TL;DR

The issue can be mitigated by ensuring consistent memory format handling when using torch.compile with F.pad on channels_last inputs.

Guidance

• Verify the memory format of the input tensor x before and after compilation to ensure it matches the expected torch.channels_last format.
• Check the documentation of torch.compile and F.pad to see if there are any known issues or limitations related to memory formats and compilation.
• Consider using the torch.memory_format attribute to explicitly set the memory format of the compiled tensor to match the eager tensor.
• Test the code with different backends, such as "eager" or other available options, to see if the issue is specific to the "aot_eager_decomp_partition" backend.

Example

import torch
import torch.nn.functional as F

def fn(x):
    return F.pad(x, (1, 2, 2, 1), mode="reflect")

x = torch.randn(2, 3, 4, 5).to(memory_format=torch.channels_last)

eager = fn(x.clone())
compiled = torch.compile(fn, backend="aot_eager_decomp_partition")(x.clone())

# Explicitly set the memory format of the compiled tensor
compiled = compiled.to(memory_format=torch.channels_last)

print("eager stride   =", eager.stride())
print("compiled stride=", compiled.stride())
print("eager channels_last   =", eager.is_contiguous(memory_format=torch.channels_last))
print("compiled channels_last=", compiled.is_contiguous(memory_format=torch.channels_last))

Notes

The issue seems to be related to the interaction between torch.compile and F.pad when using the torch.channels_last memory format. The provided code snippet and example may not fully resolve the issue, but they can help identify the root cause and potential workarounds.

Recommendation

Apply workaround: Explicitly set the memory format of the compiled tensor to match the eager tensor, as shown in the example code snippet. This may help ensure consistent memory format handling and mitigate the issue.