itamarst
commented
4 years ago It sounds like unified memory addressing might make this... tricky? Or perhaps impossible.
On the other hand: CuPy has memory allocation hooks, and already has an existing memory profiler. So package-specific integration might work.
For CUDA it's possible we can track allocations by intercepting
cudaMalloc()and friends. This would need to be tracked and reported separately than CPU memory.https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html