Open
Description
Relates to #1163 & #1042. Several functions were updated with a Union[bool, Tensor] in #1042 to support receiving a signed: Tensor parameter from IntQuantTensor and passing it to all necessary downstream functions while compiling them all with the JIT. However, when considered in the context of #1163, it would be much better to not create an IntQuantTensor with a Tensor, but rather as a bool, but PyTorch's JIT doesn't allow us to do that.
Currently, requirements from the JIT require Tensor creation, which creates additional complexity in #1042, at the cost of slow QuantTensor creation (#1163). This should be revisited.
Metadata
Metadata
Assignees
Labels
No labels