david/flash-attention

réplica de https://github.com/Dao-AILab/flash-attention

Autor	SHA1 Mensaxe	Data
Tri Dao	393882bc08 [LayerNorm] Implement LN with parallel residual, support dim 8k	hai 1 ano
Tri Dao	6738d9477d [LayerNorm] Implement RMS Norm	hai 1 ano
Tri Dao	ae137ed17a [LayerNorm] Fuse LayerScale	%!s(int64=2) %!d(string=hai) anos
Tri Dao	8c6609ae1a [LayerNorm] Support all dimensions up to 6k (if divisible by 8)	%!s(int64=2) %!d(string=hai) anos
Tri Dao	fa6d1ce44f Add fused_dense and dropout_add_layernorm CUDA extensions	%!s(int64=2) %!d(string=hai) anos