Skip to content

[FRONTEND][ONNX] Extend axes for layer_norm when gamma/beta are multi-dimensional #18143

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 14, 2025
Merged

[FRONTEND][ONNX] Extend axes for layer_norm when gamma/beta are multi-dimensional #18143

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ec6dbc9
Extend axes for layer_norm when gamma/beta are multi-dimensional
ConvolutedDog Jul 13, 2025
d6387c9
Add testcase for layer_norm when gamma/beta are multi-dimensional
ConvolutedDog Jul 13, 2025
5ee8005
fix lint
ConvolutedDog Jul 13, 2025
0f56685
fix lint
ConvolutedDog Jul 13, 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
11 changes: 11 additions & 0 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 @@ -48,6 +48,7 @@
from tvm.ir import IRModule
from tvm.ir.supply import NameSupply
from tvm.tir.generic import cast
from tvm.topi.utils import get_const_tuple

from ..common import autopad

Expand Down Expand Up @@ -2488,9 +2489,19 @@ def _impl_v17(cls, bb, inputs, attr, params):
axis = attr.get("axis", -1)
epsilon = attr.get("epsilon", 1e-05)

gamma_shape = get_const_tuple(scale.struct_info.shape)

if bias is None:
seq_len = data.struct_info.shape[1].value
bias = relax.const([0.0] * seq_len, dtype="float32")
else:
beta_shape = get_const_tuple(bias.struct_info.shape)
if gamma_shape != beta_shape:
raise ValueError("gamma and beta shapes do not match")

axis = list(axis) if isinstance(axis, (list, tuple)) else [axis]
if len(axis) < len(gamma_shape):
axis.extend(range(axis[-1] + 1, axis[-1] + 1 + len(gamma_shape) - len(axis)))

output = relax.op.nn.layer_norm(data, scale, bias, axis, epsilon)
# Onnx layernorm has 3 outputs but only the first is used.
Expand Down
64 changes: 55 additions & 9 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 @@ -1282,43 +1282,89 @@ def test_mean_variance_norm():


def test_layer_norm():
layer_norm_node = helper.make_node("LayerNormalization", ["a", "b", "c"], ["d"], epsilon=1e-12)
layer_norm_node = helper.make_node(
"LayerNormalization", ["input", "scale", "bias"], ["Y"], epsilon=1e-12
)

graph = helper.make_graph(
[layer_norm_node],
"layer_norm_test",
inputs=[
helper.make_tensor_value_info("a", TensorProto.FLOAT, [32, 32]),
helper.make_tensor_value_info("b", TensorProto.FLOAT, [32]),
helper.make_tensor_value_info("c", TensorProto.FLOAT, [32]),
helper.make_tensor_value_info("input", TensorProto.FLOAT, [32, 32]),
helper.make_tensor_value_info("scale", TensorProto.FLOAT, [32]),
helper.make_tensor_value_info("bias", TensorProto.FLOAT, [32]),
],
outputs=[
helper.make_tensor_value_info("d", TensorProto.FLOAT, [32, 32]),
helper.make_tensor_value_info("Y", TensorProto.FLOAT, [32, 32]),
],
)

model = helper.make_model(graph, producer_name="layer_norm_test")
check_correctness(model)

# Test case with no bias that is an optional input
layer_norm_node = helper.make_node("LayerNormalization", ["a", "b"], ["d"], epsilon=1e-12)
layer_norm_node = helper.make_node(
"LayerNormalization", ["input", "scale"], ["Y"], epsilon=1e-12
)

graph = helper.make_graph(
[layer_norm_node],
"layer_norm_test",
inputs=[
helper.make_tensor_value_info("a", TensorProto.FLOAT, [32, 32]),
helper.make_tensor_value_info("b", TensorProto.FLOAT, [32]),
helper.make_tensor_value_info("input", TensorProto.FLOAT, [32, 32]),
helper.make_tensor_value_info("scale", TensorProto.FLOAT, [32]),
],
outputs=[
helper.make_tensor_value_info("d", TensorProto.FLOAT, [32, 32]),
helper.make_tensor_value_info("Y", TensorProto.FLOAT, [32, 32]),
],
)

model = helper.make_model(graph, producer_name="layer_norm_test")
check_correctness(model)


def test_layer_norm_with_nd_gamma_beta():
layer_norm_node = helper.make_node(
"LayerNormalization", ["input", "scale", "bias"], ["Y"], axis=1, epsilon=1e-12
)

graph = helper.make_graph(
[layer_norm_node],
"layer_norm_with_nd_gamma_beta_test",
inputs=[
helper.make_tensor_value_info("input", TensorProto.FLOAT, [1, 3, 4, 4]),
helper.make_tensor_value_info("scale", TensorProto.FLOAT, [3, 4, 4]),
helper.make_tensor_value_info("bias", TensorProto.FLOAT, [3, 4, 4]),
],
outputs=[
helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 3, 4, 4]),
],
)

model = helper.make_model(graph, producer_name="layer_norm_with_nd_gamma_beta_test")
check_correctness(model)

# Test case with no bias that is an optional input
layer_norm_node = helper.make_node(
"LayerNormalization", ["input", "scale"], ["Y"], axis=1, epsilon=1e-12
)

graph = helper.make_graph(
[layer_norm_node],
"layer_norm_with_nd_gamma_beta_test",
inputs=[
helper.make_tensor_value_info("input", TensorProto.FLOAT, [32, 32]),
helper.make_tensor_value_info("scale", TensorProto.FLOAT, [32]),
],
outputs=[
helper.make_tensor_value_info("Y", TensorProto.FLOAT, [32, 32]),
],
)

model = helper.make_model(graph, producer_name="layer_norm_with_nd_gamma_beta_test")
check_correctness(model)


# TODO Enable dynamism
@pytest.mark.parametrize("dynamic", [False])
def test_skiplayernormalization(dynamic):
Expand Down