app_quant_latent.py

import torch
import spaces
import gradio as gr
import sys
import platform
import diffusers
import transformers
import psutil
import os
import time

from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
from diffusers import ZImagePipeline, AutoModel
from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
latent_history = []

# ============================================================
# LOGGING BUFFER
# ============================================================
LOGS = ""
def log(msg):
    global LOGS
    print(msg)
    LOGS += msg + "\n"
    return msg


# ============================================================
# SYSTEM METRICS — LIVE GPU + CPU MONITORING
# ============================================================
def log_system_stats(tag=""):
    try:
        log(f"\n===== 🔥 SYSTEM STATS {tag} =====")

        # ============= GPU STATS =============
        if torch.cuda.is_available():
            allocated = torch.cuda.memory_allocated(0) / 1e9
            reserved = torch.cuda.memory_reserved(0) / 1e9
            total = torch.cuda.get_device_properties(0).total_memory / 1e9
            free = total - allocated

            log(f"💠 GPU Total     : {total:.2f} GB")
            log(f"💠 GPU Allocated : {allocated:.2f} GB")
            log(f"💠 GPU Reserved  : {reserved:.2f} GB")
            log(f"💠 GPU Free      : {free:.2f} GB")

        # ============= CPU STATS ============
        cpu = psutil.cpu_percent()
        ram_used = psutil.virtual_memory().used / 1e9
        ram_total = psutil.virtual_memory().total / 1e9

        log(f"🧠 CPU Usage     : {cpu}%")
        log(f"🧠 RAM Used      : {ram_used:.2f} GB / {ram_total:.2f} GB")

    except Exception as e:
        log(f"⚠️ Failed to log system stats: {e}")


# ============================================================
# ENVIRONMENT INFO
# ============================================================
log("===================================================")
log("🔍 Z-IMAGE-TURBO DEBUGGING + LIVE METRIC LOGGER")
log("===================================================\n")

log(f"📌 PYTHON VERSION       : {sys.version.replace(chr(10),' ')}")
log(f"📌 PLATFORM             : {platform.platform()}")
log(f"📌 TORCH VERSION        : {torch.__version__}")
log(f"📌 TRANSFORMERS VERSION : {transformers.__version__}")
log(f"📌 DIFFUSERS VERSION    : {diffusers.__version__}")
log(f"📌 CUDA AVAILABLE       : {torch.cuda.is_available()}")

log_system_stats("AT STARTUP")

if not torch.cuda.is_available():
    raise RuntimeError("❌ CUDA Required")

device = "cuda"
gpu_id = 0

# ============================================================
# MODEL SETTINGS
# ============================================================
model_cache = "./weights/"
model_id = "Tongyi-MAI/Z-Image-Turbo"
torch_dtype = torch.bfloat16
USE_CPU_OFFLOAD = False

log("\n===================================================")
log("🧠 MODEL CONFIGURATION")
log("===================================================")
log(f"Model ID              : {model_id}")
log(f"Model Cache Directory : {model_cache}")
log(f"torch_dtype           : {torch_dtype}")
log(f"USE_CPU_OFFLOAD       : {USE_CPU_OFFLOAD}")

log_system_stats("BEFORE TRANSFORMER LOAD")


# ============================================================
# FUNCTION TO CONVERT LATENTS TO IMAGE
# ============================================================
def latent_to_image(latent):
    try:
        img_tensor = pipe.vae.decode(latent)
        img_tensor = (img_tensor / 2 + 0.5).clamp(0, 1)
        pil_img = T.ToPILImage()(img_tensor[0])
        return pil_img
    except Exception as e:
        log(f"⚠️ Failed to decode latent: {e}")
        return None



# ============================================================
# SAFE TRANSFORMER INSPECTION
# ============================================================
def inspect_transformer(model, name):
    log(f"\n🔍 Inspecting {name}")
    try:
        candidates = ["transformer_blocks", "blocks", "layers", "encoder", "model"]
        blocks = None

        for attr in candidates:
            if hasattr(model, attr):
                blocks = getattr(model, attr)
                break

        if blocks is None:
            log(f"⚠️ No block structure found in {name}")
            return

        if hasattr(blocks, "__len__"):
            log(f"Total Blocks = {len(blocks)}")
        else:
            log("⚠️ Blocks exist but are not iterable")

        for i in range(min(10, len(blocks) if hasattr(blocks, "__len__") else 0)):
            log(f"Block {i} = {blocks[i].__class__.__name__}")

    except Exception as e:
        log(f"⚠️ Transformer inspect error: {e}")


