How do you profile and debug a model that is experiencing GPU memory fragmentation or out-of-memory (OOM) errors during training?
