2 ms·
I’ve been working with NVIDIA GPU VMs of late with the intention of running GPU enabled TensorFlow and PyTorch. I find working with NVIDIA software incredibly f
by julienchastang 4y ago
I’ve been working with NVIDIA GPU VMs of late with the intention of running GPU enabled TensorFlow and PyTorch. I find working with NVIDIA software incredibly frustrating. In particular, installation instructions for CUDA / cuDNN simply do not work. The biggest culprit is missing or incorrectly versioned shared object libraries. Forums are loaded with developers having similar problems never with any viable solution. I’ve had better luck with NVIDIA distributed docker containers. There at least I can make it as far as running GPU enabled TensorFlow sample code.
- ftufek 4y agoI'd recommend using docker or conda environments. Much easier to setup and to replicate.
- bick_nyers 4y agoMy guess is SR-IOV will still be locked to Quadro cards this gen, which is necessary for effectively sharing GPU resources across MULTIPLE VMs.
- survirtual 4y agoI was having this same issue until I just got into the habit of using a conda env. There are community conda repos that make getting all versions of needed libs much simpler (condaforge is one). Then it’s just a matter of managing envs for various projects. Smooth sailing.