app.py

import gradio as gr
import torch
from diffusers import AutoPipelineForText2Image
import time

USE_TORCH_COMPILE = True

dtype = torch.float16
device = torch.device("cuda:0")

pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", variant="fp16", torch_dtype=dtype)
pipeline.vae.register_to_config(force_upcast=False)
pipeline.to(device)


if USE_TORCH_COMPILE:
    pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)
    
def generate(num_images_per_prompt: int = 1):
    print("Start...")
    print("Version", torch.__version__)

    for _ in range(3):
        prompt = 77 * "a"
        num_inference_steps = 40
        start_time = time.time()
        pipeline(prompt, num_images_per_prompt=num_images_per_prompt, num_inference_steps=num_inference_steps).images
        end_time = time.time()
    
        print(f"For {num_inference_steps} steps", end_time - start_time)
        print("Avg per step", (end_time - start_time) / num_inference_steps)


with gr.Blocks(css="style.css") as demo:
    batch_size = gr.Slider(
        label="Batch size",
        minimum=0,
        maximum=16,
        step=1,
        value=1,
    )
    btn = gr.Button("Benchmark!").style(
        margin=False,
        rounded=(False, True, True, False),
        full_width=False,
    )

    btn.click(fn=generate, inputs=[batch_size])

demo.launch()