#memory-allocator

PyTorch CUDA pool teardown can re-enter half-destroyed state

Review of pending allocator work also flags null returns that can produce CUDA tensors backed by address zero.