How to use precision FP16 in Pytorch extension?

[hminle]

Hi,

Thank for the great cuda lib.
I see in your binding code, there is a function to choose from FP16 and FP32.

But I don't know how to set FP16 with Pytorch extension.

[Tom94]

For the PyTorch bindings, all components take torch.float inputs and result in torch.half outputs by default.

The Encoding object can additionally be configured to output torch.float, by specifying the dtype in the constructor.