Skip to content

Latest commit

 

History

History
731 lines (581 loc) · 33.5 KB

README.md

File metadata and controls

731 lines (581 loc) · 33.5 KB
Raw

Functionary

Functionary is a language model that can interpret and execute functions/plugins.

The model determines when to execute functions, whether in parallel or serially, and can understand their outputs. It only triggers functions as needed. Function definitions are given as JSON Schema Objects, similar to OpenAI GPT function calls.

Documentation and more examples: functionary.meetkai.com

Changelog: (click to expand)

Getting Started

Functionary can be deployed using either our vLLM or SGLang servers. Choose either one depending on your preferences.

Installation

vLLM

pip install -e .[vllm]

SGLang

pip install -e .[sglang] --find-links https://flashinfer.ai/whl/cu121/torch2.4/flashinfer/

Running the server

Small Model

vLLM

python3 server_vllm.py --model "meetkai/functionary-small-v3.2" --host 0.0.0.0 --port 8000 --max-model-len 8192

SGLang

python3 server_sglang.py --model-path "meetkai/functionary-small-v3.2" --host 0.0.0.0 --port 8000 --context-length 8192

Medium Model

Our medium models require: 4xA6000 or 2xA100 80GB to run, need to use: tensor-parallel-size or tp (SGLang)

vLLM

# vllm requires to run this first: https://github.com/vllm-project/vllm/issues/6152
export VLLM_WORKER_MULTIPROC_METHOD=spawn

python server_vllm.py --model "meetkai/functionary-medium-v3.1" --host 0.0.0.0 --port 8000 --max-model-len 8192 --tensor-parallel-size 2

SGLang

python server_sglang.py --model-path "meetkai/functionary-medium-v3.1" --host 0.0.0.0 --port 8000 --context-length 8192 --tp 2

LoRA Support (Currently Only in vLLM)

Similar to LoRA in vLLM, our server supports serving LoRA adapters both at startup and dynamically.

To serve a LoRA adapter at startup, run the server with the --lora-modules argument:

python server_vllm.py --model {BASE_MODEL} --enable-lora --lora-modules {name}={path} {name}={path} --host 0.0.0.0 --port 8000

To serve a LoRA adapter dynamically, use the /v1/load_lora_adapter endpoint:

python server_vllm.py --model {BASE_MODEL} --enable-lora --host 0.0.0.0 --port 8000
# Load a LoRA adapter dynamically
curl -X POST http://localhost:8000/v1/load_lora_adapter \
  -H "Content-Type: application/json" \
  -d '{
    "lora_name": "my_lora",
    "lora_path": "/path/to/my_lora_adapter"
  }'
# Example chat request to lora adapter
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "my_lora",
    "messages": [...],
    "tools": [...],
    "tool_choice": "auto"
  }'
# Unload a LoRA adapter dynamically
curl -X POST http://localhost:8000/v1/unload_lora_adapter \
  -H "Content-Type: application/json" \
  -d '{
    "lora_name": "my_lora"
  }'

Grammar Sampling (Only in vLLM)

We also offer our own function-calling grammar sampling feature which constrains the LLM's generation to always follow the prompt template, and ensures 100% accuracy for function name. The parameters are generated using the efficient lm-format-enforcer, which ensures that the parameters follow the schema of the tool called. To enable grammar sampling, run the vLLM server with the command-line argument --enable-grammar-sampling:

python3 server_vllm.py --model "meetkai/functionary-medium-v3.1" --max-model-len 8192 --tensor-parallel-size 2 --enable-grammar-sampling

Note: Grammar Sampling support is applicable only for the V2, V3.0, V3.2 models. There is no such support for V1 and V3.1 models.

Text-Generation-Inference (TGI)

We also provide a service that performs inference on Functionary models using Text-Generation-Inference (TGI). Follow these steps to get started:

  1. Install Docker following their installation instructions.

  2. Install the Docker SDK for Python

pip install docker
  1. Start up the Functionary TGI server

At start-up, the Functionary TGI server tries to connect to an existing TGI endpoint. In this case, you can run the following:

python3 server_tgi.py --model <REMOTE_MODEL_ID_OR_LOCAL_MODEL_PATH> --endpoint <TGI_SERVICE_ENDPOINT>

If the TGI endpoint does not exist, the Functionary TGI server will start a new TGI endpoint container with the address provided in the endpoint CLI argument via the installed Docker Python SDK. Run the following commands for remote and local models respectively:

python3 server_tgi.py --model <REMOTE_MODEL_ID> --remote_model_save_folder <PATH_TO_SAVE_AND_CACHE_REMOTE_MODEL> --endpoint <TGI_SERVICE_ENDPOINT>
python3 server_tgi.py --model <LOCAL_MODEL_PATH> --endpoint <TGI_SERVICE_ENDPOINT>
  1. Make either OpenAI-compatible or raw HTTP requests to the Functionary TGI server.

