easy5 pts · first answer countsgpu-computinggpu-memory

First thing to try after OOM

A fine-tuning run goes out of memory at batch size 32. What is usually the quickest change that keeps the effective batch size the same?

A.Batch 64 with no accumulationB.Batch 32 on the CPU insteadC.Batch 8 with gradient accumulation of 4D.Batch 32 with a higher learning rate

Discussion 0

Join free to post in the discussion.