{
  "manifest_version": "1.0.0",
  "template": {
    "id": "069ed2a5-4f4a-4581-8b2d-e7bedc7fd96a",
    "slug": "vllm",
    "name": "vLLM | High-Throughput LLM Serving",
    "description": "Self-host vLLM on Railway with an OpenAI-compatible API",
    "url": "https://railway.com/deploy/vllm",
    "upstream": {
      "image": "vllm/vllm-openai-cpu:v0.20.2"
    }
  },
  "services": [
    {
      "name": "vllm",
      "source": {
        "image": "vllm/vllm-openai-cpu:v0.20.2"
      },
      "needs_volume": true,
      "volume_mount_path": "/data",
      "http": true
    }
  ],
  "required_inputs": [
    {
      "key": "PORT",
      "service": "vllm",
      "description": "HTTP listening port",
      "secret": false,
      "strategy": "default",
      "default": "8000"
    },
    {
      "key": "HF_HOME",
      "service": "vllm",
      "description": "Hugging Face cache on volume",
      "secret": false,
      "strategy": "default",
      "default": "/data"
    },
    {
      "key": "HF_TOKEN",
      "service": "vllm",
      "description": "optional — required for gated models",
      "secret": true,
      "strategy": "default",
      "default": "<required-for-gated-models>"
    },
    {
      "key": "VLLM_DTYPE",
      "service": "vllm",
      "description": "tensor dtype for CPU backend",
      "secret": false,
      "strategy": "default",
      "default": "bfloat16"
    },
    {
      "key": "VLLM_MODEL",
      "service": "vllm",
      "description": "Hugging Face model id. For example, qwen model is provided",
      "secret": false,
      "strategy": "default",
      "default": "Qwen/Qwen2.5-0.5B-Instruct"
    },
    {
      "key": "VLLM_API_KEY",
      "service": "vllm",
      "description": "Bearer token for /v1/* endpoints",
      "secret": true,
      "strategy": "generate",
      "generate": "random_base64_32"
    },
    {
      "key": "VLLM_MAX_MODEL_LEN",
      "service": "vllm",
      "description": "max context window in tokens",
      "secret": false,
      "strategy": "default",
      "default": "4096"
    },
    {
      "key": "VLLM_CPU_KVCACHE_SPACE",
      "service": "vllm",
      "description": "GiB reserved for KV cache",
      "secret": false,
      "strategy": "default",
      "default": "4"
    }
  ],
  "deploy": {
    "mcp": {
      "server": "railway",
      "tool": "deploy_template",
      "args": {
        "template_code": "vllm"
      }
    },
    "cli": "railway deploy --template vllm",
    "api": {
      "method": "POST",
      "path": "/graphql/v2",
      "body": {
        "query": "mutation templateDeploy($input: TemplateDeployV2Input!) { templateDeployV2(input: $input) { projectId workflowId } }",
        "variables": {
          "input": {
            "templateId": "069ed2a5-4f4a-4581-8b2d-e7bedc7fd96a",
            "serializedConfig": {
              "buckets": {},
              "services": {
                "886818f4-32dd-4294-921d-507dcf4ff0b7": {
                  "icon": "https://cdn.jsdelivr.net/gh/homarr-labs/dashboard-icons/svg/vllm.svg",
                  "name": "vllm",
                  "deploy": {
                    "startCommand": "/bin/sh -c 'python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port ${PORT} --model ${VLLM_MODEL} --dtype ${VLLM_DTYPE} --max-model-len ${VLLM_MAX_MODEL_LEN} --download-dir /data'",
                    "healthcheckPath": null,
                    "restartPolicyType": "ON_FAILURE",
                    "restartPolicyMaxRetries": 10
                  },
                  "source": {
                    "image": "vllm/vllm-openai-cpu:v0.20.2"
                  },
                  "variables": {
                    "PORT": {
                      "isOptional": false,
                      "description": "HTTP listening port",
                      "defaultValue": "8000"
                    },
                    "HF_HOME": {
                      "isOptional": false,
                      "description": "Hugging Face cache on volume",
                      "defaultValue": "/data"
                    },
                    "HF_TOKEN": {
                      "description": "optional — required for gated models",
                      "defaultValue": "<required-for-gated-models>"
                    },
                    "VLLM_DTYPE": {
                      "isOptional": false,
                      "description": "tensor dtype for CPU backend",
                      "defaultValue": "bfloat16"
                    },
                    "VLLM_MODEL": {
                      "isOptional": false,
                      "description": "Hugging Face model id. For example, qwen model is provided",
                      "defaultValue": "Qwen/Qwen2.5-0.5B-Instruct"
                    },
                    "VLLM_API_KEY": {
                      "isOptional": false,
                      "description": "Bearer token for /v1/* endpoints",
                      "defaultValue": "{{VLLM_API_KEY}}"
                    },
                    "VLLM_MAX_MODEL_LEN": {
                      "isOptional": false,
                      "description": "max context window in tokens",
                      "defaultValue": "4096"
                    },
                    "VLLM_CPU_KVCACHE_SPACE": {
                      "isOptional": false,
                      "description": "GiB reserved for KV cache",
                      "defaultValue": "4"
                    }
                  },
                  "networking": {
                    "serviceDomains": {
                      "<hasDomain>": {}
                    }
                  },
                  "volumeMounts": {
                    "886818f4-32dd-4294-921d-507dcf4ff0b7": {
                      "mountPath": "/data"
                    }
                  }
                }
              }
            }
          }
        }
      }
    }
  },
  "post_deploy": {},
  "resources": {
    "expected_services": 1,
    "needs_volume": true
  },
  "generated_at": "2026-09-22T04:14:41.730Z",
  "generator_version": "0.1.0",
  "status": "validated",
  "validated_at": "2026-09-20T12:04:01.177Z",
  "success_rate_30d": 1,
  "validation": {
    "last_run_id": "run_e4e61715d0f44876aaf3",
    "checks": [
      {
        "name": "workflow_completed",
        "passed": true
      },
      {
        "name": "all_services_deployed",
        "passed": true
      },
      {
        "name": "stays_up",
        "passed": true
      }
    ],
    "typical_ready_seconds": 94,
    "typical_build_seconds": 0,
    "typical_start_seconds": 21,
    "slowest_service": "vllm"
  }
}