fix(lora): restore checker-scrambled base weights once - #2520
Open
Arist12 wants to merge 2 commits into
Open
Conversation
Arist12
force-pushed
the
fix/lora-disaggregated-base-weight-check
branch
5 times, most recently
from
August 17, 2026 04:50
468cad9 to
b2e48cb
Compare
Arist12
marked this pull request as ready for review
August 17, 2026 04:54
Arist12
requested review from
Shi-Dong,
Zhichenzzz,
fzyzcjy,
maocheng23,
yueming-yuan and
yushengsu-thu
as code owners
August 17, 2026 04:54
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Arist12
marked this pull request as draft
August 17, 2026 06:47
Arist12
marked this pull request as ready for review
August 19, 2026 01:19
Arist12
marked this pull request as draft
August 19, 2026 01:44
Arist12
changed the title
Arist12
marked this pull request as ready for review
August 20, 2026 15:42
Arist12
changed the title
Arist12
force-pushed
the
fix/lora-disaggregated-base-weight-check
branch
from
August 21, 2026 18:27
bf9ee36 to
e4824a3
Compare
Arist12
force-pushed
the
fix/lora-disaggregated-base-weight-check
branch
from
August 21, 2026 18:58
e4824a3 to
7ecf535
Compare
Registering a new 400s 4-GPU E2E is a CI-budget decision that belongs with the reviewer, not with a weight-sync fix. The gating logic is already covered by the fast tests in this PR, and the disaggregated transport was validated by hand on 8x MI350X.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Part of #2705.
Problem
--check-weight-update-equalrandomizes rollout-engine base weights at startup. LoRA normally skips frozen-base synchronization, so the first check sees a corrupted base; repeatedly sending the base is unnecessary. Multi-LoRA slot parameters must also stay out of the base converter.Change
Restore the frozen base only on the first checked update for tensor and distributed transports, then return to adapter-only updates. Recognize both
.adapter.and.adapters.<slot>.parameter names when filtering the base path.Bridge-built architectures whose base names require Bridge conversion must include #2683 before this forced base sync is exercised; the Qwen2.5 validation below does not cover that dependency.
Validation