Repository navigation
[Bugfix][Model] Fix Ling3 SwiGLU clamping - #65088
Merged
vllm-bot merged 3 commits intoNov 5, 2026
Merged
Conversation
Assisted-by: OpenAI Codex Signed-off-by: zexplorerhj <zhjoneson@163.com>
zexplorerhj
requested review from
mgoin,
pavanimajety and
zyongye
as code owners
November 2, 2026 06:38
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
jeejeelee
approved these changes
Nov 2, 2026
Member
|
/ci run |
jeejeelee
enabled auto-merge (squash)
November 2, 2026 08:24
github-actions
Bot
added
the
ready
|
✅ Triggered Buildkite CI #99120 for commit |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Overview
While investigating an accuracy issue during training–inference alignment for Ling RL, we found that routed experts applied the SwiGLU gate limit before SiLU, whereas the training/reference implementation applies it after SiLU. This mismatch can affect RL accuracy through inconsistent training and rollout computations, even though its impact is not obvious in standalone inference evaluations.
Use the existing
swiglustepactivation for Ling layers with a configured limit and forward that limit through the common MoE activation paths.Claims
clamp(SiLU(gate), max=L) * clamp(up, -L, L)instead ofSiLU(clamp(gate, max=L)) * clamp(up, -L, L).SWIGLUSTEPactivation paths.siluand retain the defaultSWIGLUSTEPlimit of7.0when no limit is supplied.