ITADN

[BUG] It seems that SuperOfflload is unable to overlap the CPU (param update) and GPU (backwawrd).

#7966Openarminzhu 创建于 2026-04-09
bugtraining
A
arminzhucommented
**Describe the bug** @xylian86 It seems that SuperOfflload is unable to overlap the CPU (param update) and GPU (backwawrd). The CUDA stream may not be configured correctly. Its TSG (Tokens per Second per GPU) is far inferior to that of the ZeRO-2/ZeRO-3, and very close to ZeRO-Offload. It's worth noting that I found gradient offloading always occurs after backpropagation has finished.
1 条评论