Spaces:
Running
Running
| import os | |
| import gradio as gr | |
| import pandas as pd | |
| import plotly.express as px | |
| from apscheduler.schedulers.background import BackgroundScheduler | |
| from src.assets.css_html_js import custom_css | |
| from src.assets.text_content import ( | |
| TITLE, | |
| INTRODUCTION_TEXT, | |
| ABOUT_TEXT, | |
| EXAMPLE_CONFIG_TEXT, | |
| CITATION_BUTTON_LABEL, | |
| CITATION_BUTTON_TEXT, | |
| ) | |
| from src.utils import ( | |
| restart_space, | |
| load_dataset_repo, | |
| process_model_name, | |
| process_model_type, | |
| ) | |
| HARDWARE_NAMES = ["A100-80GB", "RTX4090-24GB"] | |
| HARDWARES_EMOJIS = ["🖥️", "💻"] | |
| LLM_PERF_LEADERBOARD_REPO = "optimum/llm-perf-leaderboard" | |
| LLM_PERF_DATASET_REPO = "optimum/llm-perf-dataset" | |
| OPTIMUM_TOKEN = os.environ.get("OPTIMUM_TOKEN", None) | |
| ALL_COLUMNS_MAPPING = { | |
| "backend.name": "Backend 🏭", | |
| "backend.torch_dtype": "Dtype 📥", | |
| "optimizations": "Optimizations 🛠️", | |
| "quantization": "Quantization 🗜️", | |
| # | |
| "weight_class": "Class 🏋️", | |
| "model_type": "Type 🤗", | |
| # | |
| "generate.peak_memory(MB)": "Memory (MB) ⬇️", | |
| "generate.throughput(tokens/s)": "Throughput (tokens/s) ⬆️", | |
| "generate.energy_consumption(tokens/kWh)": "Energy (tokens/kWh) ⬇️", | |
| "best_score": "Best Score (%) ⬆️", | |
| # | |
| "best_scored_model": "Best Scored LLM 🏆", | |
| } | |
| ALL_COLUMNS_DATATYPES = [ | |
| "str", | |
| "str", | |
| "str", | |
| "str", | |
| # | |
| "str", | |
| "str", | |
| # | |
| "number", | |
| "number", | |
| "number", | |
| "str", | |
| # | |
| "markdown", | |
| ] | |
| NO_DUPLICATES_COLUMNS = [ | |
| "backend.name", | |
| "backend.torch_dtype", | |
| "optimizations", | |
| "quantization", | |
| # | |
| "weight_class", | |
| "model_type", | |
| ] | |
| SORTING_COLUMN = ["best_score", "generate.latency(s)", "generate.peak_memory(MB)"] | |
| SORTING_ASCENDING = [False, True, True] | |
| llm_perf_dataset_repo = load_dataset_repo(LLM_PERF_DATASET_REPO, OPTIMUM_TOKEN) | |
| def get_benchmark_df(benchmark="Succeeded-1xA100-80GB"): | |
| if llm_perf_dataset_repo: | |
| llm_perf_dataset_repo.git_pull() | |
| # load data | |
| benchmark_df = pd.read_csv(f"./llm-perf-dataset/reports/{benchmark}.csv") | |
| clusters_df = pd.read_csv("./llm-perf-dataset/Clustered-Open-LLM-Leaderboard.csv") | |
| # merge on model | |
| merged_df = benchmark_df.merge( | |
| clusters_df, left_on="model", right_on="best_scored_model" | |
| ) | |
| # transpose energy consumption | |
| merged_df["generate.energy_consumption(tokens/kWh)"] = 1 / merged_df[ | |
| "generate.energy_consumption(kWh/token)" | |
| ].fillna(1) | |
| # fix nan values | |
| merged_df[merged_df["generate.energy_consumption(tokens/kWh)"] == 1] = "N/A" | |
| # add optimizations | |
| merged_df["optimizations"] = merged_df["backend.bettertransformer"].apply( | |
| lambda x: "BetterTransformer" if x else "None" | |
| ) | |
| # add quantization scheme | |
| merged_df["quantization"] = merged_df["backend.quantization_strategy"].apply( | |
| lambda x: "BnB.4bit" if x == "bnb" else ("GPTQ.4bit" if x == "gptq" else "None") | |
| ) | |
| # sort | |
| merged_df.sort_values(by=SORTING_COLUMN, ascending=SORTING_ASCENDING, inplace=True) | |
| # drop duplicates | |
| merged_df.drop_duplicates(subset=NO_DUPLICATES_COLUMNS, inplace=True) | |
| return merged_df | |
| def get_benchmark_table(bench_df): | |
| copy_df = bench_df.copy() | |
| # filter | |
| copy_df = copy_df[list(ALL_COLUMNS_MAPPING.keys())] | |
| # rename | |
| copy_df.rename(columns=ALL_COLUMNS_MAPPING, inplace=True) | |
| # transform | |
| copy_df["Type 🤗"] = copy_df["Type 🤗"].apply(process_model_type) | |
| copy_df["Best Scored LLM 🏆"] = copy_df["Best Scored LLM 🏆"].apply( | |
| process_model_name | |
| ) | |
| # process quantization | |
| copy_df["Best Score (%) ⬆️"] = copy_df.apply( | |
