Open erceguder opened 1 year ago
Hmm is this really necessary? Couldn't we just do the device conversion after loading?
Surely we can do as such. Yet, this might be useful when the processor_id
argument passed to the Processor
class constructor is a variable, and we want to utilize cuda for checkpoint-based processors. In this case, the proposed change avoids the need for explicitly type-checking whether the processor needs a checkpoint, to load it on cuda.
Hey, I've basically added a device argument for the Processor class constructor to allow for
cuda
, although defaults tocpu
. Directly built upon this PR. Backward compatible.