Finetuning with Adapter

LLaMA-Adapter is a form of prefix-tuning that prepends a learnable adaption-prompt to the inputs of the attention blocks in LLaMA. In total, there are only 1.2M parameters to update during finetuning, which significantly reduces the memory footprint and speeds up training.

We are able to demonstrate instruction-finetuning Lit-LLaMA 7B on the Alpaca dataset on a single GTX 3090 (24GB) GPU. If using 8 GPUs, finetuning can be completed in under 1 hour.

If you are new to LLaMA-Adapter and are interest to learn more about how it works before proceeding with the finetuning guide below, you might find our article Understanding Parameter-Efficient Finetuning of Large Language Models: From Prefix Tuning to LLaMA-Adapters helpful.

Preparation

The steps here only need to be done once:

Follow the instructions in the README to install the dependencies.
Download and convert the weights and save them in the ./checkpoints folder as described here.
If you want to utilize more than one GPU, you should pip install deepspeed.
Download the data and generate the Alpaca instruction tuning dataset:
```
python scripts/prepare_alpaca.py
```
or prepare your own dataset.

Running the finetuning

python finetune_adapter.py

The finetuning requires at least one GPU with ~24 GB memory (GTX 3090). You can speed up training by setting the devices variable in the script to utilize more GPUs if available. Depending on the available GPU memory, you can also tune the micro_batch_size parameter to utilize the GPU efficiently.

For example, the following settings will let you finetune the model in under 1 hour using DeepSpeed Zero-2:

devices = 8
micro_batch_size = 8

This script will save checkpoints periodically to the folder out/.

Note All scripts support argument customization

Test the model

You can test the finetuned model with your own instructions by running:

python generate_adapter.py \
    --prompt "Recommend a movie to watch on the weekend." \
    --quantize llm.int8

Output:

A good movie to watch on the weekend would be The Lion King, since it's a classic family film that everyone can enjoy...

If your GPU supports bfloat16, the script will automatically use it. Together with --quantize llm.int8, this brings the memory consumption down to ~8 GB.

Tune on your dataset

With only a few modifications, you can prepare and train on your own instruction dataset.

Create a json file in which each row holds one instruction-response pair. A row has an entry for 'instruction', 'input', and 'output', where 'input' is optional an can be the empty string if the instruction doesn't require a context. Below is an example json file:
```
[
    {
        "instruction": "Arrange the given numbers in ascending order.",
        "input": "2, 4, 0, 8, 3",
        "output": "0, 2, 3, 4, 8"
    },
    ...
]
```
Make a copy of scripts/prepare_alpaca.py and name it what you want:
```
cp scripts/prepare_alpaca.py scripts/prepare_mydata.py
```
Modify scripts/prepare_mydata.py to read the json data file.
Run the script to generate the preprocessed, tokenized train-val split:
```
python scripts/prepare_mydata.py --destination_path data/mydata/
```
Run finetune_adapter.py by passing in the location of your data (and optionally other parameters):
```
python finetune_adapter.py --data_dir data/mydata/ --out_dir out/myexperiment
```

Troubleshooting

If you run into a CUDA error "Expected is_sm80 to be true, but got false", uncomment the line torch.backends.cuda.enable_flash_sdp(False) in the script below (see https://github.com/Lightning-AI/lit-llama/issues/101).