Inferencing types at saving is not feasible sometimes

[aguschin]

The type inference is nice. But it requires running the model on the machine packaging the model. This is a problem if I want to package a model requiring a GPU, on a machine which does not have a GPU (e.g. in CI/CD), or with a GPU too small to run the model with the parameters we want to have in production.