Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

TorchVision Main branch fails with `slow_conv2d: grad_weight must be contiguous`

See original GitHub issue

🐛 Describe the bug

It looks like a breakage caused by the 20220512 PyTorch Core Nightly.

Several CI jobs fail with:

test_classification_model[cpu-convnext_base]
Traceback (most recent call last):
  File "/root/project/test/test_models.py", line 628, in test_classification_model
    _check_input_backprop(model, x)
  File "/root/project/test/test_models.py", line 181, in _check_input_backprop
    out[0].sum().backward()
  File "/root/project/env/lib/python3.10/site-packages/torch/_tensor.py", line 399, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "/root/project/env/lib/python3.10/site-packages/torch/autograd/__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: slow_conv2d: grad_weight must be contiguous

So far I’ve identified the following PRs as potential causes:

Given that this network uses LayerNorm changes from the following PRs are also like to have caused it:

cc @mingfeima @VitalyFedyunin