This solves a huge pain-point. I do wonder if there is some lower level of abstraction to PyTorch that is more explicit about GPU memory allocation. Often, hun

This solves a huge pain-point. I do wonder if there is some lower level of abstraction to PyTorch that is more explicit about GPU memory allocation. Often, hunting down why GPU memory usage keeps increasing after every batch feels like a needle-in-a-haystack problem. I would in fact be happy to use the C++ API (or another) to PyTorch, if it is more transparent about memory allocation.