fix int4

app.py

@@ -68,14 +68,10 @@ def get_benchmark_df(benchmark="1xA100-80GB"):
     bench_df["optimizations"] = bench_df[
         ["backend.bettertransformer", "backend.load_in_8bit", "backend.load_in_4bit"]
     ].apply(
-        lambda x: ", ".join(
-            [
-                optimization if value == True else "None"
-                for optimization, value in zip(
-                    ["BetterTransformer", "LLM.int8", "NF4"], x.values
-                )
-            ]
-        ),
+        lambda x: "BetterTransformer"
+        if x[0] == True
+        else ("LLM.int8" if x[1] == True else ("LLM.fp4" if x[2] == True else "None")),
+        axis=1,
     )
 
     return bench_df
@@ -99,7 +95,7 @@ def get_benchmark_table(bench_df):
 
 def get_benchmark_plot(bench_df):
     # untill falcon gets fixed / natively supported
-    bench_df = bench_df[bench_df["generate.latency(s)"] < 200]
+    bench_df = bench_df[bench_df["generate.latency(s)"] < 100]
 
     fig = px.scatter(
         bench_df,
@@ -117,7 +113,6 @@ def get_benchmark_plot(bench_df):
             "generate.throughput(tokens/s)",
         ],
         symbol_sequence=["triangle-up", "circle"],
-        # as many distinct colors as there are model_type,backend.name couples
         color_discrete_sequence=px.colors.qualitative.Light24,
     )
 
@@ -254,8 +249,8 @@ with demo:
         with gr.Column(scale=2):
             optimizations_checkboxes = gr.CheckboxGroup(
                 label="Optimizations 🛠️",
-                choices=["None", "BetterTransformer", "LLM.int8", "NF4"],
-                value=["None", "BetterTransformer", "LLM.int8", "NF4"],
+                choices=["None", "BetterTransformer", "LLM.int8", "LLM.fp4"],
+                value=["None", "BetterTransformer", "LLM.int8", "LLM.fp4"],
                 info="☑️ Select the optimizations",
                 elem_id="optimizations-checkboxes",
             )