The PyTorch docs only have documentation on how to tweak its memory management for CUDA allocations. It uses the caching allocator for it’s memory management which maintains the allocated memory blocks and reuses them instead of freeing them immediately.