nos

The open-source platform for running AI workloads on k8s in an optimized way, both in terms of hardware utilization and workload performance.

Maintainers

Name	Email	Url
Michele Zanotti	m.zanotti@nebuly.com
Diego Fiori	d.fiori@nebuly.com

Source Code

• https://github.com/nebuly-ai/nos

Values

Key	Type	Default	Description
allowDefaultNamespace	bool	false	If true allows to deploy nos chart in the default namespace
gpuPartitioner.affinity	object	{}	Sets the affinity config of the GPU Partitioner Pod.
gpuPartitioner.batchWindowIdleSeconds	int	10	Idle seconds before the GPU partitioner processes the current batch if no new pending Pods are created, and the timeout has not been reached. Higher values make the GPU partitioner will potentially take into account more pending Pods when deciding the GPU partitioning plan, but the partitioning will be performed less frequently
gpuPartitioner.batchWindowTimeoutSeconds	int	60	Timeout of the window used by the GPU partitioner for batching pending Pods. Higher values make the GPU partitioner will potentially take into account more pending Pods when deciding the GPU partitioning plan, but the partitioning will be performed less frequently
gpuPartitioner.devicePlugin.config.name	string	"nos-device-plugin-configs"	Name of the ConfigMap containing the NVIDIA Device Plugin configuration files. It must be equal to the value "devicePlugin.config.name" of the Helm chart used for deploying the NVIDIA GPU Operator.
gpuPartitioner.devicePlugin.config.namespace	string	"nebuly-nvidia"	Namespace of the ConfigMap containing the NVIDIA Device Plugin configuration files. It must be equal to the namespace where the Nebuly NVIDIA Device Plugin has been deployed to.
gpuPartitioner.devicePlugin.configUpdateDelaySeconds	int	5	Duration of the delay between when the new partitioning config is computed and when it is sent to the NVIDIA device plugin. Since the config is provided to the plugin as a mounted ConfigMap, this delay is required to ensure that the updated ConfigMap is propagated to the mounted volume.
gpuPartitioner.enabled	bool	true	Enable or disable the nos gpu partitioner
gpuPartitioner.fullnameOverride	string	""
gpuPartitioner.gpuAgent	object	-	Configuration of the GPU Agent component of the GPU Partitioner.
gpuPartitioner.gpuAgent.image.pullPolicy	string	"IfNotPresent"	Sets the GPU Agent Docker image pull policy.
gpuPartitioner.gpuAgent.image.repository	string	"ghcr.io/nebuly-ai/nos-gpu-agent"	Sets the GPU Agent Docker image.
gpuPartitioner.gpuAgent.image.tag	string	""	Overrides the GPU Agent image tag whose default is the chart appVersion.
gpuPartitioner.gpuAgent.logLevel	int	0	The level of log of the GPU Agent. Zero corresponds to info, while values greater or equal than 1 corresponds to higher debug levels. Must be >= 0.
gpuPartitioner.gpuAgent.reportConfigIntervalSeconds	int	10	Interval at which the mig-agent will report to k8s status of the GPUs of the Node
gpuPartitioner.gpuAgent.resources	object	{"limits":{"cpu":"100m","memory":"128Mi"}}	Sets the resource requests and limits of the GPU Agent container.
gpuPartitioner.gpuAgent.tolerations	list	[{"effect":"NoSchedule","key":"kubernetes.azure.com/scalesetpriority","operator":"Equal","value":"spot"}]	Sets the tolerations of the GPU Agent Pod.
gpuPartitioner.image.pullPolicy	string	"IfNotPresent"	Sets the GPU Partitioner Docker image pull policy.
gpuPartitioner.image.repository	string	"ghcr.io/nebuly-ai/nos-gpu-partitioner"	Sets the GPU Partitioner Docker image.
gpuPartitioner.image.tag	string	""	Overrides the GPU Partitioner image tag whose default is the chart appVersion.
gpuPartitioner.knownMigGeometries	list	-	List that associates GPU models to the respective allowed MIG configurations
gpuPartitioner.kubeRbacProxy	object	-	Configuration of the Kube RBAC Proxy, which runs as sidecar of all the GPU Partitioner components Pods.
gpuPartitioner.leaderElection.enabled	bool	true	Enables/Disables the leader election of the GPU Partitioner controller manager.
gpuPartitioner.logLevel	int	0	The level of log of the GPU Partitioner. Zero corresponds to info, while values greater or equal than 1 corresponds to higher debug levels. Must be >= 0.
gpuPartitioner.migAgent	object	-	Configuration of the MIG Agent component of the GPU Partitioner.
gpuPartitioner.migAgent.image.pullPolicy	string	"IfNotPresent"	Sets the MIG Agent Docker image pull policy.
gpuPartitioner.migAgent.image.repository	string	"ghcr.io/nebuly-ai/nos-mig-agent"	Sets the MIG Agent Docker image.
gpuPartitioner.migAgent.image.tag	string	""	Overrides the MIG Agent image tag whose default is the chart appVersion.
gpuPartitioner.migAgent.logLevel	int	0	The level of log of the MIG Agent. Zero corresponds to info, while values greater or equal than 1 corresponds to higher debug levels. Must be >= 0.
