[Bugfix] Allocate less memory in non-batched CUTLASS MoE #21121

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Merged

DarkLight1337 merged 2 commits into vllm-project:main from neuralmagic:cutlass-moe-less-mem

Jul 18, 2025

Contributor

ElizaWszola commented Jul 17, 2025 •

edited by github-actions bot

Loading

The way we obtain N from _moe_problem_size() caused the function to allocate up to twice as much memory in cutlass_moe's workspace_shapes() as needed. The fix brings down the size of workspaces to what actually reflects the size needed to store cutlass_moe's intermediate tensors.

Testing

Tested for correctness with offline inference for nm-testing/DeepSeek-Coder-V2-Lite-Instruct-FP8 and

pytest tests/kernels/moe/test_cutlass_moe.py -k test_cutlass_moe_8_bit_EP_large
pytest tests/kernels/moe/test_cutlass_moe.py -k test_cutlass_moe_8_bit_cuda_graph


          Allocate less memory in CUTLASS MoE

ba131c3

Signed-off-by: ElizaWszola <ewszola@redhat.com>

github-actions bot commented Jul 17, 2025

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

gemini-code-assist bot reviewed

View reviewed changes

Contributor

gemini-code-assist bot left a comment

Code Review

The pull request reduces memory allocation in non-batched CUTLASS MoE by optimizing the calculation of N in _moe_problem_size(), leading to smaller workspace sizes. No specific issues were identified requiring immediate attention based on the provided changes and review criteria.

mgoin reviewed

View reviewed changes

vllm/model_executor/layers/fused_moe/cutlass_moe.py

Comment on lines -288 to +289

-                          workspace1 = (M * topk, max(2 * N, K))
-                          workspace2 = (M * topk, N)
+                          workspace1 = (M * topk, max(N, K))
+                          workspace2 = (M * topk, N // 2)

Member

mgoin Jul 17, 2025

Is there a reason why this is the inverse of TritonExperts?

vllm/vllm/model_executor/layers/fused_moe/fused_moe.py

Lines 1623 to 1624 in 9fb2d22

    
           workspace1 = (M, topk, max(N // 2, K)) 
        
           workspace2 = (M, topk, max(N, K))

Contributor Author

ElizaWszola Jul 17, 2025

Looks like the variables are just flipped.

In Triton (workspace13 is the same tensor as workspace1):
https://github.com/vllm-project/vllm/blob/9fb2d22032cee577a189f8c4cddd88a3c190cb0c/vllm/model_executor/layers/fused_moe/fused_moe.py#L1702C1-L1707C72

In CUTLASS:
https://github.com/vllm-project/vllm/blob/9fb2d22032cee577a189f8c4cddd88a3c190cb0c/vllm/model_executor/layers/fused_moe/cutlass_moe.py#L168C1-L170C55

The CUTLASS names seem to make more sense given what we assign the resized workspace tensors to

mgoin added the bug label

mgoin approved these changes

View reviewed changes

Member

mgoin left a comment

LGTM, thanks Eliza!

mgoin added the ready label


          Merge branch 'main' into cutlass-moe-less-mem

1712fc4

Signed-off-by: ElizaWszola <ewszola@redhat.com>

DarkLight1337 merged commit 4adc66f into vllm-project:main

67 checks passed

WorldExplored pushed a commit to nadathurv/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

f0a42b0

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

WorldExplored pushed a commit to nadathurv/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

c78b349

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: WorldExplored <srreyansh.sethi@gmail.com>

WorldExplored pushed a commit to nadathurv/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

794909d

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: WorldExplored <srreyansh.sethi@gmail.com>

hj-mistral pushed a commit to hj-mistral/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

f5f1fef

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: Himanshu Jaju <hj@mistral.ai>

LyrisZhong pushed a commit to LyrisZhong/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

cc3bd92

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

avigny pushed a commit to avigny/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

28cc61a

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: avigny <47987522+avigny@users.noreply.github.com>

x22x22 pushed a commit to x22x22/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

11098c0

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: x22x22 <wadeking@qq.com>

Pradyun92 pushed a commit to Pradyun92/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

6ebadc9

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

npanpaliya pushed a commit to odh-on-pz/vllm-upstream that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

42b628a

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

jinzhen-lin pushed a commit to jinzhen-lin/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

b485ee3

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>

paulpak58 pushed a commit to paulpak58/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

20052f3

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: Paul Pak <paulpak58@gmail.com>

taneem-ibrahim pushed a commit to taneem-ibrahim/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

3cf4c8c

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

diegocastanibm pushed a commit to diegocastanibm/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

1243fb8

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: Diego-Castan <diego.castan@ibm.com>

epwalsh pushed a commit to epwalsh/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

a7dc747

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

googlercolin pushed a commit to googlercolin/vllm that referenced this pull request


          [Bugfix] Allocate less memory in non-batched CUTLASS MoE (vllm-projec…

be65787

…t#21121)

Signed-off-by: ElizaWszola <ewszola@redhat.com>

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug ready