0" />
 		<edge from-layer="2580" from-port="1" to-layer="2581" to-port="1" />
-		<edge from-layer="2581" from-port="2" to-layer="2584" to-port="0" />
-		<edge from-layer="2582" from-port="0" to-layer="2583" to-port="0" />
-		<edge from-layer="2583" from-port="1" to-layer="2584" to-port="1" />
-		<edge from-layer="2584" from-port="2" to-layer="2586" to-port="0" />
-		<edge from-layer="2585" from-port="0" to-layer="2586" to-port="1" />
-		<edge from-layer="2586" from-port="2" to-layer="2587" to-port="1" />
+		<edge from-layer="2581" from-port="2" to-layer="2583" to-port="0" />
+		<edge from-layer="2582" from-port="0" to-layer="2583" to-port="1" />
+		<edge from-layer="2583" from-port="2" to-layer="2585" to-port="0" />
+		<edge from-layer="2584" from-port="0" to-layer="2585" to-port="1" />
+		<edge from-layer="2585" from-port="2" to-layer="2586" to-port="0" />
+		<edge from-layer="2586" from-port="1" to-layer="2587" to-port="1" />
 		<edge from-layer="2587" from-port="2" to-layer="2588" to-port="1" />
 		<edge from-layer="2588" from-port="2" to-layer="2589" to-port="1" />
 		<edge from-layer="2589" from-port="2" to-layer="2590" to-port="0" />
@@ -53803,12 +53517,12 @@
 		<edge from-layer="2604" from-port="1" to-layer="2607" to-port="0" />
 		<edge from-layer="2605" from-port="0" to-layer="2606" to-port="0" />
 		<edge from-layer="2606" from-port="1" to-layer="2607" to-port="1" />
-		<edge from-layer="2607" from-port="2" to-layer="2610" to-port="0" />
-		<edge from-layer="2608" from-port="0" to-layer="2609" to-port="0" />
-		<edge from-layer="2609" from-port="1" to-layer="2610" to-port="1" />
-		<edge from-layer="2610" from-port="2" to-layer="2612" to-port="0" />
-		<edge from-layer="2611" from-port="0" to-layer="2612" to-port="1" />
-		<edge from-layer="2612" from-port="2" to-layer="2613" to-port="1" />
+		<edge from-layer="2607" from-port="2" to-layer="2609" to-port="0" />
+		<edge from-layer="2608" from-port="0" to-layer="2609" to-port="1" />
+		<edge from-layer="2609" from-port="2" to-layer="2611" to-port="0" />
+		<edge from-layer="2610" from-port="0" to-layer="2611" to-port="1" />
+		<edge from-layer="2611" from-port="2" to-layer="2612" to-port="0" />
+		<edge from-layer="2612" from-port="1" to-layer="2613" to-port="1" />
 		<edge from-layer="2613" from-port="2" to-layer="2614" to-port="1" />
 		<edge from-layer="2614" from-port="2" to-layer="2616" to-port="0" />
 		<edge from-layer="2614" from-port="2" to-layer="2620" to-port="0" />
@@ -53824,12 +53538,12 @@
 		<edge from-layer="2624" from-port="1" to-layer="2627" to-port="0" />
 		<edge from-layer="2625" from-port="0" to-layer="2626" to-port="0" />
 		<edge from-layer="2626" from-port="1" to-layer="2627" to-port="1" />
-		<edge from-layer="2627" from-port="2" to-layer="2630" to-port="0" />
-		<edge from-layer="2628" from-port="0" to-layer="2629" to-port="0" />
-		<edge from-layer="2629" from-port="1" to-layer="2630" to-port="1" />
-		<edge from-layer="2630" from-port="2" to-layer="2632" to-port="0" />
-		<edge from-layer="2631" from-port="0" to-layer="2632" to-port="1" />
-		<edge from-layer="2632" from-port="2" to-layer="2633" to-port="1" />
+		<edge from-layer="2627" from-port="2" to-layer="2629" to-port="0" />
+		<edge from-layer="2628" from-port="0" to-layer="2629" to-port="1" />
+		<edge from-layer="2629" from-port="2" to-layer="2631" to-port="0" />
+		<edge from-layer="2630" from-port="0" to-layer="2631" to-port="1" />
+		<edge from-layer="2631" from-port="2" to-layer="2632" to-port="0" />
+		<edge from-layer="2632" from-port="1" to-layer="2633" to-port="1" />
 		<edge from-layer="2633" from-port="2" to-layer="2634" to-port="1" />
 		<edge from-layer="2634" from-port="2" to-layer="2635" to-port="1" />
 		<edge from-layer="2635" from-port="2" to-layer="2636" to-port="0" />
@@ -53846,19 +53560,19 @@
 		<edge from-layer="2645" from-port="2" to-layer="2647" to-port="0" />
 		<edge from-layer="2646" from-port="0" to-layer="2647" to-port="1" />
 		<edge from-layer="2647" from-port="2" to-layer="2648" to-port="0" />
-		<edge from-layer="2648" from-port="1" to-layer="2700" to-port="0" />
 		<edge from-layer="2648" from-port="1" to-layer="2679" to-port="0" />
+		<edge from-layer="2648" from-port="1" to-layer="2700" to-port="0" />
 		<edge from-layer="2648" from-port="1" to-layer="2659" to-port="0" />
 		<edge from-layer="2649" from-port="0" to-layer="2650" to-port="0" />
 		<edge from-layer="2650" from-port="1" to-layer="2653" to-port="0" />
 		<edge from-layer="2651" from-port="0" to-layer="2652" to-port="0" />
 		<edge from-layer="2652" from-port="1" to-layer="2653" to-port="1" />
-		<edge from-layer="2653" from-port="2" to-layer="2656" to-port="0" />
-		<edge from-layer="2654" from-port="0" to-layer="2655" to-port="0" />
-		<edge from-layer="2655" from-port="1" to-layer="2656" to-port="1" />
-		<edge from-layer="2656" from-port="2" to-layer="2658" to-port="0" />
-		<edge from-layer="2657" from-port="0" to-layer="2658" to-port="1" />
-		<edge from-layer="2658" from-port="2" to-layer="2659" to-port="1" />
+		<edge from-layer="2653" from-port="2" to-layer="2655" to-port="0" />
+		<edge from-layer="2654" from-port="0" to-layer="2655" to-port="1" />
+		<edge from-layer="2655" from-port="2" to-layer="2657" to-port="0" />
+		<edge from-layer="2656" from-port="0" to-layer="2657" to-port="1" />
+		<edge from-layer="2657" from-port="2" to-layer="2658" to-port="0" />
+		<edge from-layer="2658" from-port="1" to-layer="2659" to-port="1" />
 		<edge from-layer="2659" from-port="2" to-layer="2660" to-port="1" />
 		<edge from-layer="2660" from-port="2" to-layer="2662" to-port="0" />
 		<edge from-layer="2661" from-port="0" to-layer="2662" to-port="1" />
@@ -53867,19 +53581,19 @@
 		<edge from-layer="2664" from-port="2" to-layer="2668" to-port="0" />
 		<edge from-layer="2665" from-port="0" to-layer="2667" to-port="0" />
 		<edge from-layer="2666" from-port="0" to-layer="2667" to-port="1" />
-		<edge from-layer="2667" from-port="2" to-layer="2668" to-port="1" />
 		<edge from-layer="2667" from-port="2" to-layer="2684" to-port="1" />
+		<edge from-layer="2667" from-port="2" to-layer="2668" to-port="1" />
 		<edge from-layer="2668" from-port="2" to-layer="2685" to-port="0" />
 		<edge from-layer="2669" from-port="0" to-layer="2670" to-port="0" />
 		<edge from-layer="2670" from-port="1" to-layer="2673" to-port="0" />
 		<edge from-layer="2671" from-port="0" to-layer="2672" to-port="0" />
 		<edge from-layer="2672" from-port="1" to-layer="2673" to-port="1" />
-		<edge from-layer="2673" from-port="2" to-layer="2676" to-port="0" />
