This website requires JavaScript.
Explore
Help
Register
Sign In
squall
/
flash-attention
Watch
1
Star
0
Fork
0
You've already forked flash-attention
Code
Issues
Pull Requests
Actions
Packages
Projects
Releases
Wiki
Activity
b17c6fe235
flash-attention
/
csrc
History
Tri Dao
b17c6fe235
Reduce smem usage for Q and dO in the backward pass
...
From 4KB per buffer to 2KB per buffer. This saves us 8KB of smem (each Q and dO have 2 buffers)
2022-06-03 16:59:11 -07:00
..
flash_attn
Reduce smem usage for Q and dO in the backward pass
2022-06-03 16:59:11 -07:00