easy5 pts · first answer countsgpu-computingdistributed-trainingcheckpointing

Eight copies of the same checkpoint

In a DDP job on 8 GPUs, every rank writes model.pt to the same shared path at the same time. What is the usual fix?

A.Each rank saves to a random file nameB.Save only at the end of trainingC.Save on the CPU of every node twiceD.Only rank 0 saves the DDP model checkpoint

Discussion 0

Join free to post in the discussion.

Eight copies of the same checkpoint

Discussion 0

Related questions