trending/miles
1
0
Fork 0
mirror of https://github.com/radixark/miles.git synced 2026-10-02 07:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
3d3fb97bacfac5cc3776ee80382ca30cbca24f7b
miles/tests/snapshots/charts/miles-run/typical.yaml
T

1674 lines
56 KiB
YAML
Raw Blame History

---
# Source: miles-run/templates/colocate-pairing.yaml
apiVersion: v1
kind: ServiceAccount
metadata:
name: "myrun-miles-run-colocate-pairing"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "colocate-pairing"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
---
# Source: miles-run/templates/platform-rbac.yaml
apiVersion: v1
kind: ServiceAccount
metadata:
name: "myrun-miles-run-platform-read"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "platform-read"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
---
# Source: miles-run/templates/platform-rbac.yaml
apiVersion: v1
kind: ServiceAccount
metadata:
name: "myrun-miles-run-platform-read-delete"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "platform-read-delete"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
---
# Source: miles-run/templates/uninstall-manifest.yaml
apiVersion: v1
kind: ConfigMap
metadata:
name: "myrun-miles-run-uninstall-manifest"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "uninstall"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
data:
uninstall-job.yaml: |
apiVersion: batch/v1
kind: Job
metadata:
name: "myrun-miles-run-uninstall"
namespace: "myns"
labels:
miles.radixark.io/uninstall-of: "myrun"
spec:
backoffLimit: 2
ttlSecondsAfterFinished: 3600
activeDeadlineSeconds: 1020
template:
metadata:
labels:
miles.radixark.io/uninstall-of: "myrun"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
serviceAccountName: "miles-uninstaller"
restartPolicy: Never
containers:
- name: uninstall
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
command:
- sh
- -c
- "sleep 120 && helm uninstall myrun --namespace myns --ignore-not-found"
---
# Source: miles-run/templates/colocate-pairing.yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
name: "myrun-miles-run-colocate-pairing"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "colocate-pairing"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
rules:
- apiGroups: [""]
resources: ["pods"]
verbs: ["get", "list", "watch", "patch", "update"]
---
# Source: miles-run/templates/platform-rbac.yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
name: "myrun-miles-run-platform-read"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "platform-read"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
rules:
- apiGroups: [""]
resources: ["pods"]
verbs: ["get", "list", "watch"]
---
# Source: miles-run/templates/platform-rbac.yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
name: "myrun-miles-run-platform-read-delete"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "platform-read-delete"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
rules:
- apiGroups: [""]
resources: ["pods"]
verbs: ["get", "list", "watch", "delete"]
- apiGroups: ["batch"]
resources: ["jobs"]
verbs: ["create"]
---
# Source: miles-run/templates/colocate-pairing.yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
name: "myrun-miles-run-colocate-pairing"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "colocate-pairing"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
roleRef:
apiGroup: rbac.authorization.k8s.io
kind: Role
name: "myrun-miles-run-colocate-pairing"
subjects:
- kind: ServiceAccount
name: "myrun-miles-run-colocate-pairing"
namespace: "myns"
---
# Source: miles-run/templates/platform-rbac.yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
name: "myrun-miles-run-platform-read"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "platform-read"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
roleRef:
apiGroup: rbac.authorization.k8s.io
kind: Role
name: "myrun-miles-run-platform-read"
subjects:
- kind: ServiceAccount
name: "myrun-miles-run-platform-read"
namespace: "myns"
---
# Source: miles-run/templates/platform-rbac.yaml
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
name: "myrun-miles-run-platform-read-delete"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "platform-read-delete"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
roleRef:
apiGroup: rbac.authorization.k8s.io
kind: Role
name: "myrun-miles-run-platform-read-delete"
