Skip to content

fix: align Relax ONNX ArgMax/ArgMin NaN handling with ONNX Runtime #19671

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
mvanhorn wants to merge 1 commit into
base: main
Choose a base branch
from
+98 −6
Open

fix: align Relax ONNX ArgMax/ArgMin NaN handling with ONNX Runtime #19671

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
23 changes: 17 additions & 6 deletions python/tvm/relax/frontend/onnx/onnx_frontend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4203,6 +4203,17 @@ def _argreduce_select_last_index(bb, data, axis, keepdims, op):
return relax.op.subtract(offset, flipped_idx)


def _argreduce_sanitize_nan(bb, data, *, for_min):
"""Match ONNX Runtime ArgMax/ArgMin behavior by making NaN win comparisons."""
dtype = data.struct_info.dtype
if not _relax_dtype_is_floating_point(dtype):
return data
replacement = -_np.inf if for_min else _np.inf
return bb.emit(
relax.op.where(relax.op.isnan(data), relax.const(replacement, dtype), data)
)
Comment on lines +4206 to +4214

@gemini-code-assist gemini-code-assist Bot Jun 4, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

medium

Optimization & Defensive Programming Opportunity

  1. Defensive Check: Added a check if not hasattr(data.struct_info, "dtype") to prevent potential AttributeError if data.struct_info is None or does not have a dtype attribute.
  2. Constant Folding: If data is a relax.Constant, we can perform the NaN replacement directly at compile-time using NumPy. This avoids emitting unnecessary isnan, where, and const operators in the Relax graph, keeping the imported IR cleaner and more efficient.
def _argreduce_sanitize_nan(bb, data, *, for_min):
    """Match ONNX Runtime ArgMax/ArgMin behavior by making NaN win comparisons."""
    if not hasattr(data.struct_info, "dtype"):
        return data
    dtype = data.struct_info.dtype
    if not _relax_dtype_is_floating_point(dtype):
        return data
    replacement = -_np.inf if for_min else _np.inf
    if isinstance(data, relax.Constant):
        np_data = data.data.numpy()
        if _np.any(_np.isnan(np_data)):
            np_data = _np.where(_np.isnan(np_data), replacement, np_data)
            return relax.const(np_data, dtype)
        return data
    return bb.emit(
        relax.op.where(relax.op.isnan(data), relax.const(replacement, dtype), data)
    )



class ArgMax(OnnxOpConverter):
"""Converts an onnx ArgMax node into an equivalent Relax expression."""

Expand All @@ -4218,19 +4229,19 @@ def _check_attrs(cls, data, attr, shift_axis=True):

@classmethod
def _impl_v1(cls, bb, inputs, attr, params):
data = inputs[0]
data = _argreduce_sanitize_nan(bb, inputs[0], for_min=False)
axis, keepdims = cls._check_attrs(data, attr, False)
return relax.op.argmax(data, axis, keepdims)

@classmethod
def _impl_v11(cls, bb, inputs, attr, params):
data = inputs[0]
data = _argreduce_sanitize_nan(bb, inputs[0], for_min=False)
axis, keepdims = cls._check_attrs(data, attr)
return relax.op.argmax(data, axis, keepdims)

@classmethod
def _impl_v12(cls, bb, inputs, attr, params):
data = inputs[0]
data = _argreduce_sanitize_nan(bb, inputs[0], for_min=False)
axis, keepdims = cls._check_attrs(data, attr)
select_last_index = attr.get("select_last_index", False)
if select_last_index:
Expand All @@ -4253,19 +4264,19 @@ def _check_attrs(cls, data, attr, shift_axis=True):

@classmethod
def _impl_v1(cls, bb, inputs, attr, params):
data = inputs[0]
data = _argreduce_sanitize_nan(bb, inputs[0], for_min=True)
axis, keepdims = cls._check_attrs(data, attr, False)
return relax.op.argmin(data, axis, keepdims)

@classmethod
def _impl_v11(cls, bb, inputs, attr, params):
data = inputs[0]
data = _argreduce_sanitize_nan(bb, inputs[0], for_min=True)
axis, keepdims = cls._check_attrs(data, attr)
return relax.op.argmin(data, axis, keepdims)

