Awesome
<h1 align="center"> mistral.rs </h1> <h3 align="center"> Blazingly fast LLM inference. </h3> <p align="center"> | <a href="https://ericlbuehler.github.io/mistral.rs/mistralrs/"><b>Rust Documentation</b></a> | <a href="https://github.com/EricLBuehler/mistral.rs/blob/master/mistralrs-pyo3/API.md"><b>Python Documentation</b></a> | <a href="https://discord.gg/SZrecqK8qw"><b>Discord</b></a> | <a href="https://matrix.to/#/#mistral.rs:matrix.org"><b>Matrix</b></a> | </p>Mistral.rs 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 🚀
-
Deploy with our easy to use APIs
Quick examples
After following installation instructions
-
🔥🧠AnyMoE: Build a memory-efficient MoE model from anything, in seconds
./mistralrs_server -i toml -f toml-selectors/anymoe_lora.toml
-
🦙 Run the Llama 3.1 model
./mistralrs_server -i plain -m meta-llama/Meta-Llama-3.1-8B-Instruct -a llama
-
φ³ Run the new Phi 3.5/3.1/3 model with 128K context window
./mistralrs_server -i plain -m microsoft/Phi-3.5-mini-instruct -a phi3
-
🌀 Run the Phi 3.5 MoE model with 128K context window: documentation and guide here
./mistralrs_server -i plain -m microsoft/Phi-3.5-MoE-instruct -a phi3.5moe
-
φ³ 📷 Run the Phi 3 vision model: documentation and guide here
<img src="https://www.nhmagazine.com/content/uploads/2019/05/mtwashingtonFranconia-2-19-18-108-Edit-Edit.jpg" alt="Mount Washington" width = "400" height = "267"> <h6><a href = "https://www.nhmagazine.com/mount-washington/">Credit</a></h6>./mistralrs_server --port 1234 vision-plain -m microsoft/Phi-3.5-vision-instruct -a phi3v
-
Other models: see a support matrix and how to run them
Mistal.rs supports several model categories:
- text
- vision (see the docs)
Description
Fast:
- Apple silicon support with the Metal framework.
- CPU inference with
mkl
,accelerate
support and optimized backend. - CUDA support with flash attention and cuDNN.
- Continuous batching and PagedAttention support.
- Prefix caching.
- Device mapping: load and run some layers on the device and the rest on the CPU.
Quantization:
- Details
- GGML: 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit, with ISQ support.
- GPTQ: 2-bit, 3-bit, 4-bit and 8-bit
- HQQ: 4-bit and 8 bit, with ISQ support
- 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.
- Extremely fast due to working in parallel
Easy:
- Lightweight OpenAI API compatible HTTP server.
- Python API.
- Grammar support with Regex and Yacc.
- Use a simple model topology to configure ISQ and device mapping for per-layer with a single YAML file (examples here)
Powerful:
- 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
- AnyMoE: Build a memory-efficient MoE model from anything, in seconds
- PagedAttention: docs
- Various sampling and penalty techniques:
- Top K
- Top P
- Min P
- Dry Penalty
- Frequency and Presence Penalty
- Please suggest more by raising an issue!
- Tool calling: docs
- Prompt chunking (only without PagedAttention for now): handle larger prompts where the activation size would cause an OOM by sending chunks
- Custom logits processor API in Rust: example
This is a demo of interactive mode with streaming running Phi 3 128k mini with quantization via ISQ to Q4K.
