mistralrs-cuda 0.1.5

Fast and easy LLM serving.

mistral.rs PyO3 Bindings: mistralrs

mistralrs is a Python package which provides an API for mistral.rs. We build mistralrs with the maturin build manager.

Installation from PyPi

0. Install Rust: https://rustup.rs/

   curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
   source $HOME/.cargo/env
   

1. mistralrs depends on the openssl library.

To install it on Ubuntu:

sudo apt install libssl-dev
sudo apt install pkg-config

2. Install it!

• CUDA

  pip install mistralrs-cuda

• Metal

  pip install mistralrs-metal

• Apple Accelerate

  pip install mistralrs-accelerate

• Intel MKL

  pip install mistralrs-mkl

• Without accelerators

  pip install mistralrs

All installations will install the mistralrs package. The suffix on the package installed by pip only controls the feature activation.

Installation from source

1. Install required packages

   • openssl (ex., sudo apt install libssl-dev)
   • pkg-config (ex., sudo apt install pkg-config)

2. Install Rust: https://rustup.rs/

   curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
   source $HOME/.cargo/env
   

3. Set HF token correctly (skip if already set or your model is not gated, or if you want to use the token_source parameters in Python or the command line.)

   mkdir ~/.cache/huggingface
   touch ~/.cache/huggingface/token
   echo <HF_TOKEN_HERE> > ~/.cache/huggingface/token
   

4. Download the code

   git clone https://github.com/EricLBuehler/mistral.rs.git
   cd mistral.rs
   

5. cd into the correct directory for building mistralrs: cd mistralrs-pyo3

6. Install maturin, our Rust + Python build system: Maturin requires a Python virtual environment such as venv or conda to be active. The mistralrs package will be installed into that environment.

   pip install maturin[patchelf]
   

7. Install mistralrs Install mistralrs by executing the following in this directory where features such as cuda or flash-attn may be specified with the --features argument just like they would be for cargo run.

   The base build command is:

   maturin develop -r
   

   • To build for CUDA:

   maturin develop -r --features cuda
   

   • To build for CUDA with flash attention:

   maturin develop -r --features "cuda flash-attn"
   

   • To build for Metal:

   maturin develop -r --features metal
   

   • To build for Accelerate:

   maturin develop -r --features accelerate
   

   • To build for MKL:

   maturin develop -r --features mkl
   

Please find API docs here and the type stubs here, which are another great form of documentation.

We also provide a cookbook here!

Example

from mistralrs import ModelKind, MistralLoader, ChatCompletionRequest

kind = ModelKind.QuantizedGGUF
loader = MistralLoader(
    model_id="mistralai/Mistral-7B-Instruct-v0.1",
    kind=kind,
    no_kv_cache=False,
    repeat_last_n=64,
    quantized_model_id="TheBloke/Mistral-7B-Instruct-v0.1-GGUF",
    quantized_filename="mistral-7b-instruct-v0.1.Q4_K_M.gguf",
)
runner = loader.load()
res = runner.send_chat_completion_request(
    ChatCompletionRequest(
        model="mistral",
        messages=[
            {"role": "user", "content": "Tell me a story about the Rust type system."}
        ],
        max_tokens=256,
        frequency_penalty=1.0,
        top_p=0.1,
        temperature=0.1,
    )
)
print(res)