xformers
xformers copied to clipboard
Is this a mistake in requirement of llama_inference?
https://github.com/facebookresearch/xformers/blob/main/examples/llama_inference/requirements.txt
torch>=2.2.0 this should be a mistake
we should use FlashAttention >=2.2.0
Hi, I believe this is on-purpose. There is a bug in PyTorch 2.1 which prevents CUDA GRaphs from working properly with NCCL collectives. It will be fixed in 2.2, but in the meantime you can use the nightly (which are already tagged as 2.2)
Hi, I believe this is on-purpose. There is a bug in PyTorch 2.1 which prevents CUDA GRaphs from working properly with NCCL collectives. It will be fixed in 2.2, but in the meantime you can use the nightly (which are already tagged as 2.2)
ok,thanks