[GH-ISSUE #120] [low latency] why Stronger acquire scope for low-latency kernels ? #108

New issue

Closed

opened 2026-10-26 09:54:25 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 09:54:25 +00:00

Owner

Originally created by @elevenxiang on GitHub (Apr 14, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/120

why need change Stronger acquire scope for low-latency kernels ?

does it have potential data mismatch issue ?

Thanks

Originally created by @elevenxiang on GitHub (Apr 14, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/120 Hi why need change Stronger acquire scope for low-latency kernels ? does it have potential data mismatch issue ? Thanks

bsakaguchi closed this issue

2026-10-26 09:54:25 +00:00

bsakaguchi commented

2026-10-26 09:54:26 +00:00

Author

Owner

@LyricZhao commented on GitHub (Apr 14, 2025):

Do you mean while ((num_recv_tokens = ld_acquire_sys_global(rdma_recv_count + local_expert_idx * num_ranks + src_rank)) == 0);?

Because the network card is belonged to the system scope. This is more likely to be "semantically correct", changing it to GPU scope works as well.

@LyricZhao commented on GitHub (Apr 14, 2025): Do you mean `while ((num_recv_tokens = ld_acquire_sys_global(rdma_recv_count + local_expert_idx * num_ranks + src_rank)) == 0);`? Because the network card is belonged to the system scope. This is more likely to be "semantically correct", changing it to GPU scope works as well.

bsakaguchi commented

2026-10-26 09:54:26 +00:00

Author

Owner

@elevenxiang commented on GitHub (Apr 15, 2025):

Do you mean while ((num_recv_tokens = ld_acquire_sys_global(rdma_recv_count + local_expert_idx * num_ranks + src_rank)) == 0);?

Because the network card is belonged to the system scope. This is more likely to be "semantically correct", changing it to GPU scope works as well.

ok, thanks

@elevenxiang commented on GitHub (Apr 15, 2025): > Do you mean `while ((num_recv_tokens = ld_acquire_sys_global(rdma_recv_count + local_expert_idx * num_ranks + src_rank)) == 0);`? > > Because the network card is belonged to the system scope. This is more likely to be "semantically correct", changing it to GPU scope works as well. ok, thanks

bsakaguchi referenced this issue

2026-10-26 10:00:41 +00:00

[PR #108] [MERGED] Super tiny fix shape typo #434