debug_vae.py

"""
Debug script: print shapes through VAE encoder for different temporal inputs.

Usage:
  python debug_vae2.py --vae_path /path/to/Wan2.2_VAE.pth --height 384 --width 640
"""

import sys
import os
import argparse
import torch

print("[0] Script started", flush=True)

_REPO_ROOT = os.environ.get("DIFFSYNTH_ROOT", ".")
if _REPO_ROOT not in sys.path:
    sys.path.insert(0, _REPO_ROOT)

print("[1] Importing diffsynth modules...", flush=True)
from diffsynth.models.wan_video_vae import (
    WanVideoVAE, WanVideoVAE38,
    Down_ResidualBlock, AvgDown3D, Resample, Resample38,
    CausalConv3d, ResidualBlock, Encoder3d, Encoder3d_38,
)
print("[2] Import done", flush=True)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--vae_path", type=str, required=True)
    parser.add_argument("--height", type=int, default=384)
    parser.add_argument("--width", type=int, default=640)
    parser.add_argument("--device", type=str, default="cuda")
    parser.add_argument("--test_frames", type=str, default="1,4,5,8,9,13,17,21")
    args = parser.parse_args()

    temporal_lengths = [int(x) for x in args.test_frames.split(",")]
    print(f"[3] Will test frames: {temporal_lengths}", flush=True)

    print(f"[4] Loading checkpoint from {args.vae_path} ...", flush=True)
    state_dict = torch.load(args.vae_path, map_location="cpu", weights_only=False)
    print(f"[5] Checkpoint loaded, type={type(state_dict)}", flush=True)

    if "model_state" in state_dict:
        state_dict = state_dict["model_state"]
        print("[5.1] Extracted model_state", flush=True)

    keys = list(state_dict.keys())
    print(f"[6] Total keys: {len(keys)}", flush=True)
    print(f"    First 5 keys: {keys[:5]}", flush=True)

    has_model_prefix = any(k.startswith("model.") for k in keys)
    is_vae38 = any("avg_shortcut" in k for k in keys)
    print(f"    has_model_prefix={has_model_prefix}, is_vae38={is_vae38}", flush=True)

    if is_vae38 or any("avg_shortcut" in k for k in keys):
        print("[7] Creating WanVideoVAE38...", flush=True)
        vae = WanVideoVAE38()
    else:
        print("[7] Creating WanVideoVAE...", flush=True)
        vae = WanVideoVAE()

    if not has_model_prefix:
        print("[8] Adding 'model.' prefix to keys...", flush=True)
        state_dict = {f"model.{k}": v for k, v in state_dict.items()}

    print("[9] Loading state dict...", flush=True)
    missing, unexpected = vae.load_state_dict(state_dict, strict=False)
    print(f"    Missing: {len(missing)}, Unexpected: {len(unexpected)}", flush=True)
    if missing:
        for k in missing[:10]:
            print(f"      missing: {k}", flush=True)
    if unexpected:
        for k in unexpected[:10]:
            print(f"      unexpected: {k}", flush=True)

    print(f"[10] Moving VAE to {args.device}...", flush=True)
    vae = vae.to(args.device).eval()
    print("[11] VAE ready", flush=True)

    encoder = vae.model.encoder
    print(f"[12] Encoder type: {type(encoder).__name__}", flush=True)

    if hasattr(encoder, 'downsamples'):
        for i, layer in enumerate(encoder.downsamples):
            print(f"    downsamples[{i}]: {type(layer).__name__}", flush=True)
            if isinstance(layer, Down_ResidualBlock):
                avg = layer.avg_shortcut
                print(f"      avg_shortcut: factor_t={avg.factor_t}, factor_s={avg.factor_s}", flush=True)
                # list sub-modules
                for j, sub in enumerate(layer.downsamples):
                    print(f"      downsamples[{j}]: {type(sub).__name__}", flush=True)

    for t in temporal_lengths:
        print(f"\n{'='*50}", flush=True)
        print(f"Testing t={t}", flush=True)
        print(f"{'='*50}", flush=True)

        video = torch.randn(3, t, args.height, args.width, dtype=torch.float32)
        print(f"  Input video shape: {video.shape}", flush=True)

        try:
            with torch.no_grad():
                latent = vae.encode([video], device=args.device)
            print(f"  Output latent shape: {latent.shape}", flush=True)
        except RuntimeError as e:
            print(f"  FAILED: {e}", flush=True)

    print("\n[DONE]", flush=True)


if __name__ == "__main__":
    main()