app.py

import gradio as gr
import torch
from diffusers import DiffusionPipeline
from accelerate import init_empty_weights, load_checkpoint_and_dispatch

# Detección y configuración del dispositivo para compatibilidad con GPU o CPU
if torch.cuda.is_available():
    device = "cuda"  # Para GPUs NVIDIA
elif hasattr(torch.backends, "mps") and torch.backends.mps.is_built():
    device = "mps"  # Para GPUs Apple Silicon (M1/M2) y otras GPUs con soporte Metal
elif hasattr(torch.backends, "rocm") and torch.backends.rocm.is_available():
    device = "rocm"  # Para GPUs AMD con ROCm, si está disponible
else:
    device = "cpu"  # En caso de no tener GPU disponible

# Definir el tipo de dato, usando bfloat16 si es compatible, si no, usar float32
dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32

# Inicializar el modelo solo una vez y cargarlo en RAM y GPU/CPU
pipe = None

def load_model():
    global pipe
    if pipe is None:
        # Inicializar ZeroGPU antes de cargar el modelo
        init_empty_weights()

        # Cargar el modelo y configurarlo para usar el dispositivo adecuado
        pipe = DiffusionPipeline.from_pretrained(
            "black-forest-labs/FLUX.1-schnell", 
            torch_dtype=dtype
        )
        # Despachar los pesos al dispositivo adecuado (GPU o CPU)
        pipe = load_checkpoint_and_dispatch(
            pipe, 
            "black-forest-labs/FLUX.1-schnell", 
            device_map="auto",  # Automatiza el uso de RAM, GPU o CPU
            offload_folder=None  # Evita que se almacenen los pesos temporalmente en el disco
        )
        pipe.to(device)

MAX_SEED = torch.iinfo(torch.int32).max

def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, num_images=1):
    load_model()  # Asegurarse de que el modelo esté cargado antes de la inferencia
    
    if randomize_seed:
        seed = torch.randint(0, MAX_SEED, (1,)).item()
    generator = torch.Generator(device).manual_seed(seed)
    
    images = []
    for _ in range(num_images):
        image = pipe(
                prompt=prompt, 
                width=width,
                height=height,
                num_inference_steps=num_inference_steps, 
                generator=generator,
                guidance_scale=0.0
        ).images[0]
        images.append(image)
    
    return images, 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)
        
        # Usamos gr.Gallery para mostrar múltiples imágenes
        results = gr.Gallery(label="Results", show_label=False, elem_id="image-gallery")
        
        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=2048,  # Ajusta el tamaño máximo según sea necesario
                    step=32,
                    value=1024,
                )
                
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=2048,  # Ajusta el tamaño máximo según sea necesario
                    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,
                )
            
            # Control para el número de imágenes a generar
            num_images = gr.Slider(
                label="Number of images",
                minimum=1,
                maximum=10,  # Ajusta el número máximo de imágenes según sea necesario
                step=1,
                value=1,
            )
        
        gr.Examples(
            examples = examples,
            fn = infer,
            inputs = [prompt],
            outputs = [results, seed],
            cache_examples="lazy"
        )

    # Conectar el botón y el campo de texto a la función infer
    run_button.click(
        fn=infer,
        inputs=[prompt, seed, randomize_seed, width, height, num_inference_steps, num_images],
        outputs=[results, seed]
    )
    
    # Crear un enlace público con share=True
    demo.launch(share=True)