-		<edge from-layer="2674" from-port="0" to-layer="2675" to-port="0" />
-		<edge from-layer="2675" from-port="1" to-layer="2676" to-port="1" />
-		<edge from-layer="2676" from-port="2" to-layer="2678" to-port="0" />
-		<edge from-layer="2677" from-port="0" to-layer="2678" to-port="1" />
-		<edge from-layer="2678" from-port="2" to-layer="2679" to-port="1" />
+		<edge from-layer="2673" from-port="2" to-layer="2675" to-port="0" />
+		<edge from-layer="2674" from-port="0" to-layer="2675" to-port="1" />
+		<edge from-layer="2675" from-port="2" to-layer="2677" to-port="0" />
+		<edge from-layer="2676" from-port="0" to-layer="2677" to-port="1" />
+		<edge from-layer="2677" from-port="2" to-layer="2678" to-port="0" />
+		<edge from-layer="2678" from-port="1" to-layer="2679" to-port="1" />
 		<edge from-layer="2679" from-port="2" to-layer="2681" to-port="0" />
 		<edge from-layer="2680" from-port="0" to-layer="2681" to-port="1" />
 		<edge from-layer="2681" from-port="2" to-layer="2683" to-port="0" />
@@ -53895,12 +53609,12 @@
 		<edge from-layer="2691" from-port="1" to-layer="2694" to-port="0" />
 		<edge from-layer="2692" from-port="0" to-layer="2693" to-port="0" />
 		<edge from-layer="2693" from-port="1" to-layer="2694" to-port="1" />
-		<edge from-layer="2694" from-port="2" to-layer="2697" to-port="0" />
-		<edge from-layer="2695" from-port="0" to-layer="2696" to-port="0" />
-		<edge from-layer="2696" from-port="1" to-layer="2697" to-port="1" />
-		<edge from-layer="2697" from-port="2" to-layer="2699" to-port="0" />
-		<edge from-layer="2698" from-port="0" to-layer="2699" to-port="1" />
-		<edge from-layer="2699" from-port="2" to-layer="2700" to-port="1" />
+		<edge from-layer="2694" from-port="2" to-layer="2696" to-port="0" />
+		<edge from-layer="2695" from-port="0" to-layer="2696" to-port="1" />
+		<edge from-layer="2696" from-port="2" to-layer="2698" to-port="0" />
+		<edge from-layer="2697" from-port="0" to-layer="2698" to-port="1" />
+		<edge from-layer="2698" from-port="2" to-layer="2699" to-port="0" />
+		<edge from-layer="2699" from-port="1" to-layer="2700" to-port="1" />
 		<edge from-layer="2700" from-port="2" to-layer="2701" to-port="1" />
 		<edge from-layer="2701" from-port="2" to-layer="2703" to-port="0" />
 		<edge from-layer="2702" from-port="0" to-layer="2703" to-port="1" />
@@ -53927,12 +53641,12 @@
 		<edge from-layer="2721" from-port="1" to-layer="2724" to-port="0" />
 		<edge from-layer="2722" from-port="0" to-layer="2723" to-port="0" />
 		<edge from-layer="2723" from-port="1" to-layer="2724" to-port="1" />
-		<edge from-layer="2724" from-port="2" to-layer="2727" to-port="0" />
-		<edge from-layer="2725" from-port="0" to-layer="2726" to-port="0" />
-		<edge from-layer="2726" from-port="1" to-layer="2727" to-port="1" />
-		<edge from-layer="2727" from-port="2" to-layer="2729" to-port="0" />
-		<edge from-layer="2728" from-port="0" to-layer="2729" to-port="1" />
-		<edge from-layer="2729" from-port="2" to-layer="2730" to-port="1" />
+		<edge from-layer="2724" from-port="2" to-layer="2726" to-port="0" />
+		<edge from-layer="2725" from-port="0" to-layer="2726" to-port="1" />
+		<edge from-layer="2726" from-port="2" to-layer="2728" to-port="0" />
+		<edge from-layer="2727" from-port="0" to-layer="2728" to-port="1" />
+		<edge from-layer="2728" from-port="2" to-layer="2729" to-port="0" />
+		<edge from-layer="2729" from-port="1" to-layer="2730" to-port="1" />
 		<edge from-layer="2730" from-port="2" to-layer="2731" to-port="1" />
 		<edge from-layer="2731" from-port="2" to-layer="2732" to-port="1" />
 		<edge from-layer="2732" from-port="2" to-layer="2733" to-port="0" />
@@ -53954,12 +53668,12 @@
 		<edge from-layer="2747" from-port="1" to-layer="2750" to-port="0" />
 		<edge from-layer="2748" from-port="0" to-layer="2749" to-port="0" />
 		<edge from-layer="2749" from-port="1" to-layer="2750" to-port="1" />
-		<edge from-layer="2750" from-port="2" to-layer="2753" to-port="0" />
-		<edge from-layer="2751" from-port="0" to-layer="2752" to-port="0" />
-		<edge from-layer="2752" from-port="1" to-layer="2753" to-port="1" />
-		<edge from-layer="2753" from-port="2" to-layer="2755" to-port="0" />
-		<edge from-layer="2754" from-port="0" to-layer="2755" to-port="1" />
-		<edge from-layer="2755" from-port="2" to-layer="2756" to-port="1" />
+		<edge from-layer="2750" from-port="2" to-layer="2752" to-port="0" />
+		<edge from-layer="2751" from-port="0" to-layer="2752" to-port="1" />
+		<edge from-layer="2752" from-port="2" to-layer="2754" to-port="0" />
+		<edge from-layer="2753" from-port="0" to-layer="2754" to-port="1" />
+		<edge from-layer="2754" from-port="2" to-layer="2755" to-port="0" />
+		<edge from-layer="2755" from-port="1" to-layer="2756" to-port="1" />
 		<edge from-layer="2756" from-port="2" to-layer="2757" to-port="1" />
 		<edge from-layer="2757" from-port="2" to-layer="2763" to-port="0" />
 		<edge from-layer="2757" from-port="2" to-layer="2759" to-port="0" />
@@ -53975,12 +53689,12 @@
 		<edge from-layer="2767" from-port="1" to-layer="2770" to-port="0" />
 		<edge from-layer="2768" from-port="0" to-layer="2769" to-port="0" />
 		<edge from-layer="2769" from-port="1" to-layer="2770" to-port="1" />
-		<edge from-layer="2770" from-port="2" to-layer="2773" to-port="0" />
-		<edge from-layer="2771" from-port="0" to-layer="2772" to-port="0" />
-		<edge from-layer="2772" from-port="1" to-layer="2773" to-port="1" />
-		<edge from-layer="2773" from-port="2" to-layer="2775" to-port="0" />
-		<edge from-layer="2774" from-port="0" to-layer="2775" to-port="1" />
-		<edge from-layer="2775" from-port="2" to-layer="2776" to-port="1" />
+		<edge from-layer="2770" from-port="2" to-layer="2772" to-port="0" />
+		<edge from-layer="2771" from-port="0" to-layer="2772" to-port="1" />
+		<edge from-layer="2772" from-port="2" to-layer="2774" to-port="0" />
+		<edge from-layer="2773" from-port="0" to-layer="2774" to-port="1" />
+		<edge from-layer="2774" from-port="2" to-layer="2775" to-port="0" />
+		<edge from-layer="2775" from-port="1" to-layer="2776" to-port="1" />
 		<edge from-layer="2776" from-port="2" to-layer="2777" to-port="1" />
 		<edge from-layer="2777" from-port="2" to-layer="2778" to-port="1" />
 		<edge from-layer="2778" from-port="2" to-layer="2779" to-port="0" />
@@ -54004,12 +53718,12 @@
 		<edge from-layer="2793" from-port="1" to-layer="2796" to-port="0" />
 		<edge from-layer="2794" from-port="0" to-layer="2795" to-port="0" />
 		<edge from-layer="2795" from-port="1" to-layer="2796" to-port="1" />
-		<edge from-layer="2796" from-port="2" to-layer="2799" to-port="0" />
