EricLBuehler /

Blazingly fast LLM inference.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Blazingly fast LLM inference.

| Rust Documentation | Python Documentation | Discord | is a fast LLM inference platform supporting inference on a variety of devices, quantization, and easy-to-use application with an Open-AI API compatible HTTP server and Python bindings.

Please submit requests for new models here.

Get started fast πŸš€

  1. Install

  2. Get models

  3. Deploy with our easy to use APIs

Quick examples

  • πŸ¦™ Run the Llama 3 model

    After following installation instructions

    ./mistralrs_server -i plain -m meta-llama/Meta-Llama-3-8B-Instruct -a llama
  • φ³ Run the Phi 3 model with 128K context window

    After following installation instructions

    ./mistralrs_server -i plain -m microsoft/Phi-3-mini-128k-instruct -a phi3
  • φ³ πŸ“· Run the Phi 3 vision model: documentation and guide here

    Mount Washington

    After following installation instructions

    ./mistralrs_server --port 1234 vision-plain -m microsoft/Phi-3-vision-128k-instruct -a phi3v
  • Other models: see supported models and how to run them



  • Quantized model support: 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit for faster inference and optimized memory usage.
  • Continuous batching.
  • Prefix caching.
  • Device mapping: load and run some layers on the device and the rest on the CPU.

Accelerator support:

  • Apple silicon support with the Metal framework.
  • CPU inference with mkl, accelerate support and optimized backend.
  • CUDA support with flash attention and cuDNN.


  • Lightweight OpenAI API compatible HTTP server.
  • Python API.
  • Grammar support with Regex and Yacc.
  • ISQ (In situ quantization): run .safetensors models directly from Hugging Face Hub by quantizing them after loading instead of creating a GGUF file.
    • This loads the ISQ-able weights on CPU before quantizing with ISQ and then moving to the device to avoid memory spikes.
    • Provides methods to further reduce memory spikes.


  • Fast LoRA support with weight merging.
  • First X-LoRA inference platform with first class support.
  • Speculative Decoding: Mix supported models as the draft model or the target model
  • Dynamic LoRA adapter swapping at runtime with adapter preloading: examples and docs

This is a demo of interactive mode with streaming running Mistral GGUF:


Supported models:

  • Mistral 7B (v0.1 and v0.2)
  • Gemma
  • Llama, including Llama 3
  • Mixtral 8x7B
  • Phi 2
  • Phi 3
  • Qwen 2

Please see this section for details on quantization and LoRA support.

APIs and Integrations

Rust Crate

Rust multithreaded/async API for easy integration into any application.

  • Docs
  • Examples
  • To install: Add mistralrs = { git = "" }
Python API

Python API for

from mistralrs import Runner, Which, ChatCompletionRequest

runner = Runner(

res = runner.send_chat_completion_request(
        messages=[{"role":"user", "content":"Tell me a story about the Rust type system."}],
HTTP Server

OpenAI API compatible API server

Llama Index integration

Supported accelerators

  • CUDA:
    • Enable with cuda feature: --features cuda
    • Flash attention support with flash-attn feature, only applicable to non-quantized models: --features flash-attn
    • cuDNNsupport with cudnn feature: --features cudnn
  • Metal:
    • Enable with metal feature: --features metal
  • CPU:
    • Intel MKL with mkl feature: --features mkl
    • Apple Accelerate with accelerate feature: --features accelerate

Enabling features is done by passing --features ... to the build system. When using cargo run or maturin develop, pass the --features flag before the -- separating build flags from runtime flags.

  • To enable a single feature like metal: cargo build --release --features metal.
  • To enable multiple features, specify them in quotes: cargo build --release --features "cuda flash-attn cudnn".


Device Completion T/s Llama.cpp Completion T/s Model Quant
A10 GPU, CUDA 78 78 mistral-7b 4_K_M
Intel Xeon 8358 CPU, AVX 6 19 mistral-7b 4_K_M
Raspberry Pi 5 (8GB), Neon 2 3 mistral-7b 2_K
A100 GPU, CUDA 119 119 mistral-7b 4_K_M

Please submit more benchmarks via raising an issue!

Installation and Build

Note: You can use our Docker containers here. Learn more about running Docker containers:

  1. Install required packages

    • OpenSSL (Example on Ubuntu: sudo apt install libssl-dev)
    • Linux only: pkg-config (Example on Ubuntu: sudo apt install pkg-config)
  2. Install Rust:

    Example on Ubuntu:

    curl --proto '=https' --tlsv1.2 -sSf | sh
    source $HOME/.cargo/env
  3. Optional: 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.)

