Langchain Chat Model + LLamaCPP

A well tested 🧪 working solution for Integrating LLamaCPP with langchain. Fully compatible with ChatModel, and LangGraph integration. Provide a direct interface to the LlamaCPP library, without any additional wrapper layers, to maintain full configurability and control over the LlamaCPP functionality.

If you find this project useful, please give it a star ⭐!

Support:

• ✅ invoke
• ✅ ainvoke
• ✅ stream
• ✅ astream
• ✅ Structured output (JSON mode)
• ✅ Tool/Function calling
• ✅ LLamaProxy

Quick Install

pip

pip install langchain-llamacpp-chat-model

# When using llama_proxy
pip install langchain-llamacpp-chat-model langchain-llamacpp-chat-model[llama_proxy]

poetry

poetry add langchain-llamacpp-chat-model

# When using llama_proxy
poetry add langchain-llamacpp-chat-model langchain-llamacpp-chat-model[llama_proxy]

Usage

Using Llama

Llama instance allow to create a chat model for a single llama model

import os
from langchain_core.pydantic_v1 import BaseModel, Field
from llama_cpp import Llama
from langchain_llamacpp_chat_model import LlamaChatModel
from langchain_core.tools import tool

model_path = os.path.join(
    os.path.expanduser("~/.cache/lm-studio/models"),
    "lmstudio-community/Meta-Llama-3-8B-Instruct-GGUF/Meta-Llama-3-8B-Instruct-Q4_K_M.gguf",
)

llama = Llama(
    model_path=model_path,
    _gpu_layers=-1,
    chat_format="chatml-function-calling", # https://llama-cpp-python.readthedocs.io/en/latest/#function-calling
)
chat_model = LlamaChatModel(llama=llama)

Invoke

result = chat_model.invoke("Tell me a joke about cats")
print(
    result.content
)  # Why was the cat sitting on the computer? Because it wanted to keep an eye on the mouse!

Stream

stream = chat_model.stream("Tell me a joke about cats")
final_content = ""
for token in stream:
    final_content += token.content

print(
    final_content
)  # Why was the cat sitting on the computer? Because it wanted to keep an eye on the mouse!

Strucuted Output

class Joke(BaseModel):
    setup: str = Field(description="The setup of the joke")
    punchline: str = Field(description="The punchline to the joke")


structured_llm = chat_model.with_structured_output(Joke)

result = structured_llm.invoke("Tell me a joke about cats")

assert isinstance(result, Joke)
print(result.setup)  # Why was the cat sitting on the computer?
print(result.punchline)  # Because it wanted to keep an eye on the mouse!

⚠️ Ensure temperature is set to 0 (or near 0). The open-source models (tested with Phi3 and LLama3) do not perform well when calling functions, as their behavior can be unpredictable.

Function calling

@tool
def magic_number_tool(input: int) -> int:
    """Applies a magic function to an input."""
    return input + 2


llm_with_tool = chat_model.bind_tools(
    [magic_number_tool], tool_choice="magic_number_tool"
)

result = llm_with_tool.invoke("What is the magic mumber of 2?")

assert result.tool_calls[0]["name"] == "magic_number_tool"

⚠️ Ensure temperature is set to 0 (or near 0). The open-source models (tested with Phi3 and LLama3) do not perform well when calling functions, as their behavior can be unpredictable.

Using LlamaProxy

LLamaProxy allow to define multiple models and use one of them by specifying model_name. Very useful for a server environment.

import os
from llama_cpp.server.app import LlamaProxy, ModelSettings
from langchain_llamacpp_chat_model.proxy import LlamaProxyChatModel

llama3_model_path = os.path.join(
    os.path.expanduser("~/.cache/lm-studio/models"),
    "lmstudio-community/Meta-Llama-3-8B-Instruct-GGUF/Meta-Llama-3-8B-Instruct-Q4_K_M.gguf",
)
phi3_model_path = os.path.join(
    os.path.expanduser("~/.cache/lm-studio/models"),
    "microsoft/Phi-3-mini-4k-instruct-gguf/Phi-3-mini-4k-instruct-q4.gguf",
)

llama_proxy = LlamaProxy(
    models=[
        ModelSettings(model=llama3_model_path, model_alias="llama3"),
        ModelSettings(model=phi3_model_path, model_alias="phi3"),
    ]
)
llama3_chat_model = LlamaProxyChatModel(llama_proxy=llama_proxy, model="llama3")
phi3_chat_model = LlamaProxyChatModel(llama_proxy=llama_proxy, model="phi3")


# Invoke
# --------------------------------------------------------
llama3_result = llama3_chat_model.invoke("Tell me a joke about cats")
print(llama3_result.content)

phi3_result = llama3_chat_model.invoke("Tell me a joke about cats")
print(phi3_result.content)

# Stream
# --------------------------------------------------------
stream = llama3_chat_model.stream("Tell me a joke about cats")
final_content = ""
for token in stream:
    final_content += token.content

print(final_content)