mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #566] Feature request: Enable multi-NIC RDMA support for GPU = 1:2 (CX8) systems in Hybrid-EP / GIN #347
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepEP#347
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @skeyyk on GitHub (Jan 18, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/566
On the hybrid-ep branch, on B300 systems equipped with CX8 NICs, where each GPU is connected to two RDMA NICs, inter-node communication during training appears to utilize only one NIC per GPU. As a result, the effective RDMA bandwidth per GPU does not scale with the available NIC resources, even though the hardware topology supports concurrent usage of multiple NICs. This behavior is consistently observed on B300 + CX8 platforms across training runs.
@Autumn1998 commented on GitHub (Jan 30, 2026):
Yes, this is a known issue for now. We are working on switching to a more mature backend instead of relying on the low-level API, and we will then try to support this case.
@Brook017 commented on GitHub (Apr 2, 2026):
@Autumn1998 Does it mean underlying doca api will no longer be used?
@Autumn1998 commented on GitHub (Apr 9, 2026):
We will still fix bugs, but there will be no new features.