Function Calling with Llama 3.3 – Complete Guide

Llama 3.3 was trained with function-calling support but the implementation differs from OpenAI’s format. On our dedicated GPU hosting getting reliable tool use requires both the right vLLM flags and correct message formatting.

Contents

• vLLM flags for tool parsing
• Request format
• Parsing the response
• Reliability tips

vLLM Flags

python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Llama-3.3-70B-Instruct \
  --enable-auto-tool-choice \
  --tool-call-parser llama3_json \
  --chat-template /path/to/llama3_chat_template.jinja

--tool-call-parser llama3_json teaches vLLM how to extract tool calls from Llama’s native output. Without it, tool calls come through as free text.

Request

response = client.chat.completions.create(
  model="llama3.3",
  messages=[{"role": "user", "content": "What's the weather in London?"}],
  tools=[{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Get current weather for a location",
      "parameters": {
        "type": "object",
        "properties": {
          "location": {"type": "string"},
          "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
        },
        "required": ["location"]
      }
    }
  }],
  tool_choice="auto"
)

Parsing

Llama 3.3 emits tool calls in JSON. vLLM parses them into the standard OpenAI structure:

if response.choices[0].message.tool_calls:
    for tc in response.choices[0].message.tool_calls:
        name = tc.function.name
        args = json.loads(tc.function.arguments)
        result = execute(name, **args)

Reliability

• Describe tools precisely – Llama follows descriptions closely
• Keep tool schemas small – 5-10 tools per call is the sweet spot
• Include an explicit example in the system prompt for rare tools
• At 70B you get near-OpenAI-level reliability. At 8B, function calling is noticeably weaker – keep tools simple

See tool use with Qwen Coder.