[GH-ISSUE #321] question about megamoe #162

New issue

Closed

opened 2026-10-26 21:13:56 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 21:13:56 +00:00

Owner

Originally created by @Thunderbrook on GitHub (Apr 28, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/321

hello~ I have some questions about the two comments in the code.

"Only use first 16 lanes for address" — I see that stmatrix.x4 is used, which seems to require all 32 lanes to provide addresses. link0
What does the phrase 'One warp per row' mean in there? I personally feel that one warp processes two tokens. link1

thx~

Originally created by @Thunderbrook on GitHub (Apr 28, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/321 hello~ I have some questions about the two comments in the code. 1. "Only use first 16 lanes for address" — I see that stmatrix.x4 is used, which seems to require all 32 lanes to provide addresses. [link0](https://github.com/deepseek-ai/DeepGEMM/blob/main/deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh#L1163) 2. What does the phrase 'One warp per row' mean in there? I personally feel that one warp processes two tokens. [link1](https://github.com/deepseek-ai/DeepGEMM/blob/main/deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh#L1186) thx~

bsakaguchi closed this issue

2026-10-26 21:13:57 +00:00

bsakaguchi commented

2026-10-26 21:13:59 +00:00

Author

Owner

@zheanxu commented on GitHub (Apr 30, 2026):

Thanks for your careful review. You're right — both comments are outdated after recent code changes. We'll update them to reflect the current behavior. Appreciate you pointing this out!

@zheanxu commented on GitHub (Apr 30, 2026): Thanks for your careful review. You're right — both comments are outdated after recent code changes. We'll update them to reflect the current behavior. Appreciate you pointing this out!

bsakaguchi referenced this issue

2026-10-26 21:15:36 +00:00

[PR #162] [CLOSED] Fix wrong condition, which always evaluates to True. #272