Does ~ or logical_not work on TensorList right now?

No.

import torch

a = [torch.rand(2, 2, device="cuda") for _ in range(2)]

print(torch.logical_not(a[0])) # fine
print(torch.logical_not(a)) # TypeError: logical_not(): argument 'input' (position 1) must be Tensor, not list
print(~a[0]) # TypeError: ~ (operator.invert) is only implemented on integer and Boolean-type tensors


a = [torch.rand(2, 2, device="cuda").to(torch.bool) for _ in range(2)]
print(torch.logical_not(a[0])) # fine
print(~a[0]) # fine

print(torch.logical_not(a)) # TypeError: logical_not(): argument 'input' (position 1) must be Tensor, not list
print(~a) # TypeError: bad operand type for unary ~: 'list'

Should we improve the error message? Some things I am noticing:

1. If the user gets this error message, they did not actually use the - builtin operator. E.g. a user didn't do tensors = [...]; other_tensors = -tensors. So the "the - operator" part of the error message is confusing
2. The error message assumes that the user applied operations to a bool tensor. Maybe we should mention that the API being called is "_foreach_neg"
3. The error message shows up after a user calls _foreach_neg. But the error message suggests workarounds for a single tensor. This may be misunderstood to "just use ~ on your tensorlist"

Adding all the top together, maybe something like the following would make more sense?

_foreach_neg: There is a bool tensor in the passed-in TensorList.
Negation on a bool tensor is not supported. If you are trying to invert a mask, please
use the `~` or `logical_not()` operator on the individual tensors instead

Definitely not a blocker for this PR, but one thing we do with the base OpInfos is attempt to programmatically acquire the method and inplace variants of an op, which this could do, too, if they weren't specified.

done

These names could probably also be auto-generated by default from a string like 'exp', and the reference op also acquired from the same string automatically.

Done