dingqingy comments

Repositories
Issues
Comments

Results 2 comments of


                                            dingqingy

Context parallel on Turing GPUs?

Thanks everyone for your timely response! I tested the SDPA backend locally, and it seems like memory-efficient attention backend only works fp16 and fp32 but not bf16, causing no kernel...

Context parallel on Turing GPUs?

Thanks, I just realized that the warning is indeed emitted earlier (just not part of the error trace that I didn't notice). I understand the memory-efficient sdpa backend work with...