what should be the convention of variable dtype in deepmd-kit modules.

[wanghan-iapcm]

Summary

how to define variable dtype should be made more clear for developers. see details.

DeePMD-kit Version

devel-c2944eb

Backend and its version

pytorch, python

Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

No response

Details

There are at least two cases:

1. model attributes: in this case it seems that the dtype of the variable is defined by the precision parameter of the constructor, but I find some exceptions like
   

   deepmd-kit/deepmd/pt/model/task/polarizability.py

   Lines 117 to 123 in c2944eb

   	self.scale = torch.tensor(
   	self.scale, dtype=env.GLOBAL_PT_FLOAT_PRECISION, device=env.DEVICE
   	).view(ntypes, 1)
   	self.shift_diag = shift_diag
   	self.constant_matrix = torch.zeros(
   	ntypes, dtype=env.GLOBAL_PT_FLOAT_PRECISION, device=env.DEVICE
   	)

   
   where the dtype is defined globally.
2. model output: in this case it seems that the dtype should be either
   a. consistent with the attributes' dtype or
   b. consistent with the input dtype.
3. but I find exception like
   

   deepmd-kit/deepmd/pt/model/task/fitting.py

   Lines 466 to 470 in c2944eb

   	outs = torch.zeros(
   	(nf, nloc, net_dim_out),
   	dtype=env.GLOBAL_PT_FLOAT_PRECISION,
   	device=descriptor.device,
   	) # jit assertion

I would suggest we made a decision for the two cases, and all developers should follow the same convention.