[PR #198] [MERGED] Make various updates and fixes #287

New issue

Closed

opened 2026-10-26 21:15:49 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:15:49 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/198
Author: @RayWang96
Created: 9/25/2025
Status: ✅ Merged
Merged: 9/25/2025
Merged by: @LyricZhao

Base: main ← Head: perf_update_ray_wang

📝 Commits (1)

ec6bbdc Make various updates and fixes

📊 Changes

45 files changed (+3316 additions, -1095 deletions)

View changed files

➕ csrc/apis/attention.hpp (+77 -0)
➕ csrc/apis/einsum.hpp (+115 -0)
📝 csrc/apis/gemm.hpp (+113 -1)
📝 csrc/apis/layout.hpp (+2 -2)
📝 csrc/indexing/main.cu (+8 -1)
📝 csrc/jit/device_runtime.hpp (+22 -1)
📝 csrc/jit_kernels/heuristics/common.hpp (+14 -12)
📝 csrc/jit_kernels/heuristics/sm100.hpp (+20 -2)
📝 csrc/jit_kernels/heuristics/sm90.hpp (+39 -13)
➕ csrc/jit_kernels/impls/epilogue.hpp (+12 -0)
📝 csrc/jit_kernels/impls/runtime_utils.hpp (+64 -17)
📝 csrc/jit_kernels/impls/sm100_bf16_gemm.hpp (+4 -5)
➕ csrc/jit_kernels/impls/sm100_bmk_bnk_mn.hpp (+137 -0)
📝 csrc/jit_kernels/impls/sm100_fp8_gemm_1d1d.hpp (+16 -6)
📝 csrc/jit_kernels/impls/sm100_fp8_gemm_1d2d.hpp (+11 -4)
📝 csrc/jit_kernels/impls/sm90_bf16_gemm.hpp (+6 -5)
➕ csrc/jit_kernels/impls/sm90_bmk_bnk_mn.hpp (+131 -0)
➕ csrc/jit_kernels/impls/sm90_fp8_gemm_1d1d.hpp (+214 -0)
📝 csrc/jit_kernels/impls/sm90_fp8_gemm_1d2d.hpp (+12 -4)
➕ csrc/jit_kernels/impls/smxx_cublaslt.hpp (+151 -0)

...and 25 more files

📄 Description

This PR is submitted by NVIDIA. Optimized some kernels on SM100 and added some kernels for DeepSeek V3 inference.

_{🔄 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/198 **Author:** [@RayWang96](https://github.com/RayWang96) **Created:** 9/25/2025 **Status:** ✅ Merged **Merged:** 9/25/2025 **Merged by:** [@LyricZhao](https://github.com/LyricZhao) **Base:** `main` ← **Head:** `perf_update_ray_wang` --- ### 📝 Commits (1) - [`ec6bbdc`](https://github.com/deepseek-ai/DeepGEMM/commit/ec6bbdccdc8a48ae40a7998ada20692183d4c2d5) Make various updates and fixes ### 📊 Changes **45 files changed** (+3316 additions, -1095 deletions) <details> <summary>View changed files</summary> ➕ `csrc/apis/attention.hpp` (+77 -0) ➕ `csrc/apis/einsum.hpp` (+115 -0) 📝 `csrc/apis/gemm.hpp` (+113 -1) 📝 `csrc/apis/layout.hpp` (+2 -2) 📝 `csrc/indexing/main.cu` (+8 -1) 📝 `csrc/jit/device_runtime.hpp` (+22 -1) 📝 `csrc/jit_kernels/heuristics/common.hpp` (+14 -12) 📝 `csrc/jit_kernels/heuristics/sm100.hpp` (+20 -2) 📝 `csrc/jit_kernels/heuristics/sm90.hpp` (+39 -13) ➕ `csrc/jit_kernels/impls/epilogue.hpp` (+12 -0) 📝 `csrc/jit_kernels/impls/runtime_utils.hpp` (+64 -17) 📝 `csrc/jit_kernels/impls/sm100_bf16_gemm.hpp` (+4 -5) ➕ `csrc/jit_kernels/impls/sm100_bmk_bnk_mn.hpp` (+137 -0) 📝 `csrc/jit_kernels/impls/sm100_fp8_gemm_1d1d.hpp` (+16 -6) 📝 `csrc/jit_kernels/impls/sm100_fp8_gemm_1d2d.hpp` (+11 -4) 📝 `csrc/jit_kernels/impls/sm90_bf16_gemm.hpp` (+6 -5) ➕ `csrc/jit_kernels/impls/sm90_bmk_bnk_mn.hpp` (+131 -0) ➕ `csrc/jit_kernels/impls/sm90_fp8_gemm_1d1d.hpp` (+214 -0) 📝 `csrc/jit_kernels/impls/sm90_fp8_gemm_1d2d.hpp` (+12 -4) ➕ `csrc/jit_kernels/impls/smxx_cublaslt.hpp` (+151 -0) _...and 25 more files_ </details> ### 📄 Description This PR is submitted by NVIDIA. Optimized some kernels on SM100 and added some kernels for DeepSeek V3 inference. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:15:49 +00:00