-		<edge from-layer="2797" from-port="0" to-layer="2798" to-port="0" />
-		<edge from-layer="2798" from-port="1" to-layer="2799" to-port="1" />
-		<edge from-layer="2799" from-port="2" to-layer="2801" to-port="0" />
-		<edge from-layer="2800" from-port="0" to-layer="2801" to-port="1" />
-		<edge from-layer="2801" from-port="2" to-layer="2802" to-port="1" />
+		<edge from-layer="2796" from-port="2" to-layer="2798" to-port="0" />
+		<edge from-layer="2797" from-port="0" to-layer="2798" to-port="1" />
+		<edge from-layer="2798" from-port="2" to-layer="2800" to-port="0" />
+		<edge from-layer="2799" from-port="0" to-layer="2800" to-port="1" />
+		<edge from-layer="2800" from-port="2" to-layer="2801" to-port="0" />
+		<edge from-layer="2801" from-port="1" to-layer="2802" to-port="1" />
 		<edge from-layer="2802" from-port="2" to-layer="2803" to-port="1" />
 		<edge from-layer="2803" from-port="2" to-layer="2805" to-port="0" />
 		<edge from-layer="2804" from-port="0" to-layer="2805" to-port="1" />
@@ -54025,12 +53739,12 @@
 		<edge from-layer="2813" from-port="1" to-layer="2816" to-port="0" />
 		<edge from-layer="2814" from-port="0" to-layer="2815" to-port="0" />
 		<edge from-layer="2815" from-port="1" to-layer="2816" to-port="1" />
-		<edge from-layer="2816" from-port="2" to-layer="2819" to-port="0" />
-		<edge from-layer="2817" from-port="0" to-layer="2818" to-port="0" />
-		<edge from-layer="2818" from-port="1" to-layer="2819" to-port="1" />
-		<edge from-layer="2819" from-port="2" to-layer="2821" to-port="0" />
-		<edge from-layer="2820" from-port="0" to-layer="2821" to-port="1" />
-		<edge from-layer="2821" from-port="2" to-layer="2822" to-port="1" />
+		<edge from-layer="2816" from-port="2" to-layer="2818" to-port="0" />
+		<edge from-layer="2817" from-port="0" to-layer="2818" to-port="1" />
+		<edge from-layer="2818" from-port="2" to-layer="2820" to-port="0" />
+		<edge from-layer="2819" from-port="0" to-layer="2820" to-port="1" />
+		<edge from-layer="2820" from-port="2" to-layer="2821" to-port="0" />
+		<edge from-layer="2821" from-port="1" to-layer="2822" to-port="1" />
 		<edge from-layer="2822" from-port="2" to-layer="2824" to-port="0" />
 		<edge from-layer="2823" from-port="0" to-layer="2824" to-port="1" />
 		<edge from-layer="2824" from-port="2" to-layer="2826" to-port="0" />
@@ -54046,12 +53760,12 @@
 		<edge from-layer="2834" from-port="1" to-layer="2837" to-port="0" />
 		<edge from-layer="2835" from-port="0" to-layer="2836" to-port="0" />
 		<edge from-layer="2836" from-port="1" to-layer="2837" to-port="1" />
-		<edge from-layer="2837" from-port="2" to-layer="2840" to-port="0" />
-		<edge from-layer="2838" from-port="0" to-layer="2839" to-port="0" />
-		<edge from-layer="2839" from-port="1" to-layer="2840" to-port="1" />
-		<edge from-layer="2840" from-port="2" to-layer="2842" to-port="0" />
-		<edge from-layer="2841" from-port="0" to-layer="2842" to-port="1" />
-		<edge from-layer="2842" from-port="2" to-layer="2843" to-port="1" />
+		<edge from-layer="2837" from-port="2" to-layer="2839" to-port="0" />
+		<edge from-layer="2838" from-port="0" to-layer="2839" to-port="1" />
+		<edge from-layer="2839" from-port="2" to-layer="2841" to-port="0" />
+		<edge from-layer="2840" from-port="0" to-layer="2841" to-port="1" />
+		<edge from-layer="2841" from-port="2" to-layer="2842" to-port="0" />
+		<edge from-layer="2842" from-port="1" to-layer="2843" to-port="1" />
 		<edge from-layer="2843" from-port="2" to-layer="2844" to-port="1" />
 		<edge from-layer="2844" from-port="2" to-layer="2846" to-port="0" />
 		<edge from-layer="2845" from-port="0" to-layer="2846" to-port="1" />
@@ -54078,12 +53792,12 @@
 		<edge from-layer="2864" from-port="1" to-layer="2867" to-port="0" />
 		<edge from-layer="2865" from-port="0" to-layer="2866" to-port="0" />
 		<edge from-layer="2866" from-port="1" to-layer="2867" to-port="1" />
-		<edge from-layer="2867" from-port="2" to-layer="2870" to-port="0" />
-		<edge from-layer="2868" from-port="0" to-layer="2869" to-port="0" />
-		<edge from-layer="2869" from-port="1" to-layer="2870" to-port="1" />
-		<edge from-layer="2870" from-port="2" to-layer="2872" to-port="0" />
-		<edge from-layer="2871" from-port="0" to-layer="2872" to-port="1" />
-		<edge from-layer="2872" from-port="2" to-layer="2873" to-port="1" />
+		<edge from-layer="2867" from-port="2" to-layer="2869" to-port="0" />
+		<edge from-layer="2868" from-port="0" to-layer="2869" to-port="1" />
+		<edge from-layer="2869" from-port="2" to-layer="2871" to-port="0" />
+		<edge from-layer="2870" from-port="0" to-layer="2871" to-port="1" />
+		<edge from-layer="2871" from-port="2" to-layer="2872" to-port="0" />
+		<edge from-layer="2872" from-port="1" to-layer="2873" to-port="1" />
 		<edge from-layer="2873" from-port="2" to-layer="2874" to-port="1" />
 		<edge from-layer="2874" from-port="2" to-layer="2875" to-port="1" />
 		<edge from-layer="2875" from-port="2" to-layer="2876" to-port="0" />
@@ -54105,12 +53819,12 @@
 		<edge from-layer="2890" from-port="1" to-layer="2893" to-port="0" />
 		<edge from-layer="2891" from-port="0" to-layer="2892" to-port="0" />
 		<edge from-layer="2892" from-port="1" to-layer="2893" to-port="1" />
-		<edge from-layer="2893" from-port="2" to-layer="2896" to-port="0" />
-		<edge from-layer="2894" from-port="0" to-layer="2895" to-port="0" />
-		<edge from-layer="2895" from-port="1" to-layer="2896" to-port="1" />
-		<edge from-layer="2896" from-port="2" to-layer="2898" to-port="0" />
-		<edge from-layer="2897" from-port="0" to-layer="2898" to-port="1" />
-		<edge from-layer="2898" from-port="2" to-layer="2899" to-port="1" />
+		<edge from-layer="2893" from-port="2" to-layer="2895" to-port="0" />
+		<edge from-layer="2894" from-port="0" to-layer="2895" to-port="1" />
+		<edge from-layer="2895" from-port="2" to-layer="2897" to-port="0" />
+		<edge from-layer="2896" from-port="0" to-layer="2897" to-port="1" />
+		<edge from-layer="2897" from-port="2" to-layer="2898" to-port="0" />
+		<edge from-layer="2898" from-port="1" to-layer="2899" to-port="1" />
 		<edge from-layer="2899" from-port="2" to-layer="2900" to-port="1" />
 		<edge from-layer="2900" from-port="2" to-layer="2902" to-port="0" />
 		<edge from-layer="2900" from-port="2" to-layer="2906" to-port="0" />
@@ -54126,12 +53840,12 @@
 		<edge from-layer="2910" from-port="1" to-layer="2913" to-port="0" />
 		<edge from-layer="2911" from-port="0" to-layer="2912" to-port="0" />
 		<edge from-layer="2912" from-port="1" to-layer="2913" to-port="1" />
-		<edge from-layer="2913" from-port="2" to-layer="2916" to-port="0" />
