bartorch.set_copy_inputs

bartorch.set_copy_inputs()

Set whether a BART command works on a private copy of each input tensor.

Some commands write into their inputs. With False a command is handed the tensor’s own memory and may modify it; a CUDA tensor given to a command that runs on the host is copied regardless. Operators never copy their inputs and are not affected.

Parameters:

copy (bool) – True, the default, copies every input.