I think it'd be better to add annotations to arguments in cwrap, or use the before_call attribute, instead of reimplementing the whole function.

This is the cleanest way I could figure out. Can't think of any other way to modify the dims using just before_call. Does this look better?

This looks better. Can you please add a test for it?

A test asserting that transpose with positive and negative dimensions give the same result is good enough?

Yeah. Just do a for loop from 0 to nDim, and use positive and negative values to ensure that they yield the same result.

Great. Added a test.

@pytorchbot test this please

The builds are failing for CUDA. Maybe you should add LIBRARY_STATE in the beginning of the function calls so that it can be CUDA-compatible?

Adding LIBRARY_STATE is not fixing this. For CUDA, it is trying to pass the LIBRARY_STATE as the first argument to my function THTensor_(transpose_neg), which does not accept that.

I think either this needs to go to the backend libs as a wrapper like newTranspose or maybe a new implementation of the fix using before_call etc.

@pytorchbot test this please

Thanks a lot Hardik!