Add VAE tiled clip comparator

tools/compare_vae_tiled_clip.py

@@ -0,0 +1,51 @@
+"""Compare direct/upstream VAE tiled decode for one temporal latent clip."""
+
+import argparse
+import sys
+from pathlib import Path
+
+import torch
+from safetensors.torch import load_file
+
+from h3_blackwell_runtime.vae_decoder import MiniMaxH3VideoVAE as DirectVAE
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--latent", type=Path, required=True)
+parser.add_argument("--start", type=int, default=0)
+parser.add_argument("--tokens", type=int, default=8)
+parser.add_argument("--comfy-path", default="/opt/ComfyUI")
+args = parser.parse_args()
+
+sys.path.insert(0, args.comfy_path)
+from h3_blackwell_runtime.upstream_vae import MiniMaxH3VideoVAE as UpstreamVAE  # noqa: E402
+
+
+def stats(name: str, a: torch.Tensor, b: torch.Tensor) -> None:
+    diff = (a.float() - b.float()).abs()
+    frame_rows = []
+    for index in range(diff.shape[2]):
+        frame = diff[:, :, index]
+        frame_rows.append((index, float(frame.max()), float(frame.mean()), float(torch.quantile(frame.flatten(), 0.99))))
+    print({
+        "stage": name,
+        "shape": tuple(a.shape),
+        "max": float(diff.max()),
+        "mean": float(diff.mean()),
+        "top_mean": sorted(frame_rows, key=lambda row: row[2], reverse=True)[:8],
+    }, flush=True)
+
+
+state = torch.load(args.latent, map_location="cuda", weights_only=False)
+latent = state["latent"].to("cuda") if isinstance(state, dict) else state.to("cuda")
+
+direct = DirectVAE.from_safetensors("/vae/minimax_h3_video_vae_fp16.safetensors", device="cuda", tiling=True).eval()
+upstream = UpstreamVAE(tiling=True).to("cuda").eval()
+upstream.load_state_dict(load_file("/vae/minimax_h3_video_vae_fp16.safetensors", device="cuda"), strict=True)
+
+with torch.inference_mode():
+    z_d = latent[:, :, args.start:args.start + args.tokens].to(next(direct.parameters()).dtype)
+    z_u = latent[:, :, args.start:args.start + args.tokens].to(next(upstream.parameters()).dtype)
+    z_d = z_d * direct.latents_std.view(1, -1, 1, 1, 1).to(z_d) + direct.latents_mean.view(1, -1, 1, 1, 1).to(z_d)
+    z_u = z_u * upstream.latents_std.view(1, -1, 1, 1, 1).to(z_u) + upstream.latents_mean.view(1, -1, 1, 1, 1).to(z_u)
+    stats("tiled_clip", direct.tiled_decode(z_d), upstream.tiled_decode(z_u))