    • Note: you can install huggingface-cli as documented here.
    huggingface-cli login
  4. Download the code

    git clone
  5. Build or install

    • Base build command

      cargo build --release
    • Build with CUDA support

      cargo build --release --features cuda
    • Build with CUDA and Flash Attention V2 support

      cargo build --release --features "cuda flash-attn"
    • Build with Metal support

      cargo build --release --features metal
    • Build with Accelerate support

      cargo build --release --features accelerate
    • Build with MKL support

      cargo build --release --features mkl
    • Install with cargo install for easy command line usage

      Pass the same values to --features as you would for cargo build

      cargo install --path mistralrs-server --features cuda
  6. The build process will output a binary misralrs-server at ./target/release/mistralrs-server which may be copied into the working directory with the following command:

    Example on Ubuntu:

    cp ./target/release/mistralrs-server ./mistralrs_server
  7. Installing Python support

    You can install Python support by following the guide here.

Getting models

There are 2 ways to run a model with

  • From Hugging Face Hub (easiest)
  • From local files
    • Running a GGUF model fully locally

Getting models from Hugging Face Hub can automatically download models from HF Hub. To access gated models, you should provide a token source. They may be one of:

  • literal:<value>: Load from a specified literal
  • env:<value>: Load from a specified environment variable
  • path:<value>: Load from a specified file
  • cache: default: Load from the HF token at ~/.cache/huggingface/token or equivalent.
  • none: Use no HF token

This is passed in the following ways:

  • Command line:
./mistralrs_server --token-source none -i plain -m microsoft/Phi-3-mini-128k-instruct -a phi3
  • Python:

Here is an example of setting the token source.

If token cannot be loaded, no token will be used (i.e. effectively using none).

Loading models from local files:

You can also instruct to load models fully locally by modifying the *_model_id arguments or options:

./mistralrs_server --port 1234 plain -m . -a mistral

Throughout, any model ID argument or option may be a local path and should contain the following files for each model ID option:

  • --model-id (server) or model_id (python/rust) or --tok-model-id (server) or tok_model_id (python/rust):
    • config.json
    • tokenizer_config.json
    • tokenizer.json (if not specified separately)
    • .safetensors files.
  • --quantized-model-id (server) or quantized_model_id (python/rust):
    • Specified .gguf or .ggml file.
  • --x-lora-model-id (server) or xlora_model_id (python/rust):
    • xlora_classifier.safetensors
    • xlora_config.json
    • Adapters .safetensors and adapter_config.json files in their respective directories
  • --adapters-model-id (server) or adapters_model_id (python/rust):
    • Adapters .safetensors and adapter_config.json files in their respective directories

Running GGUF models locally

To run GGUF models fully locally, the only mandatory arguments are the quantized model ID and the quantized filename.

Chat template

The chat template can be automatically detected and loaded from the GGUF file if no other chat template source is specified including the tokenizer model ID.

you do not need to specify the tokenizer model ID argument and instead should pass a path to the chat template JSON file (examples here, you will need to create your own by specifying the chat template and bos/eos tokens) as well as specifying a local model ID. For example:

./mistralrs-server --chat-template <chat_template> gguf -m . -f Phi-3-mini-128k-instruct-q4_K_M.gguf

If you do not specify a chat template, then the --tok-model-id/-t tokenizer model ID argument is expected where the tokenizer_config.json file should be provided. If that model ID contains a tokenizer.json, then that will be used over the GGUF tokenizer.


The following tokenizer model types are currently supported. If you would like one to be added, please raise an issue. Otherwise, please consider using the method demonstrated in examples below, where the tokenizer is sourced from Hugging Face.

Supported GGUF tokenizer types

  • llama (sentencepiece)
  • gpt2 (BPE)

Run with the CLI uses subcommands to control the model type. They are generally of format <XLORA/LORA>-<QUANTIZATION>. Please run ./mistralrs_server --help to see the subcommands.

Additionally, for models without quantization, the model architecture should be provided as the --arch or -a argument in contrast to GGUF models which encode the architecture in the file.

Architecture for plain models

Note: for plain models, you can specify the data type to load and run in. This must be one of f32, f16, bf16 or auto to choose based on the device. This is specified in the --dype/-d parameter after the model architecture (plain).

  • mistral
  • gemma
  • mixtral
  • llama
  • phi2
  • phi3
  • qwen2

Architecture for vision models

Note: for vision models, you can specify the data type to load and run in. This must be one of f32, f16, bf16 or auto to choose based on the device. This is specified in the --dype/-d parameter after the model architecture (vision-plain).

  • phi3v

Interactive mode:

You can launch interactive mode, a simple chat application running in the terminal, by passing -i:

./mistralrs_server -i plain -m microsoft/Phi-3-mini-128k-instruct -a phi3

More quick examples:

  • X-LoRA with no quantization

To start an X-LoRA server with the exactly as presented in the paper:

./mistralrs_server --port 1234 x-lora-plain -o orderings/xlora-paper-ordering.json -x lamm-mit/x-lora
  • LoRA with a model from GGUF

To start an LoRA server with adapters from the X-LoRA paper (you should modify the ordering file to use only one adapter, as the adapter static scalings are all 1 and so the signal will become distorted):

./mistralrs_server --port 1234 lora-gguf -o orderings/xlora-paper-ordering.json -m TheBloke/zephyr-7B-beta-GGUF -f zephyr-7b-beta.Q8_0.gguf -a lamm-mit/x-lora

Normally with a LoRA model you would use a custom ordering file. However, for this example we use the ordering from the X-LoRA paper because we are using the adapters from the X-LoRA paper.

