Module Version Hex Docs

This package implements Low-Rank Adaptation (LoRA), a popular method for fine-tuning large language models.


This package can be installed by adding lorax to your list of dependencies in mix.exs:

def deps do
    {:lorax, "~> 0.1.0"}

Fine-tuning an LLM with LoRA

In general,

  1. Import your model
  2. Inject trainable LoRA parameters
  3. Train LoRA model
  4. Download LoRA only params
{:ok, model_info} = Bumblebee.load_model({:hf, "gpt2"})
%{model: gpt2_model, params: gpt2_params} = model_info

lora_model =
  |> Axon.freeze()
  |> Lorax.inject(%Lorax.Config{
      r: 2,
      alpha: 4,
      dropout: 0.05

lora_merged_params =
  Axon.build(lora_model, mode: :train)
  |> Axon.Loop.trainer(custom_loss_fn, Polaris.Optimizers.adam(learning_rate: 3.0e-4))
  |> Axon.Loop.run(train_batch_stream, gpt2_params, epochs: 3, iterations: 1000, compiler: EXLA)

lora_params = lora_merged_params
  |> Lorax.Params.filter(gpt2_params)
  |> Lorax.Params.kino_download()

In practice, every model has some unique architecture that you need to account for. For more detailed guides, see

  1. Finetuning LLMs with LoRA
  2. Running LLMs with LoRA

Default Settings

The default config applies LoRA to all query, key, value matrices. r = 1, alpha = 2.

The LoRA paper demonstrated that adapting only the query and value matrices with r = 1 achieved effective fine-tuning results. However, for larger language models, people often choose much higher values of r and sometimes target all linear layers.

These settings works well for fine-tuning smaller LLMs (~1b param models)

Lora Config
- r value  = at least 2
- alpha value = r * 2

- learning_rate of 3.0e-4 with an adam optimizer

Text Generation
- multinomial sampling
- p = 0.06 or 0.08 for more variety (or if you experience repetitive results)

For more details on configuring LoRA hyperparameters, see this post by Sebastian Raschka.


  1. GPU Memory Requirements: Although LoRA reduces the GPU requirements for fine-tuning, larger LLMs beyond GPT2 still demand GPUs with substantial vRAM. Inadequate memory management can lead to cuda OOM crashes.

  2. Fine-Tuning Speed: The training speed of this library isn't on par with Huggingface's PEFT library. Further optimizations can be done to close the gap.

Note: For minor fine-tuning tasks without a GPU, the BinaryBackend is a viable option, often resulting in smoother training runs. Future updates will focus on minimizing GPU memory usage by reducing the amount of tensors stored during training, and potentially a QLoRA implementation one day.