Skip to content

fix: compute local block padding for dynamic splash indexer mask in CP #4950

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
RexBearIU wants to merge 1 commit into
base: main
Choose a base branch
from
+4 −2
Open

fix: compute local block padding for dynamic splash indexer mask in CP #4950

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
6 changes: 4 additions & 2 deletions src/maxtext/layers/attention_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2111,8 +2111,10 @@ def wrap_flash_attention(
if indexer_mask is not None:
# Convert additive float mask (0.0=attend, negative=masked) to boolean mask for Tokamax splash kernel
indexer_mask = indexer_mask == 0.0
pad_q = mask_shape[0] - indexer_mask.shape[-2]
pad_kv = mask_shape[1] - indexer_mask.shape[-1]
padded_q_len = ((indexer_mask.shape[-2] + sa_config.block_q - 1) // sa_config.block_q) * sa_config.block_q
padded_kv_len = ((indexer_mask.shape[-1] + sa_config.block_kv - 1) // sa_config.block_kv) * sa_config.block_kv
pad_q = padded_q_len - indexer_mask.shape[-2]
pad_kv = padded_kv_len - indexer_mask.shape[-1]
if pad_q > 0 or pad_kv > 0:
pad_width = [(0, 0)] * (indexer_mask.ndim - 2) + [(0, pad_q), (0, pad_kv)]
indexer_mask = jnp.pad(
Expand Down
Loading