[GH-ISSUE #140] VLLM Integration: Assertion error csrc/jit_kernels/impls/smxx_layout.hpp:136 'mn % 4 == 0 and num_groups == 1' #80

Closed
opened 2026-10-26 21:12:03 +00:00 by bsakaguchi · 1 comment
Owner

Originally created by @yewentao256 on GitHub (Jul 24, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/140

Purpose

This error will appear/disappear with the same code using vllm with different version of DeepGEMM

So could someone give me more context about this? Is this a bug inside DeepGEMM, or the caller should update the way of using it.

Ways to reproduce

https://github.com/vllm-project/vllm/pull/21559, use this unit test code.

If we downgrade DeepGEMM to https://github.com/RayWang96/DeepGEMM/tree/multi_arch_support with commit cc416ee4fa, the unit will pass.

If we are using main branch:

E       RuntimeError: Failed: Assertion error csrc/jit_kernels/impls/smxx_layout.hpp:136 'mn % 4 == 0 and num_groups == 1'

../../../vllm/utils/deep_gemm.py:111: RuntimeError
Originally created by @yewentao256 on GitHub (Jul 24, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/140 ## Purpose This error will appear/disappear with the same code using vllm with different version of DeepGEMM So could someone give me more context about this? Is this a bug inside DeepGEMM, or the caller should update the way of using it. ## Ways to reproduce https://github.com/vllm-project/vllm/pull/21559, use this unit test code. If we downgrade DeepGEMM to https://github.com/RayWang96/DeepGEMM/tree/multi_arch_support with commit cc416ee4faf0533a9263c2de814e5565f56ca1cc, the unit will pass. If we are using main branch: ```bash E RuntimeError: Failed: Assertion error csrc/jit_kernels/impls/smxx_layout.hpp:136 'mn % 4 == 0 and num_groups == 1' ../../../vllm/utils/deep_gemm.py:111: RuntimeError ```
Author
Owner

@LyricZhao commented on GitHub (Jul 30, 2025):

After #141, I believe it is fixed. Thanks! Feel free to reopen it if the issue remains.

<!-- gh-comment-id:3135031492 --> @LyricZhao commented on GitHub (Jul 30, 2025): After #141, I believe it is fixed. Thanks! Feel free to reopen it if the issue remains.
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepGEMM#80
No description provided.