app.py

from diffusers import (
    StableDiffusionXLPipeline,
    EulerDiscreteScheduler,
    UNet2DConditionModel,
)
import torch
import os
from huggingface_hub import hf_hub_download


from PIL import Image
import gradio as gr
import time
from safetensors.torch import load_file
from sfast.compilers.diffusion_pipeline_compiler import compile, CompilationConfig

# Constants
BASE = "stabilityai/stable-diffusion-xl-base-1.0"
REPO = "ByteDance/SDXL-Lightning"
# 1-step
CHECKPOINT = "sdxl_lightning_1step_unet_x0.safetensors"

# {
#     "1-Step": ["sdxl_lightning_1step_unet_x0.safetensors", 1],
#     "2-Step": ["sdxl_lightning_2step_unet.safetensors", 2],
#     "4-Step": ["sdxl_lightning_4step_unet.safetensors", 4],
#     "8-Step": ["sdxl_lightning_8step_unet.safetensors", 8],
# }


TORCH_COMPILE = os.environ.get("TORCH_COMPILE", "0") == "1"
# check if MPS is available OSX only M1/M2/M3 chips

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch_device = device
torch_dtype = torch.float16

print(f"TORCH_COMPILE: {TORCH_COMPILE}")
print(f"device: {device}")

# Load model.
unet = UNet2DConditionModel.from_config(BASE, subfolder="unet").to(device, torch_dtype)
unet.load_state_dict(load_file(hf_hub_download(REPO, CHECKPOINT), device="cuda"))
pipe = StableDiffusionXLPipeline.from_pretrained(
    BASE, unet=unet, torch_dtype=torch_dtype, variant="fp16"
).to(device)

# Ensure sampler uses "trailing" timesteps and "sample" prediction type.
pipe.scheduler = EulerDiscreteScheduler.from_config(
    pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample"
)
pipe.set_progress_bar_config(disable=True)
config = CompilationConfig.Default()
try:
    import xformers

    config.enable_xformers = True
except ImportError:
    print("xformers not installed, skip")
try:
    import triton

    config.enable_triton = True
except ImportError:
    print("Triton not installed, skip")
# CUDA Graph is suggested for small batch sizes and small resolutions to reduce CPU overhead.
# But it can increase the amount of GPU memory used.
# For StableVideoDiffusionPipeline it is not needed.
config.enable_cuda_graph = True

pipe = compile(pipe, config)


def predict(prompt, seed=1231231):
    generator = torch.manual_seed(seed)
    last_time = time.time()
    results = pipe(
        prompt=prompt,
        generator=generator,
        num_inference_steps=1,
        guidance_scale=0.0,
        # width=768,
        # height=768,
        # original_inference_steps=params.lcm_steps,
        output_type="pil",
    )
    print(f"Pipe took {time.time() - last_time} seconds")
    nsfw_content_detected = (
        results.nsfw_content_detected[0]
        if "nsfw_content_detected" in results
        else False
    )
    if nsfw_content_detected:
        gr.Warning("NSFW content detected.")
        return Image.new("RGB", (512, 512))
    return results.images[0]


css = """
#container{
    margin: 0 auto;
    max-width: 40rem;
}
#intro{
    max-width: 100%;
    margin: 0 auto;
}
"""
with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="container"):
        gr.Markdown(
            """
# SDXL-Lightning- Text To Image 1-Step
**Model**: https://huggingface.co/ByteDance/SDXL-Lightning
            """,
            elem_id="intro",
        )
        with gr.Row():
            with gr.Row():
                prompt = gr.Textbox(
                    placeholder="Insert your prompt here:", scale=5, container=False
                )
                generate_bt = gr.Button("Generate", scale=1)

        image = gr.Image(type="filepath")
        with gr.Accordion("Advanced options", open=False):
            seed = gr.Slider(
                randomize=True, minimum=0, maximum=12013012031030, label="Seed", step=1
            )
        with gr.Accordion("Run with diffusers"):
            gr.Markdown(
                """## Running SDXL-Lightning with `diffusers`
```py
import torch
from diffusers import (
    StableDiffusionXLPipeline,
    UNet2DConditionModel,
    EulerDiscreteScheduler,
)
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

base = "stabilityai/stable-diffusion-xl-base-1.0"
repo = "ByteDance/SDXL-Lightning"
ckpt = "sdxl_lightning_1step_unet_x0.safetensors"  # Use the correct ckpt for your step setting!

# Load model.
unet = UNet2DConditionModel.from_config(base, subfolder="unet").to(
    "cuda", torch.float16
)
unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device="cuda"))
pipe = StableDiffusionXLPipeline.from_pretrained(
    base, unet=unet, torch_dtype=torch.float16, variant="fp16"
).to("cuda")

# Ensure sampler uses "trailing" timesteps and "sample" prediction type.
pipe.scheduler = EulerDiscreteScheduler.from_config(
    pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample"
)

# Ensure using the same inference steps as the loaded model and CFG set to 0.
pipe("A girl smiling", num_inference_steps=1, guidance_scale=0).images[0].save(
    "output.png"
)
```
            """
            )

        inputs = [prompt, seed]
        generate_bt.click(fn=predict, inputs=inputs, outputs=image, show_progress=False)
        prompt.input(fn=predict, inputs=inputs, outputs=image, show_progress=False)
        seed.change(fn=predict, inputs=inputs, outputs=image, show_progress=False)

demo.queue()
demo.launch()