app.py

import torch
from diffusers import UniPCMultistepScheduler
from diffusers import WanPipeline, AutoencoderKLWan
from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
from huggingface_hub import hf_hub_download
from PIL import Image
import numpy as np
import gradio as gr
import spaces
import gc

# --- INITIAL SETUP ---
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")

model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
print("Loading VAE...")
vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)

print("Loading WanPipeline in bfloat16...")
# This will use ZeroGPU/accelerate with meta devices
pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)

flow_shift = 1.0
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)

# Move the base pipeline to the GPU. ZeroGPU will manage this.
print("Moving pipeline to device (ZeroGPU will handle offloading)...")
pipe.to(device)

# --- LORA SETUP ---
CAUSVID_LORA_REPO = "Kijai/WanVideo_comfy"
CAUSVID_LORA_FILENAME = "Wan21_CausVid_14B_T2V_lora_rank32.safetensors"
BASE_LORA_NAME = "causvid_lora"
CUSTOM_LORA_NAME = "custom_lora"

print("Downloading base LoRA...")
try:
    causvid_path = hf_hub_download(repo_id=CAUSVID_LORA_REPO, filename=CAUSVID_LORA_FILENAME)
    print("✅ Base LoRA downloaded.")
except Exception as e:
    causvid_path = None
    print(f"⚠️ Could not download base LoRA: {e}")

print("Initialization complete. Gradio is starting...")

@spaces.GPU()
def generate(prompt, negative_prompt, width=1024, height=1024, num_inference_steps=30, lora_id=None, progress=gr.Progress(track_tqdm=True)):
    
    # --- DYNAMIC LORA MANAGEMENT FOR EACH RUN ---
    active_adapters = []
    adapter_weights = []

    # 1. Load the Base LoRA directly onto the correct device
    if causvid_path:
        try:
            print(f"Loading base LoRA '{BASE_LORA_NAME}'...")
            pipe.load_lora_weights(causvid_path, adapter_name=BASE_LORA_NAME, device_map={"":device})
            active_adapters.append(BASE_LORA_NAME)
            adapter_weights.append(1.0)
            print("✅ Base LoRA loaded to device.")
        except Exception as e:
            print(f"⚠️ Failed to load base LoRA: {e}")

    # 2. Load the Custom LoRA if provided, also directly to the device
    clean_lora_id = lora_id.strip() if lora_id else ""
    if clean_lora_id:
        try:
            print(f"Loading custom LoRA '{CUSTOM_LORA_NAME}' from '{clean_lora_id}'...")
            pipe.load_lora_weights(clean_lora_id, adapter_name=CUSTOM_LORA_NAME, device_map={"":device})
            active_adapters.append(CUSTOM_LORA_NAME)
            adapter_weights.append(1.0)
            print("✅ Custom LoRA loaded to device.")
        except Exception as e:
            print(f"⚠️ Failed to load custom LoRA '{clean_lora_id}': {e}")
            if CUSTOM_LORA_NAME in getattr(pipe.transformer, 'peft_config', {}):
                del pipe.transformer.peft_config[CUSTOM_LORA_NAME]

    # 3. Activate the successfully loaded adapters
    if active_adapters:
        print(f"Activating adapters: {active_adapters} with weights: {adapter_weights}")
        pipe.set_adapters(active_adapters, adapter_weights)
        pipe.transformer.to(device)  # Explicitly move transformer to GPU after setting adapters
    else:
        # Ensure LoRA is disabled if no adapters were loaded
        pipe.disable_lora()
    
    apply_cache_on_pipe(pipe)
    
    try:
        output = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            height=height,
            width=width,
            num_frames=1,
            num_inference_steps=num_inference_steps,
            guidance_scale=1.0,
        )
        image = output.frames[0][0]
        image = (image * 255).astype(np.uint8)
        return Image.fromarray(image)
    finally:
        # --- PROPER CLEANUP ---
        print("Unloading all LoRAs to ensure a clean state...")
        pipe.unload_lora_weights()
        gc.collect()  # Force garbage collection
        torch.cuda.empty_cache()  # Clear CUDA cache
        print("✅ LoRAs unloaded and memory cleaned.")

iface = gr.Interface(
    fn=generate,
    inputs=[
        gr.Textbox(label="Input prompt"),
        gr.Textbox(label="Negative prompt", value="Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"),
        gr.Slider(label="Width", minimum=480, maximum=1280, step=16, value=1024),
        gr.Slider(label="Height", minimum=480, maximum=1280, step=16, value=1024),
        gr.Slider(minimum=1, maximum=80, step=1, label="Inference Steps", value=10),
        gr.Textbox(label="LoRA ID (Optional, loads dynamically)"),
    ],
    outputs=gr.Image(label="output"),
)

iface.launch()