- 12 Jun, 2022 3 commits
- 04 Jun, 2022 2 commits
- 03 Jun, 2022 2 commits
- 02 Jun, 2022 6 commits
- 26 May, 2022 1 commit
-
-
Tri Dao authored
-
This speeds up the fwd by 1.5x.
From 4KB per buffer to 2KB per buffer. This saves us 8KB of smem (each Q and dO have 2 buffers)