Skip to content

[QNN-EP] Support alternate Layernorm fusion pattern in QNN preprocess #26060

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
qti-mattsinc wants to merge 2 commits into
base: main
Choose a base branch
from
+60 −29
Open

[QNN-EP] Support alternate Layernorm fusion pattern in QNN preprocess #26060

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
28ad04a
[QNN-EP] Support alternate Layernorm fusion pattern in QNN preprocess
qti-mattsinc Jan 12, 2026
ebaa0cd
Add to unit tests
qti-mattsinc Jan 12, 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
25 changes: 18 additions & 7 deletions onnxruntime/python/tools/quantization/fusions/fusion_layernorm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,14 +33,24 @@ def fuse(
| |
+-------------------------------------------------+

Or, using Mul instead of Pow:

+----------------------+
| |
| v
[Root] --> ReduceMean --> Sub --> Mul --> ReduceMean --> Add --> Sqrt --> Div --> Mul --> Add
(axis=2 or -1) | (in0=in1) (axis=2 or -1) (E-6 or E-12 or 0) ^
| |
+-------------------------------------------------+

It also handles cases of duplicated sub nodes exported from older version of PyTorch:

+----------------------+
| v
| +-------> Sub-----------------------------------------------+
| | |
| | v
[Root] --> ReduceMean --> Sub --> Pow --> ReduceMean --> Add --> Sqrt --> Div --> Mul --> Add
[Root] --> ReduceMean --> Sub --> (Pow or Mul) --> ReduceMean --> Add --> Sqrt --> Div --> Mul --> Add
| ^
| |
+----------------------+
Expand Down Expand Up @@ -70,10 +80,9 @@ def fuse(
div_node,
[
(["Sqrt", "Add", "ReduceMean", "Pow", "Sub"], [1, 0, 0, 0, 0]),
(
["Sqrt", "Add", "ReduceMean", "Pow", "Cast", "Sub"],
[1, 0, 0, 0, 0, 0],
),
(["Sqrt", "Add", "ReduceMean", "Pow", "Cast", "Sub"], [1, 0, 0, 0, 0, 0]),
(["Sqrt", "Add", "ReduceMean", "Mul", "Sub"], [1, 0, 0, 0, 0]),
(["Sqrt", "Add", "ReduceMean", "Mul", "Cast", "Sub"], [1, 0, 0, 0, 0, 0]),
],
output_name_to_node,
)
Expand All @@ -90,8 +99,10 @@ def fuse(
# Skip fusion since epsilon value is not expected.
return

pow_node = parent_nodes[3]
if self.find_constant_input(pow_node, 2.0) != 1:
pow_or_mul_node = parent_nodes[3]
if pow_or_mul_node.op_type == "Pow" and self.find_constant_input(pow_or_mul_node, 2.0) != 1:
return
elif pow_or_mul_node.op_type == "Mul" and pow_or_mul_node.input[0] != pow_or_mul_node.input[1]:
return

mul_node = input_name_to_nodes[div_node.output[0]][0]
Expand Down
64 changes: 42 additions & 22 deletions onnxruntime/test/python/quantization/test_qnn_preprocess_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,15 +55,26 @@ def build_model(self, shape, scale_val, bias_val):
bias_const = onnx.numpy_helper.from_array(np.array(bias_val, dtype=np.float32), "bias_const")
two_const = onnx.numpy_helper.from_array(np.array(2.0, dtype=np.float32), "two_const")

