shm mount

huggingface-model/Chart.yaml

@@ -4,5 +4,5 @@ description: Helm chart for deploy Hugging Face to kubernetes cluster. See [Hugg
 
 type: application
 
-version: 0.0.20
+version: 0.0.21
 

huggingface-model/README.md

@@ -27,6 +27,8 @@ See [Hugging Face models](https://huggingface.co/models)
 | `imagePullSecrets`                | May need if used private repo as a cache for image ghcr.io/huggingface/text-generation-inference                                              | `[]`                                            |
 | `nameOverride`                    | String to partially override common.names.name                                                                                                | `""`                                            |
 | `fullnameOverride`                | String to fully override common.names.fullname                                                                                                | `""`                                            |
+| `shmVolume.enabled`               | Enable emptyDir volume for /dev/shm for model pod(s)                                                                                          | `true`                                          |
+| `shmVolume.sizeLimit`             | Set this to enable a size limit on the shm tmpfs                                                                                              | `1Gi`                                           |
 | `persistence.accessModes`         | PVC accessModes                                                                                                                               | `["ReadWriteOnce"]`                             |
 | `persistence.storageClassName`    | Kubernetes storageClass name                                                                                                                  | `gp2`                                           |
 | `persistence.storage`             | Volume size                                                                                                                                   | `100Gi`                                         |

huggingface-model/templates/application.yaml

@@ -111,9 +111,13 @@ spec:
           volumeMounts:
             - name: model-storage
               mountPath: "/usr/src/{{ .Values.model.name }}"
-          {{- with .Values.extraVolumeMounts}}
+            {{- if .Values.shmVolume.enabled }}
+            - name: shm
+              mountPath: /dev/shm
+            {{- end }}
+            {{- with .Values.extraVolumeMounts}}
             {{- toYaml . | nindent 12 }}
-          {{- end }}
+            {{- end }}
           resources:
             {{- toYaml .Values.resources | nindent 12 }}
       {{- with .Values.nodeSelector }}
@@ -132,6 +136,12 @@ spec:
         - name: model-storage
           persistentVolumeClaim:
             claimName: {{ include "huggingface-model.fullname" . }}
+        {{- if .Values.shmVolume.enabled }}
+        - name: shm
+          emptyDir:
+            medium: Memory
+            sizeLimit: {{ default "1Gi" .Values.shmVolume.sizeLimit }}
+        {{- end }}
       {{- with .Values.extraVolumes }}
         {{- toYaml . | nindent 8 }}
       {{- end }}

huggingface-model/values.yaml

@@ -68,6 +68,19 @@ nameOverride: ""
 ##
 fullnameOverride: ""
 
+## Start model pod(s) without limitations on shm memory.
+## By default docker and containerd (and possibly other container runtimes) limit `/dev/shm` to `64M`
+## ref: https://github.com/containerd/containerd/issues/3654
+##
+shmVolume:
+  ## @param shmVolume.enabled Enable emptyDir volume for /dev/shm for model pod(s)
+  ##
+  enabled: true
+  ## @param shmVolume.sizeLimit Set this to enable a size limit on the shm tmpfs
+  ## Note: the size of the tmpfs counts against container's memory limit
+  ##
+  sizeLimit: "1Gi"
+
 ## Persistence parameters
 ## ref: https://kubernetes.io/docs/concepts/storage/persistent-volumes/
 ## @param persistence.accessModes PVC accessModes