| lambda x: f"{x['Best Score (%) ⬆️']}**" | |
| if x["Quantization 🗜️"] in ["BnB.4bit", "GPTQ.4bit"] | |
| else x["Best Score (%) ⬆️"], | |
| axis=1, | |
| ) | |
| return copy_df | |
| def get_benchmark_chart(bench_df): | |
| copy_df = bench_df.copy() | |
| # filter latency bigger than 100s | |
| copy_df["generate.latency(s)"] = copy_df["generate.latency(s)"].apply( | |
| lambda x: x if isinstance(x, (int, float)) else 1000 | |
| ) | |
| copy_df = copy_df[copy_df["generate.latency(s)"] <= 100] | |
| # rename model_type | |
| copy_df["model_type"] = copy_df["model_type"].apply(process_model_type) | |
| fig = px.scatter( | |
| copy_df, | |
| y="best_score", | |
| x="generate.latency(s)", | |
| size="generate.peak_memory(MB)", | |
| color="model_type", | |
| custom_data=list(ALL_COLUMNS_MAPPING.keys()), | |
| color_discrete_sequence=px.colors.qualitative.Light24, | |
| ) | |
| fig.update_layout( | |
| title={ | |
| "text": "Latency vs. Score vs. Memory", | |
| "y": 0.95, | |
| "x": 0.5, | |
| "xanchor": "center", | |
| "yanchor": "top", | |
| }, | |
| xaxis_title="Per 1000 tokens Latency (s)", | |
| yaxis_title="Open LLM Score (%)", | |
| legend_title="LLM Type", | |
| width=1200, | |
| height=600, | |
| ) | |
| fig.update_traces( | |
| hovertemplate="<br>".join( | |
| [ | |
| f"<b>{ALL_COLUMNS_MAPPING[key]}:</b> %{{customdata[{i}]}}" | |
| for i, key in enumerate(ALL_COLUMNS_MAPPING.keys()) | |
| ] | |
| ) | |
| ) | |
| return fig | |
| def filter_query( | |
| text, | |
| backends, | |
| datatypes, | |
| optimizations, | |
| quantization_scheme, | |
| score, | |
| memory, | |
| hardware, | |
| ): | |
| raw_df = get_benchmark_df(benchmark=f"Succeeded-1x{hardware}") | |
| filtered_df = raw_df[ | |
| raw_df["best_scored_model"].str.lower().str.contains(text.lower()) | |
| & raw_df["backend.name"].isin(backends) | |
| & raw_df["backend.torch_dtype"].isin(datatypes) | |
| & ( | |
| pd.concat( | |
| [ | |
| raw_df["optimizations"].str.contains(optimization) | |
| for optimization in optimizations | |
| ], | |
| axis=1, | |
| ).any(axis="columns") | |
| if len(optimizations) > 0 | |
| else True | |
| ) | |
| & ( | |
| pd.concat( | |
| [ | |
| raw_df["quantization"] == quantization | |
| for quantization in quantization_scheme | |
| ], | |
| axis=1, | |
| ).any(axis="columns") | |
| if len(quantization_scheme) > 0 | |
| else True | |
| ) | |
| & (raw_df["best_score"] >= score) | |
| & (raw_df["forward.peak_memory(MB)"] <= memory) | |
| ] | |
| filtered_table = get_benchmark_table(filtered_df) | |
| filtered_chart = get_benchmark_chart(filtered_df) | |
| return filtered_table, filtered_chart | |
| # Demo interface | |
| demo = gr.Blocks(css=custom_css) | |
| with demo: | |
| # leaderboard title | |
| gr.HTML(TITLE) | |
| # introduction text | |
| gr.Markdown(INTRODUCTION_TEXT, elem_classes="descriptive-text") | |
| with gr.Tabs(elem_classes="leaderboard-tabs"): | |
| hardware_placeholders = {} | |
| hardware_tables = {} | |
| hardware_plots = {} | |
| ####################### HARDWARE TABS ####################### | |
| for i, (hardware, emoji) in enumerate(zip(HARDWARE_NAMES, HARDWARES_EMOJIS)): | |
| # dummy placeholder of the hardware name | |
| hardware_placeholders[hardware] = gr.Textbox(value=hardware, visible=False) | |
| with gr.TabItem(f"{hardware} {emoji}", id=i): | |
| with gr.Tabs(elem_classes="hardware-tabs"): | |
| # placeholder for full dataframe | |
| hardware_df = get_benchmark_df(benchmark=f"Succeeded-1x{hardware}") | |
| with gr.TabItem("Leaderboard 🏅", id=0): | |
| gr.HTML( | |
| "👉 Scroll to the right 👉 for additional columns.", | |
| elem_id="descriptive-text", | |
| ) | |
| # Original leaderboard table | |
| hardware_tables[hardware] = gr.components.Dataframe( | |
| value=get_benchmark_table(hardware_df), | |
| headers=list(ALL_COLUMNS_MAPPING.values()), | |
| datatype=ALL_COLUMNS_DATATYPES, | |
