Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

Modules using torch.nn.utils.parametrizations.weight_norm like Conv1d layers in AudioOobleckVAE should be properly loaded to GPU during novram

Actual Behavior

weight-normalized modules are incorrectly skipped in _load_list(), causing:

RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same

this affects the Stable Audio VAE and any other model using PyTorch's weight normalization parametrizations.

Steps to Reproduce

1. Load stable-audio-open-1.0.safetensors
2. Run audio generation workflow with --novram
3. VAE decode fails with device mismatch error

Debug Logs

(salient logs shared above)

Other

in stable audio, weight normalization stores weights as parametrizations.weight.original0 and parametrizations.weight.original1, which only appear in recurse=True so:

  for name, param in m.named_parameters(recurse=True):
      if name not in params:
          if not name.startswith("parametrizations."):
              skip = True
              break

parametrizations

is the intended misspelling and is correct