[GH-ISSUE #279] Ask a question #203

Closed
opened 2026-10-26 09:56:33 +00:00 by bsakaguchi · 0 comments
Owner

Originally created by @kzlxd on GitHub (Jul 4, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/279

    // Shift cached head
    send_nvl_head += src_rdma_channel_prefix * NUM_MAX_NVL_PEERS + dst_nvl_rank;

I would like to ask why send_nvl_head is so offset, Each nvl rank needs to receive tokens from different rdma ranks. Why not offset it according to num_tokens_to_recv_from_rdma?

Originally created by @kzlxd on GitHub (Jul 4, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/279 // Shift cached head send_nvl_head += src_rdma_channel_prefix * NUM_MAX_NVL_PEERS + dst_nvl_rank; I would like to ask why send_nvl_head is so offset, Each nvl rank needs to receive tokens from different rdma ranks. Why not offset it according to num_tokens_to_recv_from_rdma?
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#203
No description provided.