[GH-ISSUE #26] Why "without costing any GPU SMs from the computation part" #25

New issue

Closed

opened 2026-10-26 09:52:43 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:52:43 +00:00

Owner

Originally created by @zhang662817 on GitHub (Feb 26, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/26

IBGDA dispatch kernels are running in backgound, Why not take up computing SMs resources?

Thanks.

Originally created by @zhang662817 on GitHub (Feb 26, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/26 ![Image](https://github.com/user-attachments/assets/c3df825b-4099-40a0-b3fa-3e158f28c7a9) IBGDA dispatch kernels are running in backgound, Why not take up computing SMs resources? Thanks.

bsakaguchi referenced this issue

2026-10-26 09:52:43 +00:00

[GH-ISSUE #25] NVlink latency #26

bsakaguchi closed this issue

2026-10-26 09:52:43 +00:00

bsakaguchi commented

2026-10-26 09:52:45 +00:00

Author

Owner

@xiaoyao9933 commented on GitHub (Feb 26, 2025):

In Low Latency mode, rdma_recv_buff is big enough to hold all limited tokens from every rank and to every expert. thus buffer backpressure is unnecessary. Therefore, the sender role's kernel exits immediately after completing one-sided transmission. Only when the hook function is invoked, the remaining receiver role kernel will be executed to finalize the copy operation. Thus, they likely mean that during computation (moe kernel) phases, there's no need for continuous polling by the receiver role kernel, reducing SM occupancy. You can invoke recv role kernel after moe kernel is finished in a lazy behaviour.

It's my personal understanding, not from deepseek team.

@xiaoyao9933 commented on GitHub (Feb 26, 2025): In Low Latency mode, rdma_recv_buff is big enough to hold all limited tokens from every rank and to every expert. thus buffer backpressure is unnecessary. Therefore, the sender role's kernel exits immediately after completing one-sided transmission. Only when the hook function is invoked, the remaining receiver role kernel will be executed to finalize the copy operation. Thus, they likely mean that during computation (moe kernel) phases, there's no need for continuous polling by the receiver role kernel, reducing SM occupancy. You can invoke recv role kernel after moe kernel is finished in a lazy behaviour. It's my personal understanding, not from deepseek team.

bsakaguchi referenced this issue

2026-10-26 09:53:39 +00:00

[GH-ISSUE #74] test_internode.py hangs at internode.cu:dispatch(...) #68

bsakaguchi referenced this issue

2026-10-26 09:56:29 +00:00

[GH-ISSUE #270] A timeout error occurred when running test_internode.py on four devices. #200

bsakaguchi referenced this issue

2026-10-26 09:56:40 +00:00

[GH-ISSUE #290] Running the test case on a 2-node H20 3E machine failed with internode.cu:1029 'unspecified launch failure' #208