feat: support Hami on k3s and k8s

Reviewed-on: #3

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
@@ -50,7 +50,7 @@ spec:
                 exit 0
               fi
               echo "⬇️ Starting download..."
-              apk add --no-cache bash aria2 wget ca-certificates
+              apk add --no-cache bash aria2 wget ca-certificates curl
               wget https://hf-mirror.com/hfd/hfd.sh -O hfd.sh && chmod +x hfd.sh
               ./hfd.sh {{ .Values.model.huggingfaceName }} --tool aria2c -x 8 --local-dir "$DEST_DIR"
 
@@ -62,7 +62,7 @@ spec:
         containers:
           - name: vllm-leader
             image: {{ .Values.vllm.image }}
-            imagePullPolicy:  IfNotPresent
+            imagePullPolicy:  {{ .Values.imagePullPolicy }}
             securityContext:
               capabilities:
                 add: [ "IPC_LOCK" ]
@@ -70,38 +70,74 @@ spec:
               # - name: HUGGING_FACE_HUB_TOKEN
               #   value: {{ .Values.vllm.huggingfaceToken }}
               - name: GLOO_SOCKET_IFNAME
-                value: eth0
+                value: {{ .Values.rdma.interface | default "eth0" | quote }}
               - name: NCCL_SOCKET_IFNAME
-                value: eth0
-              - name: NCCL_IB_DISABLE
+                value: {{ .Values.rdma.interface | default "eth0" | quote }}
+              - name: RAY_DEDUP_LOGS
                 value: "0"
               - name: NCCL_DEBUG
                 value: INFO
-              - name: NCCL_IB_HCA
-                value: mlx5_0:1
-              - name: NCCL_IB_GID_INDEX
-                value: "0" # 或 "7"，根据你的网络配置而定
-              - name: RAY_DEDUP_LOGS
+
+              # RDMA 条件配置
+              {{- if .Values.rdma.enabled }}
+              - name: NCCL_IB_DISABLE
                 value: "0"
+              - name: NCCL_IB_HCA
+                value: {{ .Values.rdma.hca | default "^mlx5" | quote }}
+              - name: NCCL_IB_GID_INDEX
+                value: {{ .Values.rdma.gidIndex | default "0" | quote }}  # 或 "7"，根据你的网络配置而定
+              {{- else }}
+              # 如果未开启 RDMA，显式禁用 IB，防止 NCCL 尝试探测报错
+              - name: NCCL_IB_DISABLE
+                value: "1"
+              {{- end }}
             command:
               - sh
               - -c
-              {{- if .Values.command }}
-              - "bash /vllm-workspace/examples/online_serving/multi-node-serving.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE); {{ .Values.command }}"
-              {{- else }}
-              - "bash /vllm-workspace/examples/online_serving/multi-node-serving.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE);
+              - |
+                # 1. 自动 RDMA 探测逻辑 (无论下方跑什么命令，先执行这段)
+                # =======================================================
+                # echo "🔍 [Init] Detecting RDMA devices..."
+                # if [ -d "/sys/class/infiniband" ] && [ "$(ls -A /sys/class/infiniband)" ]; then
+                #   echo "✅ [Init] RDMA devices found. Enabling NCCL IB."
+                #   export NCCL_IB_DISABLE=0
+                #   # 如果环境变量没指定 HCA，默认使用 ^mlx5 匹配
+                #   export NCCL_IB_HCA=${NCCL_IB_HCA:-"^mlx5"}
+                # else
+                #   echo "⚠️ [Init] No RDMA devices found. Falling back to TCP."
+                #   export NCCL_IB_DISABLE=1
+                #   export NCCL_NET_GDR_LEVEL=0
+                # fi
+                # echo "🚀 [Init] RDMA setup complete. NCCL_IB_DISABLE=$NCCL_IB_DISABLE"
+                  
+                {{- if .Values.command }}
+                bash /vllm-workspace/examples/online_serving/multi-node-serving.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE); {{ .Values.command }} --distributed-executor-backend ray
+                {{- else }}
+                bash /vllm-workspace/examples/online_serving/multi-node-serving.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE);
                 MODEL_NAME=$(basename '{{ .Values.model.huggingfaceName }}'); MODEL_PATH='{{ .Values.model.localMountPath }}/Weight/'$MODEL_NAME; 
-                python3 -m vllm.entrypoints.openai.api_server --port 8000 --model $MODEL_PATH --tensor-parallel-size {{ .Values.resources.gpuLimit }} --pipeline_parallel_size {{ .Values.workerSize }} --trust_remote_code"
-              {{- end }}
+                python3 -m vllm.entrypoints.openai.api_server --port 8000 --model $MODEL_PATH --tensor-parallel-size {{ .Values.resources.gpuLimit }} --pipeline_parallel_size {{ .Values.workerSize }} --distributed-executor-backend ray --trust_remote_code
+                {{- end }}
             resources:
               limits:
                 nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
                 memory: {{ .Values.resources.memoryLimit }}
                 ephemeral-storage: 10Gi
-                rdma/rdma_shared_device_a: 10
+                {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+                nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+                {{- end }}
+                {{- if .Values.rdma.enabled }}
+                {{ .Values.rdma.resourceName }}: {{ .Values.rdma.resourceCount | default 1 }}
+                {{- end }}
               requests:
                 ephemeral-storage: 10Gi
                 cpu: {{ .Values.resources.cpuRequest }}
+                nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
+                {{- if and .Values.resources.gpuMem (gt (int .Values.resources.gpuMem) 0) }}
+                nvidia.com/gpumem: {{ .Values.resources.gpuMem }}
+                {{- end }}
+                {{- if .Values.rdma.enabled }}
+                {{ .Values.rdma.resourceName }}: {{ .Values.rdma.resourceCount | default 1 }}
+                {{- end }}
             ports:
               - containerPort: 8000
                 name: http
@@ -150,33 +186,63 @@ spec:
             command:
               - sh
               - -c
-              - "bash /vllm-workspace/examples/online_serving/multi-node-serving.sh worker --ray_address=$(LWS_LEADER_ADDRESS)"
+              - |
+                # 1. 自动 RDMA 探测逻辑
+                # =======================================================
+                # echo "🔍 [Init] Detecting RDMA devices..."
+                # if [ -d "/sys/class/infiniband" ] && [ "$(ls -A /sys/class/infiniband)" ]; then
+                #   echo "✅ [Init] RDMA devices found. Enabling NCCL IB."
+                #   export NCCL_IB_DISABLE=0
+                #   export NCCL_IB_HCA=${NCCL_IB_HCA:-"^mlx5"}
+                # else
+                #   echo "⚠️ [Init] No RDMA devices found. Falling back to TCP."
+                #   export NCCL_IB_DISABLE=1
+                #   export NCCL_NET_GDR_LEVEL=0
+                # fi
+                bash /vllm-workspace/examples/online_serving/multi-node-serving.sh worker --ray_address=$(LWS_LEADER_ADDRESS)
             resources:
               limits:
                 nvidia.com/gpu: "{{ .Values.resources.gpuLimit }}"
                 memory: {{ .Values.resources.memoryLimit }}
                 ephemeral-storage: 10Gi
-                rdma/rdma_shared_device_a: 10
+                {{- 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 }}
               requests:
                 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 }}
             env:
-              # - name: HUGGING_FACE_HUB_TOKEN
-              #   value: {{ .Values.vllm.huggingfaceToken }}
               - name: GLOO_SOCKET_IFNAME
-                value: eth0
+                value: {{ .Values.rdma.interface | default "eth0" | quote }}
               - name: NCCL_SOCKET_IFNAME
-                value: eth0
-              - name: NCCL_IB_DISABLE
-                value: "0"
+                value: {{ .Values.rdma.interface | default "eth0" | quote }}
               - name: NCCL_DEBUG
                 value: INFO
-              - name: NCCL_IB_HCA
-                value: mlx5_0:1
-              - name: NCCL_IB_GID_INDEX
-                value: "0" # 或 "7"，根据你的网络配置而定
               - name: RAY_DEDUP_LOGS
                 value: "0"
+
+              {{- if .Values.rdma.enabled }}
+              - name: NCCL_IB_DISABLE
+                value: "0"
+              - name: NCCL_IB_HCA
+                value: {{ .Values.rdma.hca | default "^mlx5" | quote }}
+              - name: NCCL_IB_GID_INDEX
+                value: {{ .Values.rdma.gidIndex | default "0" | quote }}
+              {{- else }}
+              - name: NCCL_IB_DISABLE
+                value: "1"
+              {{- end }}
+
             volumeMounts:
               - mountPath: /dev/shm
                 name: dshm

