Comment on Do We Need a New Orchestration System for GPUs?parentComments−brucethemoose23yAnother factor is your system libs. An old PyTorch/CUDA combo, for instance, can hold 4000 GPUs back.I have been quite happy with Arch's build of PyTorch (which targets CUDA 12.1 at the moment).
Comments
Another factor is your system libs. An old PyTorch/CUDA combo, for instance, can hold 4000 GPUs back.
I have been quite happy with Arch's build of PyTorch (which targets CUDA 12.1 at the moment).