Sorry, I meant GPU memory usage. I’m just trying to verify that when the “Limit” above is reached, there’s still GPU memory available. XLA actually preallocates memory, so you can just do Nx.tensor([1]) and then see if 50% of the memory is still free.
