smpanaro's picture
Update Sonoma model with faster 8x8 conv and split einsum attention
dba673f
[
{
"metadataOutputVersion" : "3.0",
"storagePrecision" : "Mixed (Float16, Palettized (4 bits))",
"outputSchema" : [
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 8 × 8)",
"shortDescription" : "",
"shape" : "[1, 4096, 8, 8]",
"name" : "new_x",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 64 × 1 × 4096)",
"shortDescription" : "",
"shape" : "[1, 64, 1, 4096]",
"name" : "new_k_cache_0",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 64 × 1 × 4096)",
"shortDescription" : "",
"shape" : "[1, 64, 1, 4096]",
"name" : "new_k_cache_1",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 64 × 1 × 4096)",
"shortDescription" : "",
"shape" : "[1, 64, 1, 4096]",
"name" : "new_k_cache_2",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 1 × 64)",
"shortDescription" : "",
"shape" : "[1, 4096, 1, 64]",
"name" : "new_v_cache_0",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 1 × 64)",
"shortDescription" : "",
"shape" : "[1, 4096, 1, 64]",
"name" : "new_v_cache_1",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 1 × 64)",
"shortDescription" : "",
"shape" : "[1, 4096, 1, 64]",
"name" : "new_v_cache_2",
"type" : "MultiArray"
}
],
"modelParameters" : [
],
"specificationVersion" : 7,
"mlProgramOperationTypeHistogram" : {
"Concat" : 21,
"Ios16.mul" : 150,
"SliceByIndex" : 300,
"Ios16.constexprLutToDense" : 21,
"Transpose" : 3,
"Ios16.einsum" : 192,
"Ios16.conv" : 21,
"Ios16.add" : 108,
"Ios16.realDiv" : 6,
"Ios16.softmax" : 96,
"Ios16.reduceL2Norm" : 6,
"Ios16.reshape" : 21,
"Ios16.silu" : 3
},
"computePrecision" : "Mixed (Float16, Int32)",
"isUpdatable" : "0",
"availability" : {
"macOS" : "13.0",
"tvOS" : "16.0",
"visionOS" : "1.0",
"watchOS" : "9.0",
"iOS" : "16.0",
"macCatalyst" : "16.0"
},
"modelType" : {
"name" : "MLModelType_mlProgram"
},
"userDefinedMetadata" : {
"com.github.apple.coremltools.source_dialect" : "TorchScript",
"com.github.apple.coremltools.source" : "torch==2.1.0",
"com.github.apple.coremltools.version" : "8.0b1"
},
"inputSchema" : [
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 8 × 8)",
"shortDescription" : "",
"shape" : "[1, 4096, 8, 8]",
"name" : "x",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 128 × 64)",
"shortDescription" : "",
"shape" : "[128, 64]",
"name" : "cos",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 128 × 64)",
"shortDescription" : "",
"shape" : "[128, 64]",
"name" : "sin",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "0",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 512 × 1 × 64)",
"shortDescription" : "",
"shape" : "[1, 512, 1, 64]",
"name" : "mask",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "1",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 448 × 1 × 4096)?",
"shortDescription" : "",
"shape" : "[1, 448, 1, 4096]",
"name" : "k_cache_0",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "1",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 1 × 448)?",
"shortDescription" : "",
"shape" : "[1, 4096, 1, 448]",
"name" : "v_cache_0",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "1",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 448 × 1 × 4096)?",
"shortDescription" : "",
"shape" : "[1, 448, 1, 4096]",
"name" : "k_cache_1",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "1",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 1 × 448)?",
"shortDescription" : "",
"shape" : "[1, 4096, 1, 448]",
"name" : "v_cache_1",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "1",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 448 × 1 × 4096)?",
"shortDescription" : "",
"shape" : "[1, 448, 1, 4096]",
"name" : "k_cache_2",
"type" : "MultiArray"
},
{
"hasShapeFlexibility" : "0",
"isOptional" : "1",
"dataType" : "Float16",
"formattedType" : "MultiArray (Float16 1 × 4096 × 1 × 448)?",
"shortDescription" : "",
"shape" : "[1, 4096, 1, 448]",
"name" : "v_cache_2",
"type" : "MultiArray"
}
],
"generatedClassName" : "Llama_2_7b_hf_2024_08_09_09_54_41_chunk11",
"method" : "predict"
}
]