app.py

import os
import sys
import time
import json
import gc
import random
import torch
import gradio as gr
import requests
from threading import Lock, Event, Thread
from contextlib import contextmanager
from urllib.parse import urlparse

from huggingface_hub import hf_hub_download, hf_hub_url
from huggingface_hub.utils import RepositoryNotFoundError

# ===== LOGGING =====
LOG_BUFFER = []
LOG_LOCK = Lock()
def log(msg):
    with LOG_LOCK:
        t = time.strftime("%H:%M:%S")
        entry = f"{t} | {msg}"
        LOG_BUFFER.append(entry)
        if len(LOG_BUFFER) > 500:
            LOG_BUFFER.pop(0)
    return "\n".join(LOG_BUFFER)

# ===== ENV SETUP =====
os.environ["HF_HUB_DISABLE_TELEMETRY"] = "1"
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
os.environ["TRANSFORMERS_CACHE"] = "./hf_cache"
os.environ["HF_DATASETS_CACHE"] = "./hf_cache"
os.makedirs("./hf_cache", exist_ok=True)

torch.set_grad_enabled(False)
torch.set_num_threads(min(8, os.cpu_count() or 1))
torch.set_float32_matmul_precision("medium")

DEVICE = "cpu"
DTYPE = torch.float32

try:
    from diffusers import ZImagePipeline, GGUFQuantizationConfig, ZImageTransformer2DModel
    log("Loaded diffusers modules")
except ImportError as e:
    log(f"Import error: {e}")
    sys.exit(1)

# ===== DOWNLOAD CONTEXT =====
interrupt_event = Event()
pipe_cache = {}
download_lock = Lock()

# ===== MODEL LIST =====
MODEL_SPECS = {
    "Turbo Full": "Tongyi-MAI/Z-Image-Turbo",
    "Turbo Q2_K GGUF": "unsloth/Z-Image-Turbo-GGUF"
}

# ===== DOWNLOAD HELPERS =====
def list_repo_files(repo_id):
    """
    Returns a list of (filename, size) tuples by doing a dry run
    (no actual data downloaded).
    """
    try:
        infos = hf_hub_download(repo_id, dry_run=True)
        return [(info.rfilename, info.size_in_bytes) for info in infos]
    except Exception as e:
        log(f"List failed: {e}")
        return []

def download_file_chunked(repo_id, filename, target_dir, progress_updater):
    """
    Download a single file by streaming signed URL chunks.
    Supports resume by checking existing file size.
    """
    local_path = os.path.join(target_dir, filename)
    tmp_path = local_path + ".part"

    os.makedirs(os.path.dirname(local_path), exist_ok=True)
    already = 0
    if os.path.exists(tmp_path):
        already = os.path.getsize(tmp_path)

    # Get a fresh signed URL from HF for that file
    try:
        url = hf_hub_url(repo_id, filename)
    except RepositoryNotFoundError:
        # fallback to normal
        url = hf_hub_download(repo_id, filename=filename)

    headers = {}
    if already > 0:
        headers["Range"] = f"bytes={already}-"

    with requests.get(url, headers=headers, stream=True, timeout=10) as r:
        total = int(r.headers.get("Content-Length", 0)) + already
        with open(tmp_path, "ab") as f:
            downloaded = already
            for chunk in r.iter_content(chunk_size=1024*256):
                if interrupt_event.is_set():
                    return False
                if not chunk:
                    continue
                f.write(chunk)
                downloaded += len(chunk)
                progress_updater(downloaded / total)

    os.rename(tmp_path, local_path)
    return True

def parallel_download_repo(repo_id, progress: gr.Progress):
    """
    Download all files in the repo in parallel with per-file progress.
    """
    base_dir = os.path.join("./hf_cache", repo_id.replace("/", "_"))
    files = list_repo_files(repo_id)
    if not files:
        progress(1.0, desc="No files to download")
        return

    total_bytes = sum(sz for _, sz in files)
    downloaded_bytes = 0

    def file_thread(filename, size):
        nonlocal downloaded_bytes
        success = download_file_chunked(
            repo_id, filename, base_dir,
            lambda frac: progress((downloaded_bytes + frac * size) / total_bytes,
                                  desc=f"{filename} {frac*100:.1f}%")
        )
        if success:
            with download_lock:
                downloaded_bytes += size

    threads = []
    for fname, size in files:
        if interrupt_event.is_set():
            break
        # skip if fully cached already
        local_full = os.path.join(base_dir, fname)
        if os.path.exists(local_full) and os.path.getsize(local_full) == size:
            downloaded_bytes += size
            continue
        t = Thread(target=file_thread, args=(fname, size))
        t.start()
        threads.append(t)

    for t in threads:
        t.join()

