Open concretevitamin opened 7 months ago
Seems like this is because we will only select default instance types when no accelerators are specified: https://github.com/skypilot-org/skypilot/blob/ef211925c76e3e006c938edcf07678f6a9a3c45d/sky/clouds/aws.py#L438-L449
Are there any reasons we take this design?
No particular reason other than for simplicity. That comment is outdated since --cpus
acts as a filter already. It'd be great to support using most of the resource args as filters.
No particular reason other than for simplicity. That comment is outdated since
--cpus
acts as a filter already. It'd be great to support using most of the resource args as filters.
The main reason we only uses the default instance type families is to avoid weird CPU + memory combination, when only one of --cpus
or --memory
is specified. This is to make sure that most of the users will get a good default instance types.
Another reason for this is that the instances we selected have similar CPU type, i.e. intel Ice Lake. This is to avoid any compatibility issue for the user program caused by the different CPU series, e.g. ARM or AMD may not support some Intel CPU specific instruction set. We can first add those large-memory instance types with Intel CPUs.
This issue is stale because it has been open 120 days with no activity. Remove stale label or comment or this will be closed in 10 days.
Just tested on the latest master and this persists. Removing the stale label now 🫡
But both of these show >= 2048GB memory instance types:
cc @cblmemo @Michaelvll