[GH-ISSUE #110] Will the low latency kernel definitely use fp8? #102

New issue

Closed

opened 2026-10-26 09:54:19 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:54:19 +00:00

Owner

Originally created by @Qizhi697 on GitHub (Apr 7, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/110

github.com/deepseek-ai/DeepEP@42494864ba/csrc/kernels/internode_ll.cu (L251)
Will the low latency kernel definitely use fp8? When kUseFP8=false, the variable packed_recv_x_scales becomes a nullptr, which means it cannot perform addition.

Originally created by @Qizhi697 on GitHub (Apr 7, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/110 https://github.com/deepseek-ai/DeepEP/blob/42494864ba64b6520337a6d084dc209ac6be4905/csrc/kernels/internode_ll.cu#L251 Will the low latency kernel definitely use fp8? When `kUseFP8=false`, the variable `packed_recv_x_scales` becomes a `nullptr`, which means it cannot perform addition.

bsakaguchi closed this issue

2026-10-26 09:54:19 +00:00

bsakaguchi commented

2026-10-26 09:54:20 +00:00

Author

Owner

@LyricZhao commented on GitHub (Apr 8, 2025):

Will the low latency kernel definitely use fp8?

No, BF16 can be used with kUseFP8=false.

For packed_recv_x_scales == nullptr will not affect correctness.

@LyricZhao commented on GitHub (Apr 8, 2025): > Will the low latency kernel definitely use fp8? No, BF16 can be used with `kUseFP8=false`. For `packed_recv_x_scales == nullptr` will not affect correctness.

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#102

No description provided.

Rows
Columns