Fix weight_norm parametrizations skipped in _load_list causing stable audio partial-load failure

[octo-patch]

Summary

• _load_list() in model_patcher.py incorrectly sets default = True (treating a module as having random weights in child sub-modules) when a module uses torch.nn.utils.parametrizations.weight_norm
• weight_norm moves the actual weight into a parametrizations child sub-module as original0/original1, so named_parameters(recurse=True) yields parametrizations.weight.original0 while named_parameters(recurse=False) does not — triggering the false-positive check
• This causes weight-normed modules to be excluded from the partial-load list entirely, so their parameters are never moved to GPU during low-VRAM loading
• AudioOobleckVAE (Stable Audio) had a disable_offload = True workaround that forced full GPU load to paper over this bug; with the root cause fixed, the workaround is removed so the model can use partial VRAM offloading on low-memory systems

Fix

Skip parametrizations.* parameter names in the _load_list() non-leaf check:

# Before
if name not in params:
    default = True  # default random weights in non leaf modules
    break

# After
if name not in params and not name.startswith("parametrizations."):
    default = True  # default random weights in non leaf modules
    break

Test plan

• Load a Stable Audio model on a system with limited VRAM (--lowvram or --novram) and confirm it generates audio without error
• Confirm the fix does not affect regular (non-weight-normed) models that correctly have child sub-modules with random weights

Closes #11855

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 0cfe6d56-fc0b-4fd0-9823-13e6279287a9

📥 Commits

Reviewing files that changed from the base of the PR and between 1391579 and 7c1ec4a.

📒 Files selected for processing (2)

• comfy/model_patcher.py
• comfy/sd.py

💤 Files with no reviewable changes (1)

• comfy/sd.py

---

📝 Walkthrough

Walkthrough

This pull request contains two internal implementation adjustments. First, in the model patcher's weight-loading logic, parameter names beginning with "parametrizations." are now excluded from the default random weight handling for non-leaf modules. Second, the VAE initialization code no longer sets the disable_offload flag to True when AudioOobleckVAE weights are detected. Both changes are internal to their respective modules with no modifications to public APIs.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title directly describes the main fix: correcting the _load_list() logic to handle weight_norm parametrizations and resolving the Stable Audio partial-load failure.
Description check	✅ Passed	The description is comprehensive and directly related to the changeset, explaining the root cause, the fix, and the removal of the workaround.
Linked Issues check	✅ Passed	The PR fully addresses issue #11855 by fixing the _load_list() non-leaf check to skip parametrizations.* names [#11855] and removing the AudioOobleckVAE disable_offload workaround.
Out of Scope Changes check	✅ Passed	All changes are scoped to the linked issue: fixing the parametrizations check in model_patcher.py and removing the workaround in sd.py as documented in the PR objectives.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}