feat: support Hami on k3s and k8s

vllm-serve/Chart.yaml

@@ -5,4 +5,4 @@ appVersion: 1.16.0
 description: A Helm chart for deploying vLLM with NFS storage
 name: vllm-serve
 type: application
-version: 0.2.1
+version: 0.3.0

vllm-serve/templates/lws.yaml

@@ -16,7 +16,7 @@ spec:
         initContainers:
         # 模型下载作为第一个 initContainer
         - name: download-model
-          image: {{ .Values.model.download.image }}
+          image: alpine:latest
           imagePullPolicy: {{ .Values.imagePullPolicy | default "IfNotPresent" }}
           env:
             - name: HF_ENDPOINT
@@ -122,6 +122,9 @@ spec:
                 nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
                 memory: {{ .Values.resources.memoryLimit }}
                 ephemeral-storage: 10Gi
+                {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+                nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+                {{- end }}
                 {{- if .Values.rdma.enabled }}
                 {{ .Values.rdma.resourceName }}: {{ .Values.rdma.resourceCount | default 1 }}
                 {{- end }}
@@ -129,6 +132,9 @@ spec:
                 ephemeral-storage: 10Gi
                 cpu: {{ .Values.resources.cpuRequest }}
                 nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
+                {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+                nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+                {{- end }}
                 {{- if .Values.rdma.enabled }}
                 {{ .Values.rdma.resourceName }}: {{ .Values.rdma.resourceCount | default 1 }}
                 {{- end }}
@@ -199,6 +205,9 @@ spec:
                 nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
                 memory: {{ .Values.resources.memoryLimit }}
                 ephemeral-storage: 10Gi
+                {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+                nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+                {{- end }}
                 {{- if .Values.rdma.enabled }}
                 {{ .Values.rdma.resourceName }}: {{ .Values.rdma.resourceCount | default 1 }}
                 {{- end }}
@@ -206,6 +215,9 @@ spec:
                 ephemeral-storage: 10Gi
                 cpu: {{ .Values.resources.cpuRequest }}
                 nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
+                {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+                nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+                {{- end }}
                 {{- if .Values.rdma.enabled }}
                 {{ .Values.rdma.resourceName }}: {{ .Values.rdma.resourceCount | default 1 }}
                 {{- end }}

vllm-serve/templates/single.yaml

@@ -92,12 +92,18 @@ spec:
         resources:
           limits:
             nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
+            {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+            nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+            {{- end }}
             memory: {{ .Values.resources.memoryLimit }}
             ephemeral-storage: 10Gi
             cpu: {{ .Values.resources.cpuRequest }}
           requests:
             ephemeral-storage: 10Gi
             cpu: {{ .Values.resources.cpuRequest }}
+            {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+            nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+            {{- end }}
         ports:
         - containerPort: 8000
           name: http

vllm-serve/values.schema.json

@@ -593,6 +593,12 @@
           "default": 1,
           "minimum": 1
         },
+	"gpuMem": {
+          "type": "integer",
+          "description": "GPU 显存限制，单位MB, 0表示独占卡",
+          "default": 0,
+          "minimum": 0
+        },
         "cpuRequest": {
           "type": "integer",
           "description": "CPU 请求",

vllm-serve/values.yaml

@@ -32,6 +32,7 @@ model:
 
 resources:
   gpuLimit: 1
+  gpuMem: 0
   cpuRequest: 12
   memoryLimit: "16Gi"
   shmSize: "20Gi"