<!-- Mistral GGUF demo, old API --> <!-- https://github.com/EricLBuehler/mistral.rs/assets/65165915/3396abcd-8d44-4bf7-95e6-aa532db09415 -->https://github.com/EricLBuehler/mistral.rs/assets/65165915/09d9a30f-1e22-4b9a-9006-4ec6ebc6473c
Support matrix
Note: See supported models for more information
Model | Supports quantization | Supports adapters | Supports device mapping | Supported by AnyMoE |
---|---|---|---|---|
Mistral v0.1/v0.2/v0.3 | ✅ | ✅ | ✅ | ✅ |
Gemma | ✅ | ✅ | ✅ | ✅ |
Llama 2/3 | ✅ | ✅ | ✅ | ✅ |
Mixtral | ✅ | ✅ | ✅ | |
Phi 2 | ✅ | ✅ | ✅ | ✅ |
Phi 3 | ✅ | ✅ | ✅ | ✅ |
Phi 3.5 MoE | ✅ | ✅ | ||
Qwen 2 | ✅ | ✅ | ✅ | |
Phi 3 Vision | ✅ | ✅ | ✅ | |
Idefics 2 | ✅ | ✅ | ✅ | |
Gemma 2 | ✅ | ✅ | ✅ | ✅ |
Starcoder 2 | ✅ | ✅ | ✅ | ✅ |
LLaVa Next | ✅ | ✅ | ✅ | |
LLaVa | ✅ | ✅ | ✅ |
APIs and Integrations
Rust Crate
Rust multithreaded/async API for easy integration into any application.
- Docs
- Examples
- To install: Add
mistralrs = { git = "https://github.com/EricLBuehler/mistral.rs.git" }
Python API
Python API for mistral.rs.
HTTP Server
OpenAI API compatible API server
Llama Index integration (Python)
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
- Enable with
- Metal:
- Enable with
metal
feature:--features metal
- Enable with
- CPU:
- Intel MKL with
mkl
feature:--features mkl
- Apple Accelerate with
accelerate
feature:--features accelerate
- Intel MKL with
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"
.
Benchmarks
Device | Mistral.rs Completion T/s | Llama.cpp Completion T/s | Model | Quant |
---|---|---|---|---|
A10 GPU, CUDA | 86 | 83 | mistral-7b | 4_K_M |
Intel Xeon 8358 CPU, AVX | 11 | 23 | mistral-7b | 4_K_M |
Raspberry Pi 5 (8GB), Neon | 2 | 3 | mistral-7b | 2_K |
A100 GPU, CUDA | 131 | 134 | mistral-7b | 4_K_M |
RTX 6000 GPU, CUDA | 103 | 96 | mistral-7b | 4_K_M |
Note: All CUDA tests for mistral.rs conducted with PagedAttention enabled, block size = 32
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: https://docs.docker.com/engine/reference/run/
Note: You can use pre-built
mistralrs-server
binaries here
- Install the Python package here.
-
Install required packages
OpenSSL
(Example on Ubuntu:sudo apt install libssl-dev
)- <b>Linux only:</b>
pkg-config
(Example on Ubuntu:sudo apt install pkg-config
)
-
Install Rust: https://rustup.rs/
Example on Ubuntu:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh source $HOME/.cargo/env
-
<b>Optional:</b> 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
- Note: you can install
-
Download the code
git clone https://github.com/EricLBuehler/mistral.rs.git cd mistral.rs
-
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 usagePass the same values to
--features
as you would forcargo build
cargo install --path mistralrs-server --features cuda
-
-
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
-
Use our APIs and integrations
Getting models
There are 2 ways to run a model with mistral.rs:
- From Hugging Face Hub (easiest)
- From local files
- Running a GGUF model fully locally
Getting models from Hugging Face Hub
Mistral.rs 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 literalenv:<value>
: Load from a specified environment variablepath:<value>
: Load from a specified filecache
: 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 mistral.rs to load models fully locally by modifying the *_model_id
arguments or options:
./mistralrs_server --port 1234 plain -m . -a mistral
Throughout mistral.rs, 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) ormodel_id
(python/rust) or--tok-model-id
(server) ortok_model_id
(python/rust):config.json
tokenizer_config.json
tokenizer.json
(if not specified separately).safetensors
/.bin
/.pth
/.pt
files (defaults to.safetensors
)preprocessor_config.json
(required for vision models).processor_config.json
(optional for vision models).
