Function calling requires two step inferences, below is the example:

Step 1:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import json

model_id = "R1_tool_call_Distill-Qwen-1.5B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tools = [
    {
        "name": "create_contact",
        "description": "Create a new contact",
        "parameters": {
            "type": "object",
            "properties": {
                "name": {
                    "type": "string",
                    "description": "The name of the contact"
                },
                "email": {
                    "type": "string",
                    "description": "The email address of the contact"
                }
            },
            "required": ["name", "email"]
        }
    }
]


messages = [
    { "role": "user", "content": f"""You have access to these tools, use them if necessary: {tools}
    I need to create a new contact for my friend John Doe. His email is johndoe@example.com."""
    }
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)


outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
# >>
# >> <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>create_contact
# >> ```json
# >> {"name": "John Doe", "email": "johndoe@example.com"}
# >> ```<｜tool▁call▁end｜><｜tool▁calls▁end｜>

# Above is a response from assistant, you need to parse it and execute a tool on your own.

Step 2:

messages = [
    {"role": "user", "content": """You have access to these tools, use them if necessary: {tools}\n\nI need to create a new contact for my friend John Doe. His email is johndoe@example.com."""},
    {"role": "assistant", "content": None, "tool_calls": [
        {
            "type": "function",
            "function": {
                "name": "create_contact",
                "arguments": json.dumps({"name": "John Doe", "email": "johndoe@example.com"})
            }
        },
    ]},
    {"role": "tool", "name": "create_contact", "content": """{"status": "success", "message": "Contact for John Doe with email johndoe@example.com has been created successfully."}"""},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
# >>
# >> <think>Based on the user's request, I have created a contact for John Doe with his email address. The tool has successfully created the contact. I will now provide the contact information to the user.</think>

# >> The contact for John Doe has been successfully created with the email address johndoe@example.com. Please feel free to reach out to him if needed.

Limitations:

• The model sometimes refused not to think

Uploaded model

• Developed by: hiieu
• License: apache-2.0
• Finetuned from model : unsloth/DeepSeek-R1-Distill-Qwen-7B-unsloth-bnb-4bit

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.