[GH-ISSUE #603] [Low Latency][IBGDA] GPU Direct 数据正确性是如何保证的 ? #367

Closed
opened 2026-10-26 09:59:14 +00:00 by bsakaguchi · 0 comments
Owner

Originally created by @elevenxiang on GitHub (Apr 15, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/603

github.com/deepseek-ai/DeepEP@567632dd59/csrc/kernels/internode_ll.cu (L243-L268)

从这里我们看到 数据经过量化 or 拷贝之后,写入 symmetric memory, 普通的 store 指令,理论上数据都 cache 在 L1/L2

但是接下来会 IBGDA,是直接 PCIe 从 HBM 读走数据,这个正确性是如何保证的 ?

谢谢 ~

Originally created by @elevenxiang on GitHub (Apr 15, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/603 https://github.com/deepseek-ai/DeepEP/blob/567632dd59810d77b3cc05553df953cc0f779799/csrc/kernels/internode_ll.cu#L243-L268 从这里我们看到 数据经过量化 or 拷贝之后,写入 symmetric memory, 普通的 store 指令,理论上数据都 cache 在 L1/L2 但是接下来会 IBGDA,是直接 PCIe 从 HBM 读走数据,这个正确性是如何保证的 ? 谢谢 ~
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#367
No description provided.