Allow users to disable lora patching optimization introduced by #6439

[leo9800]

Summary

LoRA patching optimization introduced in PR #6439 may cause issue in some edge cases, which prevents InvokeAI to generate any reasonable image.

This patch add an configuration option to allow users to disable the change introduced in #6439 in case the issue occurs.

Related Issues / Discussions

• [bug]: (v4.2.5) txt2img failed with random noise  #6559 (Created by me, confirmed fixed with this patch)
• [bug]: Weird outcome image when using specific SD 1.5 model #6596 (Not created by me, but I doubt the cause is the same)

QA Instructions

I tested this on my server (no GPU, running InvokeAI on an Intel Ice Lake SP CPU) and problem (mentioned in #6559) is addressed.

Since it is probably an edge case, (only 2 issues are found possibly related to it) it may not always be reproducible.

Merge Plan

N/A (This patch won't touch anything big, e.g. db schema, etc)

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Documentation added / updated (if applicable)

[lstein]

It looks like the issue appears specifically when rendering on CPU, is that correct? If this is the case, then we should just be disabling the LoRA optimization whenever the rendering device is set for CPU.

[lstein]

This is a bug that appears consistently when rendering on the CPU device, so a user config change is overkill. See PR #6631 for a simpler solution.

[lstein]

I'm very sorry, but this is superseded by #6631. The latter is preferred because it does not require the user to make a configuration file change. Please continue to contribute bug reports and fixes!