Docker

If you're having trouble with dependencies, and you have nvidia-container-toolkit, you can start your environment like this:

sudo docker run --gpus all -it --ipc=host --name functionary -v ${PWD}/functionary_workspace:/workspace -p 8000:8000 nvcr.io/nvidia/pytorch:23.10-py3

OpenAI Compatible Usage

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="functionary")

client.chat.completions.create(
    model="meetkai/functionary-small-v3.2",
    messages=[{"role": "user",
            "content": "What is the weather for Istanbul?"}
    ],
    tools=[{
            "type": "function",
            "function": {
                "name": "get_current_weather",
                "description": "Get the current weather",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "location": {
                            "type": "string",
                            "description": "The city and state, e.g. San Francisco, CA"
                        }
                    },
                    "required": ["location"]
                }
            }
        }],
    tool_choice="auto"
)

Raw Usage:

Details (click to expand) 
import requests

data = {
    'model': 'meetkai/functionary-small-v3.2', # model name here is the value of argument "--model" in deploying: server_vllm.py or server.py
    'messages': [
        {
            "role": "user",
            "content": "What is the weather for Istanbul?"
        }
    ],
    'tools':[ # For functionary-7b-v2 we use "tools"; for functionary-7b-v1.4 we use "functions" = [{"name": "get_current_weather", "description":..., "parameters": ....}]
        {
            "type": "function",
            "function": {
                "name": "get_current_weather",
                "description": "Get the current weather",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "location": {
                            "type": "string",
                            "description": "The city and state, e.g. San Francisco, CA"
                        }
                    },
                    "required": ["location"]
                }
            }
        }
    ]
}

response = requests.post("http://127.0.0.1:8000/v1/chat/completions", json=data, headers={
    "Content-Type": "application/json",
    "Authorization": "Bearer xxxx"
})

# Print the response text
print(response.text)

Models Available

Model Description VRAM FP16
functionary-medium-v3.2 128k context, code interpreter, using our own prompt template 160GB
functionary-small-v3.2 / GGUF 128k context, code interpreter, using our own prompt template 24GB
functionary-medium-v3.1 / GGUF 128k context, code interpreter, using original Meta's prompt template 160GB
functionary-small-v3.1 / GGUF 128k context, code interpreter, using original Meta's prompt template 24GB
functionary-medium-v3.0 / GGUF 8k context, based on meta-llama/Meta-Llama-3-70B-Instruct 160GB
functionary-small-v2.5 / GGUF 8k context, code interpreter 24GB
functionary-small-v2.4 / GGUF 8k context, code interpreter 24GB
functionary-medium-v2.4 / GGUF 8k context, code interpreter, better accuracy 90GB
functionary-small-v2.2 / GGUF 8k context 24GB
functionary-medium-v2.2 / GGUF 8k context 90GB
functionary-7b-v2.1 / GGUF 8k context 24GB
functionary-7b-v2 / GGUF Parallel function call support. 24GB
functionary-7b-v1.4 / GGUF 4k context, better accuracy (deprecated) 24GB
functionary-7b-v1.1 4k context (deprecated) 24GB
functionary-7b-v0.1 2k context (deprecated) Not recommended, use 2.1 onwards 24GB

Compatibility information

  • v1 models are compatible with both OpenAI-python v0 and v1.
  • v2 models are designed for compatibility with OpenAI-python v1.

The difference between OpenAI-python v0 and v1 you may refer to the official documentation here

The Differences Between Related Projects

Feature/Project Functionary NexusRaven Gorilla Glaive GPT-4-1106-preview
Single Function Call ✅ ✅ ✅ ✅ ✅
Parallel Function Calls ✅ ✅ ✅ ❌ ✅
Following Up on Missing Function Arguments ✅ ❌ ❌ ❌ ✅
Multi-turn ✅ ❌ ❌ ✅ ✅
Generate Model Responses Grounded in Tools Execution Results ✅ ❌ ❌ ❌ ✅
Chit-Chat ✅ ❌ ✅ ✅ ✅
Code Interpreter ✅ ❌ ❌ ❌ ✅

You can find more details of the features in here

Llama.cpp Inference

Llama.cpp Inference using Huggingface Tokenizer

Example for inference using LLama-cpp-python can be found in: llama_cpp_inference.py.

