deepseek-ai/DeepGEMM

Fork 0

mirror of https://github.com/deepseek-ai/DeepGEMM.git synced 2026-11-05 00:47:26 +00:00

[PR #270] [MERGED] fix: use SM90ArchSpec instead of SM100ArchSpec in sm90_bf16_k_grouped_gemm #336

New issue

Closed

opened 2026-10-26 21:16:24 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:16:24 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/270
Author: @yurekami
Created: 12/31/2025
Status: ✅ Merged
Merged: 1/6/2026
Merged by: @zheanxu

Base: main ← Head: fix/sm90-archspec-bug

📝 Commits (1)

6be0eb3 fix: use SM90ArchSpec instead of SM100ArchSpec in sm90_bf16_k_grouped_gemm

📊 Changes

1 file changed (+4 additions, -4 deletions)

View changed files

📝 csrc/jit_kernels/impls/sm90_bf16_gemm.hpp (+4 -4)

📄 Description

Summary

This PR fixes a copy-paste error in sm90_bf16_k_grouped_gemm where SM100ArchSpec was incorrectly used instead of SM90ArchSpec for TMA descriptor block size calculations.

Problem

In csrc/jit_kernels/impls/sm90_bf16_gemm.hpp, the function sm90_bf16_k_grouped_gemm (lines 259-274) was using SM100ArchSpec to calculate block sizes for TMA descriptors:

SM100ArchSpec::get_ab_load_block_m(...)
SM100ArchSpec::get_ab_load_block_n(...)
SM100ArchSpec::get_cd_store_block_m(...)
SM100ArchSpec::get_cd_store_block_n(...)

Since this file is the SM90 (Hopper) implementation, it should use SM90ArchSpec consistently, like all other functions in the same file (sm90_bf16_gemm, sm90_m_grouped_bf16_gemm_contiguous, sm90_bf16_m_grouped_gemm_masked, etc.).

Fix

Replace all SM100ArchSpec references with SM90ArchSpec in the sm90_bf16_k_grouped_gemm function.

Testing

This change affects block size calculations for TMA descriptors on SM90 GPUs. The fix ensures consistent architecture-specific behavior matching the function's intended target (SM90/Hopper).

Fixes #242

_{🔄 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/270 **Author:** [@yurekami](https://github.com/yurekami) **Created:** 12/31/2025 **Status:** ✅ Merged **Merged:** 1/6/2026 **Merged by:** [@zheanxu](https://github.com/zheanxu) **Base:** `main` ← **Head:** `fix/sm90-archspec-bug` --- ### 📝 Commits (1) - [`6be0eb3`](https://github.com/deepseek-ai/DeepGEMM/commit/6be0eb31d9bf897b2396157a410fc06d8e5b0cdc) fix: use SM90ArchSpec instead of SM100ArchSpec in sm90_bf16_k_grouped_gemm ### 📊 Changes **1 file changed** (+4 additions, -4 deletions) <details> <summary>View changed files</summary> 📝 `csrc/jit_kernels/impls/sm90_bf16_gemm.hpp` (+4 -4) </details> ### 📄 Description ## Summary This PR fixes a copy-paste error in `sm90_bf16_k_grouped_gemm` where `SM100ArchSpec` was incorrectly used instead of `SM90ArchSpec` for TMA descriptor block size calculations. ## Problem In `csrc/jit_kernels/impls/sm90_bf16_gemm.hpp`, the function `sm90_bf16_k_grouped_gemm` (lines 259-274) was using `SM100ArchSpec` to calculate block sizes for TMA descriptors: ```cpp SM100ArchSpec::get_ab_load_block_m(...) SM100ArchSpec::get_ab_load_block_n(...) SM100ArchSpec::get_cd_store_block_m(...) SM100ArchSpec::get_cd_store_block_n(...) ``` Since this file is the SM90 (Hopper) implementation, it should use `SM90ArchSpec` consistently, like all other functions in the same file (`sm90_bf16_gemm`, `sm90_m_grouped_bf16_gemm_contiguous`, `sm90_bf16_m_grouped_gemm_masked`, etc.). ## Fix Replace all `SM100ArchSpec` references with `SM90ArchSpec` in the `sm90_bf16_k_grouped_gemm` function. ## Testing This change affects block size calculations for TMA descriptors on SM90 GPUs. The fix ensures consistent architecture-specific behavior matching the function's intended target (SM90/Hopper). Fixes #242 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:16:24 +00:00

closed this issue
added the
pull-request
label

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#336

No description provided.

Rows
Columns