Skip to content

Eliminate duplicated softmax recomputation in indexer loss. #4454

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

Merged
copybara-service merged 1 commit into from
Jul 17, 2026
+2 −0
Merged

Eliminate duplicated softmax recomputation in indexer loss. #4454

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
2 changes: 2 additions & 0 deletions src/maxtext/layers/attention_mla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1077,6 +1077,8 @@ def calculate_indexer_loss(

# Aggregate heads: [b, h, t, s] -> [b, t, s]
attention_probs = jnp.sum(attention_probs, axis=1)
# Force materialization and prevent fusion across this point to reuse the intermediate tensor
attention_probs = jax.lax.optimization_barrier(attention_probs)
Comment thread
JHCuc3m marked this conversation as resolved.
# L1 normalize aggregated target distribution
attention_probs = attention_probs / (jnp.sum(attention_probs, axis=-1, keepdims=True) + EPS)

Expand Down
Loading