[PR #276] [MERGED] Use TMA to optimize internode dispatch. #497

New issue

Closed

opened 2026-10-27 21:03:15 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:03:15 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/276
Author: @sphish
Created: 7/3/2025
Status: ✅ Merged
Merged: 7/4/2025
Merged by: @LyricZhao

Base: main ← Head: internode-dispatch-opt

📝 Commits (10+)

e4512c9 Add TMA buffer allocation
5997b20 Use TMA for forwarders and NVL receivers
8f33008 Use lane 31 to operate TMA.
e53131c Change rdma buffer layout.
5ee1488 Use TMA to transfer scales also.
bdda383 Increase the NVL recv buffer size.
bb7ac6c Disable early stopping.
2521ce3 Apply similar optimizations on receiver warps.
6fb34d4 Prevent warp divergence.
cd87868 Disable aggressive ptx by default.

📊 Changes

3 files changed (+118 additions, -90 deletions)

View changed files

📝 csrc/kernels/internode.cu (+116 -88)
📝 setup.py (+1 -1)
📝 tests/test_internode.py (+1 -1)

📄 Description

No description provided

_{🔄 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/276 **Author:** [@sphish](https://github.com/sphish) **Created:** 7/3/2025 **Status:** ✅ Merged **Merged:** 7/4/2025 **Merged by:** [@LyricZhao](https://github.com/LyricZhao) **Base:** `main` ← **Head:** `internode-dispatch-opt` --- ### 📝 Commits (10+) - [`e4512c9`](https://github.com/deepseek-ai/DeepEP/commit/e4512c9a03a976890d7341db8f3b23e25cd4533f) Add TMA buffer allocation - [`5997b20`](https://github.com/deepseek-ai/DeepEP/commit/5997b20311e86668503ebe49f2caed558c909294) Use TMA for forwarders and NVL receivers - [`8f33008`](https://github.com/deepseek-ai/DeepEP/commit/8f33008d59f326cef5931646de3c1e7bc6c33d46) Use lane 31 to operate TMA. - [`e53131c`](https://github.com/deepseek-ai/DeepEP/commit/e53131c6c995fb5a6d264c8828192b48e073c37d) Change rdma buffer layout. - [`5ee1488`](https://github.com/deepseek-ai/DeepEP/commit/5ee148881afff5bcf957cdbe5bef8e93646b92af) Use TMA to transfer scales also. - [`bdda383`](https://github.com/deepseek-ai/DeepEP/commit/bdda3833b22516ed1dc8e9403c0f6124e502aae2) Increase the NVL recv buffer size. - [`bb7ac6c`](https://github.com/deepseek-ai/DeepEP/commit/bb7ac6ce6dfde6851ccd47aee77355c2b3fdd68f) Disable early stopping. - [`2521ce3`](https://github.com/deepseek-ai/DeepEP/commit/2521ce30d5b5932abaa1bb5bb9169fe0225bd24f) Apply similar optimizations on receiver warps. - [`6fb34d4`](https://github.com/deepseek-ai/DeepEP/commit/6fb34d41d123a55eb1c09f7ff24270c6f18f81c0) Prevent warp divergence. - [`cd87868`](https://github.com/deepseek-ai/DeepEP/commit/cd87868158842ad4b9121d4fa11ae59991871676) Disable aggressive ptx by default. ### 📊 Changes **3 files changed** (+118 additions, -90 deletions) <details> <summary>View changed files</summary> 📝 `csrc/kernels/internode.cu` (+116 -88) 📝 `setup.py` (+1 -1) 📝 `tests/test_internode.py` (+1 -1) </details> ### 📄 Description _No description provided_ --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:03:15 +00:00