Skip to content

Fix reduction operator shape inference at opset 18 and later #32376

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
Kayvan Zahiri (Kayvan-Zahiri) wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix reduction operator shape inference at opset 18 and later #32376

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e21e686
Fix reduction operator shape inference at opset 18 and later
Kayvan-Zahiri Sep 2, 2026
24c2da2
Include ReduceProd in the axes initializer forwarding
Kayvan-Zahiri Sep 2, 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
18 changes: 17 additions & 1 deletion onnxruntime/python/tools/symbolic_shape_infer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -508,7 +508,20 @@ def _onnx_infer_single_node(self, node):
# (2) opset version >= 9. In older version, initializer is required in graph input by onnx spec.
# (3) The initializer is not in graph input. The means the node input is "constant" in inference.
initializers = []
if (get_opset(self.out_mp_) >= 9) and node.op_type in ["Unsqueeze"]:
if (get_opset(self.out_mp_) >= 9) and node.op_type in [
"Unsqueeze",
# Since opset 18 the reduction operators take axes as an input rather
# than an attribute. Without the axes value onnx shape inference cannot
# compute the output shape and returns nothing.
"ReduceL1",
"ReduceL2",
"ReduceLogSum",
"ReduceLogSumExp",
"ReduceMax",
"ReduceMin",
"ReduceProd",
"ReduceSumSquare",
]:
initializers = [
self.initializers_[name]
for name in node.input
Expand Down Expand Up @@ -1627,6 +1640,9 @@ def _infer_ReduceSum(self, node): # noqa: N802
shape = self._get_shape(node, 0)
output_shape = []
axes = [handle_negative_axis(a, len(shape)) for a in axes]
if not axes and not get_attribute(node, "noop_with_empty_axes", 0):
# An empty axes input reduces every axis unless noop_with_empty_axes is set.
axes = list(range(len(shape)))
for i, d in enumerate(shape):
if i in axes:
if keep_dims:
Expand Down
69 changes: 69 additions & 0 deletions onnxruntime/test/python/onnxruntime_test_python_symbolic_shape_infer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -725,6 +725,75 @@ def test_qlinear_binary(self):
]
self._check_shapes(graph, inferred.graph, expected_shapes)

def test_reduce_ops_opset18_axes_input(self):
# Since opset 18 axes is an input. Without the initializer the output shape is lost.
for op_type in [
"ReduceL1",
"ReduceL2",
"ReduceLogSum",
"ReduceLogSumExp",
"ReduceMax",
"ReduceMin",
"ReduceProd",
"ReduceSumSquare",
]:
with self.subTest(op_type=op_type):
initializers = [numpy_helper.from_array(numpy.array([1], dtype=numpy.int64), "axes")]
graph = helper.make_graph(
[
helper.make_node(op_type, ["input", "axes"], ["temp"], keepdims=0),
helper.make_node("Identity", ["temp"], ["output"]),
],
f"{op_type}_Opset18_Test",
[
helper.make_tensor_value_info("input", TensorProto.FLOAT, [2, 3, 4]),
],
[
helper.make_tensor_value_info("output", TensorProto.FLOAT, [2, 4]),
],
initializers,
)
model = helper.make_model(graph, producer_name=f"{op_type}_Opset18_Test_Model")
model.opset_import[0].version = 18

inferred = SymbolicShapeInference.infer_shapes(model, auto_merge=True)
expected_shapes = [
helper.make_tensor_value_info("temp", TensorProto.FLOAT, [2, 4]),
helper.make_tensor_value_info("output", TensorProto.FLOAT, [2, 4]),
]
self._check_shapes(graph, inferred.graph, expected_shapes)

def test_reduce_sum_empty_axes(self):
# An empty axes input reduces every axis unless noop_with_empty_axes is set.
for noop, expected in [(0, []), (1, [2, 3, 4])]:
with self.subTest(noop_with_empty_axes=noop):
initializers = [numpy_helper.from_array(numpy.array([], dtype=numpy.int64), "axes")]
graph = helper.make_graph(
[
helper.make_node(
"ReduceSum", ["input", "axes"], ["temp"], keepdims=0, noop_with_empty_axes=noop
),
helper.make_node("Identity", ["temp"], ["output"]),
],
"ReduceSum_EmptyAxes_Test",
[
helper.make_tensor_value_info("input", TensorProto.FLOAT, [2, 3, 4]),
],
[
helper.make_tensor_value_info("output", TensorProto.FLOAT, expected),
],
initializers,
)
model = helper.make_model(graph, producer_name="ReduceSum_EmptyAxes_Test_Model")
model.opset_import[0].version = 18

inferred = SymbolicShapeInference.infer_shapes(model, auto_merge=True)
expected_shapes = [
helper.make_tensor_value_info("temp", TensorProto.FLOAT, expected),
helper.make_tensor_value_info("output", TensorProto.FLOAT, expected),
]
self._check_shapes(graph, inferred.graph, expected_shapes)


class TestSymbolicShapeInferenceForSlice(unittest.TestCase):
def check_slice_of_concat(self, input_dims, start, end, step, expected_output_dim):
Expand Down
Loading