app.py

# float16 +32 
import os
import random
import numpy as np
import torch
import gradio as gr
from diffusers import StableDiffusionPipeline
import paramiko
from huggingface_hub import login

# Hugging Face Token
HF_TOKEN = os.getenv('HF_TOKEN', '').strip()
if not HF_TOKEN:
    raise ValueError("HUGGING_TOKEN is not set. Please set the token as an environment variable.")

# Hugging Face Login
login(token=HF_TOKEN)

# Konfiguration
STORAGE_DOMAIN = os.getenv('STORAGE_DOMAIN', '').strip()  # SFTP Server Domain
STORAGE_USER = os.getenv('STORAGE_USER', '').strip()  # SFTP User
STORAGE_PSWD = os.getenv('STORAGE_PSWD', '').strip()  # SFTP Passwort
STORAGE_PORT = int(os.getenv('STORAGE_PORT', '22').strip())  # SFTP Port
STORAGE_SECRET = os.getenv('STORAGE_SECRET', '').strip()  # Secret Token

# Modell-Konfiguration
available_models = {
    "sd3-medium": "stabilityai/stable-diffusion-3-medium-diffusers",
    "sd2-base": "stabilityai/stable-diffusion-2-1-base"
}

# SFTP-Funktion
def upload_to_sftp(local_file, remote_path):
    try:
        transport = paramiko.Transport((STORAGE_DOMAIN, STORAGE_PORT))
        transport.connect(username=STORAGE_USER, password=STORAGE_PSWD)
        sftp = paramiko.SFTPClient.from_transport(transport)
        sftp.put(local_file, remote_path)
        sftp.close()
        transport.close()
        print(f"File {local_file} successfully uploaded to {remote_path}")
        return True
    except Exception as e:
        print(f"Error during SFTP upload: {e}")
        return False

# Modell laden Funktion
def load_model(model_name, precision):
    device = "cuda" if torch.cuda.is_available() else "cpu"
    repo = available_models.get(model_name, available_models["sd3-medium"])
    
    try:
        # Wähle Präzision basierend auf Auswahl
        if precision == "float16":
            torch_dtype = torch.float16
        else:  # float32
            torch_dtype = torch.float32
            
        pipe = StableDiffusionPipeline.from_pretrained(
            repo, 
            torch_dtype=torch_dtype
        ).to(device)
        
        # Wenn auf CPU und Speicheroptimierung gewünscht
        if device == "cpu":
            pipe.enable_sequential_cpu_offload()
            
        return pipe
    except Exception as e:
        raise RuntimeError(f"Failed to load the model. Ensure the token has access to the repo. Error: {e}")

# Maximalwerte
MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1344

# Globale Pipe-Variable
pipe = None

# Inferenz-Funktion
def infer(prompt, width, height, guidance_scale, num_inference_steps, seed, randomize_seed, model_name, precision):
    global pipe
    
    # Prüfe, ob Modell neu geladen werden muss
    if pipe is None:
        pipe = load_model(model_name, precision)
    
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    
    generator = torch.manual_seed(seed)
    image = pipe(
        prompt, 
        guidance_scale=guidance_scale, 
        num_inference_steps=num_inference_steps, 
        width=width, 
        height=height, 
        generator=generator
    ).images[0]
    
    # Speichere Bild lokal
    local_file = f"/tmp/generated_image_{seed}.png"
    image.save(local_file)
    
    # Hochladen zu SFTP
    remote_path = f"/uploads/generated_image_{seed}.png"
    if upload_to_sftp(local_file, remote_path):
        os.remove(local_file)
        return f"Image uploaded to {remote_path}", seed
    else:
        return "Failed to upload image", seed

# Modell neu laden
def reload_model(model_name, precision):
    global pipe
    pipe = load_model(model_name, precision)
    return f"Model loaded: {model_name} with {precision} precision"

# Gradio-App
with gr.Blocks() as demo:
    gr.Markdown("### Stable Diffusion - Test App")
    
    with gr.Row():
        with gr.Column():
            # Modell Auswahl
            model_name = gr.Radio(
                choices=list(available_models.keys()),
                value="sd3-medium",
                label="Model"
            )
            
            # Präzision Auswahl
            precision = gr.Radio(
                choices=["float16", "float32"],
                value="float16",
                label="Precision"
            )
            
            reload_button = gr.Button("Load/Reload Model")
            model_status = gr.Textbox(label="Model Status")
            
            # Modell laden Button
            reload_button.click(
                reload_model,
                inputs=[model_name, precision],
                outputs=[model_status]
            )
    
    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(label="Prompt", placeholder="Enter your prompt here")
            width = gr.Slider(256, MAX_IMAGE_SIZE, step=64, value=512, label="Width")
            height = gr.Slider(256, MAX_IMAGE_SIZE, step=64, value=512, label="Height")
            guidance_scale = gr.Slider(0.0, 10.0, step=0.1, value=7.5, label="Guidance Scale")
            num_inference_steps = gr.Slider(1, 50, step=1, value=25, label="Inference Steps")
            seed = gr.Number(value=42, label="Seed")
            randomize_seed = gr.Checkbox(value=False, label="Randomize Seed")
            generate_button = gr.Button("Generate Image")
            output = gr.Text(label="Output")
    
    generate_button.click(
        infer,
        inputs=[
            prompt, width, height, guidance_scale, 
            num_inference_steps, seed, randomize_seed,
            model_name, precision
        ],
        outputs=[output, seed]
    )

demo.launch()