Storing Quasi-Constants in a Namespace

I want to store a small set of useful-to-have constants in my model:

ZERO = torch.tensor( float(0.0) )
ONE = torch.tensor( float(0.0) )
INF = torch.tensor( float('inf') )
NAN = torch.tensor( float('nan') )
TRUE = torch.tensor( bool(True) )

These are often needed when e.g. perfoming operations like torch.where(mask, tensor, ZERO)

What is the best way to store these kinds of constants in my model? I have the follwoing desiderata:

  1. They are cast to the correct device/dtype when using
  2. They are collected in a namespace model.constants
  3. I do not need to copy-paste the whole code in each model every time
  4. The list is extendable on a per-model basis
  5. It is fully compatible with JIT

Currently, I am doing

class model(nn.Module):
    ZERO: torch.Tensor
    def __init__(self):
          self.register_buffer('ZERO', torch.tensor(0.0))

Which satisfies (1, 4, 5) but violates (2, 3). Any better ideas?