pytorch - 💡(How to fix) Fix [MPS] Conv3d backward fails with "required rank 4 tensor to use channels_last format" when combined with view/permute/Linear [1 comments, 2 participants]

  RuntimeError: required rank 4 tensor to use channels_last format

The error occurs during backward(), not during the forward pass.

Key observations

• B=1 (BT=28): works. B≥2 (BT≥56): fails.
• Adding .contiguous() before any layer (Linear, einsum, conv2) does not fix it.
• All tensors are contiguous and not channels_last_3d during the forward pass.
• Each operation in isolation works fine — the bug only manifests in the combined backward graph.
• PYTORCH_ENABLE_MPS_FALLBACK=1 does not help.
• PYTORCH_MPS_PREFER_METAL=1 does not help (and causes additional Metal command buffer errors).
• Also reproduced on PyTorch nightly 2.12.0.dev20260323 — same failure.

Fix Plan

To fix the issue, we need to ensure that the tensor layout is compatible with the channels_last format required by the mps backend. We can achieve this by adding a to method call with the memory_format argument set to torch.channels_last_3d after the permute and reshape operations.

Here are the concrete steps:

• Add to method call with memory_format argument after permute and reshape operations:

x = x.view(B, T, C, -1).permute(0, 3, 2, 1).to(memory_format=torch.channels_last_3d)
x = torch.einsum('BNCT -> BNTC', x)
x = proj(x)
x = torch.einsum('BNTC -> BNCT', x)
x = x.permute(0, 3, 2, 1).reshape(BT, C, r, r, r).to(memory_format=torch.channels_last_3d)

• Verify that the fix works by running the modified code and checking that the backward pass completes without errors.

Verification

To verify that the fix worked, run the modified code and check that the backward pass completes without errors. You can do this by adding a try-except block around the backward call and checking that no exception is raised:

try:
    x.sum().backward()
    print("Backward pass completed successfully")
except RuntimeError as e:
    print(f"Error: {e}")

If the fix is successful, the backward pass should complete without errors, and the message "Backward pass completed successfully" should be printed.

Extra Tips

• Make sure to use the to method with the memory_format argument to ensure that the tensor layout is compatible with the channels_last format required by the mps backend.
• If you encounter similar issues in the future, try adding to method calls with the memory_format argument to ensure that the tensor layout is compatible with the required format.