IlyasMoutawwakil's picture
updated the llm-perf
134a499
raw
history blame
13.8 kB
import os
import gradio as gr
import pandas as pd
import plotly.express as px
from huggingface_hub.file_download import hf_hub_download
from src.utils import process_model_name, process_model_arch
from src.assets.css_html_js import custom_css
from src.assets.text_content import (
TITLE,
ABOUT_TEXT,
INTRODUCTION_TEXT,
EXAMPLE_CONFIG_TEXT,
CITATION_BUTTON_LABEL,
CITATION_BUTTON_TEXT,
)
HF_TOKEN = os.environ.get("HF_TOKEN", None)
LLM_PERF_DATASET_REPO = "optimum/llm-perf-dataset"
MACHINE_TO_HARDWARE = {"hf-dgx-01": "A100-80GB πŸ–₯️"}
ALL_COLUMNS_MAPPING = {
# model
"Model": "Model πŸ€—",
"Arch": "Arch πŸ›οΈ",
"Size": "Size πŸ‹οΈ",
# deployment settings
"backend.name": "Backend 🏭",
"backend.torch_dtype": "Dtype πŸ“₯",
"optimizations": "Optimizations πŸ› οΈ",
"quantization": "Quantization πŸ—œοΈ",
# throughput measurements
"decode.throughput(tokens/s)": "Decode Throughput (tokens/s) ⬆️",
"generate.throughput(tokens/s)": "E2E Throughput (tokens/s) ⬆️",
# latency measurements
"forward.latency(s)": "Prefill Latency (s) ⬇️",
"generate.latency(s)": "E2E Latency (s) ⬇️",
# memory measurements
"generate.max_memory_allocated(MB)": "Allocated Memory (MB) ⬇️",
"generate.max_memory_reserved(MB)": "Reserved Memory (MB) ⬇️",
"generate.max_memory_used(MB)": "Used Memory (MB) ⬇️",
# energy measurements
"generate.energy_consumption(tokens/kWh)": "Energy (tokens/kWh) ⬇️",
# quality measurements
"Score": "Avg Score (%) ⬆️",
}
SORTING_COLUMN = ["Score", "generate.throughput(tokens/s)"]
SORTING_ASCENDING = [False, True]
ALL_COLUMNS_DATATYPES = [
# open llm
"markdown",
"markdown",
"number",
# deployment settings
"str",
"str",
"str",
"str",
# measurements
"number",
"number",
"number",
"number",
"number",
"number",
"number",
"number",
"number",
"number",
]
def get_benchmark_df(machine="hf-dgx-01"):
# download data
hf_hub_download(
repo_id="optimum/llm-perf-dataset",
filename="open-llm.csv",
local_dir="dataset",
repo_type="dataset",
token=HF_TOKEN,
)
hf_hub_download(
repo_id="optimum/llm-perf-dataset",
filename=f"{machine}/full-report.csv",
local_dir="dataset",
repo_type="dataset",
token=HF_TOKEN,
)
open_llm = pd.read_csv("dataset/open-llm.csv")
full_report = pd.read_csv(f"dataset/{machine}/full-report.csv")
# merge on model
merged_df = open_llm.merge(full_report, left_on="Model", right_on="model")
# transpose energy consumption
merged_df["generate.energy_consumption(tokens/kWh)"] = (
1 / merged_df["generate.energy_consumption(kWh/token)"].fillna(1)
).astype(int)
# fix nan values
merged_df.loc[
merged_df["generate.energy_consumption(tokens/kWh)"] == 1,
"generate.energy_consumption(tokens/kWh)",
] = pd.NA
# add optimizations column
merged_df["optimizations"] = merged_df[
["backend.to_bettertransformer", "backend.use_flash_attention_2"]
].apply(
lambda x: "BetterTransformer"
if x["backend.to_bettertransformer"]
else ("FlashAttentionV2" if x["backend.use_flash_attention_2"] else "None"),
axis=1,
)
# add quantization scheme
merged_df["quantization"] = merged_df["backend.quantization_scheme"].apply(
lambda x: "BnB.4bit" if x == "bnb" else ("GPTQ.4bit" if x == "gptq" else "None")
)
# add decode throughput
merged_df["decode.throughput(tokens/s)"] = (
1000 / (merged_df["generate.latency(s)"] - merged_df["forward.latency(s)"])
).round(2)
# sort by metric
merged_df.sort_values(by=SORTING_COLUMN, ascending=SORTING_ASCENDING, inplace=True)
# filter columns
merged_df = merged_df[list(ALL_COLUMNS_MAPPING.keys())]
# rename columns
merged_df.rename(columns=ALL_COLUMNS_MAPPING, inplace=True)
return merged_df
def get_benchmark_table(bench_df):
copy_df = bench_df.copy()
# transform
copy_df["Model πŸ€—"] = copy_df["Model πŸ€—"].apply(process_model_name)
copy_df["Arch πŸ›οΈ"] = copy_df["Arch πŸ›οΈ"].apply(process_model_arch)
# process quantization
copy_df["Avg Score (%) ⬆️"] = copy_df.apply(
lambda x: f"{x['Avg Score (%) ⬆️']}**"
if x["Quantization πŸ—œοΈ"] in ["BnB.4bit", "GPTQ.4bit"]
else x["Avg Score (%) ⬆️"],
axis=1,
)
return copy_df
def get_benchmark_chart(bench_df):
copy_df = bench_df.copy()
# transform
copy_df["Arch πŸ›οΈ"] = copy_df["Arch πŸ›οΈ"].apply(process_model_arch)
# filter latency bigger than 100s
# copy_df = copy_df[copy_df["E2E Latency (s) ⬇️"] <= 100]
fig = px.scatter(
