7aceb9f91c
* Add --use-flash-attention flag. This is useful on AMD systems, as FA builds are still 10% faster than Pytorch cross-attention.
36 KiB
36 KiB
7aceb9f91c