Implicit data type promotion in torch.cat is undocumented about pytorch HOT 1 OPEN

I also found some description in the Developer FAQ

While type promotion can be convenient, it can also be confusing to users and a source of error. Therefore, we've decided to limit type promotion in PyTorch to the following classes of operations:

• unary pointwise operations where the codomain of the input cannot be represented in the input's dtype
• binary pointwise operations
• reductions where the codomain of the input cannot be represented in the input's dtype

While some operations outside of these classes implement type promotion today, PyTorch's current plan is only to add type promotion support to unary pointwise, binary pointwise, and reduction operations. Other operations supporting type promotion will not be changed to support backwards compatibility.

IMO, for backwards compatibility the behavior shouldn't be changed. But it's better to document the behavior as the wiki suggested it can also be confusing to users and a source of error.

Thanks

@bdhirsh @svekars @brycebortree

from pytorch.