Open LTJer opened 1 month ago
can you please clarify if you are asking about multithreaded training with the model on CPU vs using multiple dataloader processes when training with the model on GPU?
from what I've read online PyTorch defaults to one thread per physical core for parallel CPU ops. you can check the output of torch.get_num_threads to see if it's what you expect - if you're in an HPC environment you may want to double check what resources you are requesting
for multiple dataloader processes, the alignn config key num_workers sets the num_workers
argument of the pytorch DataLoader
Hi
How can I assign multiple cpus to the training process?
Thanks a lot.