[Performance] Remove input pads in cutlass_mla and optimize v_proj output handling #25184
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
alexm-redhat
requested review from
WoosukKwon,
comaniac,
njhill,
robertgshaw2-redhat and
ywang96
as code owners
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces performance optimizations for MLA by pre-padding inputs and optimizing the v_proj output handling. The changes are well-aligned with the performance improvement goals. I have identified two areas for improvement. First, in vllm/v1/attention/backends/mla/common.py, the _v_up_proj method uses an unsafe resize_ on a tensor view, which should be refactored for safety and clarity. Second, vllm/v1/attention/backends/mla/cutlass_mla.py contains temporary, commented-out code that should be removed before merging to improve maintainability. Addressing these points will enhance the code's quality and robustness.
alexm-redhat
force-pushed
the
opt_mla_2
branch
from
be624a7 to
5add586
Compare
alexm-redhat
force-pushed
the
opt_mla_2
branch
from
5d49947 to
678db9b
Compare
mgoin
reviewed
Sep 18, 2025
|
This pull request has merge conflicts that must be resolved before it can be |
Open
1 task
alexm-redhat
force-pushed
the
opt_mla_2
branch
from
9d254a8 to
53e50fd
Compare
mgoin
approved these changes
Sep 22, 2025
…tput reshape Signed-off-by: Alexander Matveev <amatveev@redhat.com>
alexm-redhat
force-pushed
the
opt_mla_2
branch
from
53e50fd to
8bf1e11
Compare
FeiDaLI
pushed a commit
to FeiDaLI/vllm
that referenced
this pull request
Sep 25, 2025
…tput handling (vllm-project#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com>
charlifu
pushed a commit
to ROCm/vllm
that referenced
this pull request
Sep 25, 2025
…tput handling (vllm-project#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com> Signed-off-by: charlifu <charlifu@amd.com>
yewentao256
pushed a commit
that referenced
this pull request
Oct 3, 2025
…tput handling (#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com> Signed-off-by: yewentao256 <zhyanwentao@126.com>
gjc0824
pushed a commit
to gjc0824/vllm
that referenced
this pull request
Oct 10, 2025
…tput handling (vllm-project#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com> Signed-off-by: gaojc <1055866782@qq.com>
xuebwang-amd
pushed a commit
to xuebwang-amd/vllm
that referenced
this pull request
Oct 10, 2025
…tput handling (vllm-project#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com> Signed-off-by: xuebwang-amd <xuebwang@amd.com>
choprahetarth
pushed a commit
to Tandemn-Labs/vllm
that referenced
this pull request
Oct 11, 2025
…tput handling (vllm-project#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com>
lywa1998
pushed a commit
to lywa1998/vllm
that referenced
this pull request
Oct 20, 2025
…tput handling (vllm-project#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com>
xuebwang-amd
pushed a commit
to xuebwang-amd/vllm
that referenced
this pull request
Oct 24, 2025
…tput handling (vllm-project#25184) Signed-off-by: Alexander Matveev <amatveev@redhat.com> Signed-off-by: xuebwang-amd <xuebwang@amd.com>
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.
This PR removes the need to pad cutlass mla inputs (q_nope and q_pe) to max_heads==128 by pre-padding the buffers that are used by previous operations. Also, the PR improves the way v_proj handles the output by reusing the output buffer earlier inside torch.bmm. For DeepSeekR1 on 8xB200 batch_size==32, decode iteration TPOT performance imporves from 18.87 to 18.25ms, about 3.3%.
Verified correctness with: lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-R1-0528,tensor_parallel_size=8 --tasks gsm8k --num_fewshot 5 --batch_size auto