[jit] torch.isfinite is broken

[driazati]

def fn(x):
    print(torch.isfinite(x))


s = torch.jit.script(fn)
fn(torch.randn(2, 2))
s(torch.randn(2, 2))

results in

tensor([[True, True],
        [True, True]])
graph(%x.1 : Tensor):
  %4 : None = prim::Constant() # ../test.py:14:0
  %2 : float = prim::ImplicitTensorToNum(%x.1) # ../test.py:15:10
  %3 : bool = aten::isfinite(%2) # ../test.py:15:10
   = prim::Print(%3) # ../test.py:15:4
  return (%4)                                                                                                                                   
Traceback (most recent call last):
  File "../test.py", line 21, in <module>
    s(torch.randn(2, 2))
RuntimeError: Cannot input a tensor of dimension other than 0 as a scalar argument
The above operation failed in interpreter, with the following stack trace:
at ../test.py:15:10
def fn(x):
    print(torch.isfinite(x))
          ~~~~~~~~~~~~~~ <--- HERE

It crashes at runtime and the signature is wrong since the actual op is not bound in, it's instead done here

pytorch/torch/csrc/jit/register_prim_ops.cpp

Line 2941 in 0d9dc46

DEFINE_UNARY_FLOAT_OP(aten::isfinite, std::isfinite(a), bool),

cc @suo

[eellison]

This is an example of a general source of bugs in TorchScript, which is that if you register an operator that should be overloaded with both Scalar and Tensor for an argument for only the Scalar case it will lead to bad implicit casting and a runtime error. The fix is to just expose the op that has a Tensor input.

See also: #27512

#20664

[zetyquickly]

This is an example of a general source of bugs in TorchScript, which is that if you register an operator that should be overloaded with both Scalar and Tensor for an argument for only the Scalar case it will lead to bad implicit casting and a runtime error. The fix is to just expose the op that has a Tensor input.

See also: #27512

#20664

Thanks a lot for your reply,
Does expose means to copypaste code of function or that means to specify input types to a function call?

[eellison]

Hmm actually, looks like we're not correctly resolving torch.isfinite to torch.functional.isfinite.

A workaround is to call torch.functional.isfinite(x)

[zetyquickly]

Hmm actually, looks like we're not correctly resolving torch.isfinite to torch.functional.isfinite.

A workaround is to call torch.functional.isfinite(x)

I've tried, it won't work. F.isfinite(x) is returning 'bool' while scripting and Tensor in Python

Variable 'bool_tensor' is annotated with type Tensor but is being assigned to a value of type bool:
at box_regression.py:84:8
        """
            deltas (Tensor): some tensor
        """
        bool_tensor : torch.Tensor = torch.functional.isfinite(deltas)
        ~~~~~~~~~~~ <--- HERE

[eellison]

Hmm, actually you're right, all of the functions in torch/functional.py aren't working in JIT unless they are a 1-1 mapping with the C++ functions. Also those only defined in torch/tensor.py don't work...

[zetyquickly]

Hmm, actually you're right, all of the functions in torch/functional.py aren't working in JIT unless they are a 1-1 mapping with the C++ functions. Also those only defined in torch/tensor.py don't work...

How can we solve it, what type of changes could we propose to repository?

[eellison]

FYI: #28918

[eellison]

This should be fixed now.