Skip to content

Make onnx export SDPA match aten behavior #2479

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
titaiwangms merged 1 commit into from
Aug 7, 2025
Merged

Make onnx export SDPA match aten behavior #2479

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 onnxscript/function_libs/torch_lib/ops/nn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2076,6 +2076,11 @@ def _aten_scaled_dot_product_attention_bool_mask_onnx(
op.Add(op.MatMul(query_scaled, key_transposed_scaled), attn_mask),
axis=-1,
)
# When using scaled dot product attention with a boolean mask, the softmax operation might return NaN values
# due to the presence of -inf in an entire row (padding tokens), resulting in 0/0 (NaN) in the softmax output.
# This is because there's no safe/masked softmax imp in ONNX, so we need to handle NaN values explicitly to match
# the behavior of PyTorch with boolean masks.
attn_weight = op.Where(op.IsNaN(attn_weight), zero, attn_weight)
attn_weight, _ = op.Dropout(attn_weight, dropout_p)
Copy link
Collaborator

@justinchuby justinchuby Aug 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@titaiwangms we should probably conditionally skip this line (even though there is a rewrite rule already)

Copy link
Contributor

@titaiwangms titaiwangms Aug 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

OK

Copy link
Collaborator

@gramalingam gramalingam Aug 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If you fix this, can you also please add a reference to pytorch/pytorch#103749 in the comments for the previous line fixing NaN?

Copy link
Contributor

@titaiwangms titaiwangms Aug 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We skip when dropout_p is 0?

Copy link
Contributor

@titaiwangms titaiwangms Aug 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

#2482

return op.MatMul(attn_weight, value)

Expand Down
Loading