Hugging Face Deploy Inference Endpoint and Wait

arazzo: 1.0.1
info:
  title: Hugging Face Deploy Inference Endpoint and Wait
  summary: Create a dedicated Inference Endpoint, then poll its status until it is running.
  description: >-
    A deployment flow over the Inference Endpoints management API. The workflow
    creates a new dedicated endpoint for a model on the chosen cloud provider and
    hardware, then polls the endpoint's status, looping while it is still
    pending or initializing and exiting once it reaches the running state. The
    poll step branches back to itself until the endpoint is ready. Every step
    spells out its request inline so the flow can be read and executed without
    opening the underlying OpenAPI description.
  version: 1.0.0
sourceDescriptions:
- name: inferenceEndpointsApi
  url: ../openapi/hugging-face-inference-endpoints-api.yml
  type: openapi
workflows:
- workflowId: deploy-inference-endpoint
  summary: Create a dedicated Inference Endpoint and wait until it is running.
  description: >-
    Provisions a new Inference Endpoint for a model and polls its status until
    it reports the running state, returning the inference URL.
  inputs:
    type: object
    required:
    - hfToken
    - namespace
    - name
    - repository
    properties:
      hfToken:
        type: string
        description: Hugging Face access token used as a Bearer credential.
      namespace:
        type: string
        description: User or organization namespace that will own the endpoint.
      name:
        type: string
        description: The endpoint name to create.
      repository:
        type: string
        description: The model repository to deploy (e.g. gpt2).
      revision:
        type: string
        description: Git revision of the model to deploy.
        default: main
      task:
        type: string
        description: The pipeline task the endpoint will serve.
        default: text-generation
      vendor:
        type: string
        description: Cloud vendor to deploy on.
        default: aws
      region:
        type: string
        description: Cloud region to deploy in.
        default: us-east-1
      instanceType:
        type: string
        description: Compute instance type identifier.
      instanceSize:
        type: string
        description: Compute instance size identifier.
  steps:
  - stepId: createEndpoint
    description: >-
      Create a new dedicated Inference Endpoint for the model. A 201 returns the
      endpoint record; a 409 means an endpoint with this name already exists.
    operationId: createEndpoint
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.hfToken
    - name: namespace
      in: path
      value: $inputs.namespace
    requestBody:
      contentType: application/json
      payload:
        name: $inputs.name
        type: public
        provider:
          vendor: $inputs.vendor
          region: $inputs.region
        compute:
          accelerator: cpu
          instanceType: $inputs.instanceType
          instanceSize: $inputs.instanceSize
          scaling:
            minReplica: 1
            maxReplica: 1
        model:
          repository: $inputs.repository
          revision: $inputs.revision
          task: $inputs.task
          framework: pytorch
    successCriteria:
    - condition: $statusCode == 201
    outputs:
      endpointName: $response.body#/name
      initialState: $response.body#/status/state
    onSuccess:
    - name: created
      type: goto
      stepId: pollStatus
      criteria:
      - condition: $statusCode == 201
  - stepId: pollStatus
    description: >-
      Poll the endpoint status. While the state is pending or initializing the
      step loops back to itself; once the state is running it proceeds to the
      end of the workflow.
    operationId: getEndpoint
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.hfToken
    - name: namespace
      in: path
      value: $inputs.namespace
    - name: endpoint_name
      in: path
      value: $steps.createEndpoint.outputs.endpointName
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      state: $response.body#/status/state
      inferenceUrl: $response.body#/status/url
    onSuccess:
    - name: stillStarting
      type: goto
      stepId: pollStatus
      criteria:
      - context: $response.body
        condition: $.status.state == 'pending' || $.status.state == 'initializing'
        type: jsonpath
    - name: ready
      type: end
      criteria:
      - context: $response.body
        condition: $.status.state == 'running'
        type: jsonpath
  outputs:
    endpointName: $steps.createEndpoint.outputs.endpointName
    state: $steps.pollStatus.outputs.state
    inferenceUrl: $steps.pollStatus.outputs.inferenceUrl