vllm-serve/templates/single.yaml

@@ -51,7 +51,7 @@ spec:
               exit 0
             fi
             echo "⬇️ Starting download..."
-            apk add --no-cache bash aria2 wget ca-certificates
+            apk add --no-cache bash aria2 wget ca-certificates curl
             wget https://hf-mirror.com/hfd/hfd.sh -O hfd.sh && chmod +x hfd.sh
             ./hfd.sh {{ .Values.model.huggingfaceName }} --tool aria2c -x 8 --local-dir "$DEST_DIR"
 
@@ -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

@@ -440,7 +440,142 @@
             "tencent/DRIVE-RL",
             "tencent/DRIVE-SFT",
             "Qwen/Qwen2.5-14B",
-            "BAAI/bge-base-zh-v1.5"
+            "BAAI/bge-base-zh-v1.5",
+            "01-ai/Yi-1.5-34B",
+            "01-ai/Yi-1.5-6B",
+            "01-ai/Yi-1.5-9B",
+            "01-ai/Yi-34B-Chat",
+            "ByteDance/Dolphin-1.5",
+            "ByteDance/Ouro-1.4B",
+            "ByteDance/Ouro-1.4B-Thinking",
+            "ByteDance/Ouro-2.6B",
+            "ByteDance/Ouro-2.6B-Thinking",
+            "OpenGVLab/InternVL3-38B",
+            "OpenGVLab/InternVL3-78B",
+            "OpenGVLab/InternVL3_5-1B",
+            "OpenGVLab/InternVL3_5-30B-A3B-Instruct",
+            "OpenGVLab/InternVL3_5-38B",
+            "OpenGVLab/InternVL3_5-8B",
+            "OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview",
+            "OpenGVLab/Mini-InternVL2-2B-DA-Medical",
+            "OpenGVLab/SDLM-32B-D4",
+            "Qwen/CodeQwen1.5-7B-Chat",
+            "Qwen/Qwen1.5-1.8B",
+            "Qwen/Qwen1.5-MoE-A2.7B-Chat-GPTQ-Int4",
+            "Qwen/Qwen2-0.5B",
+            "Qwen/Qwen2-Math-72B-Instruct",
+            "Qwen/Qwen2.5-0.5B",
+            "Qwen/Qwen2.5-1.5B",
+            "Qwen/Qwen2.5-1.5B-Instruct",
+            "Qwen/Qwen2.5-1.5B-Instruct-GGUF",
+            "Qwen/Qwen2.5-14B-Instruct",
+            "Qwen/Qwen2.5-14B-Instruct-1M",
+            "Qwen/Qwen2.5-32B-Instruct",
+            "Qwen/Qwen2.5-32B-Instruct-AWQ",
+            "Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4",
+            "Qwen/Qwen2.5-3B-Instruct",
+            "Qwen/Qwen2.5-3B-Instruct-GGUF",
+            "Qwen/Qwen2.5-72B-Instruct",
+            "Qwen/Qwen2.5-7B",
+            "Qwen/Qwen2.5-7B-Instruct",
+            "Qwen/Qwen2.5-7B-Instruct-1M",
+            "Qwen/Qwen2.5-7B-Instruct-GGUF",
+            "Qwen/Qwen2.5-Coder-0.5B",
+            "Qwen/Qwen2.5-Coder-1.5B",
+            "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+            "Qwen/Qwen2.5-Coder-14B-Instruct",
+            "Qwen/Qwen2.5-Coder-32B-Instruct",
+            "Qwen/Qwen2.5-Coder-3B-Instruct-GGUF",
+            "Qwen/Qwen2.5-Coder-7B-Instruct",
+            "Qwen/Qwen2.5-Coder-7B-Instruct-GGUF",
+            "Qwen/Qwen2.5-Math-72B",
+            "Qwen/Qwen2.5-Math-7B",
+            "Qwen/Qwen2.5-VL-32B-Instruct",
+            "Qwen/Qwen2.5-VL-72B-Instruct-AWQ",
+            "Qwen/Qwen2.5-VL-7B-Instruct",
+            "Qwen/Qwen3-0.6B",
+            "Qwen/Qwen3-0.6B-Base",
+            "Qwen/Qwen3-1.7B",
+            "Qwen/Qwen3-1.7B-FP8",
+            "Qwen/Qwen3-14B-FP8",
+            "Qwen/Qwen3-14B-GGUF",
+            "Qwen/Qwen3-30B-A3B-Base",
+            "Qwen/Qwen3-30B-A3B-Thinking-2507-FP8",
+            "Qwen/Qwen3-Next-80B-A3B-Thinking-FP8",
+            "Qwen/Qwen3-VL-4B-Instruct-FP8",
