How does FlashAttention tile SRAM memory to avoid materializing the full N x N attention matrix?
#187
|
What is the block tiling algorithm in online softmax? |
Answered by
Ultron09
Aug 28, 2026
Replies: 1 comment
|
It maintains running max and sum statistics per block, computing attention incrementally without HBM memory traffic. |
0 replies
Answer selected by
Ultron09
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
It maintains running max and sum statistics per block, computing attention incrementally without HBM memory traffic.