ORT_ENABLE_ALL changes Constant DequantizeLinear Reshape output for int8 weights

[ALinrunrun]

Describe the issue

ONNX Runtime CPUExecutionProvider produces different results for a constant DequantizeLinear -> Reshape pattern when graph optimization is enabled.

The model contains an int8 constant tensor:

[12, -45, 7, -3]

with scale 0.02, followed by DequantizeLinear and Reshape.

With ORT_DISABLE_ALL, the output is correct:

[[0.24, -0.9], [0.14, -0.06]]

With ORT_ENABLE_ALL, the negative dequantized values become 0:

[[0.24, 0.0], [0.14, 0.0]]

This indicates that an optimization or constant-folding path changes the semantics of the model.

To reproduce

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

def make_model():
    weights = np.array([12, -45, 7, -3], dtype=np.int8)
    scale = np.float32(0.02)
    shape = np.array([2, 2], dtype=np.int64)

    nodes = [
        helper.make_node("Constant", [], ["wq"], value=numpy_helper.from_array(weights)),
        helper.make_node("Constant", [], ["scale"], value=numpy_helper.from_array(scale)),
        helper.make_node("Constant", [], ["shape"], value=numpy_helper.from_array(shape)),
        helper.make_node("DequantizeLinear", ["wq", "scale"], ["wf"]),
        helper.make_node("Reshape", ["wf", "shape"], ["y"]),
    ]

    graph = helper.make_graph(
        nodes,
        "g",
        [],
        [helper.make_tensor_value_info("y", TensorProto.FLOAT, [2, 2])],
    )

    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 21)])
    model.ir_version = 10
    return model.SerializeToString()

def run(level):
    options = ort.SessionOptions()
    options.graph_optimization_level = level
    sess = ort.InferenceSession(
        make_model(),
        sess_options=options,
        providers=["CPUExecutionProvider"],
    )
    return sess.run(None, {})[0]

disabled = run(ort.GraphOptimizationLevel.ORT_DISABLE_ALL)
enabled = run(ort.GraphOptimizationLevel.ORT_ENABLE_ALL)

print("disable_all:", disabled)
print("enable_all:", enabled)
print("max_abs_diff:", float(np.max(np.abs(disabled - enabled))))
print("PASS=", np.allclose(disabled, enabled))

sys.exit(0 if not np.allclose(disabled, enabled) else 1)

Urgency

Expected output

disable_all: [[ 0.24 -0.9 ]
 [ 0.14 -0.06]]
enable_all: [[ 0.24 -0.9 ]
 [ 0.14 -0.06]]
max_abs_diff: 0.0
PASS=True

Actual output

disable_all: [[ 0.24 -0.9 ]
 [ 0.14 -0.06]]
enable_all: [[0.24 0.  ]
 [0.14 0.  ]]
max_abs_diff: 0.8999999761581421
PASS=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