ivanfioravanti / mlx-moe

Scripts to create your own moe models using mlx

mlx-moe

This repository contains scripts to create MoE models from pretrained llama/mistral models.

Installation

Install the required dependencies from requirements.txt.

pip install -r requirements.txt

Using the Script

Configure Model Paths: Open the moe.py script and modify the EXPERT_MODEL_PATHS list at the top of the script with the paths to your model weights and configurations. The first path in this list should be your base model.

Execute the Script: Run the script to create the MoE model from the command line.

python moe.py

Output: The script will generate a new directory, mlx_moe, containing the tokenizer, updated weights, and configuration for the MoE model.

MoE Gate Fine-tuning

After merging the MoE model, fine-tuning is crucial for optimal performance. The gate weights are initialized uniformly, and fine-tuning them is necessary for the model to learn how to effectively utilize the different experts.

Quantization: Given the potentially large size of the MoE model, you may want to quantize the model to reduce its size. Before starting to fine-tune the model, you can quantize it using the quantize.py script. This script will quantize the model and save it to the mlx_model directory.

python quantize.py --model mlx_moe -q

Download the Training Dataset: Use the download_dataset.py script to download the training dataset. This script will download the dataset and save it to the data directory.

python download_dataset.py WizardLM/WizardLM_evol_instruct_70k

Fine-tuning Script: Use the lora.py script for this purpose. This script applies the LoRA to fine-tune the gate and Q,V projection weights of the MoE model. Adjust the script accordingly based on your needs.

python lora.py

Inference: Use the inference.py script to run inference on the fine-tuned model. This script will load the model and tokenizer from the mlx_model directory and run inference on the given input prompt.

python inference.py --model mlx_model --adapter-file adapters.npz --prompt "[INST] write a quick sort in python. [/INST]"

Fuse model

you can fuse the model with lora adapters using fuse.py script, and you can use llama.cpp/gptq/aws to quantize the fused model and use it in your application.

python fuse.py --model mlx_model --adapter-file adapters.npz --de-quantize

To see supported options run:

python fuse.py --help