[PR #191] [MERGED] Use TMA instead of LD/ST for intra-node normal kernels #456

New issue

Closed

opened 2026-10-26 10:00:58 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 10:00:58 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/191
Author: @LyricZhao
Created: 6/6/2025
Status: ✅ Merged
Merged: 6/6/2025
Merged by: @LyricZhao

Base: main ← Head: intra-tma

📝 Commits (10+)

047d9b8 Update CMake files
a5c9c10 Use TMA instead of LD/ST for intranode dispatch
ada6e4e Use TMA instead of LD/ST for intranode combine
8d7e20b Adjust configs
18785fe Test default configs as well
031a5c4 More warps for combine
d0f0d40 Add inter-thread fence
80b4624 Enable more warps
2f29a08 Do not use TMA for senders
04b06f7 Update configs

📊 Changes

6 files changed (+232 additions, -89 deletions)

View changed files

📝 csrc/CMakeLists.txt (+5 -3)
📝 csrc/kernels/configs.cuh (+0 -2)
📝 csrc/kernels/intranode.cu (+136 -70)
📝 csrc/kernels/utils.cuh (+61 -0)
📝 deep_ep/buffer.py (+5 -5)
📝 tests/test_intranode.py (+25 -9)

📄 Description

Better performance for intranode EP kernels on H800/H20.

_{🔄 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/DeepEP/pull/191 **Author:** [@LyricZhao](https://github.com/LyricZhao) **Created:** 6/6/2025 **Status:** ✅ Merged **Merged:** 6/6/2025 **Merged by:** [@LyricZhao](https://github.com/LyricZhao) **Base:** `main` ← **Head:** `intra-tma` --- ### 📝 Commits (10+) - [`047d9b8`](https://github.com/deepseek-ai/DeepEP/commit/047d9b8bfbea1098c722c570d59f7d583abbf701) Update CMake files - [`a5c9c10`](https://github.com/deepseek-ai/DeepEP/commit/a5c9c10b1444e6ed4ab3027540ce64f71e6aa75a) Use TMA instead of LD/ST for intranode dispatch - [`ada6e4e`](https://github.com/deepseek-ai/DeepEP/commit/ada6e4eae22d5681c8fe0ce198f264f919d4714f) Use TMA instead of LD/ST for intranode combine - [`8d7e20b`](https://github.com/deepseek-ai/DeepEP/commit/8d7e20b997141ddbfae316de3b98e0da6d3a365f) Adjust configs - [`18785fe`](https://github.com/deepseek-ai/DeepEP/commit/18785fe44c209267e2a8758810fbe012b30045b6) Test default configs as well - [`031a5c4`](https://github.com/deepseek-ai/DeepEP/commit/031a5c4920082605fc7c0bec0f081d4388930e17) More warps for combine - [`d0f0d40`](https://github.com/deepseek-ai/DeepEP/commit/d0f0d407673fd15193bc3d4cbeb8ffefb43bb22d) Add inter-thread fence - [`80b4624`](https://github.com/deepseek-ai/DeepEP/commit/80b46242948f8fbd9a9532a6020cc7d9442d1f75) Enable more warps - [`2f29a08`](https://github.com/deepseek-ai/DeepEP/commit/2f29a0833d975ca4e559b8fd624fa5b0b033290d) Do not use TMA for senders - [`04b06f7`](https://github.com/deepseek-ai/DeepEP/commit/04b06f7a299c8be3e97ec1dee89e481ceb08ac10) Update configs ### 📊 Changes **6 files changed** (+232 additions, -89 deletions) <details> <summary>View changed files</summary> 📝 `csrc/CMakeLists.txt` (+5 -3) 📝 `csrc/kernels/configs.cuh` (+0 -2) 📝 `csrc/kernels/intranode.cu` (+136 -70) 📝 `csrc/kernels/utils.cuh` (+61 -0) 📝 `deep_ep/buffer.py` (+5 -5) 📝 `tests/test_intranode.py` (+25 -9) </details> ### 📄 Description Better performance for intranode EP kernels on H800/H20. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 10:00:58 +00:00