allow quantized models on plot

app.py

@@ -82,6 +82,11 @@ def get_benchmark_df(benchmark="1xA100-80GB"):
         lambda x: TRUE_WEIGHT_CLASSES[x] if x in TRUE_WEIGHT_CLASSES else x
     )
 
+    # convert peak memory to int
+    merged_df["forward.peak_memory(MB)"] = merged_df["forward.peak_memory(MB)"].apply(
+        lambda x: int(x)
+    )
+
     # add optimizations
     merged_df["optimizations"] = merged_df[
         ["backend.bettertransformer", "backend.load_in_8bit", "backend.load_in_4bit"]
@@ -101,6 +106,8 @@ def get_benchmark_df(benchmark="1xA100-80GB"):
         axis=1,
     )
 
+    merged_df["quantized"] = merged_df["optimizations"].str.contains("LLM.int8|LLM.fp4")
+
     # create composite score
     score_distance = 100 - merged_df["best_score"]
     # normalize latency between 0 and 100
@@ -108,19 +115,16 @@ def get_benchmark_df(benchmark="1xA100-80GB"):
     merged_df["tradeoff"] = (score_distance**2 + latency_distance**2) ** 0.5
     merged_df["tradeoff"] = merged_df["tradeoff"].round(2)
 
-    # add * to quantized models
-    merged_df.loc[
-        merged_df["optimizations"].str.contains("LLM.int8|LLM.fp4"), "best_score"
-    ] = merged_df.loc[
-        merged_df["optimizations"].str.contains("LLM.int8|LLM.fp4"), "best_score"
-    ].apply(
-        lambda x: f"{x}*"
-    )
-
     return merged_df
 
 
 def get_benchmark_table(bench_df):
+    # add * to quantized models score
+    bench_df["best_score"] = bench_df.apply(
+        lambda x: f"{x['best_score']}**" if x["quantized"] else x["best_score"],
+        axis=1,
+    )
+
     # sort
     bench_df.sort_values(by=SORTING_COLUMN, ascending=True, inplace=True)
     # filter
@@ -132,6 +136,7 @@ def get_benchmark_table(bench_df):
     bench_df["Best Scored Model 🏆"] = bench_df["Best Scored Model 🏆"].apply(
         process_model_name
     )
+
     return bench_df