Bug
Tests fail to collect on with:
▎ ImportError: libtorch_cuda.so: undefined symbol: ncclDevCommDestroy
▎ torch 2.10.0 requires NCCL 2.19+, which is unavailable on the runner. Downgrading torch is not straightforward since the prebuild flash-attn wheels
(mjun0812/flash-attention-prebuild-wheels v0.7.12) are only built for torch 2.10.
To Reproduce
See example here
Repository version
v0.5.2