Fix some fp8 scaled checkpoints no longer working.

[comfyanonymous]

No description provided.

[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: 2d464589-26d7-4031-af79-8ec73a4f4d05

📥 Commits

Reviewing files that changed from the base of the PR and between 076639f and f85a7ef.

📒 Files selected for processing (1)

• comfy/sd.py

---

📝 Walkthrough

Walkthrough

The change modifies comfy/sd.py in the load_diffusion_model_state_dict function to conditionally re-apply quantization conversion after normalizing the diffusion model prefix. Previously, comfy.utils.convert_old_quants() executed only once at the function start before prefix handling. The update adds a second conditional execution of this conversion step after prefix stripping, but only when custom_operations is None. This allows the quantization conversion logic to process the normalized state dictionary and associated metadata, potentially ensuring consistent quant handling across different prefix configurations.

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 inconclusive)

Check name	Status	Explanation	Resolution
Description check	❓ Inconclusive	No description was provided by the author, making it impossible to verify relevance to the changeset.	Add a description explaining why the quant conversion needs to be re-applied after prefix stripping and how this fixes the fp8 checkpoint loading issue.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the specific issue being fixed (fp8 scaled checkpoints not working), which aligns with the code change that re-applies quant conversion after prefix handling.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ 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.}