IlyasMoutawwakil's picture
fix
77d2cc4
import os
import subprocess
import importlib.util
NVIDIA_AVAILABLE = (
subprocess.run(
"nvidia-smi",
shell=True,
stdout=subprocess.DEVNULL,
stderr=subprocess.DEVNULL,
).returncode
== 0
)
if NVIDIA_AVAILABLE:
DEVICES = ["cpu", "cuda"]
if importlib.util.find_spec("optimum_benchmark") is None:
os.system(
"pip install optimum-benchmark[onnxruntime-gpu,openvino,neural-compressor,diffusers,peft]@git+https://github.com/huggingface/optimum-benchmark.git"
)
os.system("pip uninstall onnxruntime onnxruntime-gpu -y")
os.system("pip install onnxruntime-gpu")
else:
DEVICES = ["cpu"]
if importlib.util.find_spec("optimum_benchmark") is None:
os.system(
"pip install optimum-benchmark[onnxruntime,openvino,neural-compressor,diffusers,peft]@git+https://github.com/huggingface/optimum-benchmark.git"
)
BACKENDS = ["pytorch", "onnxruntime", "openvino", "neural-compressor"]
BENCHMARKS = ["inference", "training"]
import random
import gradio as gr
from optimum_benchmark.task_utils import (
TASKS_TO_AUTOMODELS,
infer_task_from_model_name_or_path,
)
from run import run_benchmark
from config_store import (
get_training_config,
get_inference_config,
get_neural_compressor_config,
get_onnxruntime_config,
get_openvino_config,
get_pytorch_config,
)
with gr.Blocks() as demo:
# add image
gr.Markdown(
"""<img src="https://huggingface.co/spaces/optimum/optimum-benchmark-ui/resolve/main/huggy_bench.png" style="display: block; margin-left: auto; margin-right: auto; width: 30%;">"""
)
# title text
gr.Markdown("<h1 style='text-align: center'>πŸ€— Optimum-Benchmark UI πŸ‹οΈ</h1>")
# explanation text
gr.HTML(
"<h3 style='text-align: center'>"
"Zero code Gradio interface of <a href='https://github.com/huggingface/optimum-benchmark.git'>Optimum-Benchmark</a><br>"
"</h3>"
"<p style='text-align: center'>"
"Note: <a href='https://huggingface.co/spaces/optimum/optimum-benchmark-ui?duplicate=true'>Duplicate this space</a> and change its hardware to enable CUDA device<br>"
"or <a href='https://huggingface.co/spaces/optimum/optimum-benchmark-ui?docker=true'>Run with Docker</a> locally to target your own hardware."
"</p>"
)
model = gr.Textbox(
label="model",
value="distilbert-base-uncased-finetuned-sst-2-english",
info="Model to run the benchmark on. Press enter to infer the task automatically.",
)
task = gr.Dropdown(
label="task",
value="text-classification",
choices=list(TASKS_TO_AUTOMODELS.keys()),
info="Task to run the benchmark on. Can be infered automatically by submitting a model.",
)
device = gr.Dropdown(
value="cpu",
label="device",
choices=DEVICES,
info="Device to run the benchmark on. make sure to duplicate the space if you wanna run on CUDA devices.",
)
experiment = gr.Textbox(
label="experiment_name",
value=f"awesome-experiment-{random.randint(0, 100000)}",
info="Name of the experiment. Will be used to create a folder where results are stored.",
)
model.submit(fn=infer_task_from_model_name_or_path, inputs=model, outputs=task)
with gr.Row():
with gr.Column():
with gr.Row():
backend = gr.Dropdown(
label="backend",
choices=BACKENDS,
value=BACKENDS[0],
info="Backend to run the benchmark on.",
)
with gr.Row() as backend_configs:
with gr.Accordion(label="backend options", open=False, visible=True):
pytorch_config = get_pytorch_config()
with gr.Accordion(label="backend config", open=False, visible=False):
onnxruntime_config = get_onnxruntime_config()
with gr.Accordion(label="backend config", open=False, visible=False):
openvino_config = get_openvino_config()
with gr.Accordion(label="backend config", open=False, visible=False):
neural_compressor_config = get_neural_compressor_config()
# hide backend configs based on backend
backend.change(
inputs=backend,
outputs=backend_configs.children,
fn=lambda value: [gr.update(visible=value == key) for key in BACKENDS],
)
with gr.Column():
with gr.Row():
benchmark = gr.Dropdown(
label="benchmark",
choices=BENCHMARKS,
value=BENCHMARKS[0],
info="Type of benchmark to run.",
)
with gr.Row() as benchmark_configs:
with gr.Accordion(label="benchmark Config", open=False, visible=True):
inference_config = get_inference_config()
with gr.Accordion(label="benchmark Config", open=False, visible=False):
training_config = get_training_config()
# hide benchmark configs based on benchmark
benchmark.change(
inputs=benchmark,
outputs=benchmark_configs.children,
fn=lambda value: [gr.update(visible=value == key) for key in BENCHMARKS],
)
baseline = gr.Checkbox(
value=False,
label="compare_to_baseline",
info="Check this box to compare your chosen configuration to the baseline configuration.",
)
button = gr.Button(value="Run Benchmark", variant="primary")
with gr.Accordion(label="", open=True):
html_output = gr.HTML()
table_output = gr.Dataframe(visible=False)
button.click(
fn=run_benchmark,
inputs={
experiment,
baseline,
model,
task,
device,
backend,
benchmark,
*pytorch_config,
*openvino_config,
*onnxruntime_config,
*neural_compressor_config,
*inference_config,
*training_config,
},
outputs=[html_output, button, table_output],
queue=True,
)
button.click(
fn=lambda: f"awesome-experiment-{random.randint(0, 100000)}",
inputs=[],
outputs=experiment,
queue=True,
)
demo.queue().launch()