# ============================================================
# LOAD TRANSFORMER — WITH LIVE STATS
# ============================================================
log("\n===================================================")
log("🔧 LOADING TRANSFORMER BLOCK")
log("===================================================")

log("📌 Logging memory before load:")
log_system_stats("START TRANSFORMER LOAD")

try:
    quant_cfg = DiffusersBitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch_dtype,
        bnb_4bit_use_double_quant=True,
    )

    transformer = AutoModel.from_pretrained(
        model_id,
        cache_dir=model_cache,
        subfolder="transformer",
        quantization_config=quant_cfg,
        torch_dtype=torch_dtype,
        device_map=device,
    )
    log("✅ Transformer loaded successfully.")

except Exception as e:
    log(f"❌ Transformer load failed: {e}")
    transformer = None

log_system_stats("AFTER TRANSFORMER LOAD")

if transformer:
    inspect_transformer(transformer, "Transformer")


# ============================================================
# LOAD TEXT ENCODER
# ============================================================
log("\n===================================================")
log("🔧 LOADING TEXT ENCODER")
log("===================================================")

log_system_stats("START TEXT ENCODER LOAD")

try:
    quant_cfg2 = TransformersBitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch_dtype,
        bnb_4bit_use_double_quant=True,
    )

    text_encoder = AutoModel.from_pretrained(
        model_id,
        cache_dir=model_cache,
        subfolder="text_encoder",
        quantization_config=quant_cfg2,
        torch_dtype=torch_dtype,
        device_map=device,
    )
    log("✅ Text encoder loaded successfully.")

except Exception as e:
    log(f"❌ Text encoder load failed: {e}")
    text_encoder = None

log_system_stats("AFTER TEXT ENCODER LOAD")

if text_encoder:
    inspect_transformer(text_encoder, "Text Encoder")


# ============================================================
# BUILD PIPELINE
# ============================================================
log("\n===================================================")
log("🔧 BUILDING PIPELINE")
log("===================================================")

log_system_stats("START PIPELINE BUILD")

try:
    pipe = ZImagePipeline.from_pretrained(
        model_id,
        transformer=transformer,
        text_encoder=text_encoder,
        torch_dtype=torch_dtype,
        attn_implementation="kernels-community/vllm-flash-attn3",
    )
    pipe.to(device)
    log("✅ Pipeline built successfully.")

except Exception as e:
    log(f"❌ Pipeline build failed: {e}")
    pipe = None

log_system_stats("AFTER PIPELINE BUILD")





from PIL import Image
import torch

# --------------------------
# Helper: Safe latent extractor
# --------------------------
def safe_get_latents(pipe, height, width, generator, device, LOGS):
    """
    Attempts multiple ways to get latents.
    Returns a valid tensor even if pipeline hides UNet.
    """
    # Try official prepare_latents
    try:
        if hasattr(pipe, "unet") and hasattr(pipe.unet, "in_channels"):
            num_channels = pipe.unet.in_channels
            latents = pipe.prepare_latents(
                batch_size=1,
                num_channels=num_channels,
                height=height,
                width=width,
                dtype=torch.float32,
                device=device,
                generator=generator
            )
            LOGS.append("✅ Latents extracted using official prepare_latents.")
            return latents
    except Exception as e:
        LOGS.append(f"⚠️ Official latent extraction failed: {e}")

    # Try hidden internal attribute
    try:
        if hasattr(pipe, "_default_latents"):
            LOGS.append("⚠️ Using hidden _default_latents.")
            return pipe._default_latents
    except:
        pass

    # Fallback: raw Gaussian tensor
    try:
        LOGS.append("⚠️ Using raw Gaussian latents fallback.")
        return torch.randn(
            (1, 4, height // 8, width // 8),
            generator=generator,
            device=device,
            dtype=torch.float32
        )
    except Exception as e:
        LOGS.append(f"⚠️ Gaussian fallback failed: {e}")

    LOGS.append("❗ Using CPU hard fallback latents.")
    return torch.randn((1, 4, height // 8, width // 8))


# --------------------------
# Main generation function
# --------------------------
@spaces.GPU
def generate_image(prompt, height, width, steps, seed, guidance_scale=0.0):
    LOGS = []
    latents = None
    image = None
    gallery = []

    # placeholder image if all fails
    placeholder = Image.new("RGB", (width, height), color=(255, 255, 255))
    print(prompt)

    try:
        generator = torch.Generator(device).manual_seed(int(seed))

        # -------------------------------
        # Try advanced latent extraction
        # -------------------------------
        try:
            latents = safe_get_latents(pipe, height, width, generator, device, LOGS)

            output = pipe(
                prompt=prompt,
                height=height,
                width=width,
                num_inference_steps=steps,
                guidance_scale=guidance_scale,
                generator=generator,
                latents=latents
            )

            image = output.images[0]
            gallery = [image]
            LOGS.append("✅ Advanced latent pipeline succeeded.")

