app.py

import os
import math
import gradio as gr
import pandas as pd
import plotly.express as px
from apscheduler.schedulers.background import BackgroundScheduler

from src.assets.text_content import (
    TITLE,
    INTRODUCTION_TEXT,
    SINGLE_A100_TEXT,
    CITATION_BUTTON_LABEL,
    CITATION_BUTTON_TEXT,
)
from src.utils import (
    change_tab,
    restart_space,
    load_dataset_repo,
    make_clickable_model,
    # make_clickable_score,
    # num_to_str,
)
from src.assets.css_html_js import custom_css, custom_js


LLM_PERF_LEADERBOARD_REPO = "optimum/llm-perf-leaderboard"
LLM_PERF_DATASET_REPO = "optimum/llm-perf-dataset"
OPTIMUM_TOKEN = os.environ.get("OPTIMUM_TOKEN", None)

COLUMNS_MAPPING = {
    "model": "Model 🤗",
    "backend.name": "Backend 🏭",
    "backend.torch_dtype": "Load Dtype 📥",
    "optimizations": "Optimizations 🛠️",
    #
    "perf": "Open LLM-Perf Score ⬆️",
    #
    "generate.throughput(tokens/s)": "Throughput (tokens/s) ⬆️",
    "score": "Open LLM Score ⬆️",
    "forward.peak_memory(MB)": "Peak Memory (MB) ⬇️",
    "num_params": "#️⃣ Parameters (M) 📏",
}
COLUMNS_DATATYPES = [
    "markdown",
    "str",
    "str",
    "str",
    #
    "number",
    "number",
    #
    "number",
    "number",
    "number",
]
SORTING_COLUMN = ["Open LLM-Perf Score ⬆️"]


llm_perf_dataset_repo = load_dataset_repo(LLM_PERF_DATASET_REPO, OPTIMUM_TOKEN)


def get_benchmark_df(benchmark="1xA100-80GB"):
    if llm_perf_dataset_repo:
        llm_perf_dataset_repo.git_pull()

    # load and merge
    bench_df = pd.read_csv(f"./llm-perf-dataset/reports/{benchmark}.csv")
    scores_df = pd.read_csv(f"./llm-perf-dataset/reports/open-llm-leaderboard.csv")
    bench_df = bench_df.merge(scores_df, on="model", how="left")

    # filter out models with no score
    bench_df = bench_df[bench_df["score"].notna()]

    # create composite score
    score_distance = 100 - bench_df["score"]
    latency_distance = bench_df["generate.latency(s)"]
    bench_df["perf"] = 100 / (1 + (score_distance**2 + latency_distance**2) ** 0.5)
    bench_df["perf"] = bench_df["perf"].round(2)

    # add optimizations
    bench_df["optimizations"] = bench_df[
        ["backend.bettertransformer", "backend.load_in_8bit", "backend.load_in_4bit"]
    ].apply(
        lambda x: ", ".join(
            filter(
                lambda x: x != "",
                [
                    "BetterTransformer" if x[0] == True else "",
                    "LLM.int8" if x[1] == True else "",
                    "LLM.fp4" if x[2] == True else "",
                ],
            ),
        )
        if any([x[0] == True, x[1] == True, x[2] == True])
        else "None",
        axis=1,
    )

    return bench_df


def get_benchmark_table(bench_df):
    # filter
    bench_df = bench_df[list(COLUMNS_MAPPING.keys())]
    # rename
    bench_df.rename(columns=COLUMNS_MAPPING, inplace=True)
    # sort
    bench_df.sort_values(by=SORTING_COLUMN, ascending=True, inplace=True)
    # transform
    bench_df["Model 🤗"] = bench_df["Model 🤗"].apply(make_clickable_model)
    bench_df["#️⃣ Parameters (M) 📏"] = bench_df["#️⃣ Parameters (M) 📏"].apply(
        lambda x: int(x / (1024 * 1024))
    )

    return bench_df


def get_benchmark_plot(bench_df):
    # untill falcon gets fixed / natively supported
    bench_df = bench_df[bench_df["generate.latency(s)"] < 150]

