app.py

import gradio as gr
import numpy as np
import random
from diffusers import StableDiffusionPipeline, LCMScheduler
import torch

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

transformer = Transformer2DModel.from_pretrained(
  "PixArt-alpha/PixArt-XL-2-1024-MS",
  subfolder="transformer",
  torch_dtype=torch.float16
)
transformer = PeftModel.from_pretrained(
  transformer,
  "jasperai/flash-pixart"
)


if torch.cuda.is_available():
    torch.cuda.max_memory_allocated(device=device)
    pipe = PixArtAlphaPipeline.from_pretrained(
      "PixArt-alpha/PixArt-XL-2-1024-MS",
      transformer=transformer,
      torch_dtype=torch.float16
    )
    pipe.enable_xformers_memory_efficient_attention()
    pipe = pipe.to(device)
else: 
    pipe = PixArtAlphaPipeline.from_pretrained(
      "PixArt-alpha/PixArt-XL-2-1024-MS",
      transformer=transformer,
      torch_dtype=torch.float16
    )
    pipe = pipe.to(device)

pipe.scheduler = LCMScheduler.from_pretrained(
    "PixArt-alpha/PixArt-XL-2-1024-MS",
    subfolder="scheduler",
    timestep_spacing="trailing",
)

pipe.load_lora_weights(adapter_id)
pipe.fuse_lora()

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024
NUM_INFERENCE_STEPS = 4

def infer(prompt, seed, randomize_seed, num_inference_steps):

    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
        
    generator = torch.Generator().manual_seed(seed)
    
    image = pipe(
        prompt = prompt, 
        guidance_scale = 0, 
        num_inference_steps = num_inference_steps, 
        generator = generator
    ).images[0] 
    
    return image

examples = [
    "The image showcases a freshly baked bread, possibly focaccia, with rosemary sprigs and red pepper flakes sprinkled on top. It's sliced and placed on a wire cooling rack, with a bowl of mixed peppercorns beside it.",
    "A raccoon reading a book in a lush forest.",
    "A serene landscape showcases a winding road alongside a vast, turquoise lake, flanked by majestic snow-capped mountains under a partly cloudy sky.",
]

css="""
#col-container {
    margin: 0 auto;
    max-width: 512px;
}
"""

if torch.cuda.is_available():
    power_device = "GPU"
else:
    power_device = "CPU"

with gr.Blocks(css=css) as demo:
    
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""
        # ⚡ FlashDiffusion: FlashPixart ⚡
        This is an interactive demo of [Flash Diffusion](https://huggingface.co/jasperai/flash-pixart), a diffusion distillation method proposed in [ADD ARXIV]() *by Clément Chadebec, Onur Tasar and Benjamin Aubin.*
        This model is a **66.5M** LoRA distilled version of [Pixart-α](https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS) model that is able to generate 1024x1024 images in **4 steps**. 
        Currently running on {power_device}.
        """)
        
        with gr.Row():
            
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                container=False,
            )
            
            run_button = gr.Button("Run", scale=0)
        
        result = gr.Image(label="Result", show_label=False)

        with gr.Accordion("Advanced Settings", open=False):
            
            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )
            
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
            
        
        gr.Examples(
            examples = examples,
            inputs = [prompt]
        )

    run_button.click(
        fn = infer,
        inputs = [prompt, seed, randomize_seed, NUM_INFERENCE_STEPS],
        outputs = [result]
    )

demo.queue().launch()