torch.roll converts nn.Parameter into regular torch.Tensor about pytorch HOT 3 CLOSED

Paramters generally should not be directly changed in the forward pass. For best practise, you may use self.register_buffer instead.

Anyway, if you really want to change the data in a parameter, use the following code:

def forward(self, x: torch.Tensor) -> torch.Tensor:
        n_tokens = x.size(1)
        with torch.no_grad():
            self.cache.data = torch.roll(self.cache, -n_tokens, dims=1)
        self.cache[:, -n_tokens:, :] = x
        return self.cache

By the way, I'm not sure what you want to do with MyModule(torch.zeros(2, 3, 4))(torch.zeros(2, 2, 4)). Generally the returned tensor/parameter cannot be called.

from pytorch.

This seems to be the expected behavior, you would expect an out of place op on a Parameter (e.g. torch.roll) to return a tensor, and you attempt to reassign this to self.cache, which fails with the expected error message

from pytorch.

Thank you for your reply. As you mentioned, register_buffer might be the correct way to go.
The goal of the cache is to append historical elements alongside the current tensor.

from pytorch.