Skip to content

Fix aten_gru: add linear_before_reset=1 to match PyTorch GRU semantics #2853

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
Copilot wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Fix aten_gru: add linear_before_reset=1 to match PyTorch GRU semantics #2853

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
906fdc5
Initial plan
Copilot Mar 14, 2026
170a407
fix: add linear_before_reset=1 to GRU op calls in aten_gru
Copilot Mar 14, 2026
0226bb3
fix: add explanatory comment for linear_before_reset=1 in aten_gru
Copilot Mar 15, 2026
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
7 changes: 7 additions & 0 deletions onnxscript/function_libs/torch_lib/ops/core.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4348,6 +4348,11 @@ def aten_gru(
# Extract hidden_size from hx shape: [num_layers * num_directions, batch, hidden_size]
hidden_size_attr = hx.shape[2]

# linear_before_reset=1 matches PyTorch's GRU formulation where the linear
# transformation is applied before multiplying by the reset gate:
# ht = g(Xt*(Wh^T) + rt (.) (Ht-1*(Rh^T) + Rbh) + Wbh)
# The ONNX default (linear_before_reset=0) uses a different equation and
# would produce numerically incorrect results.
if B is not None:
Y, Y_h = op.GRU(
current_input,
Expand All @@ -4357,6 +4362,7 @@ def aten_gru(
initial_h=layer_h,
direction=direction,
hidden_size=hidden_size_attr,
linear_before_reset=1,
)
else:
Y, Y_h = op.GRU(
Expand All @@ -4366,6 +4372,7 @@ def aten_gru(
initial_h=layer_h,
direction=direction,
hidden_size=hidden_size_attr,
linear_before_reset=1,
)

# Y shape: [seq_length, num_directions, batch_size, hidden_size]
Expand Down
Loading