Repository navigation
feat(peft): expand_shared_outer export flag for shared-outer MoE LoRA - #5588
Merged
adityavavreNVDA merged 4 commits intoAug 20, 2026
Merged
Conversation
Shared-outer MoE LoRA exports the shared adapter side as a single `[1, ...]` tensor under an expert-agnostic name (SGLang contract). vLLM's 2D `pack_moe` path for text MoE models expects per-expert 2D tensors, so it can't find the shared factor and crashes at `lora.optimize()` (`lora_b is None`). So this PR adds `expand_shared_outer` to the adapter export path. When set, the shared 2D factor is replicated under per-expert 2D names (reused, not cloned); training keeps the shared parameter. Default `False` preserves the existing layout. Signed-off-by: Hollow Man <hollowman@opensuse.org>
yaoyu-33
previously approved these changes
Aug 14, 2026
yaoyu-33
added
area:peft
Contributor
|
/ok to test 4ce1b2c |
Contributor
|
@HollowMan6 Please check the CI failures |
Signed-off-by: Hollow Man <hollowman@opensuse.org>
Contributor
Author
|
@kamran-nvidia Fixed, also make |
Contributor
|
/ok to test 09f6355 |
kamran-nvidia
approved these changes
Aug 17, 2026
8 tasks done
cuichenx
approved these changes
Aug 20, 2026
3 of 4 tasks
3 of 5 tasks
This branch was previously 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 ?
Related to #3958
Shared-outer MoE LoRA exports the shared adapter side as a single
[1, ...]tensor under an expert-agnostic name (SGLang contract). vLLM's 2Dpack_moepath for text MoE models expects per-expert 2D tensors, so it can't find the shared factor and crashes atlora.optimize()(lora_b is None). So this PR addsexpand_shared_outerto the adapter export path. When set, the shared 2D factor is replicated under per-expert 2D names (reused, not cloned); training keeps the shared parameter. DefaultFalsepreserves the existing layout.Changelog
expand_shared_outerparam with logicsGitHub Actions CI
See the CI section in the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.
Before your PR is "Ready for review"
Pre checks:
If you haven't finished some of the above items you can still open "Draft" PR.
Additional Information
✨ Presented to you with Mind Lab - A Lab for Experiential Intelligence.