-		<edge from-layer="2914" from-port="0" to-layer="2915" to-port="0" />
-		<edge from-layer="2915" from-port="1" to-layer="2916" to-port="1" />
-		<edge from-layer="2916" from-port="2" to-layer="2918" to-port="0" />
-		<edge from-layer="2917" from-port="0" to-layer="2918" to-port="1" />
-		<edge from-layer="2918" from-port="2" to-layer="2919" to-port="1" />
+		<edge from-layer="2913" from-port="2" to-layer="2915" to-port="0" />
+		<edge from-layer="2914" from-port="0" to-layer="2915" to-port="1" />
+		<edge from-layer="2915" from-port="2" to-layer="2917" to-port="0" />
+		<edge from-layer="2916" from-port="0" to-layer="2917" to-port="1" />
+		<edge from-layer="2917" from-port="2" to-layer="2918" to-port="0" />
+		<edge from-layer="2918" from-port="1" to-layer="2919" to-port="1" />
 		<edge from-layer="2919" from-port="2" to-layer="2920" to-port="1" />
 		<edge from-layer="2920" from-port="2" to-layer="2921" to-port="1" />
 		<edge from-layer="2921" from-port="2" to-layer="2922" to-port="0" />
@@ -54155,12 +53869,12 @@
 		<edge from-layer="2936" from-port="1" to-layer="2939" to-port="0" />
 		<edge from-layer="2937" from-port="0" to-layer="2938" to-port="0" />
 		<edge from-layer="2938" from-port="1" to-layer="2939" to-port="1" />
-		<edge from-layer="2939" from-port="2" to-layer="2942" to-port="0" />
-		<edge from-layer="2940" from-port="0" to-layer="2941" to-port="0" />
-		<edge from-layer="2941" from-port="1" to-layer="2942" to-port="1" />
-		<edge from-layer="2942" from-port="2" to-layer="2944" to-port="0" />
-		<edge from-layer="2943" from-port="0" to-layer="2944" to-port="1" />
-		<edge from-layer="2944" from-port="2" to-layer="2945" to-port="1" />
+		<edge from-layer="2939" from-port="2" to-layer="2941" to-port="0" />
+		<edge from-layer="2940" from-port="0" to-layer="2941" to-port="1" />
+		<edge from-layer="2941" from-port="2" to-layer="2943" to-port="0" />
+		<edge from-layer="2942" from-port="0" to-layer="2943" to-port="1" />
+		<edge from-layer="2943" from-port="2" to-layer="2944" to-port="0" />
+		<edge from-layer="2944" from-port="1" to-layer="2945" to-port="1" />
 		<edge from-layer="2945" from-port="2" to-layer="2946" to-port="1" />
 		<edge from-layer="2946" from-port="2" to-layer="2948" to-port="0" />
 		<edge from-layer="2947" from-port="0" to-layer="2948" to-port="1" />
@@ -54176,12 +53890,12 @@
 		<edge from-layer="2956" from-port="1" to-layer="2959" to-port="0" />
 		<edge from-layer="2957" from-port="0" to-layer="2958" to-port="0" />
 		<edge from-layer="2958" from-port="1" to-layer="2959" to-port="1" />
-		<edge from-layer="2959" from-port="2" to-layer="2962" to-port="0" />
-		<edge from-layer="2960" from-port="0" to-layer="2961" to-port="0" />
-		<edge from-layer="2961" from-port="1" to-layer="2962" to-port="1" />
-		<edge from-layer="2962" from-port="2" to-layer="2964" to-port="0" />
-		<edge from-layer="2963" from-port="0" to-layer="2964" to-port="1" />
-		<edge from-layer="2964" from-port="2" to-layer="2965" to-port="1" />
+		<edge from-layer="2959" from-port="2" to-layer="2961" to-port="0" />
+		<edge from-layer="2960" from-port="0" to-layer="2961" to-port="1" />
+		<edge from-layer="2961" from-port="2" to-layer="2963" to-port="0" />
+		<edge from-layer="2962" from-port="0" to-layer="2963" to-port="1" />
+		<edge from-layer="2963" from-port="2" to-layer="2964" to-port="0" />
+		<edge from-layer="2964" from-port="1" to-layer="2965" to-port="1" />
 		<edge from-layer="2965" from-port="2" to-layer="2967" to-port="0" />
 		<edge from-layer="2966" from-port="0" to-layer="2967" to-port="1" />
 		<edge from-layer="2967" from-port="2" to-layer="2969" to-port="0" />
@@ -54197,12 +53911,12 @@
 		<edge from-layer="2977" from-port="1" to-layer="2980" to-port="0" />
 		<edge from-layer="2978" from-port="0" to-layer="2979" to-port="0" />
 		<edge from-layer="2979" from-port="1" to-layer="2980" to-port="1" />
-		<edge from-layer="2980" from-port="2" to-layer="2983" to-port="0" />
-		<edge from-layer="2981" from-port="0" to-layer="2982" to-port="0" />
-		<edge from-layer="2982" from-port="1" to-layer="2983" to-port="1" />
-		<edge from-layer="2983" from-port="2" to-layer="2985" to-port="0" />
-		<edge from-layer="2984" from-port="0" to-layer="2985" to-port="1" />
-		<edge from-layer="2985" from-port="2" to-layer="2986" to-port="1" />
+		<edge from-layer="2980" from-port="2" to-layer="2982" to-port="0" />
+		<edge from-layer="2981" from-port="0" to-layer="2982" to-port="1" />
+		<edge from-layer="2982" from-port="2" to-layer="2984" to-port="0" />
+		<edge from-layer="2983" from-port="0" to-layer="2984" to-port="1" />
+		<edge from-layer="2984" from-port="2" to-layer="2985" to-port="0" />
+		<edge from-layer="2985" from-port="1" to-layer="2986" to-port="1" />
 		<edge from-layer="2986" from-port="2" to-layer="2987" to-port="1" />
 		<edge from-layer="2987" from-port="2" to-layer="2989" to-port="0" />
 		<edge from-layer="2988" from-port="0" to-layer="2989" to-port="1" />
@@ -54229,12 +53943,12 @@
 		<edge from-layer="3007" from-port="1" to-layer="3010" to-port="0" />
 		<edge from-layer="3008" from-port="0" to-layer="3009" to-port="0" />
 		<edge from-layer="3009" from-port="1" to-layer="3010" to-port="1" />
-		<edge from-layer="3010" from-port="2" to-layer="3013" to-port="0" />
-		<edge from-layer="3011" from-port="0" to-layer="3012" to-port="0" />
-		<edge from-layer="3012" from-port="1" to-layer="3013" to-port="1" />
-		<edge from-layer="3013" from-port="2" to-layer="3015" to-port="0" />
-		<edge from-layer="3014" from-port="0" to-layer="3015" to-port="1" />
-		<edge from-layer="3015" from-port="2" to-layer="3016" to-port="1" />
+		<edge from-layer="3010" from-port="2" to-layer="3012" to-port="0" />
+		<edge from-layer="3011" from-port="0" to-layer="3012" to-port="1" />
+		<edge from-layer="3012" from-port="2" to-layer="3014" to-port="0" />
+		<edge from-layer="3013" from-port="0" to-layer="3014" to-port="1" />
+		<edge from-layer="3014" from-port="2" to-layer="3015" to-port="0" />
+		<edge from-layer="3015" from-port="1" to-layer="3016" to-port="1" />
 		<edge from-layer="3016" from-port="2" to-layer="3017" to-port="1" />
 		<edge from-layer="3017" from-port="2" to-layer="3018" to-port="1" />
 		<edge from-layer="3018" from-port="2" to-layer="3019" to-port="0" />
@@ -54256,12 +53970,12 @@
 		<edge from-layer="3033" from-port="1" to-layer="3036" to-port="0" />
 		<edge from-layer="3034" from-port="0" to-layer="3035" to-port="0" />
 		<edge from-layer="3035" from-port="1" to-layer="3036" to-port="1" />
-		<edge from-layer="3036" from-port="2" to-layer="3039" to-port="0" />