m_rm0_node = onnx.helper.make_node("ReduceMean", ["l2_seq_output", "axes_const"], ["m_rm0_out"])
m_sub_node = onnx.helper.make_node("Sub", ["l2_seq_output", "m_rm0_out"], ["m_sub_out"])
m_pow_node = onnx.helper.make_node("Pow", ["m_sub_out", "two_const"], ["m_pow_out"])
m_rm1_node = onnx.helper.make_node("ReduceMean", ["m_pow_out", "axes_const"], ["m_rm1_out"])
m_add0_node = onnx.helper.make_node("Add", ["m_rm1_out", "eps_const"], ["m_add0_out"])
m_sqrt_node = onnx.helper.make_node("Sqrt", ["m_add0_out"], ["m_sqrt_out"])
m_div_node = onnx.helper.make_node("Div", ["m_sub_out", "m_sqrt_out"], ["m_div_out"])
m_mul_node = onnx.helper.make_node("Mul", ["m_div_out", "scale_const"], ["m_mul_out"])
m_add1_node = onnx.helper.make_node("Add", ["m_mul_out", "bias_const"], ["output"])
m0_rm0_node = onnx.helper.make_node("ReduceMean", ["l2_seq_output", "axes_const"], ["m0_rm0_out"])
m0_sub_node = onnx.helper.make_node("Sub", ["l2_seq_output", "m0_rm0_out"], ["m0_sub_out"])
m0_pow_node = onnx.helper.make_node("Pow", ["m0_sub_out", "two_const"], ["m0_pow_out"])
m0_rm1_node = onnx.helper.make_node("ReduceMean", ["m0_pow_out", "axes_const"], ["m0_rm1_out"])
m0_add0_node = onnx.helper.make_node("Add", ["m0_rm1_out", "eps_const"], ["m0_add0_out"])
m0_sqrt_node = onnx.helper.make_node("Sqrt", ["m0_add0_out"], ["m0_sqrt_out"])
m0_div_node = onnx.helper.make_node("Div", ["m0_sub_out", "m0_sqrt_out"], ["m0_div_out"])
m0_mul_node = onnx.helper.make_node("Mul", ["m0_div_out", "scale_const"], ["m0_mul_out"])
m0_add1_node = onnx.helper.make_node("Add", ["m0_mul_out", "bias_const"], ["m0_add1_out"])

# Alternate ReduceMean sequence
m1_rm0_node = onnx.helper.make_node("ReduceMean", ["m0_add1_out", "axes_const"], ["m1_rm0_out"])
m1_sub_node = onnx.helper.make_node("Sub", ["m0_add1_out", "m1_rm0_out"], ["m1_sub_out"])
m1_mul0_node = onnx.helper.make_node("Mul", ["m1_sub_out", "m1_sub_out"], ["m1_mul0_out"])
m1_rm1_node = onnx.helper.make_node("ReduceMean", ["m1_mul0_out", "axes_const"], ["m1_rm1_out"])
m1_add0_node = onnx.helper.make_node("Add", ["m1_rm1_out", "eps_const"], ["m1_add0_out"])
m1_sqrt_node = onnx.helper.make_node("Sqrt", ["m1_add0_out"], ["m1_sqrt_out"])
m1_div_node = onnx.helper.make_node("Div", ["m1_sub_out", "m1_sqrt_out"], ["m1_div_out"])
m1_mul1_node = onnx.helper.make_node("Mul", ["m1_div_out", "scale_const"], ["m1_mul1_out"])
m1_add1_node = onnx.helper.make_node("Add", ["m1_mul1_out", "bias_const"], ["output"])

graph = onnx.helper.make_graph(
[
Expand All @@ -76,15 +87,24 @@ def build_model(self, shape, scale_val, bias_val):
l2_clip_node,
l2_expand_node,
l2_div_node,
m_rm0_node,
m_sub_node,
m_pow_node,
m_rm1_node,
m_add0_node,
m_sqrt_node,
m_div_node,
m_mul_node,
m_add1_node,
m0_rm0_node,
m0_sub_node,
m0_pow_node,
m0_rm1_node,
m0_add0_node,
m0_sqrt_node,
m0_div_node,
m0_mul_node,
m0_add1_node,
m1_rm0_node,
m1_sub_node,
m1_mul0_node,
m1_rm1_node,
m1_add0_node,
m1_sqrt_node,
m1_div_node,
m1_mul1_node,
m1_add1_node,
],
"qnn_f32_model",
[root_inp],
Expand Down Expand Up @@ -119,8 +139,8 @@ def test_all_fusions(self):

fused_model = onnx.load_model("model.qnn_pp.onnx")

# 3 fused Ops: Gelu, LpNorm, LayerNorm
self.assertEqual(len(fused_model.graph.node), 3)
# 4 fused Ops: Gelu, LpNorm, LayerNorm of two patterns
self.assertEqual(len(fused_model.graph.node), 4)
expected_op_types = {"Gelu", "LpNormalization", "LayerNormalization"}
for node in fused_model.graph.node:
self.assertIn(node.op_type, expected_op_types)
Expand Down Expand Up @@ -167,8 +187,8 @@ def test_external_data(self):

fused_model = onnx.load_model("model.qnn_pp.onnx", load_external_data=False)

# 3 fused Ops: Gelu, LpNorm, LayerNorm
self.assertEqual(len(fused_model.graph.node), 3)
# 4 fused Ops: Gelu, LpNorm, LayerNorm of two patterns
self.assertEqual(len(fused_model.graph.node), 4)
expected_op_types = {"Gelu", "LpNormalization", "LayerNormalization"}
for node in fused_model.graph.node:
self.assertIn(node.op_type, expected_op_types)
Expand Down