| elem_id="hardware-table", | |
| # show_label=False, | |
| ) | |
| with gr.TabItem("Plot 📊", id=1): | |
| gr.HTML( | |
| "👆 Hover over the points 👆 for additional information.", | |
| elem_id="descriptive-text", | |
| ) | |
| # Original leaderboard plot | |
| hardware_plots[hardware] = gr.components.Plot( | |
| value=get_benchmark_chart(hardware_df), | |
| elem_id="hardware-plot", | |
| show_label=False, | |
| ) | |
| ####################### CONTROL PANEL ####################### | |
| with gr.TabItem("Control Panel 🎛️", id=2): | |
| gr.HTML( | |
| "Use this control panel to filter the leaderboard's table and plot.", # noqa: E501 | |
| elem_id="descriptive-text", | |
| ) | |
| with gr.Row(): | |
| with gr.Column(): | |
| search_bar = gr.Textbox( | |
| label="Model 🤗", | |
| info="🔍 Search for a model name", | |
| elem_id="search-bar", | |
| ) | |
| with gr.Row(): | |
| with gr.Column(scale=1): | |
| with gr.Box(): | |
| score_slider = gr.Slider( | |
| label="Open LLM Score 📈", | |
| info="🎚️ Slide to minimum Open LLM score", | |
| value=0, | |
| elem_id="threshold-slider", | |
| ) | |
| with gr.Column(scale=1): | |
| with gr.Box(): | |
| memory_slider = gr.Slider( | |
| label="Peak Memory (MB) 📈", | |
| info="🎚️ Slide to maximum Peak Memory", | |
| minimum=0, | |
| maximum=80 * 1024, | |
| value=80 * 1024, | |
| elem_id="memory-slider", | |
| ) | |
| with gr.Column(scale=1): | |
| backend_checkboxes = gr.CheckboxGroup( | |
| label="Backends 🏭", | |
| choices=["pytorch", "onnxruntime"], | |
| value=["pytorch", "onnxruntime"], | |
| info="☑️ Select the backends", | |
| elem_id="backend-checkboxes", | |
| ) | |
| with gr.Row(): | |
| with gr.Column(scale=1): | |
| datatype_checkboxes = gr.CheckboxGroup( | |
| label="Load Dtypes 📥", | |
| choices=["float32", "float16"], | |
| value=["float32", "float16"], | |
| info="☑️ Select the load dtypes", | |
| elem_id="dtype-checkboxes", | |
| ) | |
| with gr.Column(scale=1): | |
| optimizations_checkboxes = gr.CheckboxGroup( | |
| label="Optimizations 🛠️", | |
| choices=["None", "BetterTransformer"], | |
| value=["None", "BetterTransformer"], | |
| info="☑️ Select the optimizations", | |
| elem_id="optimizations-checkboxes", | |
| ) | |
| with gr.Column(scale=1): | |
| quantization_checkboxes = gr.CheckboxGroup( | |
| label="Quantizations 🗜️", | |
| choices=["None", "BnB.4bit", "GPTQ.4bit"], | |
| value=["None", "BnB.4bit", "GPTQ.4bit"], | |
| info="☑️ Select the quantization schemes", | |
| elem_id="quantization-checkboxes", | |
| ) | |
| with gr.Row(): | |
| filter_button = gr.Button( | |
| value="Filter 🚀", | |
| elem_id="filter-button", | |
| ) | |
| for hardware in HARDWARE_NAMES: | |
| filter_button.click( | |
| filter_query, | |
| [ | |
| search_bar, | |
| backend_checkboxes, | |
| datatype_checkboxes, | |
| optimizations_checkboxes, | |
| quantization_checkboxes, | |
| score_slider, | |
| memory_slider, | |
| hardware_placeholders[hardware], | |
| ], | |
| [hardware_tables[hardware], hardware_plots[hardware]], | |
| ) | |
| ####################### ABOUT TAB ####################### | |
| with gr.TabItem("About 📖", id=3): | |
| gr.HTML(ABOUT_TEXT, elem_classes="descriptive-text") | |
| gr.Markdown(EXAMPLE_CONFIG_TEXT, elem_classes="descriptive-text") | |
| ####################### CITATION ####################### | |
| with gr.Row(): | |
| with gr.Accordion("📙 Citation", open=False): | |
| citation_button = gr.Textbox( | |
| value=CITATION_BUTTON_TEXT, | |
| label=CITATION_BUTTON_LABEL, | |
| elem_id="citation-button", | |
| ).style(show_copy_button=True) | |
| # Restart space every hour | |
| scheduler = BackgroundScheduler() | |
| scheduler.add_job( | |
| restart_space, | |
| "interval", | |
| seconds=3600, | |
| args=[LLM_PERF_LEADERBOARD_REPO, OPTIMUM_TOKEN], | |
| ) | |
| scheduler.start() | |
| # Launch demo | |
| demo.queue(concurrency_count=10).launch() | |