-		<edge from-layer="3037" from-port="0" to-layer="3038" to-port="0" />
-		<edge from-layer="3038" from-port="1" to-layer="3039" to-port="1" />
-		<edge from-layer="3039" from-port="2" to-layer="3041" to-port="0" />
-		<edge from-layer="3040" from-port="0" to-layer="3041" to-port="1" />
-		<edge from-layer="3041" from-port="2" to-layer="3042" to-port="1" />
+		<edge from-layer="3036" from-port="2" to-layer="3038" to-port="0" />
+		<edge from-layer="3037" from-port="0" to-layer="3038" to-port="1" />
+		<edge from-layer="3038" from-port="2" to-layer="3040" to-port="0" />
+		<edge from-layer="3039" from-port="0" to-layer="3040" to-port="1" />
+		<edge from-layer="3040" from-port="2" to-layer="3041" to-port="0" />
+		<edge from-layer="3041" from-port="1" to-layer="3042" to-port="1" />
 		<edge from-layer="3042" from-port="2" to-layer="3043" to-port="1" />
 		<edge from-layer="3043" from-port="2" to-layer="3049" to-port="0" />
 		<edge from-layer="3043" from-port="2" to-layer="3045" to-port="0" />
@@ -54277,12 +53991,12 @@
 		<edge from-layer="3053" from-port="1" to-layer="3056" to-port="0" />
 		<edge from-layer="3054" from-port="0" to-layer="3055" to-port="0" />
 		<edge from-layer="3055" from-port="1" to-layer="3056" to-port="1" />
-		<edge from-layer="3056" from-port="2" to-layer="3059" to-port="0" />
-		<edge from-layer="3057" from-port="0" to-layer="3058" to-port="0" />
-		<edge from-layer="3058" from-port="1" to-layer="3059" to-port="1" />
-		<edge from-layer="3059" from-port="2" to-layer="3061" to-port="0" />
-		<edge from-layer="3060" from-port="0" to-layer="3061" to-port="1" />
-		<edge from-layer="3061" from-port="2" to-layer="3062" to-port="1" />
+		<edge from-layer="3056" from-port="2" to-layer="3058" to-port="0" />
+		<edge from-layer="3057" from-port="0" to-layer="3058" to-port="1" />
+		<edge from-layer="3058" from-port="2" to-layer="3060" to-port="0" />
+		<edge from-layer="3059" from-port="0" to-layer="3060" to-port="1" />
+		<edge from-layer="3060" from-port="2" to-layer="3061" to-port="0" />
+		<edge from-layer="3061" from-port="1" to-layer="3062" to-port="1" />
 		<edge from-layer="3062" from-port="2" to-layer="3063" to-port="1" />
 		<edge from-layer="3063" from-port="2" to-layer="3064" to-port="1" />
 		<edge from-layer="3064" from-port="2" to-layer="3065" to-port="0" />
@@ -54306,12 +54020,12 @@
 		<edge from-layer="3079" from-port="1" to-layer="3082" to-port="0" />
 		<edge from-layer="3080" from-port="0" to-layer="3081" to-port="0" />
 		<edge from-layer="3081" from-port="1" to-layer="3082" to-port="1" />
-		<edge from-layer="3082" from-port="2" to-layer="3085" to-port="0" />
-		<edge from-layer="3083" from-port="0" to-layer="3084" to-port="0" />
-		<edge from-layer="3084" from-port="1" to-layer="3085" to-port="1" />
-		<edge from-layer="3085" from-port="2" to-layer="3087" to-port="0" />
-		<edge from-layer="3086" from-port="0" to-layer="3087" to-port="1" />
-		<edge from-layer="3087" from-port="2" to-layer="3088" to-port="1" />
+		<edge from-layer="3082" from-port="2" to-layer="3084" to-port="0" />
+		<edge from-layer="3083" from-port="0" to-layer="3084" to-port="1" />
+		<edge from-layer="3084" from-port="2" to-layer="3086" to-port="0" />
+		<edge from-layer="3085" from-port="0" to-layer="3086" to-port="1" />
+		<edge from-layer="3086" from-port="2" to-layer="3087" to-port="0" />
+		<edge from-layer="3087" from-port="1" to-layer="3088" to-port="1" />
 		<edge from-layer="3088" from-port="2" to-layer="3089" to-port="1" />
 		<edge from-layer="3089" from-port="2" to-layer="3091" to-port="0" />
 		<edge from-layer="3090" from-port="0" to-layer="3091" to-port="1" />
@@ -54327,12 +54041,12 @@
 		<edge from-layer="3099" from-port="1" to-layer="3102" to-port="0" />
 		<edge from-layer="3100" from-port="0" to-layer="3101" to-port="0" />
 		<edge from-layer="3101" from-port="1" to-layer="3102" to-port="1" />
-		<edge from-layer="3102" from-port="2" to-layer="3105" to-port="0" />
-		<edge from-layer="3103" from-port="0" to-layer="3104" to-port="0" />
-		<edge from-layer="3104" from-port="1" to-layer="3105" to-port="1" />
-		<edge from-layer="3105" from-port="2" to-layer="3107" to-port="0" />
-		<edge from-layer="3106" from-port="0" to-layer="3107" to-port="1" />
-		<edge from-layer="3107" from-port="2" to-layer="3108" to-port="1" />
+		<edge from-layer="3102" from-port="2" to-layer="3104" to-port="0" />
+		<edge from-layer="3103" from-port="0" to-layer="3104" to-port="1" />
+		<edge from-layer="3104" from-port="2" to-layer="3106" to-port="0" />
+		<edge from-layer="3105" from-port="0" to-layer="3106" to-port="1" />
+		<edge from-layer="3106" from-port="2" to-layer="3107" to-port="0" />
+		<edge from-layer="3107" from-port="1" to-layer="3108" to-port="1" />
 		<edge from-layer="3108" from-port="2" to-layer="3110" to-port="0" />
 		<edge from-layer="3109" from-port="0" to-layer="3110" to-port="1" />
 		<edge from-layer="3110" from-port="2" to-layer="3112" to-port="0" />
@@ -54348,12 +54062,12 @@
 		<edge from-layer="3120" from-port="1" to-layer="3123" to-port="0" />
 		<edge from-layer="3121" from-port="0" to-layer="3122" to-port="0" />
 		<edge from-layer="3122" from-port="1" to-layer="3123" to-port="1" />
-		<edge from-layer="3123" from-port="2" to-layer="3126" to-port="0" />
-		<edge from-layer="3124" from-port="0" to-layer="3125" to-port="0" />
-		<edge from-layer="3125" from-port="1" to-layer="3126" to-port="1" />
-		<edge from-layer="3126" from-port="2" to-layer="3128" to-port="0" />
-		<edge from-layer="3127" from-port="0" to-layer="3128" to-port="1" />
-		<edge from-layer="3128" from-port="2" to-layer="3129" to-port="1" />
+		<edge from-layer="3123" from-port="2" to-layer="3125" to-port="0" />
+		<edge from-layer="3124" from-port="0" to-layer="3125" to-port="1" />
+		<edge from-layer="3125" from-port="2" to-layer="3127" to-port="0" />
+		<edge from-layer="3126" from-port="0" to-layer="3127" to-port="1" />
+		<edge from-layer="3127" from-port="2" to-layer="3128" to-port="0" />
+		<edge from-layer="3128" from-port="1" to-layer="3129" to-port="1" />
 		<edge from-layer="3129" from-port="2" to-layer="3130" to-port="1" />
 		<edge from-layer="3130" from-port="2" to-layer="3132" to-port="0" />
 		<edge from-layer="3131" from-port="0" to-layer="3132" to-port="1" />
@@ -54380,12 +54094,12 @@
 		<edge from-layer="3150" from-port="1" to-layer="3153" to-port="0" />
 		<edge from-layer="3151" from-port="0" to-layer="3152" to-port="0" />
 		<edge from-layer="3152" from-port="1" to-layer="3153" to-port="1" />
-		<edge from-layer="3153" from-port="2" to-layer="3156" to-port="0" />
