LocalAI

⚠️ This project has been renamed from llama-cli to LocalAI to reflect the fact that we are focusing on a fast drop-in OpenAI API rather on the CLI interface. We think that there are already many projects that can be used as a CLI interface already, for instance llama.cpp and gpt4all. If you are were using llama-cli for CLI interactions and want to keep using it, use older versions or please open up an issue - contributions are welcome!

LocalAI is a straightforward, drop-in replacement API compatible with OpenAI for local CPU inferencing, based on llama.cpp, gpt4all and ggml, including support GPT4ALL-J which is Apache 2.0 Licensed and can be used for commercial purposes.

• OpenAI compatible API
• Supports multiple-models
• Once loaded the first time, it keep models loaded in memory for faster inference
• Provides a simple command line interface that allows text generation directly from the terminal
• Support for prompt templates
• Doesn't shell-out, but uses C bindings for a faster inference and better performance. Uses go-llama.cpp and go-gpt4all-j.cpp.

Model compatibility

It is compatible with the models supported by llama.cpp and also GPT4ALL-J.

Note: You might need to convert older models to the new format, see here for instance to run gpt4all.

Usage

LocalAI comes by default as a container image. You can check out all the available images with corresponding tags here.

The easiest way to run LocalAI is by using docker-compose:

git clone https://github.com/go-skynet/LocalAI

cd LocalAI

# copy your models to models/
cp your-model.bin models/

# (optional) Edit the .env file to set things like context size and threads
# vim .env

# start with docker-compose
docker compose up -d --build

# Now API is accessible at localhost:8080
curl http://localhost:8080/v1/models
# {"object":"list","data":[{"id":"your-model.bin","object":"model"}]}

curl http://localhost:8080/v1/completions -H "Content-Type: application/json" -d '{
     "model": "your-model.bin",            
     "prompt": "A long time ago in a galaxy far, far away",
     "temperature": 0.7
   }'

Prompt templates

The API doesn't inject a default prompt for talking to the model. You have to use a prompt similar to what's described in the standford-alpaca docs: https://github.com/tatsu-lab/stanford_alpaca#data-release.

You can use a default template for every model present in your model path, by creating a corresponding file with the `.tmpl` suffix next to your model. For instance, if the model is called `foo.bin`, you can create a sibiling file, `foo.bin.tmpl` which will be used as a default prompt, for instance this can be used with alpaca:

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{{.Input}}

### Response:

See the prompt-templates directory in this repository for templates for most popular models.

API

LocalAI provides an API for running text generation as a service, that follows the OpenAI reference and can be used as a drop-in. The models once loaded the first time will be kept in memory.

Example of starting the API with `docker`:

docker run -p 8080:8080 -ti --rm quay.io/go-skynet/local-api:latest --models-path /path/to/models --context-size 700 --threads 4

And you'll see:

┌───────────────────────────────────────────────────┐ 
│                   Fiber v2.42.0                   │ 
│               http://127.0.0.1:8080               │ 
│       (bound on host 0.0.0.0 and port 8080)       │ 
│                                                   │ 
│ Handlers ............. 1  Processes ........... 1 │ 
│ Prefork ....... Disabled  PID ................. 1 │ 
└───────────────────────────────────────────────────┘ 

Note: Models have to end up with .bin so can be listed by the /models endpoint.

You can control the API server options with command line arguments:

local-api --models-path <model_path> [--address <address>] [--threads <num_threads>]

The API takes takes the following parameters:

Parameter	Environment Variable	Default Value	Description
models-path	MODELS_PATH		The path where you have models (ending with .bin).
threads	THREADS	CPU cores	The number of threads to use for text generation.
address	ADDRESS	:8080	The address and port to listen on.
context-size	CONTEXT_SIZE	512	Default token context size.

Once the server is running, you can start making requests to it using HTTP, using the OpenAI API.

Supported OpenAI API endpoints

You can check out the OpenAI API reference.

Following the list of endpoints/parameters supported.

Chat completions

For example, to generate a chat completion, you can send a POST request to the /v1/chat/completions endpoint with the instruction as the request body:

curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
     "model": "ggml-koala-7b-model-q4_0-r2.bin",
     "messages": [{"role": "user", "content": "Say this is a test!"}],
     "temperature": 0.7
   }'

Available additional parameters: top_p, top_k, max_tokens

Completions

For example, to generate a comletion, you can send a POST request to the /v1/completions endpoint with the instruction as the request body:

curl http://localhost:8080/v1/completions -H "Content-Type: application/json" -d '{
     "model": "ggml-koala-7b-model-q4_0-r2.bin",
     "prompt": "A long time ago in a galaxy far, far away",
     "temperature": 0.7
   }'

Available additional parameters: top_p, top_k, max_tokens

List models

You can list all the models available with:

curl http://localhost:8080/v1/models

Using other models

gpt4all (https://github.com/nomic-ai/gpt4all) works as well, however the original model needs to be converted (same applies for old alpaca models, too):

wget -O tokenizer.model https://huggingface.co/decapoda-research/llama-30b-hf/resolve/main/tokenizer.model
mkdir models
cp gpt4all.. models/
git clone https://gist.github.com/eiz/828bddec6162a023114ce19146cb2b82
pip install sentencepiece
python 828bddec6162a023114ce19146cb2b82/gistfile1.txt models tokenizer.model
# There will be a new model with the ".tmp" extension, you have to use that one!

Windows compatibility

It should work, however you need to make sure you give enough resources to the container. See https://github.com/go-skynet/LocalAI/issues/2

Kubernetes

You can run the API in Kubernetes, see an example deployment in kubernetes

Build locally

Pre-built images might fit well for most of the modern hardware, however you can and might need to build the images manually.

In order to build the LocalAI container image locally you can use docker:

# build the image
docker build -t LocalAI .
docker run LocalAI

Or build the binary with make:

make build

Short-term roadmap

• Mimic OpenAI API (https://github.com/go-skynet/LocalAI/issues/10)
• Binary releases (https://github.com/go-skynet/LocalAI/issues/6)
• Upstream our golang bindings to llama.cpp (https://github.com/ggerganov/llama.cpp/issues/351)
• Multi-model support
• Have a webUI!

License

MIT

Acknowledgements

• llama.cpp
• https://github.com/tatsu-lab/stanford_alpaca
• https://github.com/cornelk/llama-go for the initial ideas
• https://github.com/antimatter15/alpaca.cpp for the light model version (this is compatible and tested only with that checkpoint model!)