torch.compile fullgraph=True

[bhack]

It seems that this section it is not compilable fullgraph:

torch._dynamo.exc.Unsupported: call_method UserDefinedObjectVariable(forward) __call__ [TensorVariable(), TensorVariable(), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int), ConstantVariable(int)] {}

from user code:
   File "/opt/conda/lib/python3.10/site-packages/spatial_correlation_sampler/spatial_correlation_sampler.py", line 105, in forward
    return SpatialCorrelationSamplerFunction.apply(input1, input2, self.kernel_size,
  File "/opt/conda/lib/python3.10/site-packages/spatial_correlation_sampler/spatial_correlation_sampler.py", line 66, in forward
    output = correlation.forward(input1, input2,

[bhack]

This is required https://pytorch.org/docs/main/notes/custom_operators.html
@ClementPinard After pytorch/pytorch#127301 (comment) we are constantly notified to wrap it as custom operators.

[ClementPinard]

Hi, there has been lots of development in the cpp-extension repo lately.
https://github.com/pytorch/extension-cpp

I do believe that a related update to the tutorial is coming.

Once it's available, I'll try to update this code to be tracable and compilable. In the mean time, I prefer to wait because I don't have much time right now for experimentations.

I hope all this will be sorted out in the next month or so, stay tuned !

[bhack]

I think that it is available at:
https://docs.google.com/document/d/1-LdJZBzlxiF0Tm-8NfbyFvRJaofdwRgLcycXGmlIpS0

[ClementPinard]

Thanks for the link, did not know it was already in the works !

How did you find it, is it an official draft ?

I'll try to find some time to read it and apply it on this repo as soon as possible.

[bhack]

Yes it is official, you missed my previous comment link 😉
https://pytorch.org/docs/main/notes/custom_operators.html

[bhack]

Just a side note.. do you think that there is a second chance to test https://github.com/limacv/CorrelationLayer with torch.compile?
It seems that in the meantime it is not aligned anymore to your API but if it could be aligned without to much effort it could be interesting to test/benchmark it with modern torch.compile on pure python code.