copy_df,
y="Avg Score (%) ⬆️",
x="E2E Latency (s) ⬇️",
size="Allocated Memory (MB) ⬇️",
color="Arch πŸ›οΈ",
custom_data=list(ALL_COLUMNS_MAPPING.values()),
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="Avg Open LLM Score (%)",
legend_title="LLM Architecture",
width=1200,
height=600,
)
fig.update_traces(
hovertemplate="<br>".join(
[
f"<b>{column}:</b> %{{customdata[{i}]}}"
for i, column in enumerate(ALL_COLUMNS_MAPPING.values())
]
)
)
return fig
def filter_query(
text,
backends,
datatypes,
optimizations,
quantization_scheme,
score,
memory,
machine,
):
raw_df = get_benchmark_df(machine=machine)
filtered_df = raw_df[
raw_df["Model πŸ€—"].str.contains(text, case=False)
& raw_df["Backend 🏭"].isin(backends)
& raw_df["Dtype πŸ“₯"].isin(datatypes)
& (
pd.concat(
[
raw_df["Optimizations πŸ› οΈ"].str.contains(optimization, case=False)
for optimization in optimizations
],
axis=1,
).any(axis="columns")
if len(optimizations) > 0
else True
)
& (
pd.concat(
[
raw_df["Quantization πŸ—œοΈ"].str.contains(quantization, case=False)
for quantization in quantization_scheme
],
axis=1,
).any(axis="columns")
if len(quantization_scheme) > 0
else True
)
& (raw_df["Avg Score (%) ⬆️"] >= score)
& (raw_df["Allocated 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"):
machine_placeholders = {}
machine_tables = {}
machine_plots = {}
####################### HARDWARE TABS #######################
for i, (machine, hardware) in enumerate(MACHINE_TO_HARDWARE.items()):
# dummy placeholder of the machine name
machine_placeholders[machine] = gr.Textbox(value=machine, visible=False)
with gr.TabItem(hardware, id=i):
with gr.Tabs(elem_classes="machine-tabs"):
# placeholder for full dataframe
machine_df = get_benchmark_df(machine=machine)
with gr.TabItem("Leaderboard πŸ…", id=0):
gr.HTML(
"πŸ‘‰ Scroll to the right πŸ‘‰ for additional columns.",
elem_id="descriptive-text",
)
# Original leaderboard table
machine_tables[machine] = gr.components.Dataframe(
value=get_benchmark_table(machine_df),
headers=list(ALL_COLUMNS_MAPPING.values()),
datatype=ALL_COLUMNS_DATATYPES,
elem_id="machine-table",
)
with gr.TabItem("Plot πŸ“Š", id=1):
gr.HTML(
"πŸ‘† Hover over the points πŸ‘† for additional information.",
elem_id="descriptive-text",
)
# Original leaderboard plot
machine_plots[machine] = gr.components.Plot(
value=get_benchmark_chart(machine_df),
elem_id="machine-plot",
show_label=False,
)
###################### CONTROL PANEL #######################
with gr.TabItem("Control Panel πŸŽ›οΈ", 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 πŸ€—",
info="πŸ” 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 πŸ“ˆ",
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) πŸ“ˆ",
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="β˜‘οΈ Select the backends",
elem_id="backend-checkboxes",
)
with gr.Row():
with gr.Column(scale=1):
datatype_checkboxes = gr.CheckboxGroup(
label="Load Dtypes πŸ“₯",
choices=["float32", "float16"],
value=["float32", "float16"],
info="β˜‘οΈ Select the load dtypes",
elem_id="dtype-checkboxes",
)
with gr.Column(scale=1):
optimizations_checkboxes = gr.CheckboxGroup(
label="Optimizations πŸ› οΈ",
choices=["None", "BetterTransformer"],
value=["None", "BetterTransformer"],
info="β˜‘οΈ Select the optimizations",
elem_id="optimizations-checkboxes",
)
with gr.Column(scale=1):
quantization_checkboxes = gr.CheckboxGroup(
label="Quantizations πŸ—œοΈ",
choices=["None", "BnB.4bit", "GPTQ.4bit"],
value=["None", "BnB.4bit", "GPTQ.4bit"],
info="β˜‘οΈ Select the quantization schemes",
elem_id="quantization-checkboxes",
)
with gr.Row():
filter_button = gr.Button(
value="Filter πŸš€",
elem_id="filter-button",
)
for machine in MACHINE_TO_HARDWARE:
filter_button.click(
filter_query,
[
search_bar,
backend_checkboxes,
datatype_checkboxes,
optimizations_checkboxes,
quantization_checkboxes,
score_slider,
memory_slider,
machine_placeholders[machine],
],
[machine_tables[machine], machine_plots[machine]],
)
####################### ABOUT TAB #######################
with gr.TabItem("About πŸ“–", 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("πŸ“™ Citation", open=False):
citation_button = gr.Textbox(
value=CITATION_BUTTON_TEXT,
label=CITATION_BUTTON_LABEL,
elem_id="citation-button",
show_copy_button=True,
)
# Launch demo
demo.launch(show_api=False)