What is the use case for this ?

What is the use case for this ?

This is needed to support pack_padded_sequence. There're also bunch of internal functions that check if it's in the tracing state, and this can unblock them from scripting.

What is the use case for this ?

This is needed to support pack_padded_sequence. There're also bunch of internal functions that check if it's in the tracing state, and this can unblock them from scripting.

You could also move that call to a function with torch.jit.ignore. I think it would be better to advocate people to reuse existing apis then add to a list of special cases

What is the use case for this ?

This is needed to support pack_padded_sequence. There're also bunch of internal functions that check if it's in the tracing state, and this can unblock them from scripting.

You could also move that call to a function with torch.jit.ignore. I think it would be better to advocate people to reuse existing apis then add to a list of special cases

I tried your approach, it unfortunately not working because of the torch.jit and torch.nn initialization order issue... I think this could considered to be an op as we already have different special ops.

You can do @torch._jit_internal.ignore (the one in jit just aliases this), but I don't see how that would help since the function would still be run (even if it returns False every time), so places that use it couldn't be exported.

@wanchaol merged this pull request in c384fbf.