AWS Trainium and Inferentia silicon accelerators

Domino supports the use of cost and energy-efficient AWS-designed silicon processors, AWS Trainum and Inferentia, to accelerate deep-learning model training and AI inference workloads. Use the AWS Neuron SDK to reuse existing code. Learn how to set up Trainium and Inferentia accelerators in your Domino. Set up involves the following:

Node group creation: Create a new node group for Trainium and Inferentia instances.
Device plugin configuration: Provide hardware-specific settings.
Hardware tier setup: Enable Domino users to use Trainium and Inferentia instances for their workloads.
Environment configuration: Set up the necessary development tools and software libraries.

Node group creation

To use AWS accelerators, create a new node group that:

Uses one of the instance types in the following table
Uses the GPU-enabled AMI for your EKS version
Has a unique node pool label identifying its accelerator type

Name	vCPU	Memory (GiB)	aws.amazon.com/neuron	Total Neuron Memory (GiB)
inf1.xlarge	4	8	1	8
inf1.2xlarge	8	16	1	8
inf1.6xlarge	24	48	4	32
inf1.24xlarge	96	192	16	128
inf2.xlarge	4	16	1	32
inf2.8xlarge	32	128	1	32
inf2.24xlarge	96	384	6	192
inf2.48xlarge	192	768	12	384
trn1.2xlarge	8	32	1	32
trn1.32xlarge	128	512	16	512

For the cluster-autoscaler to scale your Neuron-based node group successfully, you must tag those autoscaling groups with the Neuron device resource template, like the k8s.io/cluster-autoscaler/node-template/resources/aws.amazon.com/neuron tag in the table below:

Key	Value	Tag new instances
Name	inferentia-test-domino-trn1-Node	Yes
alpha.eksctl.io/cluster-name	inferentia-test	Yes
alpha.eksctl.io/eksctl-version	0.155.0	Yes
alpha.eksctl.io/nodegroup-name	domino-trn1	Yes
alpha.eksctl.io/nodegroup-type	unmanaged	Yes
aws:cloudformation:logical-id	NodeGroup	Yes
aws:cloudformation:stack-id	arn:aws:cloudformation:us-west-2:873872646799:stack/eksctl-inferentia-test-n…	Yes
aws:cloudformation:stack-name	eksctl-inferentia-test-nodegroup-domino-trn1	Yes
eksctl.cluster.k8s.io/v1alpha1/cluster-name	inferentia-test	Yes
eksctl.io/v1alpha2/nodegroup-name	domino-trn1	Yes
k8s.io/cluster-autoscaler/enabled	true	Yes
k8s.io/cluster-autoscaler/inferentia-test	owned	Yes
k8s.io/cluster-autoscaler/node-template/label/dominodatalab.com/node-pool	trainium	Yes
k8s.io/cluster-autoscaler/node-template/resources/aws.amazon.com/neuron	1	Yes
kubernetes.io/cluster/inferentia-test	owned	Yes

Example eksctl node group config

Here’s an example eksctl node group config for Neuron-based node groups:

apiVersion: eksctl.io/v1alpha5
kind: ClusterConfig

metadata:
  name: inferentia-test
  region: us-west-2

nodeGroups:
  - name: domino-trn1
    instanceType: trn1.2xlarge
    minSize: 0
    maxSize: 3
    desiredCapacity: 1
    volumeSize: 200
    volumeType: gp3
    availabilityZones: ["us-west-2a"]
    labels:
      "dominodatalab.com/node-pool": "trainium"
    tags:
      k8s.io/cluster-autoscaler/node-template/dominodatalab.com/node-pool: trainium
      k8s.io/cluster-autoscaler/node-template/resources/aws.amazon.com/neuron: 1
    iam:
      withAddonPolicies:
        ebs: true
        efs: true
  - name: domino-inf1
    instanceType: inf1.2xlarge
    minSize: 0
    maxSize: 3
    desiredCapacity: 1
    volumeSize: 200
    volumeType: gp3
    availabilityZones: ["us-west-2a"]
    labels:
      "dominodatalab.com/node-pool": "inferentia"
    tags:
      k8s.io/cluster-autoscaler/node-template/dominodatalab.com/node-pool: inferentia
      k8s.io/cluster-autoscaler/node-template/resources/aws.amazon.com/neuron: 1
    iam:
      withAddonPolicies:
        ebs: true
        efs: true
  - name: domino-inf2
    instanceType: inf2.2xlarge
    minSize: 0
    maxSize: 3
    desiredCapacity: 1
    volumeSize: 200
    volumeType: gp3
    availabilityZones: ["us-west-2a"]
    labels:
      "dominodatalab.com/node-pool": "inferentia2"
    tags:
      k8s.io/cluster-autoscaler/node-template/dominodatalab.com/node-pool: inferentia2
      k8s.io/cluster-autoscaler/node-template/resources/aws.amazon.com/neuron: 1
    iam:
      withAddonPolicies:
        ebs: true
        efs: true

