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

# ============================================================
# 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,
    )
    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")


# ============================================================
# INFERENCE
# ============================================================
@spaces.GPU
def generate_image(prompt, height, width, steps, seed):
global LOGS
LOGS = ""  # reset logs


log("===================================================")
log("🎨 RUNNING INFERENCE")
log("===================================================")
log_system_stats("BEFORE INFERENCE")

try:
    generator = torch.Generator(device).manual_seed(seed)
    latent_history = []

    # Callback to save latents and GPU info
    def save_latents(step, timestep, latents):
        latent_history.append(latents.detach().clone())
        gpu_mem = torch.cuda.memory_allocated(0)/1e9
        log(f"Step {step} - GPU Memory Used: {gpu_mem:.2f} GB")

    # Step 3: Loop over pipeline for step-wise generation
    for step, img in pipe(
        prompt=prompt,
        height=height,
        width=width,
        num_inference_steps=steps,
        guidance_scale=0.0,
        generator=generator,
        callback=save_latents,
        callback_steps=1
    ).iter():
        # Optionally: yield intermediate images or just store latents
        current_latent = latent_history[-1] if latent_history else None
        # You can process current_latent here if needed

    log("✅ Inference finished.")
    log_system_stats("AFTER INFERENCE")

    # Return final image + logs
    return img, LOGS

except Exception as e:
    log(f"❌ Inference error: {e}")
    return None, LOGS



@spaces.GPU
def generate_image(prompt, height, width, steps, seed):
global LOGS
LOGS = ""  # reset logs


log("===================================================")
log("🎨 RUNNING INFERENCE")
log("===================================================")
log_system_stats("BEFORE INFERENCE")

try:
    generator = torch.Generator(device).manual_seed(seed)
    latent_history = []

    # Callback to save latents and GPU info
    def save_latents(step, timestep, latents):
        latent_history.append(latents.detach().clone())
        gpu_mem = torch.cuda.memory_allocated(0)/1e9
        log(f"Step {step} - GPU Memory Used: {gpu_mem:.2f} GB")

    # Step-wise loop just for latent capture
    for step, _ in pipe(
        prompt=prompt,
        height=height,
        width=width,
        num_inference_steps=steps,
        guidance_scale=0.0,
        generator=generator,
        callback=save_latents,
        callback_steps=1
    ).iter():
        pass  # only capturing latents, ignoring intermediate images

    # Original final image generation
    output = pipe(
        prompt=prompt,
        height=height,
        width=width,
        num_inference_steps=steps,
        guidance_scale=0.0,
        generator=generator,
    )

    log("✅ Inference finished.")
    log_system_stats("AFTER INFERENCE")

    return output.images[0], latent_history, LOGS

except Exception as e:
    log(f"❌ Inference error: {e}")
    return None, None, LOGS


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

with gr.Blocks(title="Z-Image-Turbo Generator") as demo:
gr.Markdown("# **🚀 Z-Image-Turbo — Final Image & Latents**")

with gr.Row():
    with gr.Column(scale=1):
        prompt = gr.Textbox(label="Prompt", value="Realistic mid-aged male image")
        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").style(grid=[4], height="256px")
        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()