Commit b08de1b
committed
Update on "Use unfused SDPA for short sequences (q_len <= 128 or kv_len <= 128)"
ATT
Differential Revision: [D96044308](https://our.internmc.facebook.com/intern/diff/D96044308/)
[ghstack-poisoned]4 files changed
Lines changed: 615 additions & 317 deletions
0 commit comments