Device plugin deployment

Once your nodes have joined the cluster, deploy the Neuron device plugin DaemonSet using the following specification. You must use version 2.17.3.0 or greater for Domino workloads to be correctly processed by the device plugin. To deploy this DaemonSet:

Save the following specification to a file (such as neuron-device-plugin-ds.yaml).

Apply the specification with kubectl apply -f neuron-device-plugin-ds.yaml.

---
kind: ClusterRole
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: neuron-device-plugin
rules:
- apiGroups:
  - ""
  resources:
  - nodes
  verbs:
  - get
  - list
  - watch
- apiGroups:
  - ""
  resources:
  - events
  verbs:
  - create
  - patch
- apiGroups:
  - ""
  resources:
  - pods
  verbs:
  - update
  - patch
  - get
  - list
  - watch
- apiGroups:
  - ""
  resources:
  - nodes/status
  verbs:
  - patch
  - update
---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: neuron-device-plugin
  namespace: kube-system
---
kind: ClusterRoleBinding
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: neuron-device-plugin
  namespace: kube-system
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: neuron-device-plugin
subjects:
- kind: ServiceAccount
  name: neuron-device-plugin
  namespace: kube-system
---
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: neuron-device-plugin-daemonset
  namespace: kube-system
spec:
  revisionHistoryLimit: 10
  selector:
    matchLabels:
      name: neuron-device-plugin-ds
  template:
    metadata:
      annotations:
        scheduler.alpha.kubernetes.io/critical-pod: ""
      creationTimestamp: null
      labels:
        name: neuron-device-plugin-ds
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: node.kubernetes.io/instance-type
                operator: In
                values:
                - inf1.xlarge
                - inf1.2xlarge
                - inf1.6xlarge
                - inf1.24xlarge
                - inf2.xlarge
                - inf2.8xlarge
                - inf2.24xlarge
                - inf2.48xlarge
                - trn1.2xlarge
                - trn1.32xlarge
      containers:
      - env:
        - name: KUBECONFIG
          value: /etc/kubernetes/kubelet.conf
        - name: NODE_NAME
          valueFrom:
            fieldRef:
              apiVersion: v1
              fieldPath: spec.nodeName
        image: public.ecr.aws/neuron/neuron-device-plugin:2.17.3.0
        imagePullPolicy: Always
        name: k8s-neuron-device-plugin-ctr
        resources: {}
        securityContext:
          allowPrivilegeEscalation: false
          capabilities:
            drop:
            - ALL
        terminationMessagePath: /dev/termination-log
        terminationMessagePolicy: File
        volumeMounts:
        - mountPath: /var/lib/kubelet/device-plugins
          name: device-plugin
        - mountPath: /run
          name: infa-map
      dnsPolicy: ClusterFirst
      priorityClassName: system-node-critical
      restartPolicy: Always
      schedulerName: default-scheduler
      securityContext: {}
      serviceAccount: neuron-device-plugin
      serviceAccountName: neuron-device-plugin
      terminationGracePeriodSeconds: 30
      tolerations:
      - key: CriticalAddonsOnly
        operator: Exists
      - effect: NoSchedule
        key: aws.amazon.com/neuron
        operator: Exists
      volumes:
      - hostPath:
          path: /var/lib/kubelet/device-plugins
          type: ""
        name: device-plugin
      - hostPath:
          path: /run
          type: ""
        name: infa-map
  updateStrategy:
    rollingUpdate:
      maxSurge: 0
      maxUnavailable: 1
    type: RollingUpdate

Once the device plugin DaemonSet is deployed, run kubectl describe node to confirm that you see device plugin daemons running on your Neuron-based instances, and that they advertise aws.amazon.com/neuron resources to Kubernetes.

