[PR #377] [MERGED] Public release 26/07 #399

New issue

Closed

opened 2026-10-26 21:17:15 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:17:15 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/377
Author: @guyan364
Created: 7/14/2026
Status: ✅ Merged
Merged: 7/15/2026
Merged by: @LyricZhao

Base: main ← Head: public-release-2607

📝 Commits (1)

61ca7f1 Public release 26/07

📊 Changes

30 files changed (+1909 additions, -1142 deletions)

View changed files

📝 csrc/apis/attention.hpp (+89 -130)
📝 csrc/apis/einsum.hpp (+1 -1)
📝 csrc/apis/mega.hpp (+152 -92)
📝 csrc/jit_kernels/heuristics/mega_moe.hpp (+13 -85)
📝 csrc/jit_kernels/impls/sm100_bf16_mega_moe.hpp (+58 -7)
📝 csrc/jit_kernels/impls/sm100_fp8_fp4_mega_moe.hpp (+95 -6)
📝 csrc/jit_kernels/impls/sm100_mqa_logits.hpp (+116 -143)
📝 csrc/jit_kernels/impls/sm90_fp8_mqa_logits.hpp (+2 -1)
📝 csrc/jit_kernels/impls/smxx_layout.hpp (+1 -1)
📝 csrc/utils/layout.hpp (+10 -0)
📝 deep_gemm/__init__.py (+2 -0)
📝 deep_gemm/include/deep_gemm/impls/sm100_bf16_mega_moe.cuh (+166 -125)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh (+239 -171)
📝 deep_gemm/include/deep_gemm/impls/sm100_mqa_logits.cuh (+48 -48)
📝 deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d1d.cuh (+4 -5)
📝 deep_gemm/include/deep_gemm/layout/mega_moe.cuh (+167 -4)
📝 deep_gemm/include/deep_gemm/layout/mqa_logits.cuh (+21 -22)
📝 deep_gemm/include/deep_gemm/mma/sm100.cuh (+23 -7)
📝 deep_gemm/include/deep_gemm/ptx/ld_st.cuh (+40 -0)
📝 deep_gemm/include/deep_gemm/ptx/utils.cuh (+2 -0)

...and 10 more files

📄 Description

Support MXFP8 indexer
MegaMoE optimization, 20% faster than last version, 10%+ for E2E inference
- New scheduler, interleaving L1/L2 blocks for better overlapping
- Fuse shared experts into MegaMoE
- Now 11 kernels into 1

_{🔄 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/377 **Author:** [@guyan364](https://github.com/guyan364) **Created:** 7/14/2026 **Status:** ✅ Merged **Merged:** 7/15/2026 **Merged by:** [@LyricZhao](https://github.com/LyricZhao) **Base:** `main` ← **Head:** `public-release-2607` --- ### 📝 Commits (1) - [`61ca7f1`](https://github.com/deepseek-ai/DeepGEMM/commit/61ca7f1378038898f76b859b4c50541f27dbf891) Public release 26/07 ### 📊 Changes **30 files changed** (+1909 additions, -1142 deletions) <details> <summary>View changed files</summary> 📝 `csrc/apis/attention.hpp` (+89 -130) 📝 `csrc/apis/einsum.hpp` (+1 -1) 📝 `csrc/apis/mega.hpp` (+152 -92) 📝 `csrc/jit_kernels/heuristics/mega_moe.hpp` (+13 -85) 📝 `csrc/jit_kernels/impls/sm100_bf16_mega_moe.hpp` (+58 -7) 📝 `csrc/jit_kernels/impls/sm100_fp8_fp4_mega_moe.hpp` (+95 -6) 📝 `csrc/jit_kernels/impls/sm100_mqa_logits.hpp` (+116 -143) 📝 `csrc/jit_kernels/impls/sm90_fp8_mqa_logits.hpp` (+2 -1) 📝 `csrc/jit_kernels/impls/smxx_layout.hpp` (+1 -1) 📝 `csrc/utils/layout.hpp` (+10 -0) 📝 `deep_gemm/__init__.py` (+2 -0) 📝 `deep_gemm/include/deep_gemm/impls/sm100_bf16_mega_moe.cuh` (+166 -125) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh` (+239 -171) 📝 `deep_gemm/include/deep_gemm/impls/sm100_mqa_logits.cuh` (+48 -48) 📝 `deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d1d.cuh` (+4 -5) 📝 `deep_gemm/include/deep_gemm/layout/mega_moe.cuh` (+167 -4) 📝 `deep_gemm/include/deep_gemm/layout/mqa_logits.cuh` (+21 -22) 📝 `deep_gemm/include/deep_gemm/mma/sm100.cuh` (+23 -7) 📝 `deep_gemm/include/deep_gemm/ptx/ld_st.cuh` (+40 -0) 📝 `deep_gemm/include/deep_gemm/ptx/utils.cuh` (+2 -0) _...and 10 more files_ </details> ### 📄 Description - Support MXFP8 indexer - MegaMoE optimization, 20% faster than last version, 10%+ for E2E inference - New scheduler, interleaving L1/L2 blocks for better overlapping - Fuse shared experts into MegaMoE - Now 11 kernels into 1 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:17:15 +00:00