Mitigation: Include proper GPU counts for current VM offerings
Opened this issue · 0 comments
ryanhamel commented
The idea is to include a new default_resource that comes with scalelib by default, that will hard code the proper GPU values for known VM Sizes. Ideally, we create a new resource called something like "scalelib_gpus", and then schedulers can set their gpu resource to scalelib_gpus. I do not want this defined in the autoscale.json by default, but when a user runs az* config, they should see the definitions of scalelib_gpus for each relevant VM Size, and if they do not see their VM Size, they can simply use this output as a template. Of course, this is in addition to documenting it in the readme.