The following output is an example of a correctly configured Neuron-based node. Note the Neuron device plugin daemon present on the node, the advertised aws.amazon.com/neuron resource, and the Domino node pool label identifying the node as Trainium.

Name:               ip-192-168-42-179.us-west-2.compute.internal
Roles:              <none>
Labels:             alpha.eksctl.io/cluster-name=inferentia-test
                    alpha.eksctl.io/instance-id=i-00549360c9911f4f1
                    alpha.eksctl.io/nodegroup-name=domino-trn1
                    beta.kubernetes.io/arch=amd64
                    beta.kubernetes.io/instance-type=trn1.2xlarge
                    beta.kubernetes.io/os=linux
                    dominodatalab.com/node-pool=trainium
                    failure-domain.beta.kubernetes.io/region=us-west-2
                    failure-domain.beta.kubernetes.io/zone=us-west-2b
                    k8s.io/cloud-provider-aws=7c4bfb478ecbb2400bead13fc878a3a1
                    kubernetes.io/arch=amd64
                    kubernetes.io/hostname=ip-192-168-42-179.us-west-2.compute.internal
                    kubernetes.io/os=linux
                    node-lifecycle=on-demand
                    node.kubernetes.io/instance-type=trn1.2xlarge
                    topology.ebs.csi.aws.com/zone=us-west-2b
                    topology.kubernetes.io/region=us-west-2
                    topology.kubernetes.io/zone=us-west-2b

... <snip> ...

Capacity:
  aws.amazon.com/neuron:        1
  aws.amazon.com/neuroncore:    2
  aws.amazon.com/neurondevice:  1
  cpu:                          8
  ephemeral-storage:            209702892Ki
  hugepages-1Gi:                0
  hugepages-2Mi:                0
  memory:                       32338380Ki
  pods:                         58
  smarter-devices/fuse:         20
Allocatable:
  aws.amazon.com/neuron:        1
  aws.amazon.com/neuroncore:    2
  aws.amazon.com/neurondevice:  1
  cpu:                          7910m
  ephemeral-storage:            192188443124
  hugepages-1Gi:                0
  hugepages-2Mi:                0
  memory:                       31321548Ki
  pods:                         58
  smarter-devices/fuse:         20
System Info:
  Machine ID:                 ec240d0453aef36a07d5248a753946c5
  System UUID:                ec240d04-53ae-f36a-07d5-248a753946c5
  Boot ID:                    9d066b81-273b-406d-ad0f-6375adebca5d
  Kernel Version:             5.4.253-167.359.amzn2.x86_64
  OS Image:                   Amazon Linux 2
  Operating System:           linux
  Architecture:               amd64
  Container Runtime Version:  containerd://1.6.19
  Kubelet Version:            v1.26.7-eks-8ccc7ba
  Kube-Proxy Version:         v1.26.7-eks-8ccc7ba
ProviderID:                   aws:///us-west-2b/i-00549360c9911f4f1
Non-terminated Pods:          (11 in total)
  Namespace                   Name                                    CPU Requests  CPU Limits   Memory Requests  Memory Limits  Age
  ---------                   ----                                    ------------  ----------   ---------------  -------------  ---
  domino-compute              run-650b32d5eb80df0d60cec514-kgcc4      5610m (70%)   6600m (83%)  28174Mi (92%)    28174Mi (92%)  3h17m
  domino-platform             aws-ebs-csi-driver-node-lwp9g           30m (0%)      200m (2%)    80Mi (0%)        456Mi (1%)     3h10m
  domino-platform             aws-efs-csi-driver-node-kk7mb           20m (0%)      200m (2%)    40Mi (0%)        200Mi (0%)     3h10m
  domino-platform             docker-registry-cert-mgr-k67bb          0 (0%)        0 (0%)       0 (0%)           0 (0%)         3h10m
  domino-platform             fluentd-ljkhk                           200m (2%)     1 (12%)      600Mi (1%)       2Gi (6%)       3h10m
  domino-platform             image-cache-agent-7ztv6                 0 (0%)        0 (0%)       0 (0%)           0 (0%)         3h10m
  domino-platform             prometheus-node-exporter-vr9bc          0 (0%)        0 (0%)       0 (0%)           0 (0%)         3h10m
  domino-platform             smarter-device-manager-ncsr6            10m (0%)      100m (1%)    15Mi (0%)        15Mi (0%)      3h10m
  kube-system                 aws-node-q89xl                          25m (0%)      0 (0%)       0 (0%)           0 (0%)         3h10m
  kube-system                 kube-proxy-b2qpv                        100m (1%)     0 (0%)       0 (0%)           0 (0%)         3h10m
  kube-system                 neuron-device-plugin-daemonset-h8d7k    0 (0%)        0 (0%)       0 (0%)           0 (0%)         3h10m