-		<edge from-layer="3154" from-port="0" to-layer="3155" to-port="0" />
-		<edge from-layer="3155" from-port="1" to-layer="3156" to-port="1" />
-		<edge from-layer="3156" from-port="2" to-layer="3158" to-port="0" />
-		<edge from-layer="3157" from-port="0" to-layer="3158" to-port="1" />
-		<edge from-layer="3158" from-port="2" to-layer="3159" to-port="1" />
+		<edge from-layer="3153" from-port="2" to-layer="3155" to-port="0" />
+		<edge from-layer="3154" from-port="0" to-layer="3155" to-port="1" />
+		<edge from-layer="3155" from-port="2" to-layer="3157" to-port="0" />
+		<edge from-layer="3156" from-port="0" to-layer="3157" to-port="1" />
+		<edge from-layer="3157" from-port="2" to-layer="3158" to-port="0" />
+		<edge from-layer="3158" from-port="1" to-layer="3159" to-port="1" />
 		<edge from-layer="3159" from-port="2" to-layer="3160" to-port="1" />
 		<edge from-layer="3160" from-port="2" to-layer="3161" to-port="1" />
 		<edge from-layer="3161" from-port="2" to-layer="3162" to-port="0" />
@@ -54407,12 +54121,12 @@
 		<edge from-layer="3176" from-port="1" to-layer="3179" to-port="0" />
 		<edge from-layer="3177" from-port="0" to-layer="3178" to-port="0" />
 		<edge from-layer="3178" from-port="1" to-layer="3179" to-port="1" />
-		<edge from-layer="3179" from-port="2" to-layer="3182" to-port="0" />
-		<edge from-layer="3180" from-port="0" to-layer="3181" to-port="0" />
-		<edge from-layer="3181" from-port="1" to-layer="3182" to-port="1" />
-		<edge from-layer="3182" from-port="2" to-layer="3184" to-port="0" />
-		<edge from-layer="3183" from-port="0" to-layer="3184" to-port="1" />
-		<edge from-layer="3184" from-port="2" to-layer="3185" to-port="1" />
+		<edge from-layer="3179" from-port="2" to-layer="3181" to-port="0" />
+		<edge from-layer="3180" from-port="0" to-layer="3181" to-port="1" />
+		<edge from-layer="3181" from-port="2" to-layer="3183" to-port="0" />
+		<edge from-layer="3182" from-port="0" to-layer="3183" to-port="1" />
+		<edge from-layer="3183" from-port="2" to-layer="3184" to-port="0" />
+		<edge from-layer="3184" from-port="1" to-layer="3185" to-port="1" />
 		<edge from-layer="3185" from-port="2" to-layer="3186" to-port="1" />
 		<edge from-layer="3186" from-port="2" to-layer="3188" to-port="0" />
 		<edge from-layer="3186" from-port="2" to-layer="3192" to-port="0" />
@@ -54428,12 +54142,12 @@
 		<edge from-layer="3196" from-port="1" to-layer="3199" to-port="0" />
 		<edge from-layer="3197" from-port="0" to-layer="3198" to-port="0" />
 		<edge from-layer="3198" from-port="1" to-layer="3199" to-port="1" />
-		<edge from-layer="3199" from-port="2" to-layer="3202" to-port="0" />
-		<edge from-layer="3200" from-port="0" to-layer="3201" to-port="0" />
-		<edge from-layer="3201" from-port="1" to-layer="3202" to-port="1" />
-		<edge from-layer="3202" from-port="2" to-layer="3204" to-port="0" />
-		<edge from-layer="3203" from-port="0" to-layer="3204" to-port="1" />
-		<edge from-layer="3204" from-port="2" to-layer="3205" to-port="1" />
+		<edge from-layer="3199" from-port="2" to-layer="3201" to-port="0" />
+		<edge from-layer="3200" from-port="0" to-layer="3201" to-port="1" />
+		<edge from-layer="3201" from-port="2" to-layer="3203" to-port="0" />
+		<edge from-layer="3202" from-port="0" to-layer="3203" to-port="1" />
+		<edge from-layer="3203" from-port="2" to-layer="3204" to-port="0" />
+		<edge from-layer="3204" from-port="1" to-layer="3205" to-port="1" />
 		<edge from-layer="3205" from-port="2" to-layer="3206" to-port="1" />
 		<edge from-layer="3206" from-port="2" to-layer="3207" to-port="1" />
 		<edge from-layer="3207" from-port="2" to-layer="3208" to-port="0" />
@@ -54457,12 +54171,12 @@
 		<edge from-layer="3222" from-port="1" to-layer="3225" to-port="0" />
 		<edge from-layer="3223" from-port="0" to-layer="3224" to-port="0" />
 		<edge from-layer="3224" from-port="1" to-layer="3225" to-port="1" />
-		<edge from-layer="3225" from-port="2" to-layer="3228" to-port="0" />
-		<edge from-layer="3226" from-port="0" to-layer="3227" to-port="0" />
-		<edge from-layer="3227" from-port="1" to-layer="3228" to-port="1" />
-		<edge from-layer="3228" from-port="2" to-layer="3230" to-port="0" />
-		<edge from-layer="3229" from-port="0" to-layer="3230" to-port="1" />
-		<edge from-layer="3230" from-port="2" to-layer="3231" to-port="1" />
+		<edge from-layer="3225" from-port="2" to-layer="3227" to-port="0" />
+		<edge from-layer="3226" from-port="0" to-layer="3227" to-port="1" />
+		<edge from-layer="3227" from-port="2" to-layer="3229" to-port="0" />
+		<edge from-layer="3228" from-port="0" to-layer="3229" to-port="1" />
+		<edge from-layer="3229" from-port="2" to-layer="3230" to-port="0" />
+		<edge from-layer="3230" from-port="1" to-layer="3231" to-port="1" />
 		<edge from-layer="3231" from-port="2" to-layer="3232" to-port="1" />
 		<edge from-layer="3232" from-port="2" to-layer="3234" to-port="0" />
 		<edge from-layer="3233" from-port="0" to-layer="3234" to-port="1" />
@@ -54478,12 +54192,12 @@
 		<edge from-layer="3242" from-port="1" to-layer="3245" to-port="0" />
 		<edge from-layer="3243" from-port="0" to-layer="3244" to-port="0" />
 		<edge from-layer="3244" from-port="1" to-layer="3245" to-port="1" />
-		<edge from-layer="3245" from-port="2" to-layer="3248" to-port="0" />
-		<edge from-layer="3246" from-port="0" to-layer="3247" to-port="0" />
-		<edge from-layer="3247" from-port="1" to-layer="3248" to-port="1" />
-		<edge from-layer="3248" from-port="2" to-layer="3250" to-port="0" />
-		<edge from-layer="3249" from-port="0" to-layer="3250" to-port="1" />
-		<edge from-layer="3250" from-port="2" to-layer="3251" to-port="1" />
+		<edge from-layer="3245" from-port="2" to-layer="3247" to-port="0" />
+		<edge from-layer="3246" from-port="0" to-layer="3247" to-port="1" />
+		<edge from-layer="3247" from-port="2" to-layer="3249" to-port="0" />
+		<edge from-layer="3248" from-port="0" to-layer="3249" to-port="1" />
+		<edge from-layer="3249" from-port="2" to-layer="3250" to-port="0" />
+		<edge from-layer="3250" from-port="1" to-layer="3251" to-port="1" />
 		<edge from-layer="3251" from-port="2" to-layer="3253" to-port="0" />
 		<edge from-layer="3252" from-port="0" to-layer="3253" to-port="1" />
 		<edge from-layer="3253" from-port="2" to-layer="3255" to-port="0" />
@@ -54499,12 +54213,12 @@
 		<edge from-layer="3263" from-port="1" to-layer="3266" to-port="0" />
 		<edge from-layer="3264" from-port="0" to-layer="3265" to-port="0" />
 		<edge from-layer="3265" from-port="1" to-layer="3266" to-port="1" />
-		<edge from-layer="3266" from-port="2" to-layer="3269" to-port="0" />