Integration into Llama-cpp

Besides, functionary was also integrated into LLama-cpp-python, however the integration might not be quickly updated, so if there is something wrong or weird in the result, please use: llama_cpp_inference.py instead. Currently, v2.5 hasn't been integrated, so if you are using functionary-small-v2.5-GGUF, please use: llama_cpp_inference.py

Make sure that the latest version of llama-cpp-python is successully installed in your system. Functionary v2 is fully integrated into llama-cpp-python. You can perform inference using Functionary's GGUF models either via normal chat completion or through llama-cpp-python's OpenAI-compatible server which behaves similarly to ours.

The following is the sample code using normal chat completion:

from llama_cpp import Llama
from llama_cpp.llama_tokenizer import LlamaHFTokenizer

# We should use HF AutoTokenizer instead of llama.cpp's tokenizer because we found that Llama.cpp's tokenizer doesn't give the same result as that from Huggingface. The reason might be in the training, we added new tokens to the tokenizer and Llama.cpp doesn't handle this successfully
llm = Llama.from_pretrained(
    repo_id="meetkai/functionary-small-v2.4-GGUF",
    filename="functionary-small-v2.4.Q4_0.gguf",
    chat_format="functionary-v2",
    tokenizer=LlamaHFTokenizer.from_pretrained("meetkai/functionary-small-v2.4-GGUF"),
    n_gpu_layers=-1
)

messages = [
    {"role": "user", "content": "what's the weather like in Hanoi?"}
]
tools = [ # For functionary-7b-v2 we use "tools"; for functionary-7b-v1.4 we use "functions" = [{"name": "get_current_weather", "description":..., "parameters": ....}]
    {
        "type": "function",
        "function": {
            "name": "get_current_weather",
            "description": "Get the current weather",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "The city and state, e.g., San Francisco, CA"
                    }
                },
                "required": ["location"]
            }
        }
    }
]

result = llm.create_chat_completion(
      messages = messages,
      tools=tools,
      tool_choice="auto",
)

print(result["choices"][0]["message"])

The output would be:

{'role': 'assistant', 'content': None, 'tool_calls': [{'type': 'function', 'function': {'name': 'get_current_weather', 'arguments': '{\n  "location": "Hanoi"\n}'}}]}

For more details, please refer to the Function Calling section in llama-cpp-python. To use our Functionary GGUF models using llama-cpp-python's OpenAI-compatible server, please refer to here for more details and documentation.

Note:

  • For Functionary in llama-cpp-python, the default system messages are added automatically during the API call. Therefore, there is no need to provide the default system messages in messages.
  • Streaming feature for Functionary models in both the normal chat completion and in llama-cpp-python's OpenAI-compatible server is officially supported from v0.2.70 onwards.

Call Real Python Function

To call the real python function, get the result and extract the result to respond, you can use chatlab. The following example uses chatlab==0.16.0:

Please note that Chatlab currently doesn't support Parallel Function calls. This sample code is compatible only with Functionary Version 1.4 and may not work correctly with Functionary Version 2.0.

from chatlab import Conversation
import openai
import os
openai.api_key = "functionary" # We just need to set this something other than None
os.environ['OPENAI_API_KEY'] = "functionary" # chatlab requires us to set this too
openai.api_base = "http://localhost:8000/v1"

# now provide the function with description
def get_car_price(car_name: str):
    """this function is used to get the price of the car given the name
    :param car_name: name of the car to get the price
    """
    car_price = {
        "tang": {"price": "$20000"},
        "song": {"price": "$25000"} 
    }
    for key in car_price:
        if key in car_name.lower():
            return {"price": car_price[key]}
    return {"price": "unknown"}

chat = Conversation(model="meetkai/functionary-7b-v2")
chat.register(get_car_price)  # register this function
chat.submit("what is the price of the car named Tang?") # submit user prompt

# print the flow
for message in chat.messages:
    role = message["role"].upper()
    if "function_call" in message:
        func_name = message["function_call"]["name"]
        func_param = message["function_call"]["arguments"]
        print(f"{role}: call function: {func_name}, arguments:{func_param}")
    else:
        content = message["content"]
        print(f"{role}: {content}")

The output will look like this:

USER: what is the price of the car named Tang?
ASSISTANT: call function: get_car_price, arguments:{
  "car_name": "Tang"
}
FUNCTION: {'price': {'price': '$20000'}}
ASSISTANT: The price of the car named Tang is $20,000.

