[PR #341] Fix TMEM lane address for debug mode across all SM100 kernels #374

Open
opened 2026-10-26 21:16:59 +00:00 by bsakaguchi · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/341
Author: @yejunjin
Created: 5/28/2026
Status: 🔄 Open

Base: main ← Head: fix-tmem-lane-check


📝 Commits (1)

  • ed83aa2 Fix TMEM lane address for debug mode across all SM100 kernels

📊 Changes

10 files changed (+45 additions, -4 deletions)

View changed files

📝 deep_gemm/include/deep_gemm/epilogue/sm100_store_cd.cuh (+3 -0)
📝 deep_gemm/include/deep_gemm/epilogue/sm100_store_cd_swap_ab.cuh (+3 -0)
📝 deep_gemm/include/deep_gemm/impls/sm100_bmk_bnk_mn.cuh (+3 -0)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp4_mqa_logits.cuh (+6 -1)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp4_paged_mqa_logits.cuh (+6 -1)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh (+6 -0)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_gemm_1d1d.cuh (+3 -0)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_mqa_logits.cuh (+6 -1)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_paged_mqa_logits.cuh (+6 -1)
📝 deep_gemm/include/deep_gemm/impls/sm100_tf32_hc_prenorm_gemm.cuh (+3 -0)

📄 Description

Set explicit TMEM lane bits (22:16) under __CUDACC_DEBUG__ to satisfy the hardware's strict Warp Tensor Memory Access Check enabled by -G. In release mode the hardware auto-routes lanes, so the fix is guarded to avoid unnecessary overhead.


🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/deepseek-ai/DeepGEMM/pull/341 **Author:** [@yejunjin](https://github.com/yejunjin) **Created:** 5/28/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `fix-tmem-lane-check` --- ### 📝 Commits (1) - [`ed83aa2`](https://github.com/deepseek-ai/DeepGEMM/commit/ed83aa2694826395808c14b94dfdd03b415fe550) Fix TMEM lane address for debug mode across all SM100 kernels ### 📊 Changes **10 files changed** (+45 additions, -4 deletions) <details> <summary>View changed files</summary> 📝 `deep_gemm/include/deep_gemm/epilogue/sm100_store_cd.cuh` (+3 -0) 📝 `deep_gemm/include/deep_gemm/epilogue/sm100_store_cd_swap_ab.cuh` (+3 -0) 📝 `deep_gemm/include/deep_gemm/impls/sm100_bmk_bnk_mn.cuh` (+3 -0) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp4_mqa_logits.cuh` (+6 -1) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp4_paged_mqa_logits.cuh` (+6 -1) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh` (+6 -0) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_gemm_1d1d.cuh` (+3 -0) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_mqa_logits.cuh` (+6 -1) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_paged_mqa_logits.cuh` (+6 -1) 📝 `deep_gemm/include/deep_gemm/impls/sm100_tf32_hc_prenorm_gemm.cuh` (+3 -0) </details> ### 📄 Description Set explicit TMEM lane bits (22:16) under `__CUDACC_DEBUG__` to satisfy the hardware's strict Warp Tensor Memory Access Check enabled by `-G`. In release mode the hardware auto-routes lanes, so the fix is guarded to avoid unnecessary overhead. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
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#374
No description provided.