llama-cpp-py 0.1.33

Python wrapper for running the llama.cpp server

llama-cpp-py

Python wrapper for running the llama.cpp server with automatic or manual binary management.
Runs the server in a separate subprocess supporting both synchronous and asynchronous APIs.

Requirements

Python 3.10 or higher.

Installation

From PyPI

pip install llama-cpp-py

From source

git clone https://github.com/sergey21000/llama-cpp-py
cd llama-cpp-py
pip install -e .

Using UV

uv pip install llama-cpp-py

Quick Start

More examples in the Google Colab notebook

1. Set up environment file for llama.cpp

Creating an .llama.env file with variables for llama.cpp server

# download example env file
wget https://github.com/sergey21000/llama-cpp-py/raw/main/.llama.env

# or create manually
nano .llama.env

See example .llama.env

2. Launch the server and send requests

Launching a synchronous server based on the latest llama.cpp release version

import os
from dotenv import load_dotenv
from openai import OpenAI
from llama_cpp_py import LlamaSyncServer


# environment variables for llama.cpp
load_dotenv(dotenv_path='.llama.env')

# auto-download last release and start server
# set verbose=True to display server logs
server = LlamaSyncServer()
server.start(verbose=True)


# sending requests with OpenAI client
client = OpenAI(
	base_url=server.server_url + '/v1',
	api_key='sk-no-key-required',
)
response = client.chat.completions.create(
    model='local',
    messages=[{'role': 'user', 'content': 'Hello!'}]
)

# stopping the server
server.stop()

Launching an asynchronous server based on a specific release version

import os
import asyncio
from openai import AsyncOpenAI
from dotenv import load_dotenv
from llama_cpp_py import LlamaAsyncServer, LlamaReleaseManager


# environment variables for llama.cpp
load_dotenv(dotenv_path='.llama.env')

# a) download a release by a specific tag with the 'cuda' priority in the title
# set tag='latest' to use the latest llama.cpp release version
# optionally specify priority_patterns to prefer certain builds (e.g. 'cuda' or 'cpu')
release_manager = LlamaReleaseManager(tag='b6780', priority_patterns=['cuda'])

# b) or set a specific release url in zip format
# release_manager = LlamaReleaseManager(
#     release_zip_url='https://github.com/ggml-org/llama.cpp/releases/download/b6780/llama-b6780-bin-win-cuda-12.4-x64.zip'
# )

# c) or selecting the compiled directory llama.cpp
# release_manager = LlamaReleaseManager(release_dir='/content/llama.cpp/build/bin')
	
async def main():
    # start llama.cpp server (set verbose=True to display server logs)
    llama_server = LlamaAsyncServer(verbose=False, release_manager=release_manager)
    await llama_server.start()

    # sending requests with OpenAI client
    client = AsyncOpenAI(
        base_url=f'{llama_server.server_url}/v1',
        api_key='sk-no-key-required',
    )
    stream_response = await client.chat.completions.create(
        model='local',
        messages=[{'role': 'user', 'content': 'How are you?'}],
        stream=True,
        temperature=0.8,
        max_tokens=-1,
        extra_body=dict(
            top_k=40,
            reasoning_format='none',
            chat_template_kwargs=dict(
                enable_thinking=True,
            ),
        ),
    )
    full_response = ''
    async for chunk in stream_response:
        if (token := chunk.choices[0].delta.content) is not None:
            full_response += token
            print(token, end='', flush=True)

    # stopping the server
    await llama_server.stop()

if __name__ == '__main__':
    asyncio.run(main())

Use Context manager

import os
from openai import AsyncOpenAI
from llama_cpp_py import LlamaAsyncServer

os.environ['LLAMA_ARG_MODEL_URL'] = 'https://huggingface.co/bartowski/google_gemma-3-4b-it-GGUF/resolve/main/google_gemma-3-4b-it-Q4_K_S.gguf'

async with LlamaAsyncServer() as server:
    client = AsyncOpenAI(
        base_url=f'{server.server_url}/v1',
        api_key='sk-no-key-required',
    )
    stream_response = await client.chat.completions.create(
        model='local',
        messages=[{'role': 'user', 'content': 'Hello!'}],
        stream=True,
    )
    full_response = ''
    async for chunk in stream_response:
        if (token := chunk.choices[0].delta.content) is not None:
            full_response += token
            print(token, end='', flush=True)

Enviroment Variables

Environment variables for llama-cpp-py

# Server startup wait timeout in seconds.
# Increase if model loading takes a long time.
# (default: 300)
LLAMACPP_SERVER_TIMEOUT_WAIT=900

# llama.cpp release tag. If set to "latest", the most recent release will be downloaded.
# (default: "latest")
LLAMACPP_RELEASE_TAG=b7806

# Direct download link to the archive from the llama.cpp releases page.
# Takes higher priority than LLAMACPP_RELEASE_TAG.
# (default: "")
LLAMACPP_RELEASE_ZIP_URL=https://github.com/ggml-org/llama.cpp/releases/download/b7806/llama-b7806-bin-win-cuda-13.1-x64.zip

# Path to a precompiled llama.cpp directory.
# Takes the highest priority, overriding LLAMACPP_RELEASE_TAG and LLAMACPP_RELEASE_ZIP_URL.
# (default: "")
LLAMACPP_DIR="/content/llama.cpp/build/bin"

# Logging level for llama-cpp-py (uses loguru, default INFO).
# A separate global setup via logger.add() also works.
# (default: "")
LLAMACPP_LOG_LEVEL=DEBUG

# or set global loguru level
LOGURU_LEVEL=WARNING

[!NOTE] Function arguments override environment variables. For example:

server = LlamaSyncServer(llama_dir='/path/bin')

will take precedence over the LLAMACPP_DIR variable

Troubleshooting

If the server fails to start or behaves unexpectedly, check the following:

• Check that the model path or URL in .llama.env is correct
• Verify that the port is not already in use
• Try setting verbose=True to see server logs

llama_server = LlamaAsyncServer(verbose=True)

• Link to the llama.cpp release archive appropriate for your system via

LlamaReleaseManager(release_zip_url=url)

• Or use the path to the directory with the pre-compiled llama.cpp

LlamaReleaseManager(release_dir=path_to_binaries)

---

If the model is being downloaded from a URL and the server times out before it finishes loading, you can:

• Increase the startup timeout by setting the environment variable

import os
os.environ['TIMEOUT_WAIT_FOR_SERVER'] = 600  # default 300

(value is in seconds), or

• Pre-download the model manually and set its local path in

import os
os.environ['LLAMA_ARG_MODEL'] = 'C:\path\to\model.gguf'

---

llama.cpp binary releases are downloaded to:

• Windows

%LOCALAPPDATA%\llama-cpp-py\releases

• Linux

~/.local/share/llama-cpp-py/releases

• MacOS

~/Library/Application Support/llama-cpp-py/releases

See platformdirs examle output

Dependencies

• aiohttp - Asynchronous HTTP client, used to check llama.cpp server readiness and interact with the server in async mode.
• requests - Synchronous HTTP client, used to check llama.cpp server readiness and interact with the server in sync mode.
• tqdm - Progress bar utility, used to display download progress when fetching llama.cpp releases.
• openai-python - OpenAI-compatible client, used to provide an OpenAI-style API interface for the server.
• python-dotenv - Environment variable loader, used for configuration via .env files.
• platformdirs - Cross-platform directory management, used to determine cache and data storage locations.
• pillow - Image processing library, used for multimodal (vision) input support.
• loguru - logging