🐛 Describe the bug

conv2d on MPS produces incorrect results when the input is a channel-slice view of a tensor stored in channels_last memory format.

The same operation on CPU is correct, and forcing .contiguous() on the sliced input also makes MPS correct again.

Minimal repro:

import torch
import torch.nn.functional as F

shared = torch.randn(1, 2, 1, 2).contiguous(memory_format=torch.channels_last)
task_slice = shared[:, :1]
weight = torch.randn(1, 1, 1, 1)

cpu_view = F.conv2d(task_slice, weight)
cpu_contig = F.conv2d(task_slice.contiguous(), weight)
print("cpu view vs contig", (cpu_view - cpu_contig).abs().max().item())

if torch.backends.mps.is_available():
    shared_mps = shared.to("mps")
    weight_mps = weight.to("mps")
    task_slice_mps = shared_mps[:, :1]

    mps_view = F.conv2d(task_slice_mps, weight_mps)
    mps_contig = F.conv2d(task_slice_mps.contiguous(), weight_mps)
    torch.mps.synchronize()

    print("mps view vs contig", (mps_view.cpu() - mps_contig.cpu()).abs().max().item())
    print("mps contig vs cpu", (mps_contig.cpu() - cpu_contig).abs().max().item())

Observed output on my system:

cpu view vs contig 0.0
mps view vs contig 0.30804574489593506
mps contig vs cpu 0.0

These two calls should produce the same result:

F.conv2d(task_slice, weight)
F.conv2d(task_slice.contiguous(), weight)

Versions

Collecting environment information... PyTorch version: 2.11.0 Is debug build: False CUDA used to build PyTorch: None ROCM used to build PyTorch: N/A

OS: macOS 26.4.1 (arm64) GCC version: Could not collect Clang version: 21.0.0 (clang-2100.0.123.102) CMake version: version 4.3.1 Libc version: N/A

Python version: 3.12.13 | packaged by Anaconda, Inc. | (main, Mar 19 2026, 20:12:32) [Clang 20.1.8 ] (64-bit runtime) Python platform: macOS-26.4.1-arm64-arm-64bit Is CUDA available: False CUDA runtime version: No CUDA CUDA_MODULE_LOADING set to: N/A GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA Is XPU available: False HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True Caching allocator config: N/A

CPU: Apple M1 Pro

Versions of relevant libraries: [pip3] numpy==2.4.2 [pip3] onnx==1.21.0 [pip3] torch==2.11.0 [pip3] torchaudio==2.10.0 [pip3] torchmetrics==1.8.2 [pip3] torchvision==0.26.0 [conda] numpy 2.4.2 pypi_0 pypi [conda] torch 2.11.0 pypi_0 pypi [conda] torchaudio 2.10.0 pypi_0 pypi [conda] torchmetrics 1.8.2 pypi_0 pypi [conda] torchvision 0.26.0 pypi_0 pypi

cc @jamesr66a @kulinseth @malfet @DenisVieriu97 @jhavukainen @aditvenk