+            "baichuan-inc/Baichuan-M2-32B-GPTQ-Int4",
+            "baidu/ERNIE-4.5-0.3B-Base-PT",
+            "baidu/ERNIE-4.5-21B-A3B-Base-PT",
+            "baidu/ERNIE-4.5-VL-28B-A3B-Base-PT",
+            "baidu/ERNIE-4.5-VL-28B-A3B-Thinking",
+            "baidu/ERNIE-4.5-VL-424B-A47B-Base-PT",
+            "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
+            "deepseek-ai/DeepSeek-V3.1-Terminus",
+            "deepseek-ai/DeepSeek-V3.2-Exp",
+            "deepseek-ai/deepseek-coder-1.3b-base",
+            "deepseek-ai/deepseek-coder-1.3b-instruct",
+            "deepseek-ai/deepseek-coder-6.7b-instruct",
+            "google/codegemma-1.1-2b-GGUF",
+            "google/gemma-3-27b-pt",
+            "google/paligemma2-28b-pt-896",
+            "google/reformer-crime-and-punishment",
+            "google/reformer-enwik8",
+            "google/t5-11b-ssm-nq",
+            "google/t5-3b-ssm-nq",
+            "google/t5-large-ssm-nq",
+            "google/t5-small-ssm-nq",
+            "google/t5-xl-ssm-nq",
+            "google/t5-xxl-ssm-nq",
+            "google/t5gemma-ml-ml-ul2-it",
+            "internlm/Spatial-SSRL-7B",
+            "llava-hf/llava-1.5-13b-hf",
+            "llava-hf/llava-v1.6-34b-hf",
+            "meta-llama/Llama-3.1-405B",
+            "meta-llama/Llama-3.1-405B-Instruct",
+            "meta-llama/Llama-3.1-70B",
+            "meta-llama/Llama-3.1-70B-Instruct",
+            "meta-llama/Llama-3.1-8B",
+            "meta-llama/Llama-3.1-8B-Instruct",
+            "meta-llama/Llama-3.2-11B-Vision-Instruct",
+            "meta-llama/Llama-3.2-1B",
+            "meta-llama/Llama-3.2-1B-Instruct",
+            "meta-llama/Llama-3.2-3B",
+            "meta-llama/Llama-3.2-3B-Instruct",
+            "meta-llama/Llama-3.2-90B-Vision",
+            "meta-llama/Llama-3.2-90B-Vision-Instruct",
+            "meta-llama/Llama-3.3-70B-Instruct",
+            "microsoft/MediPhi-Instruct",
+            "microsoft/Phi-3-medium-4k-instruct-onnx-cpu",
+            "microsoft/Phi-3.5-mini-instruct",
+            "microsoft/bitnet-b1.58-2B-4T",
+            "microsoft/bitnet-b1.58-2B-4T-gguf",
+            "microsoft/kosmos-2.5",
+            "microsoft/kosmos-2.5-chat",
+            "microsoft/llava-med-v1.5-mistral-7b",
+            "mistralai/Mistral-7B-Instruct-v0.2",
+            "moonshotai/Kimi-Dev-72B",
+            "moonshotai/Kimi-K2-Base",
+            "moonshotai/Kimi-K2-Instruct",
+            "moonshotai/Kimi-K2-Instruct-0905",
+            "moonshotai/Kimi-K2-Thinking",
+            "moonshotai/Kimi-Linear-48B-A3B-Base",
+            "moonshotai/Kimi-Linear-48B-A3B-Instruct",
+            "moonshotai/Moonlight-16B-A3B",
+            "openbmb/MiniCPM4.1-8B",
+            "tencent/DeepSeek-V3.1-Terminus-W4AFP8",
+            "tencent/Hunyuan-0.5B-Pretrain",
+            "zai-org/GLM-4-9B-0414",
+            "zai-org/GLM-4.1V-9B-Base",
+            "zai-org/GLM-4.5-Air",
+            "zai-org/GLM-4.5V-FP8",
+            "zai-org/GLM-4.6",
+            "zai-org/GLM-4.6-FP8",
+            "zai-org/GLM-Z1-32B-0414",
+            "zai-org/GLM-Z1-9B-0414",
+            "zai-org/Glyph",
+            "zai-org/UI2Code_N",
+            "zai-org/WebVIA-Agent",
+            "zai-org/codegeex4-all-9b"
           ]
         }
       },
@@ -458,6 +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

