app.py

import gradio as gr
import numpy as np
import random
import spaces
import torch
from diffusers import DiffusionPipeline

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

# Load the model in FP16
pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.float16)

# Move the pipeline to GPU if available
pipe = pipe.to(device)

# Convert text encoders to full precision
pipe.text_encoder = pipe.text_encoder.to(torch.float32)
if hasattr(pipe, 'text_encoder_2'):
    pipe.text_encoder_2 = pipe.text_encoder_2.to(torch.float32)

# Enable memory efficient attention if available and on CUDA
if device == "cuda" and hasattr(pipe, 'enable_xformers_memory_efficient_attention'):
    try:
        pipe.enable_xformers_memory_efficient_attention()
        print("xformers memory efficient attention enabled")
    except Exception as e:
        print(f"Could not enable memory efficient attention: {e}")

# Compile the UNet for potential speedups if on CUDA
if device == "cuda":
    try:
        pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
        print("UNet compiled for potential speedups")
    except Exception as e:
        print(f"Could not compile UNet: {e}")

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 2048

@spaces.GPU()
def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, progress=gr.Progress(track_tqdm=True)):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)
    
    # Use full precision for text encoding
    with torch.no_grad():
        text_inputs = pipe.tokenizer(prompt, return_tensors="pt").to(device)
        text_embeddings = pipe.text_encoder(text_inputs.input_ids)[0]
    
    # Use mixed precision for the rest of the pipeline
    with torch.inference_mode(), torch.autocast(device_type='cuda', dtype=torch.float16):
        image = pipe(
            prompt_embeds=text_embeddings,
            width=width,
            height=height,
            num_inference_steps=num_inference_steps,
            generator=generator,
            guidance_scale=0.0
        ).images[0]
    
    return image, seed
 
examples = [
    "a tiny astronaut hatching from an egg on the moon",
    "a cat holding a sign that says hello world",
    "an anime illustration of a wiener schnitzel",
]

css="""
#col-container {
    margin: 0 auto;
    max-width: 520px;
}
"""

with gr.Blocks(css=css) as demo:
    
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""# FLUX.1 [schnell]
12B param rectified flow transformer distilled from [FLUX.1 [pro]](https://blackforestlabs.ai/) for 4 step generation
[[blog](https://blackforestlabs.ai/announcing-black-forest-labs/)] [[model](https://huggingface.co/black-forest-labs/FLUX.1-schnell)]
        """)
        
        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)
            
            with gr.Row():
                
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,
                )
                
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,
                )
            
            with gr.Row():
                
  
                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=4,
                )
        
        gr.Examples(
            examples = examples,
            fn = infer,
            inputs = [prompt],
            outputs = [result, seed],
            cache_examples="lazy"
        )

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn = infer,
        inputs = [prompt, seed, randomize_seed, width, height, num_inference_steps],
        outputs = [result, seed]
    )

demo.launch()