update

app.py

@@ -4,14 +4,14 @@ from src.assets import custom_css
 
 # from src.attention import create_attn_plots
 from src.content import ABOUT, CITATION_BUTTON, CITATION_BUTTON_LABEL, LOGO, TITLE
-from src.control_panel import (
+from src.leaderboard import create_leaderboard_table
+from src.llm_perf import get_llm_perf_df
+from src.map import create_lat_score_mem_plot
+from src.panel import (
     create_control_callback,
     create_control_panel,
     create_select_callback,
 )
-from src.leaderboard import create_leaderboard_table
-from src.llm_perf import get_llm_perf_df
-from src.map import create_lat_score_mem_plot
 
 # from custom_kernels import create_quant_krnl_plots
 

src/kernels.py

@@ -32,7 +32,7 @@ def get_quant_df(llm_perf_df):
     vanilla_df = copy_df[
         (copy_df["Backend 🏭"] == "pytorch")
         & (copy_df["DType 📥"] == "float16")
-        & (copy_df["Quantization 🗜️"] == "None")
+        & (copy_df["Quantization 🗜️"] == "Unquantized")
     ]
     exllamav1_df = copy_df[(copy_df["Quantization 🗜️"] == "GPTQ.4bit+ExllamaV1")]
     exllamav2_df = copy_df[(copy_df["Quantization 🗜️"] == "GPTQ.4bit+ExllamaV2")]

src/leaderboard.py

@@ -48,8 +48,7 @@ def get_leaderboard_df(llm_perf_df):
     df["Model 🤗"] = df["Model 🤗"].apply(process_model)
     # process quantization for leaderboard
     df["Open LLM Score (%)"] = df.apply(
-        lambda x: process_score(x["Open LLM Score (%)"], x["Quantization 🗜️"]),
-        axis=1,
+        lambda x: process_score(x["Open LLM Score (%)"], x["Quantization 🗜️"]), axis=1
     )
     return df
 

src/llm_perf.py

@@ -61,6 +61,7 @@ def processed_llm_perf_df(llm_perf_df):
     assert llm_perf_df["config.scenario.generate_kwargs.max_new_tokens"].nunique() == 1
     assert llm_perf_df["config.scenario.generate_kwargs.min_new_tokens"].nunique() == 1
     # fix couple stuff
+    llm_perf_df.dropna(subset=["report.decode.latency.p50"], inplace=True)
     llm_perf_df["config.name"] = llm_perf_df["config.name"].str.replace(
         "flash_attention_2", "fa2"
     )

src/map.py

@@ -21,11 +21,12 @@ SCORE_MEMORY_LATENCY_DATA = [
 def get_lat_score_mem_fig(llm_perf_df):
     copy_df = llm_perf_df.copy()
     # plot
+    # filter nan memory
     fig = px.scatter(
         copy_df,
+        size="Memory (MB)",
         x="End-to-End (s)",
         y="Open LLM Score (%)",
-        size="Memory (MB)",
         color="Architecture 🏛️",
         custom_data=SCORE_MEMORY_LATENCY_DATA,
         color_discrete_sequence=px.colors.qualitative.Light24,
@@ -41,10 +42,10 @@ def get_lat_score_mem_fig(llm_perf_df):
     fig.update_layout(
         title={
             "text": "Latency vs. Score vs. Memory",
-            "y": 0.95,
-            "x": 0.5,
             "xanchor": "center",
             "yanchor": "top",
+            "y": 0.95,
+            "x": 0.5,
         },
         xaxis_title="Time To Generate 64 Tokens (s)",
         yaxis_title="Open LLM Score (%)",
@@ -52,6 +53,8 @@ def get_lat_score_mem_fig(llm_perf_df):
         width=1200,
         height=600,
     )
+    # update x range with 95 percentile of
+    fig.update_xaxes(range=[0, copy_df["End-to-End (s)"].quantile(0.95)])
 
     return fig
 

src/{control_panel.py → panel.py}

@@ -58,14 +58,14 @@ def create_control_panel(machine: str):
                 quantization_checkboxes = gr.CheckboxGroup(
                     label="Quantizations 🗜️",
                     choices=[
-                        "None",
+                        "Unquantized",
                         "BnB.4bit",
                         "BnB.8bit",
                         "AWQ.4bit",
                         "GPTQ.4bit",
                     ],
                     value=[
-                        "None",
+                        "Unquantized",
                         "BnB.4bit",
                         "BnB.8bit",
                         "AWQ.4bit",

src/utils.py

@@ -43,7 +43,7 @@ def process_architectures(model):
 
 
 def process_score(score, quantization):
-    if quantization != "None":
+    if quantization != "Unquantized":
         return f"{score:.2f}*"
     else:
         return f"{score:.2f} "
@@ -71,7 +71,7 @@ def process_quantizations(x):
     ):
         return "AWQ.4bit"
     else:
-        return "None"
+        return "Unquantized"
 
 
 def process_kernels(x):
@@ -97,7 +97,7 @@ def process_kernels(x):
     ):
         return "AWQ.GEMV"
     else:
-        return "None"
+        return "Unquantized"
 
 
 # def change_tab(query_param):