CreateEndpointRequest

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "#/components/schemas/CreateEndpointRequest",
  "title": "CreateEndpointRequest",
  "type": "object",
  "required": [
    "name",
    "type",
    "provider",
    "compute",
    "model"
  ],
  "properties": {
    "name": {
      "type": "string",
      "description": "Unique endpoint name",
      "example": "my-text-gen-endpoint",
      "pattern": "^[a-z0-9][a-z0-9-]{0,30}[a-z0-9]$"
    },
    "type": {
      "type": "string",
      "description": "Endpoint security type",
      "enum": [
        "public",
        "protected",
        "private"
      ],
      "default": "protected",
      "example": "public"
    },
    "provider": {
      "type": "object",
      "required": [
        "vendor",
        "region"
      ],
      "properties": {
        "vendor": {
          "type": "string",
          "enum": [
            "aws",
            "azure",
            "gcp"
          ]
        },
        "region": {
          "type": "string",
          "example": "us-east-1"
        }
      },
      "example": "example_value"
    },
    "compute": {
      "type": "object",
      "required": [
        "accelerator",
        "instanceType",
        "instanceSize",
        "scaling"
      ],
      "properties": {
        "accelerator": {
          "type": "string",
          "enum": [
            "cpu",
            "gpu"
          ]
        },
        "instanceType": {
          "type": "string",
          "example": "nvidia-a10g"
        },
        "instanceSize": {
          "type": "string",
          "example": "x1"
        },
        "scaling": {
          "type": "object",
          "required": [
            "minReplica",
            "maxReplica"
          ],
          "properties": {
            "minReplica": {
              "type": "integer",
              "minimum": 0,
              "example": 0
            },
            "maxReplica": {
              "type": "integer",
              "minimum": 1,
              "example": 2
            },
            "scaleToZeroTimeout": {
              "type": "integer",
              "description": "Minutes of inactivity before scaling to zero",
              "example": 15
            }
          }
        }
      },
      "example": "example_value"
    },
    "model": {
      "type": "object",
      "required": [
        "repository",
        "task"
      ],
      "properties": {
        "repository": {
          "type": "string",
          "description": "Hugging Face model repository ID",
          "example": "meta-llama/Llama-2-7b-chat-hf"
        },
        "revision": {
          "type": "string",
          "description": "Git revision to deploy",
          "default": "main"
        },
        "task": {
          "type": "string",
          "description": "Task type for the endpoint",
          "example": "text-generation"
        },
        "framework": {
          "type": "string",
          "enum": [
            "pytorch",
            "custom"
          ],
          "default": "pytorch"
        },
        "image": {
          "type": "object",
          "properties": {
            "huggingface": {
              "type": "object"
            },
            "custom": {
              "type": "object",
              "properties": {
                "url": {
                  "type": "string",
                  "format": "uri"
                },
                "health_route": {
                  "type": "string"
                },
                "env": {
                  "type": "object",
                  "additionalProperties": {
                    "type": "string"
                  }
                }
              }
            }
          }
        }
      },
      "example": "example_value"
    }
  }
}