[Regression] 1.14RC1 side effects on graph output

[BowenBao]

Describe the issue

Discovered by PyTorch exporter, the 1.14RC1 seems to exhibit side effects on graph output that was used as one of the inputs for "quantizeLinear" operator. The output should be np.array(1.0), but instead becomes None in 1.14RC1.

The repro script and comparison between ORT 1.14RC1 and 1.13 can be found below.

To reproduce

import onnx
from onnx import numpy_helper
import numpy as np
import onnxruntime
import onnxscript
from onnxscript import opset17 as op

@onnxscript.script()
def repro(x: onnxscript.FLOAT):
    y_scale = op.Constant(value_float=1.0)
    y_zero_point = op.Constant(
        value=numpy_helper.from_array(np.array(1, dtype=np.uint8))
    )
    return op.QuantizeLinear(x, y_scale, y_zero_point), y_scale, y_zero_point

model: onnx.ModelProto = repro.to_model_proto()
onnx.save(model, "model.onnx")

session = onnxruntime.InferenceSession("model.onnx")
x = np.array([1.0, 2.0], dtype=np.float32)
print(session.run(None, {"x": x}))

""" ORT 1.14
2023-01-30 16:36:46.616346569 [W:onnxruntime:, execution_frame.cc:835 VerifyOutputSizes] Expected shape from model of {} does not match actual shape of {2} for output return_val0
[array([2, 3], dtype=uint8), None, array(1, dtype=uint8)]
"""

""" ORT 1.13
2023-01-30 16:36:27.314479278 [W:onnxruntime:, execution_frame.cc:828 VerifyOutputSizes] Expected shape from model of {} does not match actual shape of {2} for output return_val0
[array([2, 3], dtype=uint8), array(1., dtype=float32), array(1, dtype=uint8)]
"""

Urgency

No response

Platform

Linux

OS Version

Ubuntu 20.04

ONNX Runtime Installation

Other / Unknown

ONNX Runtime Version or Commit ID

ort-nightly==1.14.0.dev20230119003

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[yufenglee]

This is not specific to QuantizeLinear. Changing QuantizeLinear to Add also reproes.

import onnx
from onnx import numpy_helper
import numpy as np
import onnxruntime
import onnxscript
from onnxscript import opset17 as op

@onnxscript.script()
def repro(x: onnxscript.FLOAT):
    y_scale = op.Constant(value_float=1.0)
    y_zero_point = op.Constant(
        value=numpy_helper.from_array(np.array(1, dtype=np.uint8))
    )
    return op.Add(x, y_scale), y_scale, y_zero_point

model: onnx.ModelProto = repro.to_model_proto()
onnx.save(model, "model.onnx")

session = onnxruntime.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
x = np.array([1.0, 2.0], dtype=np.float32)
print(session.run(None, {"x": x}))

[RyanUnderhill]

Update: I think this was caused by this change: #12878 will ask Peng

[ytaous]

Looks like it's impacting 1.12.1+ by the timing when PR was merged if I am correct.

[RyanUnderhill]

Looks like it's impacting 1.12.1+ by the timing when PR was merged if I am correct.

It doesn't repro in 1.13.1. The PR was merged on 10/9/2022. Looks like 1.13.1 was forked on 9/26/2022