        except Exception as e:
            LOGS.append(f"⚠️ Latent mode failed: {e}")
            LOGS.append("🔁 Switching to standard pipeline...")

            try:
                output = pipe(
                    prompt=prompt,
                    height=height,
                    width=width,
                    num_inference_steps=steps,
                    guidance_scale=guidance_scale,
                    generator=generator,
                )
                image = output.images[0]
                gallery = [image]
                LOGS.append("✅ Standard pipeline succeeded.")

            except Exception as e2:
                LOGS.append(f"❌ Standard pipeline failed: {e2}")
                image = placeholder
                gallery = [image]

        return image, gallery, LOGS

    except Exception as e:
        LOGS.append(f"❌ Total failure: {e}")
        return placeholder, [placeholder], LOGS
@spaces.GPU
def generate_image_backup(prompt, height, width, steps, seed, guidance_scale=0.0, return_latents=False):
    """
    Robust dual pipeline:
    - Advanced latent generation first
    - Fallback to standard pipeline if latent fails
    - Always returns final image
    - Returns gallery (latents or final image) and logs
    """

    LOGS = []
    image = None
    latents = None
    gallery = []

    # Keep a placeholder original image (white) in case everything fails
    original_image = Image.new("RGB", (width, height), color=(255, 255, 255))

    try:
        generator = torch.Generator(device).manual_seed(int(seed))

        # -------------------------------
        # Try advanced latent generation
        # -------------------------------
        try:
            batch_size = 1
            num_channels_latents = getattr(pipe.unet, "in_channels", None)
            if num_channels_latents is None:
                raise AttributeError("pipe.unet.in_channels not found, fallback to standard pipeline")

            latents = pipe.prepare_latents(
                batch_size=batch_size,
                num_channels=num_channels_latents,
                height=height,
                width=width,
                dtype=torch.float32,
                device=device,
                generator=generator
            )
            LOGS.append(f"✅ Latents prepared: {latents.shape}")

            output = pipe(
                prompt=prompt,
                height=height,
                width=width,
                num_inference_steps=steps,
                guidance_scale=guidance_scale,
                generator=generator,
                latents=latents
            )
            image = output.images[0]
            gallery = [image] if image else []

            LOGS.append("✅ Advanced latent generation succeeded.")

        # -------------------------------
        # Fallback to standard pipeline
        # -------------------------------
        except Exception as e_latent:
            LOGS.append(f"⚠️ Advanced latent generation failed: {e_latent}")
            LOGS.append("🔁 Falling back to standard pipeline...")

            try:
                output = pipe(
                    prompt=prompt,
                    height=height,
                    width=width,
                    num_inference_steps=steps,
                    guidance_scale=guidance_scale,
                    generator=generator
                )
                image = output.images[0]
                gallery = [image] if image else []
                LOGS.append("✅ Standard pipeline generation succeeded.")
            except Exception as e_standard:
                LOGS.append(f"❌ Standard pipeline generation failed: {e_standard}")
                image = original_image  # Always return some image
                gallery = [image]

        # -------------------------------
        # Return all 3 outputs
        # -------------------------------
        return image, gallery, LOGS

    except Exception as e:
        LOGS.append(f"❌ Inference failed entirely: {e}")
        return original_image, [original_image], LOGS

# ============================================================
# UI
# ============================================================

with gr.Blocks(title="Z-Image- experiment - dont run")as demo:
  gr.Markdown("# **🚀 do not run Z-Image-Turbo — Final Image & Latents**")


  with gr.Row():
    with gr.Column(scale=1):
        prompt = gr.Textbox(label="Prompt", value="boat in Ocean")
        height = gr.Slider(256, 2048, value=1024, step=8, label="Height")
        width = gr.Slider(256, 2048, value=1024, step=8, label="Width")
        steps = gr.Slider(1, 50, value=20, step=1, label="Inference Steps")
        seed = gr.Number(value=42, label="Seed")
        run_btn = gr.Button("Generate Image")

    with gr.Column(scale=1):
        final_image = gr.Image(label="Final Image")
        latent_gallery = gr.Gallery(
           label="Latent Steps",
                columns=4,
              height=256,
             preview=True
              )

        logs_box = gr.Textbox(label="Logs", lines=15)

    run_btn.click(
      generate_image,
      inputs=[prompt, height, width, steps, seed],
      outputs=[final_image, latent_gallery, logs_box]
     )



demo.launch()