[Bugfix] Fix precision loss in LoRA-wrapped RowParallelLinear by fusing bias into GEMM #28972
Conversation
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
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request addresses a precision loss issue in LoRA-wrapped RowParallelLinear by fusing the bias addition into the GEMM kernel, which aligns its behavior with the non-LoRA equivalent layer. The changes correctly pass the bias to the apply method only on rank 0 to prevent redundant additions in tensor-parallel setups, and the refactoring of the bias handling logic improves code clarity. The fix appears correct and well-implemented. I have no major concerns with this change.
Collaborator
|
over LGTM, could you please address CI failure first? |
…ng bias into GEMM Signed-off-by: prashanth058 <[email protected]>
prashanth058
force-pushed
the
fix/lora-bias-precision
branch
from
b58afac to
58e30d4
Compare
jeejeelee
approved these changes
Nov 20, 2025
github-actions
bot
added
the
ready
LuminolT
pushed a commit
to LuminolT/vllm
that referenced
this pull request
Nov 21, 2025
…ng bias into GEMM (vllm-project#28972) Signed-off-by: prashanth058 <[email protected]> Signed-off-by: LuminolT <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
Issue:
LoRA-wrapped RowParallelLinear was adding bias as a separate
bfloat16operation instead of fusing it into the GEMM kernel like the unwrapped layer does. This caused precision loss because the fused kernel can accumulate in higher precision (FP32) before converting to bfloat16, while separate addition incurs additional rounding errors. The discrepancy appeared even with zero LoRA weights when comparing LoRA-wrapped vs merged weight results.Fix:
Pass bias to apply() only on rank 0 (or when skip_bias_add=False), allowing the quantization method to fuse bias addition with matrix multiplication in the GEMM kernel. This matches the unwrapped layer's behavior and eliminates precision discrepancies.