undefined min read
PyTorch Internals 07 - Tensor Lifetime, the CUDA Caching Allocator, and Memory Reuse
PyTorch GPU memory behavior is shaped by a caching allocator, so observed memory usage is not just a story about current tensor objects