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>
LoRATrainingExample
Example application that:
- downloads the
mlx-community/Mistral-7B-v0.1-hf-4bit-mlxmodel from huggingface - loads the train/valid/test data from
$SRCROOT/Data/lora(this is copied into the build but you can imagine how it might be downloaded) - adds LoRA adapters and trains the model
- let's you evaluate a prompt against the model
This roughly equates to the command line example in Tools/llm-tool and you can read more about LoRA there.
This evaluates the LoRA adapted model rather than a fused model. This doesn't persist the LoRA weights or the fused model -- it will retrain it each time the program is launched.
Troubleshooting
The mlx-community/Mistral-7B-v0.1-hf-4bit-mlx model requires a little over 4G of
memory to load an train -- this may require ~6G of physical RAM.