[PR #782] Require bucket bandwidths only for hybrid groups #780

Open
opened 2026-10-27 21:06:15 +00:00 by bsakaguchi · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/782
Author: @lvren1485
Created: 10/15/2026
Status: 🔄 Open

Base: main ← Head: bucket-optional-bandwidth


📝 Commits (1)

  • d6f0bf7 Require bucket bandwidths only for hybrid groups

📊 Changes

3 files changed (+13 additions, -4 deletions)

View changed files

📝 README.md (+5 -1)
📝 csrc/buffers/bucket.hpp (+3 -1)
📝 deep_ep/buffers/bucket.py (+5 -2)

📄 Description

BucketBuffer asserts that both NVLink and RDMA bandwidth are detected, even when the group is NVLink-only or RDMA-only. On a single node without ibstat this makes BucketBuffer unusable, and the README documents it as a known limitation.

The bandwidths are only used to plan the hybrid all-gather pipeline (csrc/kernels/bucket/all_gather.hpp), so this skips the detection unless some group actually spans both NVLink and RDMA, and moves the C++ check into the per-group loop for hybrid groups only.

Also added the missing correctness tests to the README test list.


🔄 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/782 **Author:** [@lvren1485](https://github.com/lvren1485) **Created:** 10/15/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `bucket-optional-bandwidth` --- ### 📝 Commits (1) - [`d6f0bf7`](https://github.com/deepseek-ai/DeepEP/commit/d6f0bf7fe1121fc0e8fae9c0a769cf00eb7cfb0c) Require bucket bandwidths only for hybrid groups ### 📊 Changes **3 files changed** (+13 additions, -4 deletions) <details> <summary>View changed files</summary> 📝 `README.md` (+5 -1) 📝 `csrc/buffers/bucket.hpp` (+3 -1) 📝 `deep_ep/buffers/bucket.py` (+5 -2) </details> ### 📄 Description BucketBuffer asserts that both NVLink and RDMA bandwidth are detected, even when the group is NVLink-only or RDMA-only. On a single node without `ibstat` this makes BucketBuffer unusable, and the README documents it as a known limitation. The bandwidths are only used to plan the hybrid all-gather pipeline (`csrc/kernels/bucket/all_gather.hpp`), so this skips the detection unless some group actually spans both NVLink and RDMA, and moves the C++ check into the per-group loop for hybrid groups only. Also added the missing correctness tests to the README test list. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepEP#780
No description provided.