subjects:
- kind: ServiceAccount
name: "myrun-miles-run-platform-read-delete"
namespace: "myns"
---
# Source: miles-run/templates/orchestrator.yaml
apiVersion: v1
kind: Service
metadata:
name: "myrun-miles-run-orchestrator"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "orchestrator"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
clusterIP: None
selector:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "orchestrator"
ports:
- name: placeholder
port: 1
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: v1
kind: Service
metadata:
name: "myrun-miles-run-rollout-executor"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "rollout-executor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
clusterIP: None
selector:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "rollout-executor"
ports:
- name: "rpc"
port: 8000
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: v1
kind: Service
metadata:
name: "myrun-miles-run-inference-controller"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-controller"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
clusterIP: None
selector:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-controller"
ports:
- name: "rpc"
port: 8000
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: v1
kind: Service
metadata:
name: "myrun-miles-run-inference-router-0"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-router-0"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
clusterIP: None
selector:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-router-0"
ports:
- name: "primary"
port: 8000
- name: "prometheus"
port: 9000
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: v1
kind: Service
metadata:
name: "myrun-miles-run-session-server"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "session-server"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
clusterIP: None
selector:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "session-server"
ports:
- name: "primary"
port: 8000
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: v1
kind: Service
metadata:
name: "myrun-miles-run-trainer-controller-actor"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "trainer-controller-actor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
clusterIP: None
selector:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "trainer-controller-actor"
ports:
- name: "rpc"
port: 8000
---
# Source: miles-run/templates/colocate-pairing.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
name: "myrun-miles-run-colocate-pairing"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "colocate-pairing"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
strategy:
type: Recreate
selector:
matchLabels:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "colocate-pairing"
template:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "colocate-pairing"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
serviceAccountName: "myrun-miles-run-colocate-pairing"
automountServiceAccountToken: true
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
containers:
- name: pairing
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
command:
- python
- -m
- miles.utils.external_utils.colocate_pairing
- --config
- "{\"inference_pools\":[{\"layout\":{\"gpu_offset\":0,\"num_gpus_per_inference_pod\":8,\"num_gpus_per_node\":8,\"num_inference_cells\":1,\"num_pods_per_inference_cell\":2,\"num_pods_per_trainer_cell\":4,\"num_trainer_cells\":1},\"pool_id\":\"inference-engine-all-0-0\"},{\"layout\":{\"gpu_offset\":16,\"num_gpus_per_inference_pod\":8,\"num_gpus_per_node\":8,\"num_inference_cells\":2,\"num_pods_per_inference_cell\":1,\"num_pods_per_trainer_cell\":4,\"num_trainer_cells\":1},\"pool_id\":\"inference-engine-all-0-1\"}],\"namespace\":\"rl\",\"release\":\"myrun\",\"trainer_pool_id\":\"trainer-engine-actor\"}"
env:
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
resources:
requests:
cpu: "100m"
memory: 256Mi
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
---
# Source: miles-run/templates/orchestrator.yaml
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: "myrun-miles-run-orchestrator"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "orchestrator"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
serviceName: "myrun-miles-run-orchestrator"
selector:
matchLabels:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "orchestrator"
template:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "orchestrator"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
serviceAccountName: "myrun-miles-run-platform-read-delete"
restartPolicy: Always
containers:
- name: orchestrator
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
- name: uninstall-manifest
mountPath: "/etc/miles-uninstall"
readOnly: true
command:
- python
- -m
- miles.utils.external_utils.command_utils.helm_backend.orchestrator.wrapper
- --state-file
- "/cluster-storage/myteam/miles_data/miles-runs/260101-000000-000/state/orchestrator-260101-000000-000001.state"
- --uninstall-manifest
- "/etc/miles-uninstall/uninstall-job.yaml"
- --
- "python"
- "scripts/run_qwen3_4b.py"
- "train"
- "--cluster-backend"
- "kubernetes"
env:
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
resources:
limits: {}
requests:
cpu: "2"
memory: 8Gi
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
- name: uninstall-manifest
configMap:
name: "myrun-miles-run-uninstall-manifest"
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: "myrun-miles-run-rollout-executor"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "rollout-executor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
serviceName: "myrun-miles-run-rollout-executor"
selector:
matchLabels:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "rollout-executor"
template:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "rollout-executor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "rollout-executor"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
restartPolicy: Always
containers:
- name: worker
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
command:
- "<PYTHON>"
- "-m"
- "miles.utils.workers.serving.serve"
- "--specs"
- "miles.ray.specs.entrypoint.compute_specs_from_argv"
- "--pool-id"
- "rollout-executor"
- "--"
- "--cluster-backend"
- "kubernetes"
- "--rollout-num-gpus"
- "32"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['apps.kubernetes.io/pod-index']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
ports:
- name: "rpc"
containerPort: 8000
resources:
{}
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: "myrun-miles-run-inference-controller"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-controller"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
serviceName: "myrun-miles-run-inference-controller"
selector:
matchLabels:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-controller"
template:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-controller"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "inference-controller"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
serviceAccountName: "myrun-miles-run-platform-read"
restartPolicy: Always
containers:
- name: worker
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
command:
- "<PYTHON>"
- "-m"
- "miles.utils.workers.serving.serve"
- "--specs"
- "miles.ray.specs.entrypoint.compute_specs_from_argv"
- "--pool-id"
- "inference-controller"
- "--"
- "--cluster-backend"
- "kubernetes"
- "--rollout-num-gpus"
- "32"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['apps.kubernetes.io/pod-index']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
ports:
- name: "rpc"
containerPort: 8000
resources:
{}
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: "myrun-miles-run-inference-router-0"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-router-0"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
serviceName: "myrun-miles-run-inference-router-0"
selector:
matchLabels:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-router-0"
template:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-router-0"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "inference-router-0"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
restartPolicy: Always
containers:
- name: worker
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
command:
- "<PYTHON>"
- "-m"
- "sglang_router.launch_router"
- "--policy"
- "manual"
- "--assignment-mode"
- "min_load"
- "--pd-disaggregation"
- "--log-level"
- "warn"
- "--prometheus-port"
- "9000"
- "--request-timeout-secs"
- "14400"
- "--port"
- "8000"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['apps.kubernetes.io/pod-index']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
ports:
- name: "primary"
containerPort: 8000
- name: "prometheus"
containerPort: 9000
resources:
{}
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: "myrun-miles-run-session-server"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "session-server"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 32
serviceName: "myrun-miles-run-session-server"
selector:
matchLabels:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "session-server"
template:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "session-server"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "session-server"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
restartPolicy: Always
containers:
- name: worker
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
command:
- "<PYTHON>"
- "-m"
- "miles.rollout.session.server"
- "--config-json"
- "{\"host\":\"0.0.0.0\",\"port\":8000,\"instance_id\":\"0123456789abcdef-0\",\"backend_url\":\"http://myrun-miles-run-inference-router-0-0.myrun-miles-run-inference-router-0:8000\",\"timeout\":null,\"hf_checkpoint\":\"<FIXTURE_DIR>/typical-model\",\"chat_template_path\":null,\"tito_model\":\"default\",\"apply_chat_template_kwargs\":{},\"use_rollout_routing_replay\":false,\"use_rollout_indexer_replay\":false,\"use_sampling_support_replay\":false,\"sglang_speculative_algorithm\":null,\"num_layers\":36,\"moe_router_topk\":2,\"save_debug_trajectory_data\":null,\"lora_rank\":0,\"lora_adapter_path\":null,\"lora_train_only\":false,\"use_session_server\":true,\"session_message_matcher\":\"strict\",\"pause_generation_mode\":\"retract\",\"session_sample_picker_path\":\"miles.rollout.session.v2.picker_hub.drop_same_prompt_retries\",\"session_sample_postprocessor_path\":\"miles.rollout.session.v2.postprocessor_hub.default_postprocess\"}"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['apps.kubernetes.io/pod-index']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
ports:
- name: "primary"
containerPort: 8000
resources:
{}
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
---
# Source: miles-run/templates/static-workers.yaml
apiVersion: apps/v1
kind: StatefulSet
metadata:
name: "myrun-miles-run-trainer-controller-actor"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "trainer-controller-actor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
serviceName: "myrun-miles-run-trainer-controller-actor"
selector:
matchLabels:
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "trainer-controller-actor"
template:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "trainer-controller-actor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "trainer-controller-actor"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
serviceAccountName: "myrun-miles-run-platform-read-delete"
restartPolicy: Always
containers:
- name: worker
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
command:
- "<PYTHON>"
- "-m"
- "miles.utils.workers.serving.serve"
- "--specs"
- "miles.ray.specs.entrypoint.compute_specs_from_argv"
- "--pool-id"
- "trainer-controller-actor"
- "--"
- "--cluster-backend"
- "kubernetes"
- "--rollout-num-gpus"
- "32"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['apps.kubernetes.io/pod-index']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
ports:
- name: "rpc"
containerPort: 8000
resources:
{}
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
---
# Source: miles-run/templates/inference-engines.yaml
apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
name: "myrun-miles-run-inference-engine-all-0-0"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-engine-all-0-0"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
startupPolicy: LeaderCreated
leaderWorkerTemplate:
size: 2
restartPolicy: RecreateGroupOnPodRestart
workerTemplate:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-engine-all-0-0"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "inference-engine-all-0-0"
annotations:
miles.radixark.io/meta-gpu_ids: "0,1,2,3,4,5,6,7"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
hostIPC: true
schedulingGates:
- name: "miles.radixark.io/colocate-pairing"
containers:
- name: "engine"
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
- name: dev-shm
mountPath: "/dev/shm"
command:
- "<PYTHON>"
- "-m"
- "sglang.launch_server"
- "--trust-remote-code"
- "--model-path"
- "<FIXTURE_DIR>/typical-model"
- "--host"
- "0.0.0.0"
- "--port"