--quantized-model-id
(server) orquantized_model_id
(python/rust):- Specified
.gguf
or.ggml
file.
- Specified
--x-lora-model-id
(server) orxlora_model_id
(python/rust):xlora_classifier.safetensors
xlora_config.json
- Adapters
.safetensors
andadapter_config.json
files in their respective directories
--adapters-model-id
(server) oradapters_model_id
(python/rust):- Adapters
.safetensors
andadapter_config.json
files in their respective directories
- Adapters
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.
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
Mistral.rs 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
orauto
to choose based on the device. This is specified in the--dype
/-d
parameter after the model architecture (plain
).
If you do not specify the architecture, an attempt will be made to use the model's config. If this fails, please raise an issue.
mistral
gemma
mixtral
llama
phi2
phi3
phi3.5moe
qwen2
gemma2
starcoder2
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
orauto
to choose based on the device. This is specified in the--dype
/-d
parameter after the model architecture (vision-plain
).
phi3v
idefics2
llava_next
llava
Supported GGUF architectures
Plain:
llama
phi2
phi3
starcoder2
With adapters:
llama
phi3
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
Interactive mode for vision models:
You can launch interactive mode for vision models, a simple chat application running in the terminal, by passing -i
:
./mistralrs_server --vi plain -m microsoft/Phi-3.5-vision-instruct -a phi3v
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.
Example:
./mistralrs_server --port 1234 toml -f toml-selectors/gguf.toml
Supported models
Quantization support
Model | GGUF | GGML | ISQ |
---|---|---|---|
Mistral | ✅ | ✅ | |
Gemma | ✅ | ||
Llama | ✅ | ✅ | ✅ |
Mixtral | ✅ | ✅ | |
Phi 2 | ✅ | ✅ | |
Phi 3 | ✅ | ✅ | |
Phi 3.5 MoE | ✅ | ||
Qwen 2 | ✅ | ||
Phi 3 Vision | ✅ | ||
Idefics 2 | ✅ | ||
Gemma 2 | ✅ | ||
Starcoder 2 | ✅ | ✅ | |
LLaVa Next | ✅ | ||
LLaVa | ✅ |
Device mapping support
Model category | Supported |
---|---|
Plain | ✅ |
GGUF | ✅ |
GGML | |
Vision Plain | ✅ |
X-LoRA and LoRA support
Model | X-LoRA | X-LoRA+GGUF | X-LoRA+GGML |
---|---|---|---|
Mistral | ✅ | ✅ | |
Gemma | ✅ | ||
Llama | ✅ | ✅ | ✅ |
Mixtral | ✅ | ✅ | |
Phi 2 | ✅ | ||
Phi 3 | ✅ | ✅ | |
Phi 3.5 MoE | |||
Qwen 2 | |||
Phi 3 Vision | |||
Idefics 2 | |||
Gemma 2 | ✅ | ||
Starcoder 2 | ✅ | ||
LLaVa Next | |||
LLaVa |
AnyMoE support
Model | AnyMoE |
---|---|
Mistral 7B | ✅ |
Gemma | ✅ |
Llama | ✅ |
Mixtral | |
Phi 2 | ✅ |
Phi 3 | ✅ |
Phi 3.5 MoE | |
Qwen 2 | ✅ |
Phi 3 Vision | |
Idefics 2 | |
Gemma 2 | ✅ |
Starcoder 2 | ✅ |
LLaVa Next | ✅ |
LLaVa | ✅ |
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
Mistral.rs 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.
Contributing
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.
FAQ
- 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
ormistralrs_ggml_tensors.txt
- More logging.
- If loading a GGUF or GGML model, this will output a file containing the names, shapes, and types of each tensor.
- Setting the CUDA compiler path:
- Set the
NVCC_CCBIN
environment variable during build.
- Set the
- 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
- Some Linux distributions require compiling with
- 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
orauto
to choose based on the device.
- For non-quantized models, you can specify the data type to load and run in. This must be one of
Credits
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.