Skip to content

Fix block size handling for cross attention in TPU flash attention #301

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
michelle-yooh wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix block size handling for cross attention in TPU flash attention #301

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 6 additions & 4 deletions src/maxdiffusion/models/attention_flax.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,17 +186,19 @@ def _tpu_flash_attention(
kv_max_block_size = key.shape[1]
else:
kv_max_block_size = q_max_block_size
if flash_block_sizes:
# ensure that for cross attention we override the block sizes.
if flash_block_sizes and key.shape[1] == query.shape[1]:
block_sizes = flash_block_sizes
else:
block_size_q = flash_block_sizes.block_q if flash_block_sizes else q_max_block_size
block_sizes = splash_attention_kernel.BlockSizes(
block_q=min(q_max_block_size, query.shape[2]),
block_q=block_size_q,
block_kv_compute=min(kv_max_block_size, key.shape[2]),
block_kv=min(kv_max_block_size, key.shape[2]),
block_q_dkv=min(q_max_block_size, query.shape[2]),
block_q_dkv=block_size_q,
block_kv_dkv=min(kv_max_block_size, key.shape[2]),
block_kv_dkv_compute=min(kv_max_block_size, query.shape[2]),
block_q_dq=None if attention_kernel == "tokamax_flash" else block_sizes.block_q_dq,
block_q_dq=None if attention_kernel == "tokamax_flash" else block_size_q,
block_kv_dq=None if attention_kernel == "tokamax_flash" else min(kv_max_block_size, query.shape[2]),
use_fused_bwd_kernel=True if attention_kernel == "tokamax_flash" else False,
)
Expand Down
Loading