mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #64] why internode normal kernel don't use ibgda? #62
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#62
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 @zhang662817 on GitHub (Mar 10, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/64
we know ibgda have many benefits, as described in .
Use IBRC, CPU is on the critical path， does this decrease throughput and increase latency?
Why internode kernel don't directly use ibgda by nvshmem, and low latency kernel use ibgda between nodes?
Thanks.
@zhang662817 commented on GitHub (Mar 10, 2025):
as described in https://zhuanlan.zhihu.com/p/689043376
Using IBRC, how to get these benefits in
https://docs.nvidia.com/nvshmem/api/introduction.html#gpu-initiated-communication-and-strong-scaling ?
@sphish commented on GitHub (Mar 11, 2025):
In normal kernels, the network transmission chunks are relatively large. Even with the use of IBRC, performance is sufficient, and IBRC is easier to use.
@zhang662817 commented on GitHub (Mar 11, 2025):
Thanks
@MengYu10151 commented on GitHub (Apr 9, 2025):
Hi @sphish,if we putting aside the ease of use, just forcing IBGDA to be enabled in internode will bring any defects?
IBGDA is indeed more suitable for small data, but I don't know much about the impact of large data transmission. thx!
@sphish commented on GitHub (Apr 9, 2025):
@MengYu10151 There won't be any defects, but you need to use it more carefully because IBGDA utilizes multiple QPs for each pe.