gpuPartitioner.migAgent.reportConfigIntervalSeconds	int	10	Interval at which the mig-agent will report to k8s the MIG partitioning status of the GPUs of the Node
gpuPartitioner.migAgent.resources	object	{"limits":{"cpu":"100m","memory":"128Mi"}}	Sets the resource requests and limits of the MIG Agent container.
gpuPartitioner.migAgent.tolerations	list	[{"effect":"NoSchedule","key":"kubernetes.azure.com/scalesetpriority","operator":"Equal","value":"spot"}]	Sets the tolerations of the MIG Agent Pod.
gpuPartitioner.nameOverride	string	""
gpuPartitioner.nodeSelector	object	{}	Sets the nodeSelector config of the GPU Partitioner Pod.
gpuPartitioner.podAnnotations	object	{}	Sets the annotations of the GPU Partitioner Pod.
gpuPartitioner.podSecurityContext	object	{"runAsNonRoot":true,"runAsUser":1000}	Sets the security context of the GPU partitioner Pod.
gpuPartitioner.replicaCount	int	1	Number of replicas of the gpu-manager Pod.
gpuPartitioner.resources	object	{"limits":{"cpu":"500m","memory":"128Mi"},"requests":{"cpu":"10m","memory":"64Mi"}}	Sets the resource limits and requests of the GPU partitioner container.
gpuPartitioner.scheduler.config.name	string	"nos-scheduler-config"	Name of the ConfigMap containing the k8s scheduler configuration file. If not specified or the ConfigMap does not exist, the GPU partitioner will use the default k8s scheduler profile.
gpuPartitioner.tolerations	list	[]	Sets the tolerations of the GPU Partitioner Pod.
nvidiaGpuResourceMemoryGB	int	32	Defines how many GB of memory each nvidia.com/gpu resource has.
operator.affinity	object	{}	Sets the affinity config of the operator Pod.
operator.enabled	bool	true	Enable or disable the nos operator
operator.fullnameOverride	string	""
operator.image.pullPolicy	string	"IfNotPresent"	Sets the operator Docker image pull policy.
operator.image.repository	string	"ghcr.io/nebuly-ai/nos-operator"	Sets the operator Docker repository
operator.image.tag	string	""	Overrides the operator Docker image tag whose default is the chart appVersion.
operator.kubeRbacProxy	object	-	Configuration of the Kube RBAC Proxy, which runs as sidecar of the operator Pods.
operator.leaderElection.enabled	bool	true	Enables/Disables the leader election of the operator controller manager.
operator.logLevel	int	0	The level of log of the controller manager. Zero corresponds to info, while values greater or equal than 1 corresponds to higher debug levels. Must be >= 0.
operator.nameOverride	string	""
operator.nodeSelector	object	{}	Sets the nodeSelector config of the operator Pod.
operator.podAnnotations	object	{}	Sets the annotations of the operator Pod.
operator.podSecurityContext	object	{"runAsNonRoot":true}	Sets the security context of the operator Pod.
operator.replicaCount	int	1	Number of replicas of the controller manager Pod.
operator.resources	object	{"limits":{"cpu":"500m","memory":"128Mi"},"requests":{"cpu":"10m","memory":"64Mi"}}	Sets the resource limits and requests of the operator controller manager container.
operator.securityContext	object	{"allowPrivilegeEscalation":false,"capabilities":{"drop":["ALL"]}}	Sets the security context of the operator container.
operator.tolerations	list	[]	Sets the tolerations of the operator Pod.
scheduler.affinity	object	{}	Sets the affinity config of the scheduler deployment.
scheduler.config	object	{}	Overrides the Kube Scheduler configuration
scheduler.enabled	bool	true	Enable or disable the nos scheduler
scheduler.fullnameOverride	string	""
scheduler.image.pullPolicy	string	"IfNotPresent"	Sets Docker image pull policy.
scheduler.image.repository	string	"ghcr.io/nebuly-ai/nos-scheduler"	Sets Docker image.
scheduler.image.tag	string	""	Overrides the image tag whose default is the chart appVersion.
scheduler.leaderElection.enabled	bool	true	Enables/Disables the leader election when deployed with multiple replicas.
scheduler.logLevel	int	0	The level of log of the scheduler. Zero corresponds to info, while values greater or equal than 1 corresponds to higher debug levels. Must be >= 0.
scheduler.nameOverride	string	""
scheduler.nodeSelector	object	{}	Sets the nodeSelector config of the scheduler deployment.
scheduler.podAnnotations	object	{}	Sets the annotations of the scheduler Pod.
scheduler.podSecurityContext	object	{}	Sets the security context of the scheduler Pod
scheduler.replicaCount	int	1	Number of replicas of the scheduler.
scheduler.resources	object	{"limits":{"cpu":"500m","memory":"128Mi"},"requests":{"cpu":"10m","memory":"64Mi"}}	Sets the resource limits and requests of the scheduler container.
scheduler.securityContext	object	{"privileged":false}	Sets the security context of the scheduler container
scheduler.tolerations	list	[]	Sets the tolerations of the scheduler deployment.
shareTelemetry	bool	true	If true, shares with Nebuly telemetry data collected only during the Chart installation