Repository navigation
[LoRA] split fused img_mlp.gate_up LoRAs for Qwen-Image-2.1 - #14932
Open
dkackman wants to merge 1 commit into
Open
dkackman wants to merge 1 commit into
dkackman wants to merge 1 commit into
Conversation
Copy link
Copy Markdown
Sorry, something went wrong.
ComfyUI and ai-toolkit Qwen-Image-2.1 LoRAs target the fused SwiGLU `img_mlp.gate_up` layer, which diffusers keeps as separate `gate_layer` and `proj` layers. Those keys were dropped as unexpected, so the LoRA loaded without its MLP update. Split them into the two layers, gate rows first, as the full-checkpoint converter does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
7 of 10 tasks
Copy link
Copy Markdown
Contributor
|
Hi @dkackman, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder. Please note that PRs without a linked issue are likely to be automatically closed 10 days after this notice. Once the PR links an issue (or gets the no-issue-needed label), you can ignore this message — it stays here as a comment, but it no longer applies. |
Sorry, something went wrong.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Fixes #14937
ComfyUI and ai-toolkit LoRAs for Qwen-Image-2.1 target the fused SwiGLU img_mlp.gate_up layer. diffusers keeps that layer as two separate layers, img_mlp.gate_layer and img_mlp.proj, so the converter passed the gate_up keys through, PEFT dropped them as unexpected keys, and the LoRA loaded without its MLP update and without any error.
This PR splits those LoRAs into the two layers when they're loaded, gate rows first, in the same order convert_qwen_image21_transformer_checkpoint_to_diffusers uses for full checkpoints. lora_A is shared and lora_B is split by rows, so the split is lossless. Nothing changes for the original Qwen-Image, which has no gate_up layer.
Checked against real files: converting ML-Intern-lab's fused doodle-in LoRA gives exactly the pre-split file they publish. The Outpaint-v2, grainscape and AnyAngle LoRAs map fully onto the 2.1 model's layers, with no keys dropped.
Related: #14933. That PR fixes .alpha being dropped from lora_A/lora_B checkpoints, which is a separate bug. Both PRs edit _convert_non_diffusers_qwen_lora_to_diffusers, so whichever merges second needs a small rebase.
Self-review notes
I ran the self-review skill. It found no blocking issues. I removed the dead code it flagged (an unreachable else and a redundant early return). Left for review:
Before submitting
Who can review?
@sayakpaul
🤖 Generated with Claude Code