- "8000"
- "--device"
- "cuda"
- "--random-seed"
- "<RANDOM_SEED>"
- "--enable-memory-saver"
- "--nccl-port"
- "10000"
- "--nnodes"
- "2"
- "--node-rank"
- "$(LWS_WORKER_INDEX)"
- "--dist-init-addr"
- "$(LWS_LEADER_ADDRESS):9000"
- "--gated-launch-port"
- "13000"
- "--tp-size"
- "16"
- "--skip-server-warmup"
- "--enable-draft-weights-cpu-backup"
- "--enable-metrics"
- "--disaggregation-mode"
- "prefill"
- "--load-balance-method"
- "round_robin"
- "--disaggregation-bootstrap-port"
- "11000"
- "--engine-info-bootstrap-port"
- "12000"
- "--mem-fraction-static"
- "0.7"
- "--chunked-prefill-size"
- "4096"
- "--cuda-graph-backend-prefill"
- "disabled"
- "--lora-use-virtual-experts"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['leaderworkerset.sigs.k8s.io/group-index']
- name: MILES_POD_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['leaderworkerset.sigs.k8s.io/worker-index']
- name: MILES_BASE_GPU_ID
valueFrom:
fieldRef:
fieldPath: metadata.annotations['miles.radixark.io/base-gpu-id']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "NVIDIA_VISIBLE_DEVICES"
value: "all"
- name: "NVSHMEM_DISABLE_NCCL"
value: "1"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_HABANA_VISIBLE_MODULES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_NEURON_RT_VISIBLE_CORES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_TPU_VISIBLE_CHIPS"
value: "1"
- name: "SGLANG_BATCH_INVARIANT_OPS_ENABLE_MM_FALLBACK_VARIANT"
value: "true"
- name: "SGLANG_DG_CACHE_DIR_PER_PROCESS"
value: "1"
- name: "SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION"
value: "false"
- name: "SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE"
value: "false"
- name: "SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK"
value: "false"
- name: "SGLANG_EXPOSE_OWN_ENV_VARS"
value: "1"
- name: "SGLANG_JIT_DEEPGEMM_PRECOMPILE"
value: "false"
- name: "SGLANG_MEMORY_SAVER_CUDA_GRAPH"
value: "true"
- name: "SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2"
value: "1"
ports:
- name: "primary"
containerPort: 8000
- name: "dist-init"
containerPort: 9000
- name: "nccl"
containerPort: 10000
- name: "disaggregation"
containerPort: 11000
- name: "engine-info-boo"
containerPort: 12000
- name: "gate"
containerPort: 13000
resources:
limits:
nvidia.com/gpu: 0
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
- name: dev-shm
hostPath:
path: "/dev/shm"
type: "Directory"
---
# Source: miles-run/templates/inference-engines.yaml
apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
name: "myrun-miles-run-inference-engine-all-0-1"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-engine-all-0-1"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 2
startupPolicy: LeaderCreated
leaderWorkerTemplate:
size: 1
restartPolicy: RecreateGroupOnPodRestart
workerTemplate:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "inference-engine-all-0-1"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "inference-engine-all-0-1"
annotations:
miles.radixark.io/meta-gpu_ids: "0,1,2,3,4,5,6,7"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
hostIPC: true
schedulingGates:
- name: "miles.radixark.io/colocate-pairing"
containers:
- name: "engine"
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
- name: dev-shm
mountPath: "/dev/shm"
command:
- "<PYTHON>"
- "-m"
- "sglang.launch_server"
- "--trust-remote-code"
- "--model-path"
- "<FIXTURE_DIR>/typical-model"
- "--host"
- "0.0.0.0"
- "--port"
- "8000"
- "--device"
- "cuda"
- "--random-seed"
- "<RANDOM_SEED>"
- "--enable-memory-saver"
- "--nccl-port"
- "10000"
- "--node-rank"
- "$(LWS_WORKER_INDEX)"
- "--dist-init-addr"
- "$(LWS_LEADER_ADDRESS):9000"
- "--gated-launch-port"
- "13000"
- "--tp-size"
- "8"
- "--skip-server-warmup"
- "--enable-draft-weights-cpu-backup"
- "--enable-metrics"
- "--disaggregation-mode"
- "decode"
- "--engine-info-bootstrap-port"
- "12000"
- "--mem-fraction-static"
- "0.7"
- "--chunked-prefill-size"
- "4096"
- "--cuda-graph-backend-prefill"
- "disabled"
- "--lora-use-virtual-experts"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['leaderworkerset.sigs.k8s.io/group-index']
- name: MILES_POD_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['leaderworkerset.sigs.k8s.io/worker-index']
- name: MILES_BASE_GPU_ID
valueFrom:
fieldRef:
fieldPath: metadata.annotations['miles.radixark.io/base-gpu-id']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "NVIDIA_VISIBLE_DEVICES"
