[GH-ISSUE #298] how to understand the rdma buffer #211

Closed
opened 2026-10-26 09:56:43 +00:00 by bsakaguchi · 4 comments
Owner

Originally created by @kzlxd on GitHub (Jul 14, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/298

Each rank has its own rdma_buffer_ptr, which rank's rdma buffer will be used for rdma transfers, and will only nvl rank 0's rdma buffer be selected?

Originally created by @kzlxd on GitHub (Jul 14, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/298 Each rank has its own rdma_buffer_ptr, which rank's rdma buffer will be used for rdma transfers, and will only nvl rank 0's rdma buffer be selected?
Author
Owner

@sphish commented on GitHub (Jul 15, 2025):

Which RDMA rank you want to send to/recv from, you need to use the RDMA buffer of that RDMA rank.

<!-- gh-comment-id:3071620763 --> @sphish commented on GitHub (Jul 15, 2025): Which RDMA rank you want to send to/recv from, you need to use the RDMA buffer of that RDMA rank.
Author
Owner

@kzlxd commented on GitHub (Jul 15, 2025):

During the data transfer of the kRDMAReceiver, only data from different rdma ranks were combine. There is no local rank distribution. Is it possible to understand that the sender knows which rank to send to @sphish

<!-- gh-comment-id:3071674319 --> @kzlxd commented on GitHub (Jul 15, 2025): During the data transfer of the kRDMAReceiver, only data from different rdma ranks were combine. There is no local rank distribution. Is it possible to understand that the sender knows which rank to send to @sphish
Author
Owner

@sphish commented on GitHub (Jul 15, 2025):

Is it possible to understand that the sender knows which rank to send to

Yes

<!-- gh-comment-id:3071750659 --> @sphish commented on GitHub (Jul 15, 2025): > Is it possible to understand that the sender knows which rank to send to Yes
Author
Owner

@kzlxd commented on GitHub (Jul 15, 2025):

Does the rdma recv_buffer of a rank in normal mode map to the location of the same card number on the opposite node? @sphish

<!-- gh-comment-id:3072340392 --> @kzlxd commented on GitHub (Jul 15, 2025): Does the rdma recv_buffer of a rank in normal mode map to the location of the same card number on the opposite node? @sphish
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#211
No description provided.