Skip to content

[torchlib] Fix aten_conv3d generating a 2D bias when bias is missing #2935

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
javierdejesusda wants to merge 1 commit into
base: main
Choose a base branch
from
+28 −1
Open

[torchlib] Fix aten_conv3d generating a 2D bias when bias is missing #2935

Show file tree
Hide file tree
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: 1 addition & 1 deletion 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 @@ -2127,7 +2127,7 @@ def aten_conv3d(

if bias is None:
weight_dim_0 = op.Shape(weight, start=0, end=1)
bias_shape = op.Concat(weight_dim_0, op.Constant(value_ints=[2]), axis=0)
bias_shape = op.Expand(weight_dim_0, op.Constant(value_ints=[1]))
zero = op.CastLike(0.0, input)
bias = op.Expand(zero, bias_shape)

Expand Down
27 changes: 27 additions & 0 deletions tests/function_libs/torch_lib/e2e_ops_tests.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
import math
import unittest

import onnx

@justinchuby justinchuby Jun 15, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Suggested change
import onnx

import parameterized

# TODO(pytorch/pytorch#129279): Migrate these tests to the PyTorch repo
Expand Down Expand Up @@ -1004,6 +1005,32 @@ def forward(self, x):
got = onnx_program.call_reference({"x": inputs[0]})
torch.testing.assert_close(expected, got[0])

def test_conv3d_without_bias_produces_1d_bias(self):
class Model(torch.nn.Module):
def __init__(self):
super().__init__()
self.conv = torch.nn.Conv3d(3, 4, kernel_size=2, bias=False)

def forward(self, x):
return self.conv(x)

onnx_program = torch.onnx.export(
Model().eval(), (torch.randn(1, 3, 8, 8, 8),), dynamo=True, optimize=False
)
_testing.assert_onnx_program(onnx_program)

# The bias synthesized for a bias-less conv must be 1D ([out_channels]) to
# match the ONNX Conv spec. See https://github.com/microsoft/onnxscript/issues/2931.
inferred = onnx.shape_inference.infer_shapes(onnx_program.model_proto, data_prop=True)
shape_ranks = {
value_info.name: len(value_info.type.tensor_type.shape.dim)
for value_info in inferred.graph.value_info
if value_info.type.tensor_type.HasField("shape")
}
conv_nodes = [node for node in inferred.graph.node if node.op_type == "Conv"]
self.assertEqual(len(conv_nodes), 1)
self.assertEqual(shape_ranks[conv_nodes[0].input[2]], 1)
Comment on lines +1022 to +1032

@justinchuby justinchuby Jun 15, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Ok to review this part as assert_onnx_program will already run the model and validate the shapes.

@javierdejesusda javierdejesusda Jun 16, 2026

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

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

Before I drop it: the synthesized bias is all zeros, so assert_onnx_program passes even with the rank-2 bias. The conv output is numerically identical, so only the rank check fails (2 != 1) before the fix, which is the part that actually pins this regression. If you'd rather rely on assert_onnx_program for path coverage here, I'll remove the block and the onnx import to match the convention in the suite.



if __name__ == "__main__":
unittest.main()
Loading