ORT_ENABLE_ALL produces incorrect output for QuantizeLinear-DequantizeLinear Resize(linear) pattern

[ALinrunrun]

Describe the issue

ONNX Runtime CPUExecutionProvider produces different results for a QDQ Resize(linear) pattern when graph optimization is enabled.

The model applies:

QuantizeLinear -> DequantizeLinear -> Resize(linear, half_pixel) -> QuantizeLinear -> DequantizeLinear

With ORT_DISABLE_ALL, the output matches the expected result:

[2.0, 3.0, 6.0, 6.0, 5.0, 4.0]

With ORT_ENABLE_ALL, ORT returns:

[2.0, 3.0, 5.0, 6.0, 4.0, 4.0]

This changes the model result after optimization.

To reproduce

import numpy as np
import onnxruntime as ort
from onnx import TensorProto, helper, numpy_helper

inits = [
numpy_helper.from_array(np.array([1, 1, 1, 6], dtype=np.int64), "sizes"),
numpy_helper.from_array(np.float32(1.0), "sc"),
numpy_helper.from_array(np.int8(0), "zp"),
]

nodes = [
helper.make_node("QuantizeLinear", ["X", "sc", "zp"], ["Xq"]),
helper.make_node("DequantizeLinear", ["Xq", "sc", "zp"], ["Xd"]),
helper.make_node(
"Resize",
["Xd", "", "", "sizes"],
["R"],
mode="linear",
coordinate_transformation_mode="half_pixel",
),
helper.make_node("QuantizeLinear", ["R", "sc", "zp"], ["Yq"]),
helper.make_node("DequantizeLinear", ["Yq", "sc", "zp"], ["Y"]),
]

g = helper.make_graph(
nodes,
"g",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 1, 1, 3])],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 1, 1, 6])],
initializer=inits,
)

m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 19)])
m.ir_version = 9
mb = m.SerializeToString()

x = np.array([[[[2.0, 7.0, 4.0]]]], dtype=np.float32)

def run(opt_level):
so = ort.SessionOptions()
so.graph_optimization_level = opt_level
sess = ort.InferenceSession(
mb,
sess_options=so,
providers=["CPUExecutionProvider"],
)
return sess.run(None, {"X": x})[0].flatten()

out_disable = run(ort.GraphOptimizationLevel.ORT_DISABLE_ALL)
out_enable = run(ort.GraphOptimizationLevel.ORT_ENABLE_ALL)

expected = np.array([2.0, 3.0, 6.0, 6.0, 5.0, 4.0], dtype=np.float32)

print("expected:", expected.tolist())
print("ORT_DISABLE_ALL:", out_disable.tolist())
print("ORT_ENABLE_ALL: ", out_enable.tolist())
print("DISABLE matches:", np.array_equal(out_disable, expected))
print("ENABLE matches: ", np.array_equal(out_enable, expected))

Urgency

Expected output

expected: [2.0, 3.0, 6.0, 6.0, 5.0, 4.0]
ORT_DISABLE_ALL: [2.0, 3.0, 6.0, 6.0, 5.0, 4.0]
ORT_ENABLE_ALL:  [2.0, 3.0, 6.0, 6.0, 5.0, 4.0]
DISABLE matches: True
ENABLE matches:  True

Actual output

expected: [2.0, 3.0, 6.0, 6.0, 5.0, 4.0]
ORT_DISABLE_ALL: [2.0, 3.0, 6.0, 6.0, 5.0, 4.0]
ORT_ENABLE_ALL:  [2.0, 3.0, 5.0, 6.0, 4.0, 4.0]
DISABLE matches: True
ENABLE matches:  False

Platform

Linux

OS Version

Linux-6.17.0-20-generic-x86_64-with-glibc2.39

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.25.1

ONNX Runtime API

Python

Architecture

X86

Execution Provider

Default CPU

Execution Provider Library Version

No response