app.py

import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import spaces
import gc
import tempfile
import random
import gradio as gr
from PIL import Image

# Core Diffusers imports
from diffusers import WanImageToVideoPipeline
from diffusers.utils import export_to_video

# =========================================================
# 1. ARCHITECTURAL UPGRADES (GQA & MoE)
# =========================================================

# The custom classes (GQAAttention, SparseMoEFFN) stay the same as 
# they are architectural modifications to the base model's logic.

class GQAAttention(nn.Module):
    def __init__(self, original_attn):
        super().__init__()
        self.num_heads = original_attn.num_heads
        self.head_dim = original_attn.head_dim
        self.num_kv_heads = max(1, self.num_heads // 4) 
        self.q_proj = original_attn.q_proj
        self.k_proj = original_attn.k_proj
        self.v_proj = original_attn.v_proj
        self.o_proj = original_attn.o_proj

    def forward(self, x, freqs_cis=None):
        batch, seq_len, _ = x.shape
        q = self.q_proj(x).view(batch, seq_len, self.num_heads, self.head_dim)
        k = self.k_proj(x).view(batch, seq_len, self.num_kv_heads, self.head_dim)
        v = self.v_proj(x).view(batch, seq_len, self.num_kv_heads, self.head_dim)
        k = k.repeat_interleave(self.num_heads // self.num_kv_heads, dim=2)
        v = v.repeat_interleave(self.num_heads // self.num_kv_heads, dim=2)
        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
        attn_output = F.scaled_dot_product_attention(q, k, v)
        attn_output = attn_output.transpose(1, 2).reshape(batch, seq_len, -1)
        return self.o_proj(attn_output)

class SparseMoEFFN(nn.Module):
    def __init__(self, original_ffn):
        super().__init__()
        in_dim = original_ffn.ffn[0].in_features
        self.router = nn.Linear(in_dim, 8) 
        self.experts = nn.ModuleList([
            nn.Sequential(nn.Linear(in_dim, in_dim * 2), nn.SiLU(), nn.Linear(in_dim * 2, in_dim))
            for _ in range(8)
        ])
        self.top_k = 2

    def forward(self, x):
        batch, seq, dim = x.shape
        flat_x = x.view(-1, dim)
        logits = self.router(flat_x)
        weights, selected_experts = torch.topk(logits, self.top_k)
        weights = F.softmax(weights, dim=-1)
        out = torch.zeros_like(flat_x)
        for i, expert in enumerate(self.experts):
            mask = (selected_experts == i).any(dim=-1)
            if mask.any():
                out[mask] += expert(flat_x[mask]) * weights[mask][:, :1]
        return out.view(batch, seq, dim)

# =========================================================
# 2. CONFIGURATION & PATCHING
# =========================================================

# CORRECT MODEL ID: Wan 2.1 I2V 14B is the standard for Image-to-Video
MODEL_ID = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers" 
HF_TOKEN = os.environ.get("HF_TOKEN")

def patch_model(pipe):
    print("🛠️ Patching Transformer with GQA and MoE...")
    for i, block in enumerate(pipe.transformer.blocks):
        if hasattr(block, 'attn'):
            block.attn = GQAAttention(block.attn)
        if hasattr(block, 'ffn') and i % 2 == 0:
            block.ffn = SparseMoEFFN(block.ffn)
    return pipe

# =========================================================
# 3. GENERATION
# =========================================================

@spaces.GPU(duration=600)
def generate_20s_video(image_path, prompt, duration, steps):
    if not HF_TOKEN:
        raise gr.Error("HF_TOKEN missing. Please set it in your environment variables.")

    print(f"⏳ Loading Model: {MODEL_ID}")
    pipe = WanImageToVideoPipeline.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.bfloat16,
        token=HF_TOKEN
    )
    
    # Apply architecture modifications
    pipe = patch_model(pipe)
    pipe.enable_model_cpu_offload() 
    pipe.vae.enable_tiling() 
    
    img = Image.open(image_path).convert("RGB")
    img = img.resize((832, 480)) # Maintain 16:9 for 480P
    
    # Wan formula for frames: 4n + 1
    num_frames = int(duration * 16)
    num_frames = ((num_frames - 1) // 4) * 4 + 1
        
    with torch.inference_mode():
        output = pipe(
            image=img,
            prompt=prompt + ", high quality, cinematically consistent",
            num_frames=num_frames,
            num_inference_steps=steps,
            guidance_scale=5.0,
            generator=torch.Generator("cuda").manual_seed(42)
        )

    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as f:
        video_path = f.name
    
    export_to_video(output.frames[0], video_path, fps=16)
    return video_path

# Gradio Setup
with gr.Blocks() as demo:
    gr.Markdown("# 🎬 Optimized Wan 2.1 (GQA + MoE)")
    with gr.Row():
        with gr.Column():
            img = gr.Image(type="filepath", label="Input Image")
            txt = gr.Textbox(label="Prompt", value="A futuristic city with flying cars at sunset")
            dur = gr.Slider(5, 20, value=20, label="Duration (Seconds)")
            stp = gr.Slider(10, 30, value=20, label="Steps")
            btn = gr.Button("Generate 20s Video")
        with gr.Column():
            vid = gr.Video()

    btn.click(generate_20s_video, [img, txt, dur, stp], vid)

demo.queue().launch()