Serverless Deployment using Modal.com

Serverless deployment of Functionary models is supported via the modal_server_vllm.py script. After signing up and installing Modal, follow these steps to deploy our vLLM server on Modal:

  1. Create dev environment
modal environment create dev

If you have a dev environment created already, there is no need to create another one. Just configure to it in the next step.

  1. Configure dev environment
modal config set-environment dev
  1. Serve Functionary Model
modal serve modal_server_vllm
  1. Deploy Runner
modal deploy modal_server_vllm

Use Cases

Here are a few examples of how you can use this function calling system:

Travel and Hospitality - Trip Planning

The function plan_trip(destination: string, duration: int, interests: list) can take user input such as "I want to plan a 7-day trip to Paris with a focus on art and culture" and generate an itinerary accordingly.

Details (click to expand) 
client.chat.completions.create((
    model="meetkai/functionary-7b-v2",
    messages=[
        {"role": "user", "content": 'I want to plan a 7-day trip to Paris with a focus on art and culture'},
    ], 
    tools=[
        {
            "type": "function",
            "function": {
                "name": "plan_trip",
                "description": "Plan a trip based on user's interests",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "destination": {
                            "type": "string",
                            "description": "The destination of the trip",
                        },
                        "duration": {
                            "type": "integer",
                            "description": "The duration of the trip in days",
                        },
                        "interests": {
                            "type": "array",
                            "items": {"type": "string"},
                            "description": "The interests based on which the trip will be planned",
                        },
                    },
                    "required": ["destination", "duration", "interests"],
                }
            }
        }    
    ]
)

Response will have:

{"role": "assistant", "content": null, "tool_calls": [{"type": "function", "function": {"name": "plan_trip", "arguments": '{\n  "destination": "Paris",\n  "duration": 7,\n  "interests": ["art", "culture"]\n}'}}]}

Then you need to call plan_trip function with provided arguments. If you would like a commentary from the model, then you'll call the model again with the response from the function, the model will write necessary commentary.

Real Estate - Property Valuation

A function like estimate_property_value(property_details: dict) could allow users to input details about a property (such as location, size, number of rooms, etc.) and receive an estimated market value.

Details (click to expand) 
client.chat.completions.create(
    model="meetkai/functionary-7b-v2",
    messages=[
        {
            "role": "user", 
            "content": 'What is the estimated value of a 3-bedroom house in San Francisco with 2000 sq ft area?'
        },
        {
            "role": "assistant", 
            "content": None, 
            "tool_calls": [
                {
                    "type": "function", 
                    "function": {
                        "name": "estimate_property_value", 
                        "arguments": '{\n  "property_details": {"location": "San Francisco", "size": 2000, "rooms": 3}\n}'
                    }
                }
            ]
        }
    ], 
    tools=[
        {
            "type": "function",
            "function": {
                "name": "estimate_property_value",
                "description": "Estimate the market value of a property",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "property_details": {
                            "type": "object",
                            "properties": {
                                "location": {
                                    "type": "string",
                                    "description": "The location of the property"
                                },
                                "size": {
                                    "type": "integer",
                                    "description": "The size of the property in square feet"
                                },
                                "rooms": {
                                    "type": "integer",
                                    "description": "The number of rooms in the property"
                                }
                            },
                            "required": ["location", "size", "rooms"]
                        }
                    },
                    "required": ["property_details"]
                }
            }
        }
    ],
    tool_choice="auto"
)

Response will have:

{"role": "assistant", "content": null, "tool_calls": [{"type": "function", "function": {"name": "plan_trip", "arguments": '{\n  "destination": "Paris",\n  "duration": 7,\n  "interests": ["art", "culture"]\n}'}}]}

Then you need to call plan_trip function with provided arguments. If you would like a commentary from the model, then you'll call the model again with the response from the function, the model will write necessary commentary.

Telecommunications - Customer Support

A function parse_customer_complaint(complaint: {issue: string, frequency: string, duration: string}) could help in extracting structured information from a complex, narrative customer complaint, identifying the core issue and potential solutions. The complaint object could include properties such as issue (the main problem), frequency (how often the issue occurs), and duration (how long the issue has been occurring).

