Try retryable steps for installing nvidia tools
I noticed some flakiness on trunk with this step: Install nvidia driver, nvidia-docker runtime, set GPU_FLAG
Trying retryable steps to see what would happen+whether it may alleviate the flakiness. The GitHub action does say that nodejs is required in the linux runners, so maybe it won't work
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73254
