app.py

import logging
import random
import warnings
import os
import gradio as gr
import numpy as np
import spaces
import torch
from diffusers import FluxControlNetModel
from diffusers.pipelines import FluxControlNetPipeline
from gradio_imageslider import ImageSlider
from PIL import Image
from huggingface_hub import snapshot_download

# 메모리 관리를 위한 gc 추가
import gc
gc.collect()
torch.cuda.empty_cache()

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

# Device setup with minimal memory usage
if torch.cuda.is_available():
    power_device = "GPU"
    device = "cuda" 
    dtype = torch.float16  # Use float16 for minimum memory
    # Set CUDA memory fraction to 50%
    torch.cuda.set_per_process_memory_fraction(0.5)
else:
    power_device = "CPU"
    device = "cpu"
    dtype = torch.float32

huggingface_token = os.getenv("HUGGINFACE_TOKEN")

# Minimal model configuration
model_config = {
    "low_cpu_mem_usage": True,
    "torch_dtype": dtype,
    "use_safetensors": True,
    "variant": "fp16",  # Use fp16 variant if available
}

model_path = snapshot_download(
    repo_id="black-forest-labs/FLUX.1-dev",
    repo_type="model",
    ignore_patterns=["*.md", "*..gitattributes", "*.bin"],  # Ignore unnecessary files
    local_dir="FLUX.1-dev", 
    token=huggingface_token,
)

# Load models with minimal configuration
try:
    controlnet = FluxControlNetModel.from_pretrained(
        "jasperai/Flux.1-dev-Controlnet-Upscaler",
        **model_config
    ).to(device)

    pipe = FluxControlNetPipeline.from_pretrained(
        model_path,
        controlnet=controlnet,
        **model_config
    )

    # Enable all memory optimizations
    pipe.enable_model_cpu_offload()
    pipe.enable_attention_slicing(1)
    pipe.enable_sequential_cpu_offload()
    pipe.enable_vae_slicing()
    
    # Clear memory after loading
    gc.collect()
    torch.cuda.empty_cache()
    
except Exception as e:
    print(f"Error loading models: {e}")
    raise

# Extremely reduced parameters
MAX_SEED = 1000000
MAX_PIXEL_BUDGET = 128 * 128  # Extremely reduced from 256 * 256

def check_resources():
    if torch.cuda.is_available():
        memory_allocated = torch.cuda.memory_allocated(0)
        memory_reserved = torch.cuda.memory_reserved(0)
        if memory_allocated/memory_reserved > 0.7:  # 70% threshold
            gc.collect()
            torch.cuda.empty_cache()
    return True

def process_input(input_image, upscale_factor, **kwargs):
    input_image = input_image.convert('RGB')
    
    # Reduce image size more aggressively
    w, h = input_image.size
    max_size = int(np.sqrt(MAX_PIXEL_BUDGET))
    if w > max_size or h > max_size:
        if w > h:
            new_w = max_size
            new_h = int(h * max_size / w)
        else:
            new_h = max_size
            new_w = int(w * max_size / h)
        input_image = input_image.resize((new_w, new_h), Image.LANCZOS)
    
    w, h = input_image.size
    w = w - w % 8
    h = h - h % 8
    
    return input_image.resize((w, h)), w, h, True

@spaces.GPU
def infer(
    seed,
    randomize_seed,
    input_image,
    num_inference_steps,
    upscale_factor,
    controlnet_conditioning_scale,
    progress=gr.Progress(track_tqdm=True),
):
    try:
        gc.collect()
        torch.cuda.empty_cache()
        
        if randomize_seed:
            seed = random.randint(0, MAX_SEED)
            
        input_image, w, h, _ = process_input(input_image, upscale_factor)
        
        with torch.inference_mode():
            generator = torch.Generator().manual_seed(seed)
            image = pipe(
                prompt="",
                control_image=input_image,
                controlnet_conditioning_scale=controlnet_conditioning_scale,
                num_inference_steps=num_inference_steps,
                guidance_scale=2.0,  # Reduced from 3.5
                height=h,
                width=w,
                generator=generator,
            ).images[0]
            
        gc.collect()
        torch.cuda.empty_cache()
        
        return [input_image, image, seed]

    except Exception as e:
        gr.Error(f"An error occurred: {str(e)}")
        return None

with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
    with gr.Row():
        run_button = gr.Button(value="Run")

    with gr.Row():
        with gr.Column(scale=4):
            input_im = gr.Image(label="Input Image", type="pil")
        with gr.Column(scale=1):
            num_inference_steps = gr.Slider(
                label="Steps",
                minimum=1,
                maximum=20,  # Reduced from 30
                step=1,
                value=10,    # Reduced from 20
            )
            upscale_factor = gr.Slider(
                label="Scale",
                minimum=1,
                maximum=1,   # Fixed at 1
                step=1,
                value=1,
            )
            controlnet_conditioning_scale = gr.Slider(
                label="Control Scale",
                minimum=0.1,
                maximum=0.5,  # Reduced from 1.0
                step=0.1,
                value=0.3,   # Reduced from 0.5
            )
            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=42,
            )
            randomize_seed = gr.Checkbox(label="Random Seed", value=True)

    with gr.Row():
        result = ImageSlider(label="Result", type="pil", interactive=True)

    current_dir = os.path.dirname(os.path.abspath(__file__))
    
    examples = gr.Examples(
        examples=[
            [42, False, os.path.join(current_dir, "z1.webp"), 10, 1, 0.3],
            [42, False, os.path.join(current_dir, "z2.webp"), 10, 1, 0.3],
        ],
        inputs=[
            seed,
            randomize_seed,
            input_im,
            num_inference_steps,
            upscale_factor,
            controlnet_conditioning_scale,
        ],
        fn=infer,
        outputs=result,
        cache_examples=False,  # Disable caching
    )

    gr.on(
        [run_button.click],
        fn=infer,
        inputs=[
            seed,
            randomize_seed,
            input_im,
            num_inference_steps,
            upscale_factor,
            controlnet_conditioning_scale,
        ],
        outputs=result,
        show_api=False,
    )

# Launch with minimal resources
demo.queue(max_size=1).launch(
    share=False,
    debug=True,
    show_error=True,
    max_threads=1,
    enable_queue=True,
    cache_examples=False,
    quiet=True,
)