How does TorchInductor handle aliasing and mutation, does it always perform a functionalization, or are they exposed to the backend compilers? The description above mentions the concept of TensorBox and StorageBox, are there more details on these?
How does TorchInductor handle aliasing and mutation, does it always perform a functionalization, or are they exposed to the backend compilers? The description above mentions the concept of TensorBox and StorageBox, are there more details on these?