SiLLM simplifies the process of training and running Large Language Models (LLMs) on Apple Silicon by leveraging the MLX framework. Building upon the foundation provided by MLX Examples, this project introduces additional features specifically designed to enhance LLM operations with MLX in a streamlined package.

• LLM Loading: load LLMs for chat and training in different formats (Huggingface, Torch, GGUF, MLX)
• LoRA Training: train LLMs using Low-rank Adaptation
• DPO Training: train LLMs with Direct Preference Optimization

• Web app for a seamless chat experience running on local hardware
• API server with OpenAI compatible chat endpoints
• Model architectures: Llama, Mistral, Mixtral, Phi-2, Gemma, Qwen2, Starcoder2, DBRX, Cohere Command-R
• Conversation templates: llama-2, chatml, alpaca, vicuna, gemma, phi, openchat
• Loss functions for DPO: sigmoid, hinge, IPO, DPOP
• Training loss plots using matplotlib
• Perplexity calculation

Using pip:

pip install sillm-mlx

The web app uses Chainlit to provide a frontend for conversational AI running locally on Apple Silicon hardware.

To start the web app, clone the repository and start the app using chainlit:

git clone https://github.com/armbues/SiLLM.git
cd SiLLM/app
pip install -r requirements.txt
python -m chainlit run app.py -w

Set the environment variables SILLM_MODEL_DIR and SILLM_ADAPTER_DIR to load local models/adapters.

Run the CLI scripts with the argument -h to see a print-out of all available arguments.

Simple CLI interface for chatting with an LLM in the terminal.

python -m sillm.chat /path/to/model

Running sillm.chat in the terminal with Gemma-2B-it on a MacBook Air M2 with 16GB memory:

Run an API server with basic functionality compatible with OpenAI compatible chat endpoints.

python -m sillm.server /path/to/model --port 8000

Fine-tune a model with low-rank adaptation (LoRA).

python -m sillm.lora /path/to/model -d /path/to/dataset -o /output/adapters

Fine-tune a model with LoRA and direct preference optimization (DPO).

python -m sillm.dpo /path/to/model -d /path/to/dataset -o /output/adapters

Convert a model while merging adapters or quantizing the weights.

Example of merging an adapter into a model:

python -m sillm.convert /path/to/input/model /path/to/output/model -a /path/to/adapters

Quantize a model serially (without loading it entirely into memory):

python -m sillm.quantize /path/to/input/model /path/to/output/model --bits 4

Minimal example of loading a model with SiLLM and generating a text completion:

import sillm

model = sillm.load("/path/to/model")
for s, _ in model.generate("On a beautiful Sunday morning,"):
    print(s, flush=True, end="")

LoRA training Mistral-7B-Instruct-v0.2 with the Nvidia HelpSteer dataset.

DPO training Qwen1.5-7B-Chat with the DPO Mix 7K dataset. The training consists of a supervised fine tuning (SFT) followed by direct preference optimization (DPO).

Implementation of the "Massive Multitask Language Understanding" benchmark using the MMLU dataset.

Calculating perplexity scores for a sample dataset of entry paragraphs from Wikipedia articles.

SiLLM generally supports loading LLMs of the following model architectures/families: Llama 2, Mistral, Mixtral, Gemma, Phi, Qwen 2, StarCoder2.

Here is a list of models that were successfully tested with SiLLM:

• Repetition penalty for inference
• Learning rate schedulers for training
• Merging models
• Saving models to GGUF
• Fine tuning with ORPO

This project uses the MIT License.

Big thanks to the Apple MLX team for implementing and maintaining the MLX framework that makes it possible to unlock the power of Apple Silicon and run/train LLMs on MacBooks and other Apple devices. Thank you to all the contributors of the MLX Examples project and developers sharing model implementations online. Last but not least, thank you to the larger community sharing open weights models, fine tunes, and datasets - without you all the gen AI progress would happen behind locked doors!