value: "all"
- name: "NVSHMEM_DISABLE_NCCL"
value: "1"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_HABANA_VISIBLE_MODULES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_NEURON_RT_VISIBLE_CORES"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR"
value: "1"
- name: "RAY_EXPERIMENTAL_NOSET_TPU_VISIBLE_CHIPS"
value: "1"
- name: "SGLANG_BATCH_INVARIANT_OPS_ENABLE_MM_FALLBACK_VARIANT"
value: "true"
- name: "SGLANG_DG_CACHE_DIR_PER_PROCESS"
value: "1"
- name: "SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION"
value: "false"
- name: "SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE"
value: "false"
- name: "SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK"
value: "false"
- name: "SGLANG_EXPOSE_OWN_ENV_VARS"
value: "1"
- name: "SGLANG_JIT_DEEPGEMM_PRECOMPILE"
value: "false"
- name: "SGLANG_MEMORY_SAVER_CUDA_GRAPH"
value: "true"
- name: "SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2"
value: "1"
ports:
- name: "primary"
containerPort: 8000
- name: "dist-init"
containerPort: 9000
- name: "nccl"
containerPort: 10000
- name: "engine-info-boo"
containerPort: 12000
- name: "gate"
containerPort: 13000
resources:
limits:
nvidia.com/gpu: 0
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
- name: dev-shm
hostPath:
path: "/dev/shm"
type: "Directory"
---
# Source: miles-run/templates/trainer-engines.yaml
apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
name: "myrun-miles-run-trainer-engine-actor"
namespace: "myns"
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "trainer-engine-actor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
spec:
replicas: 1
startupPolicy: LeaderCreated
leaderWorkerTemplate:
size: 4
restartPolicy: RecreateGroupOnPodRestart
workerTemplate:
metadata:
labels:
helm.sh/chart: "miles-run-0.1.0"
app.kubernetes.io/name: "miles-run"
app.kubernetes.io/instance: "myrun"
app.kubernetes.io/component: "trainer-engine-actor"
app.kubernetes.io/version: "0.1.0"
app.kubernetes.io/managed-by: "Helm"
miles.radixark.io/pool: "trainer-engine-actor"
annotations:
miles.radixark.io/meta-gpu_ids: "0,1,2,3,4,5,6,7"
spec:
enableServiceLinks: false
imagePullSecrets:
- name: "myregistry-pull-secret"
nodeSelector:
nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
hostIPC: true
containers:
- name: "trainer"
image: "myregistry.example/myteam/miles:v0.9.3-cu128"
imagePullPolicy: "IfNotPresent"
workingDir: "/root/miles"
volumeMounts:
- name: "cluster-storage"
mountPath: "/cluster-storage"
- name: "cluster-storage"
mountPath: "/root/miles"
subPath: "myuser/miles"
- name: "cluster-storage"
mountPath: "/root/Megatron-LM"
subPath: "myuser/Megatron-LM"
- name: "models"
mountPath: "/models"
readOnly: true
- name: "scratch"
mountPath: "/scratch"
- name: dev-shm
mountPath: "/dev/shm"
command:
- "bash"
- "-c"
- "mkdir -p /scratch/Qwen3-4B && rsync -a --info=progress2 /cluster-storage/models/Qwen3-4B/ /scratch/Qwen3-4B && exec <PYTHON> -m miles.utils.workers.process_supervisor --num-subprocesses 8 -- <PYTHON> -m miles.utils.workers.serving.serve --specs miles.ray.specs.entrypoint.compute_specs_from_argv --pool-id trainer-engine-actor -- --cluster-backend kubernetes --rollout-num-gpus 32"
env:
- name: MILES_CELL_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['leaderworkerset.sigs.k8s.io/group-index']
- name: MILES_POD_INDEX
valueFrom:
fieldRef:
fieldPath: metadata.labels['leaderworkerset.sigs.k8s.io/worker-index']
- name: "CUDA_DEVICE_MAX_CONNECTIONS"
value: "1"
- name: "HF_ENDPOINT"
value: "https://hf-mirror.example"
- name: "MILES_K8S_NAMESPACE"
value: "myns"
- name: "MILES_K8S_RELEASE"
value: "myrun"
- name: "NCCL_IB_HCA"
value: "mlx5_0,mlx5_1"
- name: "NCCL_SOCKET_IFNAME"
value: "bond0"
- name: "PYTHONPATH"
value: "/root/miles:/root/Megatron-LM:/sgl-workspace/sglang/python"
- name: "PYTHONUNBUFFERED"
value: "1"
ports:
- name: "master"
containerPort: 9000
- name: "rpc"
containerPort: 8000
resources:
limits:
nvidia.com/gpu: 8
volumes:
- name: "cluster-storage"
persistentVolumeClaim:
claimName: miles-shared-nvme
- name: "models"
hostPath:
path: "/gpfs/models"
type: "Directory"
- name: "scratch"
hostPath:
path: "/mnt/local-nvme/miles-staging"
type: "DirectoryOrCreate"
- name: dev-shm
hostPath:
path: "/dev/shm"
type: "Directory"
Reference in New Issue View Git Blame Copy Permalink