[PR #665] Support multiple QPs for PP send/recv #692

New issue

Open

opened 2026-10-27 21:05:19 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:05:19 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/665
Author: @jm99k56
Created: 6/15/2026
Status: 🔄 Open

Base: main ← Head: pp-multi-qps

📝 Commits (1)

2297b88 Support multiple QPs for PP send/recv

📊 Changes

5 files changed (+106 additions, -32 deletions)

View changed files

📝 csrc/elastic/buffer.hpp (+15 -6)
📝 csrc/kernels/elastic/pp_send_recv.hpp (+10 -2)
📝 deep_ep/buffers/elastic.py (+14 -2)
📝 deep_ep/include/deep_ep/impls/pp_send_recv.cuh (+62 -21)
📝 tests/elastic/test_pp.py (+5 -1)

📄 Description

In bonding setups, the current PP send/recv path only uses a single QP, which can lead to uneven traffic distribution across ports. This PR adds multi-QP support for PP send/recv so that PP traffic can be spread more evenly across available ports.

_{🔄 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/665 **Author:** [@jm99k56](https://github.com/jm99k56) **Created:** 6/15/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `pp-multi-qps` --- ### 📝 Commits (1) - [`2297b88`](https://github.com/deepseek-ai/DeepEP/commit/2297b88a85a664db4fb088c4c9044add7150cd29) Support multiple QPs for PP send/recv ### 📊 Changes **5 files changed** (+106 additions, -32 deletions) <details> <summary>View changed files</summary> 📝 `csrc/elastic/buffer.hpp` (+15 -6) 📝 `csrc/kernels/elastic/pp_send_recv.hpp` (+10 -2) 📝 `deep_ep/buffers/elastic.py` (+14 -2) 📝 `deep_ep/include/deep_ep/impls/pp_send_recv.cuh` (+62 -21) 📝 `tests/elastic/test_pp.py` (+5 -1) </details> ### 📄 Description In bonding setups, the current PP send/recv path only uses a single QP, which can lead to uneven traffic distribution across ports. This PR adds multi-QP support for PP send/recv so that PP traffic can be spread more evenly across available ports. --- <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 21:05:19 +00:00

bsakaguchi referenced this issue

2026-10-27 21:05:28 +00:00

[PR #692] Fix JIT failure on SM100+ with CUDA <= 13.0: use 64-bit size operand for st.bulk #710