ConvTranspose with output_shape has completely different output for CPU only vs ALL

[skottmckay]

🐞Describing the bug

• Make sure to only create an issue here for bugs in the coremltools Python package. If this is a bug with the Core ML Framework or Xcode, please submit your bug here: https://developer.apple.com/bug-reporting/
• Provide a clear and consise description of the bug.
  Output from ConvTranspose when output_shape is specified is significantly different depending on which devices are enabled.

To Reproduce

• Please add a minimal code example that can reproduce the error when running it.

import numpy as np
import coremltools as ct
from coremltools.converters.mil import Builder as mb

target = ct.target.iOS15

x_shape = (1, 3, 4, 4)
w_shape = (3, 3, 3, 3)

@mb.program(input_specs=[mb.TensorSpec(shape=x_shape)],
            opset_version=target,
            )
def prog(x):
    weight = mb.const(name="weight", val=np.ones(w_shape, dtype=np.float32))
    output_shape = mb.const(name="output_shape", val=np.array([1, 3, 4, 4]))
    strides = mb.const(name="strides", val=np.ones((2), dtype=np.int32))
    dilations = mb.const(name="dilations", val=np.ones((2), dtype=np.int32))
    z = mb.conv_transpose(x=x, weight=weight, strides=strides, dilations=dilations, output_shape=output_shape)
    return z

m = ct.convert(prog, minimum_deployment_target=target, compute_precision=ct.precision.FLOAT32)

m.save("ConvTranspose.mlpackage")

m_cpu = ct.models.MLModel('ConvTranspose.mlpackage', compute_units=ct.ComputeUnit.CPU_ONLY)
m_all = ct.models.MLModel('ConvTranspose.mlpackage', compute_units=ct.ComputeUnit.ALL)

x = np.ones(x_shape, dtype=np.float32)
print(m_cpu.predict({'x': x}))
print(m_all.predict({'x': x}))

Output appears to indicate the CPU only version is offset by a row.

The non-CPU output is what I expected (based on translating an ONNX ConvTranspose to the CoreML version, but I would have expected that the CPU-only version is the reference and should be considered to be correct.

{'conv_transpose_0': array([[[[ 3.,  6.,  9.,  9.],
         [ 6., 12., 18., 18.],
         [ 9., 18., 27., 27.],
         [ 9., 18., 27., 27.]],

        [[ 3.,  6.,  9.,  9.],
         [ 6., 12., 18., 18.],
         [ 9., 18., 27., 27.],
         [ 9., 18., 27., 27.]],

        [[ 3.,  6.,  9.,  9.],
         [ 6., 12., 18., 18.],
         [ 9., 18., 27., 27.],
         [ 9., 18., 27., 27.]]]], dtype=float32)}
{'conv_transpose_0': array([[[[12., 18., 18., 12.],
         [18., 27., 27., 18.],
         [18., 27., 27., 18.],
         [12., 18., 18., 12.]],

        [[12., 18., 18., 12.],
         [18., 27., 27., 18.],
         [18., 27., 27., 18.],
         [12., 18., 18., 12.]],

        [[12., 18., 18., 12.],
         [18., 27., 27., 18.],
         [18., 27., 27., 18.],
         [12., 18., 18., 12.]]]], dtype=float32)}

System environment (please complete the following information):

• coremltools version: 7.1

• OS (e.g. MacOS version or Linux type): macOS
  Model Name: Mac mini
  Model Identifier: Macmini9,1
  Model Number: MGNT3LL/A
  Chip: Apple M1

  ProductName: macOS
  ProductVersion: 14.5
  BuildVersion: 23F79

[TobyRoseman]

I cannot reproduce this issue. It works fine for me on macOS 14.3.

This is going to be an issue with the Core ML Frame work, not the coremltools Python package. Please file a bug report using The Feedback Assitant.