"""Compare direct block-0 AdaLN modulation and gates with Comfy."""

import argparse

import torch

from h3_blackwell_runtime.checkpoint import H3Checkpoint
from h3_blackwell_runtime.denoiser import H3PackedDenoiser


parser = argparse.ArgumentParser()
parser.add_argument("--capture-dir", required=True)
parser.add_argument("--model", required=True)
args = parser.parse_args()

inputs = torch.load(f"{args.capture_dir}/input.pt", map_location="cuda", weights_only=False)
expected = torch.load(f"{args.capture_dir}/block0_norm1_adaln.pt", map_location="cuda", weights_only=False)
model = H3PackedDenoiser.from_checkpoint(H3Checkpoint(args.model)).eval()
actual = model.backbone.adaln[0](inputs["timesteps"])

for name, value, reference in zip(
    ("shift", "scale", "gate_msa", "shift_mlp", "scale_mlp", "gate_mlp"),
    actual,
    (expected["shift"], expected["scale"], expected["gate_msa"], None, None, expected["gate_mlp"]),
    strict=True,
):
    if reference is None:
        continue
    delta = (value.float() - reference.float()).abs()
    print(f"{name} max_abs={delta.max().item():.6g} mean_abs={delta.mean().item():.6g}")