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, Phi-3, 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
One of the main goals of SiLLM is to enable experimentation with the inner workings of large language models and make new techniques accessible to a wider audience running on Apple Silicon hardware.
The control module incorporates techniques based on the paper Representation Engineering and the blog Refusal Ablation. Representation engineering is a method to calculate control vectors from a model's hidden states during training that can be used to influence the behavior and generated output during inference. Refusal ablation works similarly, but can be used to remove the direction represented by the vector from model weights.
Using pip:
pip install sillm-mlxThe web app uses Chainlit to provide a frontend for conversational AI running locally on Apple Silicon hardware.
sillm-llama3.mp4
To use 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 -wSet 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/modelRunning sillm.chat in the terminal with Gemma-2B-it on a MacBook Air M2 with 16GB memory:
sillm-chat.mov
Run an API server with basic functionality compatible with OpenAI compatible chat endpoints.
python -m sillm.server /path/to/model --port 8000Fine-tune a model with low-rank adaptation (LoRA).
python -m sillm.lora /path/to/model -d /path/to/dataset -o /output/adaptersFine-tune a model with LoRA and direct preference optimization (DPO).
python -m sillm.dpo /path/to/model -d /path/to/dataset -o /output/adaptersConvert 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/adaptersQuantize a model serially (without loading it entirely into memory):
python -m sillm.quantize /path/to/input/model /path/to/output/model --bits 4Minimal 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="")The repository SiLLM-examples contains Python code examples for using the SiLLM framework for training and running LLMs.
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 significant models that were successfully tested with SiLLM:
| Model Family | Models/Sizes |
|---|---|
| Llama-3 | 8B-Instruct, 70B-Instruct |
| Phi-3 | medium-4k-instruct |
| Phi-3.5 | mini-instruct, MoE-instruct |
| Gemma-2 | 2b-it, 9b-it, 27b-it |
| Mistral | 7b-instruct-v0.3, Nemo-Instruct, Small-Instruct, Large-Instruct |
| Mixtral | 8x22B-Instruct-v0.1 |
| Codestral | 22b-v0.1 |
| Qwen 2 | 7b-instruct, 72b-instruct |
| StarCoder2 | 3b, 7b, 15b |
- 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!