-		<edge from-layer="3267" from-port="0" to-layer="3268" to-port="0" />
-		<edge from-layer="3268" from-port="1" to-layer="3269" to-port="1" />
-		<edge from-layer="3269" from-port="2" to-layer="3271" to-port="0" />
-		<edge from-layer="3270" from-port="0" to-layer="3271" to-port="1" />
-		<edge from-layer="3271" from-port="2" to-layer="3272" to-port="1" />
+		<edge from-layer="3266" from-port="2" to-layer="3268" to-port="0" />
+		<edge from-layer="3267" from-port="0" to-layer="3268" to-port="1" />
+		<edge from-layer="3268" from-port="2" to-layer="3270" to-port="0" />
+		<edge from-layer="3269" from-port="0" to-layer="3270" to-port="1" />
+		<edge from-layer="3270" from-port="2" to-layer="3271" to-port="0" />
+		<edge from-layer="3271" from-port="1" to-layer="3272" to-port="1" />
 		<edge from-layer="3272" from-port="2" to-layer="3273" to-port="1" />
 		<edge from-layer="3273" from-port="2" to-layer="3275" to-port="0" />
 		<edge from-layer="3274" from-port="0" to-layer="3275" to-port="1" />
@@ -54531,12 +54245,12 @@
 		<edge from-layer="3293" from-port="1" to-layer="3296" to-port="0" />
 		<edge from-layer="3294" from-port="0" to-layer="3295" to-port="0" />
 		<edge from-layer="3295" from-port="1" to-layer="3296" to-port="1" />
-		<edge from-layer="3296" from-port="2" to-layer="3299" to-port="0" />
-		<edge from-layer="3297" from-port="0" to-layer="3298" to-port="0" />
-		<edge from-layer="3298" from-port="1" to-layer="3299" to-port="1" />
-		<edge from-layer="3299" from-port="2" to-layer="3301" to-port="0" />
-		<edge from-layer="3300" from-port="0" to-layer="3301" to-port="1" />
-		<edge from-layer="3301" from-port="2" to-layer="3302" to-port="1" />
+		<edge from-layer="3296" from-port="2" to-layer="3298" to-port="0" />
+		<edge from-layer="3297" from-port="0" to-layer="3298" to-port="1" />
+		<edge from-layer="3298" from-port="2" to-layer="3300" to-port="0" />
+		<edge from-layer="3299" from-port="0" to-layer="3300" to-port="1" />
+		<edge from-layer="3300" from-port="2" to-layer="3301" to-port="0" />
+		<edge from-layer="3301" from-port="1" to-layer="3302" to-port="1" />
 		<edge from-layer="3302" from-port="2" to-layer="3303" to-port="1" />
 		<edge from-layer="3303" from-port="2" to-layer="3304" to-port="1" />
 		<edge from-layer="3304" from-port="2" to-layer="3305" to-port="0" />
@@ -54558,12 +54272,12 @@
 		<edge from-layer="3319" from-port="1" to-layer="3322" to-port="0" />
 		<edge from-layer="3320" from-port="0" to-layer="3321" to-port="0" />
 		<edge from-layer="3321" from-port="1" to-layer="3322" to-port="1" />
-		<edge from-layer="3322" from-port="2" to-layer="3325" to-port="0" />
-		<edge from-layer="3323" from-port="0" to-layer="3324" to-port="0" />
-		<edge from-layer="3324" from-port="1" to-layer="3325" to-port="1" />
-		<edge from-layer="3325" from-port="2" to-layer="3327" to-port="0" />
-		<edge from-layer="3326" from-port="0" to-layer="3327" to-port="1" />
-		<edge from-layer="3327" from-port="2" to-layer="3328" to-port="1" />
+		<edge from-layer="3322" from-port="2" to-layer="3324" to-port="0" />
+		<edge from-layer="3323" from-port="0" to-layer="3324" to-port="1" />
+		<edge from-layer="3324" from-port="2" to-layer="3326" to-port="0" />
+		<edge from-layer="3325" from-port="0" to-layer="3326" to-port="1" />
+		<edge from-layer="3326" from-port="2" to-layer="3327" to-port="0" />
+		<edge from-layer="3327" from-port="1" to-layer="3328" to-port="1" />
 		<edge from-layer="3328" from-port="2" to-layer="3329" to-port="1" />
 		<edge from-layer="3329" from-port="2" to-layer="3331" to-port="0" />
 		<edge from-layer="3329" from-port="2" to-layer="3335" to-port="0" />
@@ -54579,12 +54293,12 @@
 		<edge from-layer="3339" from-port="1" to-layer="3342" to-port="0" />
 		<edge from-layer="3340" from-port="0" to-layer="3341" to-port="0" />
 		<edge from-layer="3341" from-port="1" to-layer="3342" to-port="1" />
-		<edge from-layer="3342" from-port="2" to-layer="3345" to-port="0" />
-		<edge from-layer="3343" from-port="0" to-layer="3344" to-port="0" />
-		<edge from-layer="3344" from-port="1" to-layer="3345" to-port="1" />
-		<edge from-layer="3345" from-port="2" to-layer="3347" to-port="0" />
-		<edge from-layer="3346" from-port="0" to-layer="3347" to-port="1" />
-		<edge from-layer="3347" from-port="2" to-layer="3348" to-port="1" />
+		<edge from-layer="3342" from-port="2" to-layer="3344" to-port="0" />
+		<edge from-layer="3343" from-port="0" to-layer="3344" to-port="1" />
+		<edge from-layer="3344" from-port="2" to-layer="3346" to-port="0" />
+		<edge from-layer="3345" from-port="0" to-layer="3346" to-port="1" />
+		<edge from-layer="3346" from-port="2" to-layer="3347" to-port="0" />
+		<edge from-layer="3347" from-port="1" to-layer="3348" to-port="1" />
 		<edge from-layer="3348" from-port="2" to-layer="3349" to-port="1" />
 		<edge from-layer="3349" from-port="2" to-layer="3350" to-port="1" />
 		<edge from-layer="3350" from-port="2" to-layer="3351" to-port="0" />
@@ -54608,12 +54322,12 @@
 		<edge from-layer="3365" from-port="1" to-layer="3368" to-port="0" />
 		<edge from-layer="3366" from-port="0" to-layer="3367" to-port="0" />
 		<edge from-layer="3367" from-port="1" to-layer="3368" to-port="1" />
-		<edge from-layer="3368" from-port="2" to-layer="3371" to-port="0" />
-		<edge from-layer="3369" from-port="0" to-layer="3370" to-port="0" />
-		<edge from-layer="3370" from-port="1" to-layer="3371" to-port="1" />
-		<edge from-layer="3371" from-port="2" to-layer="3373" to-port="0" />
-		<edge from-layer="3372" from-port="0" to-layer="3373" to-port="1" />
-		<edge from-layer="3373" from-port="2" to-layer="3374" to-port="1" />
+		<edge from-layer="3368" from-port="2" to-layer="3370" to-port="0" />
+		<edge from-layer="3369" from-port="0" to-layer="3370" to-port="1" />
+		<edge from-layer="3370" from-port="2" to-layer="3372" to-port="0" />
+		<edge from-layer="3371" from-port="0" to-layer="3372" to-port="1" />
+		<edge from-layer="3372" from-port="2" to-layer="3373" to-port="0" />
+		<edge from-layer="3373" from-port="1" to-layer="3374" to-port="1" />
 		<edge from-layer="3374" from-port="2" to-layer="3375" to-port="1" />
 		<edge from-layer="3375" from-port="2" to-layer="3377" to-port="0" />
 		<edge from-layer="3376" from-port="0" to-layer="3377" to-port="1" />
@@ -54629,12 +54343,12 @@
 		<edge from-layer="3385" from-port="1" to-layer="3388" to-port="0" />
 		<edge from-layer="3386" from-port="0" to-layer="3387" to-port="0" />
 		<edge from-layer="3387" from-port="1" to-layer="3388" to-port="1" />
-		<edge from-layer="3388" from-port="2" to-layer="3391" to-port="0" />
