Edge case if we try to patch inference-ness in ADInplaceOrView

inference_mode_propagation.py

import torch

class T(torch.Tensor):
    def __new__(cls, elem):
        return torch.Tensor._make_wrapper_subclass(cls, elem.shape, dtype=elem.dtype)

    def __init__(self, elem):
        self.elem = elem

    @classmethod
    def __torch_dispatch__(cls, func, types, args, kwargs=None):
        if func is torch.ops.aten.detach.default:
            (self,) = args
            # Return without wrapping!
            return torch.ops.aten.detach.default(self.elem)
        raise NotImplementedError(f"{func}")

# Inner tensor is non-inference
inner = torch.randn(3)
with torch.inference_mode():
    # Subclass wrapper is inference tensor
    x = T(inner)
    y = x.detach()
print(y.is_inference())  # False