# ===== PIPELINE LOADER =====
def load_pipeline(model_key):
    """
    Load the HF pipeline, using quantized GGUF if selected.
    """
    if model_key in pipe_cache:
        return pipe_cache[model_key]

    repo = MODEL_SPECS[model_key]
    repo_cache = os.path.join("./hf_cache", repo.replace("/", "_"))

    # ensure cache
    if not os.path.isdir(repo_cache) or not os.listdir(repo_cache):
        raise gr.Error("Model not downloaded; press Preload first")

    # load model
    if "GGUF" in model_key:
        # pick .gguf file
        files = [f for f in os.listdir(repo_cache) if f.endswith(".gguf")]
        if not files:
            raise gr.Error("Quantized file not found")
        gguf = os.path.join(repo_cache, files[0])
        transformer = ZImageTransformer2DModel.from_single_file(
            gguf,
            quantization_config=GGUFQuantizationConfig(compute_dtype=DTYPE),
            torch_dtype=DTYPE
        )
        pipe = ZImagePipeline.from_pretrained(
            "Tongyi-MAI/Z-Image-Turbo",
            transformer=transformer,
            torch_dtype=DTYPE,
            cache_dir="./hf_cache"
        )
    else:
        pipe = ZImagePipeline.from_pretrained(repo_cache, torch_dtype=DTYPE, local_files_only=True)

    pipe.to(DEVICE)
    pipe.vae.eval()
    pipe.text_encoder.eval()
    pipe.transformer.eval()
    pipe_cache[model_key] = pipe
    return pipe

@contextmanager
def managed_memory():
    try:
        yield
    finally:
        gc.collect()
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()

# ===== GENERATION =====
@torch.inference_mode()
@torch.no_grad()
def generate(prompt, quality_mode, seed, model_key):
    if not prompt.strip():
        raise gr.Error("Prompt cannot be empty")

    PRESETS = {
        "ultra_fast": (1,256),
        "fast": (1,256),
        "balanced": (2,256),
        "quality": (4,384),
        "ultra_quality": (4,512),
    }
    steps, size = PRESETS.get(quality_mode, (1,256))
    width = height = size

    seed = int(seed) if seed>=0 else random.randint(0,2**31-1)
    log(f"Gen: {prompt[:30]} | {quality_mode} | {model_key} | seed={seed}")

    with managed_memory():
        pipe = load_pipeline(model_key)
        gen = torch.Generator("cpu").manual_seed(seed)
        previews=[]
        start = time.time()

        def cb(ppl, step, timestep, cbk):
            if interrupt_event.is_set():
                ppl._interrupt=True
            if step % 2 == 0:
                try:
                    previews.append(ppl.image_from_latents(cbk["latents"]))
                except:
                    pass
            return cbk

        result = pipe(
            prompt=prompt,
            negative_prompt=None,
            width=width,
            height=height,
            num_inference_steps=steps,
            guidance_scale=0.0,
            generator=gen,
            callback_on_step_end=cb,
            callback_on_step_end_tensor_inputs=["latents"],
            output_type="pil"
        )
        final = result.images[0]
        previews.append(final)
        log(f"Generated in {time.time()-start:.1f}s")
    return final, seed, previews

# ===== GRADIO UI =====
with gr.Blocks(title="Z‑Image Turbo CPU Downloader + UI A‑Progress") as demo:
    gr.Markdown("## True parallel download UI + chunked progress")

    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(label="Prompt", lines=4)
            quality = gr.Radio(["ultra_fast","fast","balanced","quality","ultra_quality"], value="fast")
            seed = gr.Number(value=-1, precision=0, label="Seed")
            model_select = gr.Dropdown(list(MODEL_SPECS.keys()), value=list(MODEL_SPECS.keys())[0], label="Model")
            preload = gr.Button("PRELOAD MODELS")
            gen_btn = gr.Button("GENERATE")
            stop_btn = gr.Button("STOP")
        with gr.Column():
            out_image = gr.Image(label="Final")
            used_seed = gr.Number(label="Seed Used")
            preview = gr.Gallery(label="Preview Frames")
            logs = gr.Textbox(label="Logs", lines=25)

    def do_preload(progress=gr.Progress()):
        interrupt_event.clear()
        for key, repo in MODEL_SPECS.items():
            parallel_download_repo(repo, progress)
        return log("📦 Preload finished")

    def do_gen(prompt, quality, seed, model_key):
        interrupt_event.clear()
        img, used, previews = generate(prompt, quality, seed, model_key)
        return img, used, previews, log("🧠 Generation done")

    def do_stop():
        interrupt_event.set()
        return log("🔴 Interrupt set")

    preload.click(do_preload, outputs=logs)
    gen_btn.click(do_gen, inputs=[prompt,quality,seed,model_select],
                  outputs=[out_image,used_seed,preview,logs])
    stop_btn.click(do_stop, outputs=logs)

demo.queue()
demo.launch(server_name="0.0.0.0", server_port=7860)