6c0b66f90a
* implement LoRA / QLoRA - example of using MLX to fine-tune an LLM with low rank adaptation (LoRA) for a target task - see also https://arxiv.org/abs/2106.09685 - based on https://github.com/ml-explore/mlx-examples/tree/main/lora * add some command line flags I found useful during use - --quiet -- don't print decorator text, just the generated text - --prompt @/tmp/file.txt -- load prompt from file * user can specify path to model OR model identifier in huggingface * update mlx-swift reference Co-authored-by: Ashraful Islam <ashraful.meche@gmail.com> Co-authored-by: JustinMeans <46542161+JustinMeans@users.noreply.github.com>
1.2 KiB
1.2 KiB
LLM
This is a port of several models from:
using the Hugging Face swift transformers package to provide tokenization:
The Models.swift provides minor overrides and customization -- if you require overrides for the tokenizer or prompt customizations they can be added there.
This is set up to load models from Hugging Face, e.g. https://huggingface.co/mlx-community
The following models have been tried:
- mlx-community/Mistral-7B-v0.1-hf-4bit-mlx
- mlx-community/CodeLlama-13b-Instruct-hf-4bit-MLX
- mlx-community/phi-2-hf-4bit-mlx
- mlx-community/quantized-gemma-2b-it
Currently supported model types are:
- Llama / Mistral
- Gemma
- Phi
See Configuration.swift for more info.
See llm-tool
LoRA
Lora.swift contains an implementation of LoRA based on this example:
See llm-tool/LoraCommands.swift for an example of a driver and llm-tool for examples of how to run it.