  • With a model from GGUF

To start a server running Mistral from GGUF:

./mistralrs_server --port 1234 gguf -t mistralai/Mistral-7B-Instruct-v0.1 -m TheBloke/Mistral-7B-Instruct-v0.1-GGUF -f mistral-7b-instruct-v0.1.Q4_K_M.gguf
  • With a model from GGML

To start a server running Llama from GGML:

./mistralrs_server --port 1234 ggml -t meta-llama/Llama-2-13b-chat-hf -m TheBloke/Llama-2-13B-chat-GGML -f llama-2-13b-chat.ggmlv3.q4_K_M.bin
  • Plain model from safetensors

To start a server running Mistral from safetensors.

./mistralrs_server --port 1234 plain -m mistralai/Mistral-7B-Instruct-v0.1 -a mistral

Structured selection with a .toml file

We provide a method to select models with a .toml file. The keys are the same as the command line, with no_kv_cache and tokenizer_json being "global" keys.


./mistralrs_server --port 1234 toml -f toml-selectors/gguf.toml

Supported models supports several model categories:

Quantization support

Mistral 7B βœ…
Llama βœ… βœ…
Mixtral 8x7B βœ…
Phi 2 βœ…
Phi 3 βœ…
Qwen 2
Phi 3 Vision

Device mapping support

Model Supported
Plain βœ…
GGUF βœ…
Vision Plain

X-LoRA and LoRA support

Mistral 7B βœ… βœ…
Gemma βœ…
Llama βœ… βœ… βœ…
Mixtral 8x7B βœ… βœ…
Phi 2 βœ…
Phi 3 βœ… βœ…
Qwen 2
Phi 3 Vision

Using derivative model

To use a derivative model, select the model architecture using the correct subcommand. To see what can be passed for the architecture, pass --help after the subcommand. For example, when using a different model than the default, specify the following for the following types of models:

  • Plain: Model id
  • Quantized: Quantized model id, quantized filename, and tokenizer id
  • X-LoRA: Model id, X-LoRA ordering
  • X-LoRA quantized: Quantized model id, quantized filename, tokenizer id, and X-LoRA ordering
  • LoRA: Model id, LoRA ordering
  • LoRA quantized: Quantized model id, quantized filename, tokenizer id, and LoRA ordering
  • Vision Plain: Model id

See this section to determine if it is necessary to prepare an X-LoRA/LoRA ordering file, it is always necessary if the target modules or architecture changed, or if the adapter order changed.

It is also important to check the chat template style of the model. If the HF hub repo has a tokenizer_config.json file, it is not necessary to specify. Otherwise, templates can be found in chat_templates and should be passed before the subcommand. If the model is not instruction tuned, no chat template will be found and the APIs will only accept a prompt, no messages.

For example, when using a Zephyr model:

./mistralrs_server --port 1234 --log output.txt gguf -t HuggingFaceH4/zephyr-7b-beta -m TheBloke/zephyr-7B-beta-GGUF -f zephyr-7b-beta.Q5_0.gguf

Adapter model support: X-LoRA and LoRA

An adapter model is a model with X-LoRA or LoRA. X-LoRA support is provided by selecting the x-lora-* architecture, and LoRA support by selecting the lora-* architecture. Please find docs for adapter models here

Chat Templates and Tokenizer will attempt to automatically load a chat template and tokenizer. This enables high flexibility across models and ensures accurate and flexible chat templating. However, this behavior can be customized. Please find detailed documentation here.


Thank you for contributing! If you have any problems or want to contribute something, please raise an issue or pull request. If you want to add a new model, please contact us via an issue and we can coordinate how to do this.


  • Debugging with the environment variable MISTRALRS_DEBUG=1 causes the following things
    • If loading a GGUF or GGML model, this will output a file containing the names, shapes, and types of each tensor.
      • mistralrs_gguf_tensors.txt or mistralrs_ggml_tensors.txt
    • More logging.
  • Setting the CUDA compiler path:
    • Set the NVCC_CCBIN environment variable during build.
  • Error: recompile with -fPIE:
    • Some Linux distributions require compiling with -fPIE.
    • Set the CUDA_NVCC_FLAGS environment variable to -fPIE during build: CUDA_NVCC_FLAGS=-fPIE
  • Error CUDA_ERROR_NOT_FOUND or symbol not found when using a normal or vison model:
    • For non-quantized models, you can specify the data type to load and run in. This must be one of f32, f16, bf16 or auto to choose based on the device.


This project would not be possible without the excellent work at candle. Additionally, thank you to all contributors! Contributing can range from raising an issue or suggesting a feature to adding some new functionality.


Blazingly fast LLM inference.

License:MIT License


Language:Rust 97.9%Language:Python 1.3%Language:Cuda 0.6%Language:Dockerfile 0.1%