# Sample vLLM inference offering for the OpenNebula site-agent backend.
#
# Import it via the marketplace `import_offering` endpoint, passing this file's
# contents as `offering_data`, e.g.:
#
#   curl -X POST \
#     -H "Authorization: token <TOKEN>" \
#     -H "Content-Type: application/json" \
#     "<WALDUR_API_URL>/api/marketplace-provider-offerings/import_offering/" \
#     -d "$(python3 - <<'PY'
#   import json, yaml
#   print(json.dumps({
#       "offering_data": open("vllm-inference-offering.yaml").read(),
#       "customer": "<CUSTOMER_UUID>",
#       "category": "AI / Inference",
#       "overwrite_existing": True,
#   }))
#   PY
#   )"
#
# Imported offerings are always created in DRAFT state — activate the offering
# afterwards. Adjust `plugin_options` (template/engine/VDC IDs) and the
# `model_image_id` choices to match your OpenNebula deployment, then point the
# site agent at this offering (resource_type: vm).

offering:
  name: "vLLM Inference"
  description: >-
    Deploy a large language model as an inference service. A VM boots the vLLM
    engine with your selected model attached as a disk and serves it over an
    OpenAI-compatible API endpoint.
  type: "Marketplace.Slurm"
  shared: true
  billable: true
  category_name: "AI / Inference"
  getting_started: |
    Your inference service **{resource_name}** serves model image
    {backend_metadata_model} over an OpenAI-compatible API.

    **API base URL:** {backend_metadata_endpoint}/v1

    ```bash
    curl {backend_metadata_endpoint}/v1/models
    ```

    Point any OpenAI-compatible client at the base URL above (no API key is
    required on the internal network).

    **Chat playground:** {backend_metadata_web_ui}
  options:
    order:
      - model
      - ONEAPP_VLLM_API_PORT
      - ONEAPP_VLLM_API_WEB
      - ONEAPP_VLLM_MODEL_QUANTIZATION
      - ONEAPP_VLLM_MODEL_MAX_LENGTH
      - ONEAPP_VLLM_ENFORCE_EAGER
      - ONEAPP_VLLM_SLEEP_MODE
      - ONEAPP_VLLM_GPU_MEMORY_UTILIZATION
    options:
      model:
        type: select_string
        label: Model
        help_text: "Model to serve. The agent resolves the name to its OpenNebula image."
        required: true
        choices:
          - "Qwen_Qwen3-0.6B"
          - "deepseek-ai_DeepSeek-V3.2"
          - "utter-project_EuroLLM-9B-Instruct-2512"
        default: "Qwen_Qwen3-0.6B"
      ONEAPP_VLLM_API_PORT:
        type: integer
        label: API port
        default: "8000"
      ONEAPP_VLLM_API_WEB:
        type: select_string
        label: Deploy chat web interface
        choices:
          - "YES"
          - "NO"
        default: "YES"
      ONEAPP_VLLM_MODEL_QUANTIZATION:
        type: select_string
        label: Quantization (bits)
        choices:
          - "0"
          - "4"
        default: "0"
      ONEAPP_VLLM_MODEL_MAX_LENGTH:
        type: integer
        label: Model context length
        default: "1024"
      ONEAPP_VLLM_ENFORCE_EAGER:
        type: select_string
        label: Enforce eager mode
        choices:
          - "YES"
          - "NO"
        default: "NO"
      ONEAPP_VLLM_SLEEP_MODE:
        type: select_string
        label: Enable sleep mode
        choices:
          - "YES"
          - "NO"
        default: "NO"
      ONEAPP_VLLM_GPU_MEMORY_UTILIZATION:
        type: string
        label: GPU memory utilisation (0, 1]
        default: "0.9"

# Inference-specific backend configuration (resolved by the site agent via
# offering_plugin_options). Adjust the IDs to your deployment.
plugin_options:
  template_id: 0
  engine_image_id: 0
  parent_vdc_backend_id: "labvdc"
  # Show the in-browser inference Playground action on resources of this
  # offering (HomePort; requires the endpoint to be reachable from the browser).
  expose_inference_playground: true

# VM size comes from these FIXED components' plan amounts.
# RAM and disk use GB as the measured unit; unit_factor converts to the
# backend's native MB (the agent multiplies plan amounts by it). Plan amounts
# and prices are therefore expressed per GB.
components:
  - type: vcpu
    name: vCPU
    billing_type: fixed
    measured_unit: cores
  - type: vm_ram
    name: RAM
    billing_type: fixed
    measured_unit: GB
    unit_factor: 1024
  - type: vm_disk
    name: Disk
    billing_type: fixed
    measured_unit: GB
    unit_factor: 1024

# Switching plan triggers a VM resize: the agent powers the VM off, applies the
# new vCPU/RAM and grows the disk, then powers it back on (brief downtime; disk
# can only grow, not shrink). Plan cost per month = sum(price x amount) over the
# fixed components.
plans:
  - name: vllm-small
    description: "4 vCPU / 8 GB RAM / 20 GB disk"
    unit: month
    components:
      - component_type: vcpu
        amount: 4
        price: 5
      - component_type: vm_ram
        amount: 8
        price: 2
      - component_type: vm_disk
        amount: 20
        price: 0.1
  - name: vllm-large
    description: "8 vCPU / 32 GB RAM / 80 GB disk"
    unit: month
    components:
      - component_type: vcpu
        amount: 8
        price: 5
      - component_type: vm_ram
        amount: 32
        price: 2
      - component_type: vm_disk
        amount: 80
        price: 0.1