[GH-ISSUE #339] [BUG] Mega MoE kernel中的小错误 #171

Closed
opened 2026-10-26 21:14:06 +00:00 by bsakaguchi · 0 comments
Owner

Originally created by @Yyf1023 on GitHub (May 25, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/339

开发者们你们好,最近测试mega moe kernel中发现了一个代码中的小错误,起因是我在测试hidden=8192, num_tokens=1024时下面assertion错误:
github.com/deepseek-ai/DeepGEMM@714dd1a4a9/deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh (L1260)
我本想修改代码以支持8192,但是发现该错误大概率是由于前面对SMEM_BEFORE_BARRIER_SIZE计算错误引起,计算SMEM_BEFORE_BARRIER_SIZE时忘记考虑了SF size:
github.com/deepseek-ai/DeepGEMM@714dd1a4a9/deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh (L209-L210)
修改为如下即可

constexpr uint32_t SMEM_BEFORE_BARRIER_SIZE = 
-     SMEM_EXPERT_COUNT_SIZE + SMEM_SEND_BUFFER_SIZE + SMEM_CD_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE); 
+     SMEM_EXPERT_COUNT_SIZE + SMEM_SEND_BUFFER_SIZE + SMEM_CD_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE + SMEM_SFA_SIZE_PER_STAGE + SMEM_SFB_SIZE_PER_STAGE);

请修复这个问题

Originally created by @Yyf1023 on GitHub (May 25, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/339 开发者们你们好,最近测试mega moe kernel中发现了一个代码中的小错误,起因是我在测试hidden=8192, num_tokens=1024时下面assertion错误: https://github.com/deepseek-ai/DeepGEMM/blob/714dd1a4a980f7937a74343d19a8eba4fe321480/deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh#L1260 我本想修改代码以支持8192,但是发现该错误大概率是由于前面对SMEM_BEFORE_BARRIER_SIZE计算错误引起,计算SMEM_BEFORE_BARRIER_SIZE时忘记考虑了SF size: https://github.com/deepseek-ai/DeepGEMM/blob/714dd1a4a980f7937a74343d19a8eba4fe321480/deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh#L209-L210 修改为如下即可 ```diff constexpr uint32_t SMEM_BEFORE_BARRIER_SIZE = - SMEM_EXPERT_COUNT_SIZE + SMEM_SEND_BUFFER_SIZE + SMEM_CD_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE); + SMEM_EXPERT_COUNT_SIZE + SMEM_SEND_BUFFER_SIZE + SMEM_CD_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE + SMEM_SFA_SIZE_PER_STAGE + SMEM_SFB_SIZE_PER_STAGE); ``` 请修复这个问题
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#171
No description provided.