Commit Graph

6 Commits

Author SHA1 Message Date
Tri Dao
5d07483bbc Refactor Gmem code to store q, k, v pointers separately 2022-06-12 16:37:32 -07:00
Tri Dao
d3e6440958 Implement bwd for head dim 128 2022-06-11 17:52:36 -07:00
Tri Dao
0d854692c6 Implement fwd for head dim 128 2022-06-11 17:52:36 -07:00
Tri Dao
321c57d07d Set block size of SM75 fwd to 256 if there's no dropout
This speeds up the fwd by 1.5x.
2022-06-04 16:51:28 -07:00
Tri Dao
2712aa4c8d Support Turing mma instructions 2022-06-03 16:58:44 -07:00
Tri Dao
9dbc491aa5 Rename, add benchmarking script 2022-05-26 13:57:38 -07:00