I'm working on a module that takes a LongTensor as argument, that I register as an indices buffer (https://github.com/lantiga/pytorch/blob/idx2col/torch/nn/modules/conv.py#L706).

Calling float() on the module casts the indices buffer to FloatTensor, which I believe is not a desired behavior in general. In particular, it breaks subsequent calls to TH* functions, which expect either DoubleTensor or FloatTensor for certain arguments but always LongTensor for others.

IMO registered buffers/params with integer types should be immune to calls to float or double, which are really there to control the precision of floating point buffers/params.

I can take care of this in case the above makes sense.