model_patcher: skip synthetic quant keys in get_key_patches

[liminfei-amd]

Summary

ModelMergeSimple — and every get_key_patches-based merge/save node — crashes with
AttributeError: 'Linear' object has no attribute 'weight_scale' when the model being merged is a
quantized (fp8 / QuantizedTensor) checkpoint. 100% reproducible, no sampling/GPU needed.

Problem

A QuantizedTensor flattens into synthetic state_dict keys — *.weight plus *.weight_scale /
*.weight_scale_2 / *.input_scale and a comfy_quant marker (per core comfy/quant_ops.py QUANT_ALGOS).
get_key_patches() iterates model_state_dict() and calls get_key_weight(), which does
getattr(op, op_keys[1]). For the synthetic sub-keys that is getattr(linear, "weight_scale") — not a
module attribute (the scale lives inside the QuantizedTensor in .weight), so it throws.

Change

Skip the synthetic quant sub-keys in get_key_patches(). The suffix set is gathered from core
comfy.quant_ops.QUANT_ALGOS (so new quant types are covered automatically), plus the comfy_quant
marker. The real *.weight key keeps its existing convert_weight dequant path, so quantized merging
still works as designed.

This skips only the known, core-generated quant suffixes — not a blanket getattr(op, key, None)
swallow, which would also hide a genuinely missing weight (cf. the #11585 discussion about not masking
bugs).

Validation

A standalone harness models a quantized state_dict (*.weight + *.weight_scale + *.input_scale +
comfy_quant): the stock path raises AttributeError: 'Linear' object has no attribute 'weight_scale',
and with the fix get_key_patches returns only the real *.weight key (preserving the dequant path) with
no crash. python -m py_compile comfy/model_patcher.py is clean. The fix is pure Python and
hardware-independent.

Fixes #14382

---

AI usage disclosure: this change was prepared with AI assistance; a human reviewed and verified it and can explain every line. (Reviewed by AMD engineers before submission.)

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR extends ModelPatcher to filter quantization synthetic state_dict keys. It introduces a helper function to collect synthetic key suffixes (such as weight_scale, input_scale) by inspecting comfy.quant_ops.QUANT_ALGOS, stores these in a module-level constant, and updates ModelPatcher.get_key_patches() to skip keys matching these synthetic suffixes. This prevents non-module attributes of QuantizedTensor objects from being processed as patchable weights.

🚥 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 accurately summarizes the main change: skipping synthetic quantization keys in get_key_patches to prevent AttributeError crashes.
Linked Issues check	✅ Passed	The code change fully addresses issue #14382: prevents AttributeError by skipping synthetic quant keys, preserves dequantization path for real weights, and limits fix to known core-generated suffixes.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to fixing the quantized model merging crash: adding _collect_quant_synthetic_keys() and _QUANT_SYNTHETIC_KEYS constant, and updating get_key_patches() to skip synthetic keys.
Description check	✅ Passed	The PR description clearly documents the bug, the root cause involving QuantizedTensor synthetic keys, the solution implemented, and validation steps performed.

_{✏️ 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.}