    fig = px.scatter(
        bench_df,
        x="generate.latency(s)",
        y="score",
        color="model_type",
        symbol="backend.name",
        size="forward.peak_memory(MB)",
        custom_data=[
            "model",
            "backend.name",
            "backend.torch_dtype",
            "optimizations",
            "forward.peak_memory(MB)",
            "generate.throughput(tokens/s)",
        ],
        symbol_sequence=["triangle-up", "circle"],
        color_discrete_sequence=px.colors.qualitative.Light24,
    )

    fig.update_layout(
        title={
            "text": "Model Score vs. Latency 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="Model Type and Backend",
        width=1200,
        height=600,
    )

    fig.update_traces(
        hovertemplate="<br>".join(
            [
                "Model: %{customdata[0]}",
                "Backend: %{customdata[1]}",
                "Datatype: %{customdata[2]}",
                "Optimizations: %{customdata[3]}",
                "Peak Memory (MB): %{customdata[4]}",
                "Throughput (tokens/s): %{customdata[5]}",
                "Per 1000 Tokens Latency (s): %{x}",
                "Open LLM Score: %{y}",
            ]
        )
    )

    return fig


def filter_query(
    text,
    backends,
    datatypes,
    optimizations,
    score,
    memory,
    benchmark="1xA100-80GB",
):
    raw_df = get_benchmark_df(benchmark=benchmark)

    filtered_df = raw_df[
        raw_df["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
        )
        & (raw_df["score"] >= score)
        & (raw_df["forward.peak_memory(MB)"] <= memory)
    ]

    filtered_table = get_benchmark_table(filtered_df)
    filtered_plot = get_benchmark_plot(filtered_df)

    return filtered_table, filtered_plot


# Dataframes
single_A100_df = get_benchmark_df(benchmark="1xA100-80GB")
single_A100_table = get_benchmark_table(single_A100_df)
single_A100_plot = get_benchmark_plot(single_A100_df)

# Demo interface
demo = gr.Blocks(css=custom_css)
with demo:
    # leaderboard title
    gr.HTML(TITLE)

    # introduction text
    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")

    # control panel title
    gr.HTML("<h2>Control Panel 🎛️</h2>")

    # control panel interface
    with gr.Row():
        with gr.Column(scale=1):
            search_bar = gr.Textbox(
                label="Model 🤗",
                info="🔍 Search for a model name",
                elem_id="search-bar",
            )
        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.Row():
        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.Column(scale=1):
            datatype_checkboxes = gr.CheckboxGroup(
                label="Datatypes 📥",
                choices=["float32", "float16"],
                value=["float32", "float16"],
                info="☑️ Select the load datatypes",
                elem_id="datatype-checkboxes",
            )
        with gr.Column(scale=2):
            optimizations_checkboxes = gr.CheckboxGroup(
                label="Optimizations 🛠️",
                choices=["None", "BetterTransformer", "LLM.int8", "LLM.fp4"],
                value=["None", "BetterTransformer", "LLM.int8", "LLM.fp4"],
                info="☑️ Select the optimizations",
                elem_id="optimizations-checkboxes",
            )

    with gr.Row():
        filter_button = gr.Button(
            value="Filter 🚀",
            elem_id="filter-button",
        )

    # leaderboard tabs
    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        with gr.TabItem("🖥️ A100-80GB Leaderboard 🏆", id=0):
            gr.HTML(SINGLE_A100_TEXT)

            # Original leaderboard table
            single_A100_leaderboard = gr.components.Dataframe(
                value=single_A100_table,
                datatype=COLUMNS_DATATYPES,
                headers=list(COLUMNS_MAPPING.values()),
                elem_id="1xA100-table",
            )

        with gr.TabItem("🖥️ A100-80GB Plot 📊", id=1):
            # Original leaderboard plot
            gr.HTML(SINGLE_A100_TEXT)

            # Original leaderboard plot
            single_A100_plotly = gr.components.Plot(
                value=single_A100_plot,
                elem_id="1xA100-plot",
                show_label=False,
            )

    filter_button.click(
        filter_query,
        [
            search_bar,
            backend_checkboxes,
            datatype_checkboxes,
            optimizations_checkboxes,
            score_slider,
            memory_slider,
        ],
        [single_A100_leaderboard, single_A100_plotly],
    )

    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)

    dummy = gr.Textbox(visible=False)
    demo.load(
        change_tab,
        dummy,
        tabs,
        _js=custom_js,
    )


# 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=40).launch()