How does PyTorch memory management work on GPU?

PyTorch uses caching allocator for efficient GPU memory reuse.

Instead of freeing memory immediately, PyTorch caches blocks to reduce allocation overhead, improving performance but sometimes causing fragmentation.