@@ -27,17 +27,25 @@ model:
   huggingfaceName: "Qwen/Qwen2.5-0.5B-Instruct"  # 用户只需输入这个
   localMountPath: "/Model"                  # PVC 固定挂载路径
   huggingfaceToken: "<your-hf-token>"
-  download:                               # 启用自动下载
-    image: "docker.io/vllm/vllm-openai:latest"  # 包含 huggingface-cli 的镜像
 
 # 功能选择
 
 resources:
   gpuLimit: 1
+  gpuMem: 0
   cpuRequest: 12
   memoryLimit: "16Gi"
   shmSize: "20Gi"
 
+#  RDMA 配置部分
+rdma:
+  enabled: false  # 开关：默认关闭，防止在无 RDMA 节点报错
+  interface: eth0 # NCCL/GLOO 通信使用的网卡名称 (有 RDMA 时可能是 ib0 或 bond0)
+  resourceName: "rdma/rdma_shared_device_a" # RDMA 资源名称 (取决于你的 k8s 插件)
+  resourceCount: 5   # 每个 Pod 需要的 RDMA 设备数量
+  hca: "mlx5_0:1"    # 指定的 HCA 设备 (或者使用 ^mlx5 进行前缀匹配)
+  gidIndex: "0"      # RoCEv2 通常需要指定 GID    
+    
 svc:
   type: LoadBalancer
   port: 80