To both solve the issue where CycleCloud cannot report the proper number of GPUs, due to incorrect backend data, as well as the use case where a user needs to override the reported number of GPUs, allow the ability to define these using scalelib's built-in default_resources mechanism.
For example, by default we would ship with simply
"default_resources": [
{
"select": {},
"name": "slurm_gpus",
"value": "node.gpu_count"
}
],
However, if a user wants to / needs to define Standard_XYZ as having 8 GPUs, they can do that by simply editing /opt/azure/slurm/autoscale.json:
"default_resources": [
{
"select": {"node.vm_size": "Standard_XYZ"},
"name": "slurm_gpus",
"value": 8
},
{
"select": {},
"name": "slurm_gpus",
"value": "node.gpu_count"
}
],
To both solve the issue where CycleCloud cannot report the proper number of GPUs, due to incorrect backend data, as well as the use case where a user needs to override the reported number of GPUs, allow the ability to define these using scalelib's built-in default_resources mechanism. For example, by default we would ship with simply "default_resources": [ { "select": {}, "name": "slurm_gpus", "value": "node.gpu_count" } ],
However, if a user wants to / needs to define Standard_XYZ as having 8 GPUs, they can do that by simply editing /opt/azure/slurm/autoscale.json: "default_resources": [ { "select": {"node.vm_size": "Standard_XYZ"}, "name": "slurm_gpus", "value": 8 }, { "select": {}, "name": "slurm_gpus", "value": "node.gpu_count" } ],