Details (click to expand) 
client.chat.completions.create(
    model="meetkai/functionary-7b-v2",
    messages=[
        {"role": "user", "content": 'My internet has been disconnecting frequently for the past week'},
    ], 
    tools=[
        {
            "type": "function",
            "function": {
            "name": "parse_customer_complaint",
            "description": "Parse a customer complaint and identify the core issue",
            "parameters": {
                "type": "object",
                "properties": {
                    "complaint": {
                        "type": "object",
                        "properties": {
                            "issue": {
                                "type": "string",
                                "description": "The main problem",
                            },
                            "frequency": {
                                "type": "string",
                                "description": "How often the issue occurs",
                            },
                            "duration": {
                                "type": "string",
                                "description": "How long the issue has been occurring",
                            },
                        },
                        "required": ["issue", "frequency", "duration"],
                    },
                },
                "required": ["complaint"],
            }
        }
     }
    ],
    tool_choice="auto"
)

Response will have:

{"role": "assistant", "content": null, "tool_calls": [{"type": "function", "function": {"name": "parse_customer_complaint", "arguments": '{\n  "complaint": {"issue": "internet disconnecting", "frequency": "frequently", "duration": "past week"}\n}'}}]}

Then you need to call parse_customer_complaint function with provided arguments. If you would like a commentary from the model, then you'll call the model again with the response from the function, the model will write necessary commentary.

How it Works?

We convert function definitions to a similar text to TypeScript definitions. Then we inject these definitions as system prompts. After that, we inject the default system prompt. Then we start the conversation messages.

The prompt example can be found here: V1 (v1.4), V2 (v2, v2.1, v2.2, v2.4) and V2.llama3 (v2.5)

We don't change the logit probabilities to conform to a certain schema, but the model itself knows how to conform. This allows us to use existing tools and caching systems with ease.

Evaluation

Berkeley Function-Calling Leaderboard

We are ranked 2nd in the Berkeley Function-Calling Leaderboard (Last Updated: 2024-08-11)

Model Name Function Calling Accuracy (Name & Arguments)
meetkai/functionary-medium-v3.1 88.88%
GPT-4-1106-Preview (Prompt) 88.53%
meetkai/functionary-small-v3.2 82.82%
meetkai/functionary-small-v3.1 82.53%
FireFunction-v2 (FC) 78.82.47%

ToolSandbox

We also evaluate our models on ToolSandbox, this benchmark is much more difficult than Berkeley Function-Calling Leaderboard. This benchmark includes stateful tool execution, implicit state dependencies between tools, a built-in user simulator supporting on-policy conversational evaluation and a dynamic evaluation strategy for intermediate and final milestones over an arbitrary trajectory. The authors of this benchmark showed that there is a huge performance gap between open source models and proprietary models.

From our evaluation result, our models are comparable to best proprietary models and much better than other open source models.

Model Name Average similarity score
GPT-4o-2024-05-13 73
Claude-3-Opus-20240229 69.2
Functionary-medium-v3.1 68.87
GPT-3.5-Turbo-0125 65.6
GPT-4-0125-Preview 64.3
Claude-3-Sonnet-20240229 63.8
Functionary-small-v3.1 63.13
Gemini-1.5-Pro-001 60.4
Functionary-small-v3.2 58.56
Claude-3-Haiku-20240307 54.9
Gemini-1.0-Pro 38.1
Hermes-2-Pro-Mistral-7B 31.4
Mistral-7B-Instruct-v0.3 29.8
C4AI-Command-R-v01 26.2
Gorilla-Openfunctions-v2 25.6
C4AI-Command R+ 24.7

Function Prediction Evaluation

Evaluation function call prediction in SGD dataset. The accuracy metric measures the overall correctness of predicted function calls, including function name prediction and arguments extraction.

Dataset Model Name Function Calling Accuracy (Name & Arguments)
SGD meetkai/functionary-medium-v3.1 88.11%
SGD gpt-4o-2024-05-13 82.75%
SGD gemini-1.5-flash 79.64%
SGD c4ai-command-r-plus 45.66%

Training

See training README

Safety & Security

While its not strictly enforced, to ensure more secure function execution, one can enable grammar sampling to enforce type checking. Main safety checks needs to be done in the functions/actions themselves. Such as validation of the given input, or the ouput that will be given to the model.

Roadmap

  • OpenAPI specification based plugin support.
  • Fast inference server
    • vLLM
    • text-generation-inference
    • Streaming Support
    • function_call parameter to server
    • Grammar Sampling to ensure 100% accuracy for function and parameter names
  • Parallel function calling support
  • Python function calling support (Automatic detection of type annotations and calling them automatically)
  • Real world usage examples, such as creating agents.
  • Train Mixtral based model
  • Code interpreter support
  • Please consider opening a PR for future requests