[PR #277] Allow using few SMs for low-latency mode #498

New issue

Open

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/277
Author: @fzyzcjy
Created: 7/3/2025
Status: 🔄 Open

Base: main ← Head: feat/few_sm_deepep

📝 Commits (1)

8d1c641 more

📊 Changes

1 file changed (+6 additions, -2 deletions)

View changed files

📝 csrc/kernels/internode_ll.cu (+6 -2)

📄 Description

The code diff is surely not for merging, but for demonstration how the experiments below are done. If anyone is interested / this direction looks acceptable to be merged, I am happy to polish and further work on the code!

The code and experiment data are extracted from old experiments for my previous https://github.com/deepseek-ai/DeepEP/pull/249.

Figure 1: num-sm vs performance
As can be seen, when using 9 warpgroup - ie few SMs, the performance only slightly slow down. Thus this makes a simple overlapping between this and computation feasible.

For dispatch we may need to do extra work though, since the warp specialization may be suboptimal when there are few SMs.

_{🔄 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/277 **Author:** [@fzyzcjy](https://github.com/fzyzcjy) **Created:** 7/3/2025 **Status:** 🔄 Open **Base:** `main` ← **Head:** `feat/few_sm_deepep` --- ### 📝 Commits (1) - [`8d1c641`](https://github.com/deepseek-ai/DeepEP/commit/8d1c641851f93d736283c01bc1548512d0f96067) more ### 📊 Changes **1 file changed** (+6 additions, -2 deletions) <details> <summary>View changed files</summary> 📝 `csrc/kernels/internode_ll.cu` (+6 -2) </details> ### 📄 Description The code diff is surely not for merging, but for demonstration how the experiments below are done. If anyone is interested / this direction looks acceptable to be merged, I am happy to polish and further work on the code! The code and experiment data are extracted from old experiments for my previous https://github.com/deepseek-ai/DeepEP/pull/249. Figure 1: num-sm vs performance As can be seen, when using 9 warpgroup - ie few SMs, the performance only slightly slow down. Thus this makes a simple overlapping between this and computation feasible. ![image](https://github.com/user-attachments/assets/be50bf75-3c42-4ad4-8aec-0b9c665ebbc5) For dispatch we may need to do extra work though, since the warp specialization may be suboptimal when there are few SMs. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

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

bsakaguchi referenced this issue

2026-10-27 21:04:24 +00:00

[PR #498] [CLOSED] Fix tensor descriptions in buffer.py #597