Hardware Tier setup

Next, you need to make the node group accessible to your users by creating a Domino hardware tier that does the following:

Targets the node pool label you’ve given to your Neuron-based nodes.
Requests a suitable amount of the node vCPU and memory, allowing for necessary overhead.
Requests a custom GPU resource with the name aws.amazon.com/neuron.

See the following example:

Key	Value
Cluster Type	Kubernetes
ID	trainium
Name	Trainium
Cores Requested	5.0
Memory Requested (GiB)	26.0
Number of GPUs	1
Use custom GPU name	Yes
GPU Resource Name	aws.amazon.com/neuron
Cents Per Minute Per Run	0.0
Node Pool	trainium
Restrict to compute cluster	Options: Spark, Ray, Dask, Mpi
Maximum Simultaneous Executions
Overprovisioning Pods	0

Environment setup

The AWS Neuron SDK is designed for use with fully integrated frameworks like PyTorch and TensorFlow. When setting up a Domino environment for a new version of Neuron or the integrated framework, you should read the documentation on:

As an example, and to facilitate testing, here’s an environment definition for adding PyTorch Neuron to the Domino 5.7 Standard Environment (quay.io/domino/compute-environment-images:ubuntu20-py3.9-r4.3-domino5.7-standard):

# Configure Linux for Neuron repository updates
RUN sudo touch /etc/apt/sources.list.d/neuron.list
RUN echo "deb https://apt.repos.neuron.amazonaws.com focal main" | sudo tee -a /etc/apt/sources.list.d/neuron.list
RUN sudo wget -qO - https://apt.repos.neuron.amazonaws.com/GPG-PUB-KEY-AMAZON-AWS-NEURON.PUB | sudo apt-key add -

# Update OS packages
RUN sudo apt-get update -y

# Install Neuron Driver
RUN sudo apt-get install aws-neuronx-dkms=2.* -y

# Install Neuron Runtime
RUN sudo apt-get install aws-neuronx-collectives=2.* -y
RUN sudo apt-get install aws-neuronx-runtime-lib=2.* -y

# Install Neuron Tools
RUN sudo apt-get install aws-neuronx-tools=2.* -y

# Add PATH
RUN export PATH=/opt/aws/neuron/bin:$PATH

# pip installs
RUN python -m pip config set global.extra-index-url https://pip.repos.neuron.amazonaws.com
RUN python -m pip install --user neuronx-cc==2.* torch-neuronx torchvision

Testing Neuron devices in Domino

To test your setup, start a Jupyter workspace using a Neuron-based hardware tier and Neuron-enabled Workspace Environment. Once your workspace has started, open a Python notebook and execute a cell with the command !/opt/aws/neuron/bin/neuron-ls to see mounted Neuron devices. You can now use the Neuron framework you’ve installed to invoke the mounted accelerator.

Next steps

Refer to the Getting Started with Neuron guide for your chosen framework to get started.

​Node group creation

​Example eksctl node group config

​Device plugin deployment

​Hardware Tier setup

​Environment setup

​Testing Neuron devices in Domino

​Next steps