[PR #96] [MERGED] Adjust kNumThreads of notify_dispatch #433

New issue

Closed

opened 2026-10-26 10:00:41 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 10:00:41 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/96
Author: @songhexiang
Created: 3/28/2025
Status: ✅ Merged
Merged: 3/28/2025
Merged by: @LyricZhao

Base: main ← Head: adjust_kNumThreads_of_notify_dispatch

📝 Commits (1)

4dd1e68 For the SMs which calculate metadata in notify_dispatch, each warp in the SM is used to calculate the metadata of one channel. The default configuration is 8 warps for 10 channels, which needs two rounds of loop. Maybe the number of warps can be configured to the number of the channels so that one loop is enough.

📊 Changes

1 file changed (+1 additions, -1 deletions)

View changed files

📝 csrc/kernels/internode.cu (+1 -1)

📄 Description

For the SMs which calculate metadata in notify_dispatch, each warp in the SM is used to calculate the metadata of one channel.
The default configuration is 8 warps for 10 channels, which needs two rounds of loop.
Maybe the number of warps can be configured to the number of the channels so that one loop is enough.

_{🔄 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/96 **Author:** [@songhexiang](https://github.com/songhexiang) **Created:** 3/28/2025 **Status:** ✅ Merged **Merged:** 3/28/2025 **Merged by:** [@LyricZhao](https://github.com/LyricZhao) **Base:** `main` ← **Head:** `adjust_kNumThreads_of_notify_dispatch` --- ### 📝 Commits (1) - [`4dd1e68`](https://github.com/deepseek-ai/DeepEP/commit/4dd1e68ac81c8fb63243bcfbbcf942eae5243210) For the SMs which calculate metadata in notify_dispatch, each warp in the SM is used to calculate the metadata of one channel. The default configuration is 8 warps for 10 channels, which needs two rounds of loop. Maybe the number of warps can be configured to the number of the channels so that one loop is enough. ### 📊 Changes **1 file changed** (+1 additions, -1 deletions) <details> <summary>View changed files</summary> 📝 `csrc/kernels/internode.cu` (+1 -1) </details> ### 📄 Description For the SMs which calculate metadata in notify_dispatch, each warp in the SM is used to calculate the metadata of one channel. The default configuration is **8 warps for 10 channels**, which needs two rounds of loop. Maybe the number of warps can be configured to **the number of the channels** so that one loop is enough. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 10:00:41 +00:00