app.py

import torch
import gc
import time
import random
import os
import hashlib
import shutil
import psutil
from diffusers import DiffusionPipeline
import gradio as gr
from PIL import Image, PngImagePlugin

MODEL_ID = "tensorart/stable-diffusion-3.5-medium-turbo"
CACHE_DIR = "./hf_cache"
OUTPUT_DIR = "./outputs"
MAX_CACHE_SIZE_GB = 2

os.makedirs(CACHE_DIR, exist_ok=True)
os.makedirs(OUTPUT_DIR, exist_ok=True)

device = "cpu"
dtype = torch.float32

pipe = DiffusionPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=dtype,
    safety_checker=None,
    cache_dir=CACHE_DIR,
    low_cpu_mem_usage=True
)

pipe.to(device)
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()
pipe.set_progress_bar_config(disable=True)

def warmup():
    with torch.inference_mode():
        pipe(
            prompt="warmup",
            num_inference_steps=1,
            guidance_scale=0.0,
            width=256,
            height=256
        )
    gc.collect()

warmup()

def get_ram_usage():
    return round(psutil.virtual_memory().used / (1024 ** 3), 2)

def prune_cache():
    total_size = 0
    files = []

    for f in os.listdir(OUTPUT_DIR):
        path = os.path.join(OUTPUT_DIR, f)
        if os.path.isfile(path):
            size = os.path.getsize(path)
            total_size += size
            files.append((path, size, os.path.getmtime(path)))

    max_bytes = MAX_CACHE_SIZE_GB * 1024 * 1024 * 1024

    if total_size <= max_bytes:
        return

    files.sort(key=lambda x: x[2])

    for path, size, _ in files:
        os.remove(path)
        total_size -= size
        if total_size <= max_bytes:
            break

def build_cache_key(prompt, negative_prompt, steps, guidance, width, height, seed):
    raw = f"{prompt}|{negative_prompt}|{steps}|{guidance}|{width}|{height}|{seed}"
    return hashlib.sha256(raw.encode()).hexdigest()

def generate(prompt, negative_prompt, steps, guidance, width, height, seed):
    start_time = time.time()

    if not prompt.strip():
        return None, "Prompt cannot be empty."

    width = max(256, min(int(width), 768))
    height = max(256, min(int(height), 768))
    steps = max(1, min(int(steps), 8))
    guidance = max(0.0, min(float(guidance), 7.5))

    if seed == -1:
        seed = random.randint(0, 2**32 - 1)

    cache_key = build_cache_key(prompt, negative_prompt, steps, guidance, width, height, seed)
    cache_path = os.path.join(OUTPUT_DIR, f"{cache_key}.png")

    if os.path.exists(cache_path):
        image = Image.open(cache_path)
        duration = round(time.time() - start_time, 2)
        ram = get_ram_usage()
        return image, f"Loaded from cache | Seed: {seed} | Time: {duration}s | RAM: {ram}GB"

    generator = torch.Generator(device=device).manual_seed(seed)

    try:
        with torch.inference_mode():
            result = pipe(
                prompt=prompt,
                negative_prompt=negative_prompt,
                num_inference_steps=steps,
                guidance_scale=guidance,
                width=width,
                height=height,
                generator=generator
            )

        image = result.images[0]

        metadata = PngImagePlugin.PngInfo()
        metadata.add_text("prompt", prompt)
        metadata.add_text("negative_prompt", negative_prompt)
        metadata.add_text("steps", str(steps))
        metadata.add_text("guidance", str(guidance))
        metadata.add_text("seed", str(seed))

        image.save(cache_path, pnginfo=metadata)

        prune_cache()

        duration = round(time.time() - start_time, 2)
        ram = get_ram_usage()

        gc.collect()

        return image, f"Generated | Seed: {seed} | Time: {duration}s | RAM: {ram}GB"

    except Exception as e:
        gc.collect()
        return None, f"Error: {str(e)}"

with gr.Blocks(title="SD 3.5 Turbo - Ultimate CPU Mode") as demo:

    gr.Markdown("## Stable Diffusion 3.5 Medium Turbo - Ultimate CPU Edition")

    with gr.Row():
        prompt = gr.Textbox(label="Prompt")
    negative_prompt = gr.Textbox(label="Negative Prompt")

    with gr.Row():
        steps = gr.Slider(1, 8, value=4, step=1, label="Steps")
        guidance = gr.Slider(0.0, 7.5, value=0.0, step=0.5, label="Guidance")

    with gr.Row():
        width = gr.Slider(256, 768, value=512, step=64, label="Width")
        height = gr.Slider(256, 768, value=512, step=64, label="Height")

    seed = gr.Number(value=-1, label="Seed (-1 random)")

    generate_btn = gr.Button("Generate")

    output_image = gr.Image(type="pil")
    status = gr.Textbox(label="Status")

    generate_btn.click(
        generate,
        inputs=[prompt, negative_prompt, steps, guidance, width, height, seed],
        outputs=[output_image, status]
    )

demo.queue(max_size=10, concurrency_count=1, status_update_rate=1)
demo.launch()