@classmethod
def _impl_v12(cls, bb, inputs, attr, params):
data = inputs[0]
data = _argreduce_sanitize_nan(bb, inputs[0], for_min=True)
axis, keepdims = cls._check_attrs(data, attr)
select_last_index = attr.get("select_last_index", False)
if select_last_index:
Expand Down
81 changes: 81 additions & 0 deletions tests/python/relax/test_frontend_onnx.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2860,6 +2860,87 @@ def verify_arg_min_max(input_dim, in_dtype, op_name="ArgMax", axis=None, keepdim
verify_arg_min_max([3, 4, 4], in_dtype, "ArgMin", axis, keepdims)


def _make_arg_min_max_model(
op_name, data_shape, out_shape, axis, keepdims, select_last_index=0
):
node = helper.make_node(
op_name,
inputs=["data"],
outputs=["out"],
axis=axis,
keepdims=keepdims,
select_last_index=select_last_index,
)
graph = helper.make_graph(
[node],
"arg_min_max_nan_test",
inputs=[helper.make_tensor_value_info("data", TensorProto.FLOAT, list(data_shape))],
outputs=[helper.make_tensor_value_info("out", TensorProto.INT64, out_shape)],
)
return helper.make_model(graph, producer_name="arg_min_max_nan_test")


@pytest.mark.parametrize("op_name", ["ArgMax", "ArgMin"])
def test_arg_min_max_nan_matches_ort_indices(op_name):
data = np.array(
[
[2.0, np.nan, 7.0, 4.0, 1.0],
[np.nan, 2.0, 7.0, 4.0, 1.0],
[2.0, 4.0, 7.0, 1.0, np.nan],
],
dtype=np.float32,
)
expected = np.array([1, 0, 4], dtype=np.int64)
numpy_result = np.argmax(data, axis=1) if op_name == "ArgMax" else np.argmin(data, axis=1)
np.testing.assert_array_equal(numpy_result, expected)

model = _make_arg_min_max_model(op_name, data.shape, [3], axis=1, keepdims=0)
check_correctness(model, inputs={"data": data}, opset=12)


@pytest.mark.parametrize("op_name", ["ArgMax", "ArgMin"])
def test_arg_min_max_nan_keepdims_and_all_nan(op_name):
data = np.array(
[
[[np.nan, np.nan, np.nan], [5.0, np.nan, 1.0]],
[[2.0, 3.0, np.nan], [np.nan, -1.0, -2.0]],
],
dtype=np.float32,
)
model = _make_arg_min_max_model(op_name, data.shape, [2, 2, 1], axis=2, keepdims=1)
check_correctness(model, inputs={"data": data}, opset=12)


@pytest.mark.parametrize("op_name", ["ArgMax", "ArgMin"])
def test_arg_min_max_nan_select_last_index(op_name):
data = np.array(
[
[[np.nan, 2.0, np.nan, 1.0], [np.nan, np.nan, np.nan, np.nan]],
[[5.0, np.nan, 1.0, np.nan], [4.0, 3.0, 2.0, 1.0]],
],
dtype=np.float32,
)
model = _make_arg_min_max_model(
op_name,
data.shape,
[2, 2],
axis=2,
keepdims=0,
select_last_index=1,
)
check_correctness(model, inputs={"data": data}, opset=12)


@pytest.mark.parametrize("op_name", ["ArgMax", "ArgMin"])
def test_arg_min_max_finite_regression(op_name):
data = np.array(
[[2.0, 4.0, 7.0, 1.0, 5.0], [3.0, -2.0, 8.0, 6.0, 0.0]],
dtype=np.float32,
)
model = _make_arg_min_max_model(op_name, data.shape, [2], axis=1, keepdims=0)
check_correctness(model, inputs={"data": data}, opset=12)


@pytest.mark.parametrize("axis", [-1, 0, 1])
@pytest.mark.parametrize("largest", [True, False])
def test_topk(axis: int, largest: int):
Expand Down