LlamaChat is a macOS app that allows you to chat with LLaMa, Alpaca and GPT4All models all running locally on your Mac.
LlamaChat requires macOS 13 Ventura, and either an Intel or Apple Silicon processor.
Download a .dmg
containing the latest version ๐ here ๐.
git clone https://github.com/alexrozanski/LlamaChat.git
cd LlamaChat
open LlamaChat.xcodeproj
NOTE: model inference runs really slowly in Debug builds, so if building from source make sure that the Build Configuration
in LlamaChat > Edit Scheme... > Run
is set to Release
.
- Supported Models: LlamaChat supports LLaMA, Alpaca and GPT4All models out of the box. Support for other models including Vicuna and Koala is coming soon. We are also looking for Chinese and French speakers to add support for Chinese LLaMA/Alpaca and Vigogne.
- Flexible Model Formats: LLamaChat is built on top of llama.cpp and llama.swift. The app supports adding LLaMA models in either their raw
.pth
PyTorch checkpoints form or the.ggml
format. - Model Conversion: If raw PyTorch checkpoints are added these can be converted to
.ggml
files compatible with LlamaChat and llama.cpp within the app. - Chat History: Chat history is persisted within the app. Both chat history and model context can be cleared at any time.
- Funky Avatars: LlamaChat ships with 7 funky avatars that can be used with your chat sources.
- Advanced Source Naming: LlamaChat uses Special Magicโข to generate playful names for your chat sources.
- Context Debugging: For the keen ML enthusiasts, the current model context can be viewed for a chat in the info popover.
NOTE: LlamaChat doesn't ship with any model files and requires that you obtain these from the respective sources in accordance with their respective terms and conditions.
- Model formats: LlamaChat allows you to use the LLaMA family of models in either their raw Python checkpoint form (
.pth
) or pre-converted.ggml
file (the format used by llama.cpp, which powers LlamaChat). - Using LLaMA models: When importing LLaMA models in the
.pth
format:- You should select the appropriate parameter size directory (e.g.
7B
,13B
etc) in the conversion flow, which includes theconsolidated.NN.pth
andparams.json
files. - As per the LLaMA model release, the parent directory should contain
tokenizer.model
. E.g. to use the LLaMA-13B model, your model directory should look something like the below, and you should select the13B
directory:
- You should select the appropriate parameter size directory (e.g.
.
โ ...
โโโ 13B
โ โโโ checklist.chk.txt
โ โโโ consolidated.00.pth
โ โโโ consolidated.01.pth
โ โโโ params.json
โ ...
โโโ tokenizer.model
- Troubleshooting: If using
.ggml
files, make sure these are up-to-date. If you run into problems, you may need to use the conversion scripts from llama.cpp:- For the GPT4All model, you may need to use convert-gpt4all-to-ggml.py
- For the Alpaca model, you may need to use convert-unversioned-ggml-to-ggml.py
- You may also need to use migrate-ggml-2023-03-30-pr613.py as well. For more information check out the llama.cpp repo.
Pull Requests and Issues are welcome and much appreciated. Please make sure to adhere to the Code of Conduct at all times.
LlamaChat is fully built using Swift and SwiftUI, and makes use of llama.swift under the hood to run inference and perform model operations.
The project is mostly built using MVVM and makes heavy use of Combine and Swift Concurrency.
LlamaChat is licensed under the MIT license.