-		<edge from-layer="3389" from-port="0" to-layer="3390" to-port="0" />
-		<edge from-layer="3390" from-port="1" to-layer="3391" to-port="1" />
-		<edge from-layer="3391" from-port="2" to-layer="3393" to-port="0" />
-		<edge from-layer="3392" from-port="0" to-layer="3393" to-port="1" />
-		<edge from-layer="3393" from-port="2" to-layer="3394" to-port="1" />
+		<edge from-layer="3388" from-port="2" to-layer="3390" to-port="0" />
+		<edge from-layer="3389" from-port="0" to-layer="3390" to-port="1" />
+		<edge from-layer="3390" from-port="2" to-layer="3392" to-port="0" />
+		<edge from-layer="3391" from-port="0" to-layer="3392" to-port="1" />
+		<edge from-layer="3392" from-port="2" to-layer="3393" to-port="0" />
+		<edge from-layer="3393" from-port="1" to-layer="3394" to-port="1" />
 		<edge from-layer="3394" from-port="2" to-layer="3396" to-port="0" />
 		<edge from-layer="3395" from-port="0" to-layer="3396" to-port="1" />
 		<edge from-layer="3396" from-port="2" to-layer="3398" to-port="0" />
@@ -54650,12 +54364,12 @@
 		<edge from-layer="3406" from-port="1" to-layer="3409" to-port="0" />
 		<edge from-layer="3407" from-port="0" to-layer="3408" to-port="0" />
 		<edge from-layer="3408" from-port="1" to-layer="3409" to-port="1" />
-		<edge from-layer="3409" from-port="2" to-layer="3412" to-port="0" />
-		<edge from-layer="3410" from-port="0" to-layer="3411" to-port="0" />
-		<edge from-layer="3411" from-port="1" to-layer="3412" to-port="1" />
-		<edge from-layer="3412" from-port="2" to-layer="3414" to-port="0" />
-		<edge from-layer="3413" from-port="0" to-layer="3414" to-port="1" />
-		<edge from-layer="3414" from-port="2" to-layer="3415" to-port="1" />
+		<edge from-layer="3409" from-port="2" to-layer="3411" to-port="0" />
+		<edge from-layer="3410" from-port="0" to-layer="3411" to-port="1" />
+		<edge from-layer="3411" from-port="2" to-layer="3413" to-port="0" />
+		<edge from-layer="3412" from-port="0" to-layer="3413" to-port="1" />
+		<edge from-layer="3413" from-port="2" to-layer="3414" to-port="0" />
+		<edge from-layer="3414" from-port="1" to-layer="3415" to-port="1" />
 		<edge from-layer="3415" from-port="2" to-layer="3416" to-port="1" />
 		<edge from-layer="3416" from-port="2" to-layer="3418" to-port="0" />
 		<edge from-layer="3417" from-port="0" to-layer="3418" to-port="1" />
@@ -54682,12 +54396,12 @@
 		<edge from-layer="3436" from-port="1" to-layer="3439" to-port="0" />
 		<edge from-layer="3437" from-port="0" to-layer="3438" to-port="0" />
 		<edge from-layer="3438" from-port="1" to-layer="3439" to-port="1" />
-		<edge from-layer="3439" from-port="2" to-layer="3442" to-port="0" />
-		<edge from-layer="3440" from-port="0" to-layer="3441" to-port="0" />
-		<edge from-layer="3441" from-port="1" to-layer="3442" to-port="1" />
-		<edge from-layer="3442" from-port="2" to-layer="3444" to-port="0" />
-		<edge from-layer="3443" from-port="0" to-layer="3444" to-port="1" />
-		<edge from-layer="3444" from-port="2" to-layer="3445" to-port="1" />
+		<edge from-layer="3439" from-port="2" to-layer="3441" to-port="0" />
+		<edge from-layer="3440" from-port="0" to-layer="3441" to-port="1" />
+		<edge from-layer="3441" from-port="2" to-layer="3443" to-port="0" />
+		<edge from-layer="3442" from-port="0" to-layer="3443" to-port="1" />
+		<edge from-layer="3443" from-port="2" to-layer="3444" to-port="0" />
+		<edge from-layer="3444" from-port="1" to-layer="3445" to-port="1" />
 		<edge from-layer="3445" from-port="2" to-layer="3446" to-port="1" />
 		<edge from-layer="3446" from-port="2" to-layer="3447" to-port="1" />
 		<edge from-layer="3447" from-port="2" to-layer="3448" to-port="0" />
@@ -54709,12 +54423,12 @@
 		<edge from-layer="3462" from-port="1" to-layer="3465" to-port="0" />
 		<edge from-layer="3463" from-port="0" to-layer="3464" to-port="0" />
 		<edge from-layer="3464" from-port="1" to-layer="3465" to-port="1" />
-		<edge from-layer="3465" from-port="2" to-layer="3468" to-port="0" />
-		<edge from-layer="3466" from-port="0" to-layer="3467" to-port="0" />
-		<edge from-layer="3467" from-port="1" to-layer="3468" to-port="1" />
-		<edge from-layer="3468" from-port="2" to-layer="3470" to-port="0" />
-		<edge from-layer="3469" from-port="0" to-layer="3470" to-port="1" />
-		<edge from-layer="3470" from-port="2" to-layer="3471" to-port="1" />
+		<edge from-layer="3465" from-port="2" to-layer="3467" to-port="0" />
+		<edge from-layer="3466" from-port="0" to-layer="3467" to-port="1" />
+		<edge from-layer="3467" from-port="2" to-layer="3469" to-port="0" />
+		<edge from-layer="3468" from-port="0" to-layer="3469" to-port="1" />
+		<edge from-layer="3469" from-port="2" to-layer="3470" to-port="0" />
+		<edge from-layer="3470" from-port="1" to-layer="3471" to-port="1" />
 		<edge from-layer="3471" from-port="2" to-layer="3472" to-port="1" />
 		<edge from-layer="3472" from-port="2" to-layer="3478" to-port="0" />
 		<edge from-layer="3472" from-port="2" to-layer="3474" to-port="0" />
@@ -54730,10 +54444,10 @@
 		<edge from-layer="3482" from-port="1" to-layer="3485" to-port="0" />
 		<edge from-layer="3483" from-port="0" to-layer="3484" to-port="0" />
 		<edge from-layer="3484" from-port="1" to-layer="3485" to-port="1" />
-		<edge from-layer="3485" from-port="2" to-layer="3488" to-port="0" />
-		<edge from-layer="3486" from-port="0" to-layer="3487" to-port="0" />
-		<edge from-layer="3487" from-port="1" to-layer="3488" to-port="1" />
-		<edge from-layer="3488" from-port="2" to-layer="3489" to-port="1" />
+		<edge from-layer="3485" from-port="2" to-layer="3487" to-port="0" />
+		<edge from-layer="3486" from-port="0" to-layer="3487" to-port="1" />
+		<edge from-layer="3487" from-port="2" to-layer="3488" to-port="0" />
+		<edge from-layer="3488" from-port="1" to-layer="3489" to-port="1" />
 		<edge from-layer="3489" from-port="2" to-layer="3490" to-port="1" />
 		<edge from-layer="3490" from-port="2" to-layer="3491" to-port="1" />
 		<edge from-layer="3491" from-port="2" to-layer="3492" to-port="0" />
@@ -54755,16 +54469,12 @@
 			<weight_compression>
 				<all_layers value="False" />
 				<awq value="False" />
+				<gptq value="False" />
 				<group_size value="128" />
-				<ignored_scope>
-					<names value="[]" />
-					<patterns value="[]" />
-					<subgraphs value="[]" />
-					<types value="[]" />
-					<validate value="True" />
-				</ignored_scope>
+				<ignored_scope value="[]" />
 				<mode value="int4_sym" />
 				<ratio value="1" />
+				<scale_estimation value="False" />
 				<sensitivity_metric value="weight_quantization_error" />
 			</weight_compression>
 		</nncf>