Starting a pod with runpod/pytorch:2.4.0-py3.11-cuda12.4.1-devel-ubuntu22.04 has cuda version 12.6
I am confused what determines the cuda version of a pod I start. I would expect that when I start a docker image with a cuda version in the name that it has this cuda version bundled into the image and when I start the pod that this is the cuda version I see, but this is not the case. How can I start a pod with a predictable cuda version?
4 Replies
Ping
You have two diffrent CUDA
Host CUDA
Template CUDA
CUDA is backward compatible so if HOST CUDA is 12.6 it can supports templates with CUDA up to 12.6
hm but when i ssh into a pod and check the cuda version I should see the template cuda version right? but that is the cuda version that is unpredictable for me
You can filter the host Cuda version when you deploy a pod
Expand the menu on top of the page when you create a pod