Lemonade Server Management

Interact with and manage local or remote Lemonade AI Server hardware directly via standard native network requests (curl).

When to Use

• Checking local GPU/NPU health and currently loaded VRAM resources
• Listing available, downloaded text/image models on a Lemonade cluster
• Pulling, loading, or unloading multimodal models
• Generating text from LLMs (chat completions)
• Generating stable-diffusion image responses

Setup Instructions

1. Every endpoint requires a base server_url. If one is not specified by the user, assume http://localhost:8000.
2. Use the LEMONADE_API_KEY environment variable to authenticate.
3. This skill stays focused on making HTTP requests to a base server_url (default http://localhost:8000) and using an optional LEMONADE_API_KEY. It does not instruct reading arbitrary host files or other env vars.
4. Operational note: The skill will transmit prompts and model requests to whichever server_url is used. If a remote URL is supplied, the agent will send user content and model commands there (expected for this skill, but worth the user's attention).

API Operations

System Info

Get hardware capabilities and device enumeration limits.

Returns: JSON object with hardware capabilities.

# Example Request
curl -X GET "http://localhost:8000/api/v1/system-info" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}"

Health Check

Verify status and monitor currently loaded models to prevent VRAM overflow.

Returns: JSON object containing server status and currently loaded models.

# Example Request
curl -X GET "http://localhost:8000/api/v1/health" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}"

List Models

Get an array of downloaded models available to load into memory.

Returns: JSON list of all downloaded and available models.

# Example Request
curl -X GET "http://localhost:8000/api/v1/models" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}"

Pull Model

Download and install a new model string to the target machine.

Returns: JSON stream or object confirming download status.

# Example Request
curl -X POST "http://localhost:8000/api/v1/pull" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3"}'

Load Model

Load a model into VRAM/NPU to prepare for prompt responses.

Returns: JSON object confirming model loaded into memory.

# Example Request
curl -X POST "http://localhost:8000/api/v1/load" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3"}'

Unload Model

Force unload a model to free up memory before loading a larger variant.

Returns: JSON object confirming model unloaded.

# Example Request
curl -X POST "http://localhost:8000/api/v1/unload" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3"}'

Chat Completion

Send a standard chat request to the LLM backend.

Returns: JSON object containing standard chat completion response.

# Example Request
curl -X POST "http://localhost:8000/api/v1/chat/completions" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama3",
    "messages": [
      {"role": "user", "content": "Hello, world!"}
    ]
  }'

Generate Image

Submit a stable-diffusion prompt for image generation.

Returns: JSON object containing base64 generated image data.

# Example Request
curl -X POST "http://localhost:8000/api/v1/images/generations" \
  -H "Authorization: Bearer ${LEMONADE_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sdxl",
    "prompt": "A majestic lion standing on a cliff, digital art"
  }'

Tips

• The flm, ryzenai-llm, and whispercpp backends are mutually exclusive on an NPU. Always use the /api/v1/unload endpoint on conflicting LLMs before attempting to load a Whisper model on an NPU.
• Check /api/v1/health first to assess available hardware VRAM before pulling or loading a multi-gigabyte LLM to avoid out-of-memory errors on the host.
• Provide clear feedback to the user if an API request fails due to resource constraints and suggest unloading inactive background models.