[PR #385] [MERGED] Speed up dispatch send by refining loop unrolling #545

New issue

Closed

opened 2026-10-27 21:03:51 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:03:51 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/385
Author: @fzyzcjy
Created: 8/30/2025
Status: ✅ Merged
Merged: 9/1/2025
Merged by: @LyricZhao

Base: main ← Head: feat/loop_unroll

📝 Commits (3)

3e1b28d nits
c7d80de hack unrolled warp copy
262d20e Revert "nits"

📊 Changes

1 file changed (+15 additions, -2 deletions)

View changed files

📝 csrc/kernels/utils.cuh (+15 -2)

📄 Description

I am mainly interested in 768 tok 48 rank nvfp4 pure nvlink case and it seems to have some speed up. Do not have time to check this fp8 case now, will check later.

		dispatch_a	dispatch_b	combine_a	combine_b
baseline	128tok, 4rank	21-24	7-8	28-29	9-10
	128tok, 48rank	28-31	7-8	31-35	9-10
	768tok, 48rank	94-104	17-20	131-137	20-22
pr	128tok, 4rank	21-24	7-8	28-30	9-10
	128tok, 48rank	26-32	7-8	32-35	9-10
	768tok, 48rank	89-97 (speedup)	18-21	131-138	20-21

_{🔄 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/385 **Author:** [@fzyzcjy](https://github.com/fzyzcjy) **Created:** 8/30/2025 **Status:** ✅ Merged **Merged:** 9/1/2025 **Merged by:** [@LyricZhao](https://github.com/LyricZhao) **Base:** `main` ← **Head:** `feat/loop_unroll` --- ### 📝 Commits (3) - [`3e1b28d`](https://github.com/deepseek-ai/DeepEP/commit/3e1b28d9b17f2c1cc46403d432ca576dbf15bd45) nits - [`c7d80de`](https://github.com/deepseek-ai/DeepEP/commit/c7d80dec9814f18a02ae850df290c679bb94c7e8) hack unrolled warp copy - [`262d20e`](https://github.com/deepseek-ai/DeepEP/commit/262d20e1154e916398a9d0a4f320d3e218d1dc6b) Revert "nits" ### 📊 Changes **1 file changed** (+15 additions, -2 deletions) <details> <summary>View changed files</summary> 📝 `csrc/kernels/utils.cuh` (+15 -2) </details> ### 📄 Description I am mainly interested in 768 tok 48 rank nvfp4 pure nvlink case and it seems to have some speed up. Do not have time to check this fp8 case now, will check later. | | | dispatch_a | dispatch_b | combine_a | combine_b | | -------- | -------------- | ---------- | ---------- | --------- | --------- | | baseline | 128tok, 4rank | 21-24 | 7-8 | 28-29 | 9-10 | | | 128tok, 48rank | 28-31 | 7-8 | 31-35 | 9-10 | | | 768tok, 48rank | 94-104 | 17-20 | 131-137 | 20-22 | | pr | 128tok, 4rank | 21-24 | 7-8 | 28-30 | 9-10 | | | 128tok, 48rank | 26-32 | 7-8 | 32-35 | 9-10 | | | 768tok, 48rank | 89-97 (speedup) | 18-21 | 131-138 | 20-21 | --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:03:51 +00:00