Allow to customize AMIs used by AWS backend#1920
Conversation
| * SSH server listening on port 22 | ||
| * `user` with passwordless sudo access | ||
| * Docker is installed | ||
| * (For NVIDIA instances) NVIDIA/CUDA drivers are installed |
There was a problem hiding this comment.
And nvidia-container-toolkit
| image = image_config.nvidia if cuda else image_config.cpu | ||
| if image is None: | ||
| logger.warning("%s image not configured", "nvidia" if cuda else "cpu") | ||
| raise ComputeResourceNotFoundError() |
There was a problem hiding this comment.
(nit) Btw, alternatively we could filter out the offers that don't have an image configured. This way, we won't waste provisioning attempts and will show more realistic run plans. That would especially be useful for Azure, e.g. to filter out A10 offers if the GRID image is not configured.
Yet the current implementation provides more insight into why the offers aren't working, so I'm not sure which alternative is better and we can stick with the current one.
There was a problem hiding this comment.
As for now, I would stick to the current implementation, as I assume that the setup where custom AMIs are used, but not all (currently, all = only two) kinds are configured, is rare. In the future, if there are more kinds of images, we could revisit the logic.
3 participants
Closes: #1913