Skip to content

[torchlib] Fix sdpa dtype in attn_mask #2445

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
merged 4 commits into from
Jul 11, 2025
Merged

[torchlib] Fix sdpa dtype in attn_mask #2445

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
095f5d7
fix sdpa dtype
titaiwangms Jul 10, 2025
810a73d
Merge branch 'main' into titaiwang/fix_sdpa_dtype
titaiwangms Jul 10, 2025
18ac2da
revert masking part
titaiwangms Jul 10, 2025
64a23d8
use ir tensor
titaiwangms Jul 10, 2025
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: 3 additions & 3 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 @@ -2069,9 +2069,9 @@
query_scaled = op.Mul(query, op.Sqrt(scale))
key_transposed_scaled = op.Mul(key_transposed, op.Sqrt(scale))
# Turn the Boolean mask to float: attn_mask.masked_fill(not attn_mask, -float('inf'))
attn_mask = op.Where(
attn_mask, op.Constant(value_float=0.0), op.Constant(value_float=-float("inf"))
)
zero = op.Constant(value=ir.tensor(0.0, dtype=query.dtype))
neg_inf = op.Constant(value=ir.tensor(-float("inf"), dtype=query.dtype))
attn_mask = op.Where(attn_mask, zero, neg_inf)

Check warning on line 2074 in onnxscript/function_libs/torch_lib/ops/nn.py

View check run for this annotation

Codecov / codecov/patch

onnxscript/function_libs/torch_lib/ops/nn.py#L2072-L2074 

Added lines #L2072 - L2074 were not covered by tests
attn_weight = op.Softmax(
op.Add(op.MatMul(query_scaled, key_transposed_scaled), attn_mask),
axis=-1,
Expand Down
Loading