Skip to content

fix: replace NaN from all-masked SDPA padding rows in Gemma 4 vision #1006

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
fabiopili wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: replace NaN from all-masked SDPA padding rows in Gemma 4 vision #1006

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
5 changes: 5 additions & 0 deletions mlx_vlm/models/gemma4/vision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -225,6 +225,11 @@ def __call__(

attn_output = ensure_fused_sdpa(q, k, v, scale=1.0, mask=mask)

# Replace NaN with 0 — all-masked padding rows produce NaN via softmax
# (0/0) at sequence lengths where the fused SDPA kernel is not used.
# These rows are zeroed out by the pooler, so 0 is the correct value.
attn_output = mx.where(mx.isnan(attn_output), 0.0, attn_output)

# [B, H, L, D] -> [B, L, H*D]
attn_output = attn_output.transpose(0, 2, 1, 3).reshape(B, L, -1)

Expand Down