added columns types

app.py

@@ -12,6 +12,17 @@ LLM_PERF_LEADERBOARD_REPO = "optimum/llm-perf-leaderboard"
 LLM_PERF_DATASET_REPO = "optimum/llm-perf-dataset"
 OPTIMUM_TOKEN = os.environ.get("OPTIMUM_TOKEN")
 
+OLD_COLUMNS = ["model", "backend.name", "backend.torch_dtype", "backend.quantization",
+               "generate.latency(s)", "generate.throughput(tokens/s)"]
+
+NEW_COLUMNS = ["Model", "Backend 🏭", "Load dtype", "Quantization 🗜️",
+               "Latency (s) ⬇️", "Throughput (tokens/s) ⬆️"]
+
+COLUMNS_TYPES = ["markdown", "text", "text", "text", "number", "number"]
+
+SORTING_COLUMN = ["Throughput (tokens/s) ⬆️"]
+
+
 llm_perf_dataset_repo = load_dataset_repo(LLM_PERF_DATASET_REPO, OPTIMUM_TOKEN)
 
 
@@ -19,25 +30,23 @@ def get_vanilla_benchmark_df():
     if llm_perf_dataset_repo:
         llm_perf_dataset_repo.git_pull()
 
+    # load
     df = pd.read_csv(
         "./llm-perf-dataset/reports/cuda_1_100/inference_report.csv")
 
-    df = df[["model", "backend.name", "backend.torch_dtype", "backend.quantization",
-             "generate.latency(s)", "generate.throughput(tokens/s)"]]
+    # preprocess
+    df["Model"] = df["Model"].apply(make_clickable_model)
 
-    df["model"] = df["model"].apply(make_clickable_model)
+    # filter
+    df = df[OLD_COLUMNS]
 
+    # rename
     df.rename(columns={
-        "model": "Model",
-        "backend.name": "Backend 🏭",
-        "backend.torch_dtype": "Load dtype",
-        "backend.quantization": "Quantization 🗜️",
-        "generate.latency(s)": "Latency (s) ⬇️",
-        "generate.throughput(tokens/s)": "Throughput (tokens/s) ⬆️",
+        df_col: rename_col for df_col, rename_col in zip(OLD_COLUMNS, NEW_COLUMNS)
     }, inplace=True)
 
-    df.sort_values(by=["Throughput (tokens/s) ⬆️"],
-                   ascending=False, inplace=True)
+    # sort
+    df.sort_values(by=SORTING_COLUMN, ascending=False, inplace=True)
 
     return df
 
@@ -54,7 +63,8 @@ with demo:
             vanilla_benchmark_df = get_vanilla_benchmark_df()
             leaderboard_table_lite = gr.components.Dataframe(
                 value=vanilla_benchmark_df,
-                headers=vanilla_benchmark_df.columns.tolist(),
+                type=COLUMNS_TYPES,
+                headers=NEW_COLUMNS,
                 elem_id="vanilla-benchmark",
             )