[PR #303] [CLOSED] Eliminate redundant global stores in SM90 MQA logits kernels #354

New issue

Closed

opened 2026-10-26 21:16:36 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:16:36 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/303
Author: @Gregory-Pereira
Created: 4/11/2026
Status: ❌ Closed

Base: main ← Head: fix-279/optimize-SM90_fp8_mqa_logits

📝 Commits (4)

2f9199d Eliminate redundant global stores in SM90 MQA logits kernels
69a2f5d remove reduced writes and increased num_kv_stages = 5
515009a Halve SM90 math threads and BLOCK_KV to fit 2 threadblocks per SM, moving weights from registers to on-demand shared memory loads to stay within register budget
bf3643e keep 1 block per SM, but with halved threads (256 math) and halved BLOCK_KV (128)

📊 Changes

2 files changed (+11 additions, -13 deletions)

View changed files

📝 csrc/jit_kernels/impls/smxx_fp8_mqa_logits.hpp (+5 -2)
📝 deep_gemm/include/deep_gemm/impls/sm90_fp8_mqa_logits.cuh (+6 -11)

📄 Description

Seeks to address: #279

_{🔄 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/303 **Author:** [@Gregory-Pereira](https://github.com/Gregory-Pereira) **Created:** 4/11/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `fix-279/optimize-SM90_fp8_mqa_logits` --- ### 📝 Commits (4) - [`2f9199d`](https://github.com/deepseek-ai/DeepGEMM/commit/2f9199d8871c3af0878d6b7177ee7b06485d6bcd) Eliminate redundant global stores in SM90 MQA logits kernels - [`69a2f5d`](https://github.com/deepseek-ai/DeepGEMM/commit/69a2f5dde48dda12f07204bd9ffee7cddcd06a1c) remove reduced writes and increased num_kv_stages = 5 - [`515009a`](https://github.com/deepseek-ai/DeepGEMM/commit/515009a5c2ef260349f9e537fa272ec80ae9adb5) Halve SM90 math threads and BLOCK_KV to fit 2 threadblocks per SM, moving weights from registers to on-demand shared memory loads to stay within register budget - [`bf3643e`](https://github.com/deepseek-ai/DeepGEMM/commit/bf3643e7b116e224c14597db4fd353350658aa1f) keep 1 block per SM, but with halved threads (256 math) and halved BLOCK_KV (128) ### 📊 Changes **2 files changed** (+11 additions, -13 deletions) <details> <summary>View changed files</summary> 📝 `csrc/jit_kernels/impls/smxx_fp8_mqa_logits.hpp` (+5 -2) 📝 `deep_gemm/include/deep_gemm/impls/sm90_fp8_mqa_logits.cuh` (+6 -11) </details> ### 📄 Description Seeks to address: #279 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:16:36 +00:00