torch native functions cannot be used with inspect.signature

[ngoldbaum]

🚀 Feature

It would be nice if native functions were annotated with the necessary metadata to allow runtime introspection via the inspect module.

Motivation

Right now using e.g. inspect.signature with a function defined natively produces a ValueError:

>>> import inspect
>>> import torch
>>> inspect.signature(torch.mm)
*** ValueError: no signature found for builtin <built-in method mm of type object at 0x7ff8961cc940>

Pitch

It would be nice if the native function generation facilities in the pytorch build could create the function objects with the necessary metadata.

Cython is able to do this, so it's definitely possible:

>>> import cython
>>> import inspect
>>> scope = cython.inline("""def f(a,*args,b=False): pass """)
>>> inspect.signature(scope['f'])
<Signature (a, *args, b=False)>

Unfortunately I don't know enough about how cython does this or how python builtins can get annotated with the necessary metadata to point to further resources for doing this in pytorch.

Alternatives

Another way to do this would be for pytorch to make a public API that provides this information, since it's available in native_functions.yaml at build time. I think this is less nice since it should be possible to integrate with python's native introspection facilities.

[smessmer]

This is a good idea. I would review a PR implementing this.

[albanD]

I am not sure if this is possible to do though?
A lot of our functions are designed based on overloads. For example add:

@overload
def add(input: Union[Tensor, Number], other: Union[Tensor, Number], *, alpha: Optional[Number]=1, out: Optional[Tensor]=None) -> Tensor: ...
@overload
def add(self: Tensor, alpha: Number, other: Tensor) -> Tensor: ...
@overload
def add(self: Tensor, alpha: Number, other: Tensor, *, out: Tensor) -> Tensor: ...

So there isn't a single signature we can generate for this op.

[stuaxo]

I think you can build signatures that are build up of more than one signature, using Union

Looking at the code for @override, it keeps a registry you can lookup by module + qualified name, and line # -

https://github.com/python/cpython/blob/5c19050546e3e37a8889a0baa2954e1444e803d3/Lib/typing.py#L2550

So it may be possible to build a Union from all the signatures collected by override.

[asmeurer]

Would it at least be possible to include all those overloads in the documentation? Right now help(torch.add) and the online docs show add(input, other, *, alpha=1, out=None) -> Tensor, which is just one of the above listed overloads. These are shown in error messages, but it would be helpful to include them in the docs as well.

[stuaxo]

@asmeurer that possible, and worth adding as another issue.

__doc__ can be created when the new function is. If it's still difficult, I've noticed __doc__ can be an fstring.

[blueardour]

any update?