Context

Currently, the torch.compile backend for Torch-TensorRT requires @fake_tensor_unsupported due to a few functions which cannot handle fake tensors. This support needs to be updated in order to support fake tensors on the path to full-scale dynamic shape support.

Key Consideration: Is it possible to support fake tensors while preserving the notion of "constant" tensors in TRT?

Proposed Changes

The to_numpy function will need to be updated to take a scalar:

The create_constant function will need to be updated to not transform scalars into tensors:

Special considerations must be made for constant tensors with more than one element.