Skip to content

Latest commit

 

History

History
140 lines (97 loc) · 6.53 KB

finetune_lora.md

File metadata and controls

140 lines (97 loc) · 6.53 KB
Raw

Finetuning with LoRA / QLoRA

Low-rank adaption (LoRA) is a technique to approximate the update to the linear layers in a LLM with a low-rank matrix factorization. This significantly reduces the number of trainable parameters and speeds up training with little impact on the final performance of the model. We demonstrate this method by instruction-finetuning LitGPT StableLM 3B on the Alpaca dataset on a single RTX 3090 (24GB) GPU with CUDA 11.8.

 

Preparation

The steps here only need to be done once:

  1. Follow the instructions in the README to install the dependencies.
  2. Download and convert the weights and save them in the ./checkpoints folder. Weights can be downloaded following the instructions in the download_model_weights documentation:

LitGPT provides common datasets for finetuning, such as Alpaca, LIMA, Dolly, and more. You can optionally prepare your own dataset. For more information about dataset preparation, also see the prepare_dataset.md tutorial.

 

Running the Finetuning

litgpt finetune lora --data Alpaca

The finetuning requires at least one GPU with ~24 GB memory (RTX 3090).

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

Note

LoRA can be applied to not only query, key or value matrices, but also to projection, mlp and classification head. According to QLoRA paper (section 4): "LoRA on all linear transformer block layers are required to match full finetuning performance". By default LoRA is applied only to the query and value matrices. In order to apply LoRA to other weight matrices - change the arguments to litgpt/finetune/lora.py accordingly.

Optionally, finetuning using 4-bit quantization (as in QLoRA) can be enabled via the --quantize flag, for example using the 4-bit NormalFloat data type:

litgpt finetune lora --quantize "bnb.nf4"

and optionally with double-quantization:

litgpt finetune lora --quantize "bnb.nf4-dq"

The table below lists a comparison with different settings on a StableLM 3B model finetuned with LoRA on Alpaca for 1,000 iterations using a microbatch size of 1:

Settings Training Memory Training Time Inference Memory
Default (bf16-mixed) 26.92 GB 1.34 min 21.43 GB
--precision bf16-true 9.69 GB 1.24 min 7.30 GB
--precision bf16-true --quantize bnb.nf4 6.35 GB 1.82 min 3.20 GB
--precision bf16-true --quantize bnb.nf4-dq 6.19 GB 1.87 min 3.04 GB

The advantages of QLoRA-style quantization are more pronounced in larger models, such as Llama 2 7B. The table below summarizes the results for Llama 2 7B on Alpaca for 1,000 iterations using a microbatch size of 1:

Settings Training Memory Training Time Inference Memory
Default (bf16-mixed) OutOfMemoryError N/A 40.21 GB
--precision bf16-true 21.30 GB 2.36 min 13.52 GB
--precision bf16-true --quantize bnb.nf4 14.14 GB 3.68 min 4.57 GB
--precision bf16-true --quantize bnb.nf4-dq 13.84 GB 3.83 min 4.26 GB

For additional benchmarks and resource requirements, please see the Resource Tables.

 

Test the Model

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

litgpt generate base \
  --checkpoint_dir "out/lora/final" \
  --prompt "Recommend a movie to watch on the weekend."

Output:

I would recommend the movie The Martian (2015). It is a sci-fi movie starring Matt Damon that follows the story of...

If your GPU supports bfloat16, you can additionally pass --precision "bf16-true" to bring the memory consumption down to ~7.6 GB for StableLM-3B (versus ~15.2 GB for --precision "32-full"). In addition, you may use quantization methods, for example --precision "bf16-true" --quantize "bnb.nf4" brings the memory consumption further down to ~4.4 GB for StableLM-3B.

 

Tune on Your Dataset

You can easily train on your own instruction dataset saved in JSON format.

  1. 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 and 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"
    },
    ...
]

  2. Run litgpt/finetune/lora.py by passing in the location of your data (and optionally other parameters):

    litgpt finetune lora \
    --data JSON \
    --data.json_path data/mydata.json \
    --checkpoint_dir checkpoints/tiiuae/falcon-7b \
    --out_dir data/mydata-finetuned

 

Merging LoRA Weights (Optional)

Finetuning a model with LoRA generates a lit_model.pth.lora file. This file exclusively contains the LoRA weights, which are much smaller than the original model checkpoint to conserve storage space.

Note

LitGPT will automatically merge the checkpoint for you if you use it in any of the inference commands, such as litgpt generate or litgpt chat. Manual merging is only necessary if you want to use the checkpoint outside LitGPT.

If desired, there is the option to merge these LoRA weights manually into the original model's checkpoint, which creates a full lit_model.pth checkpoint. The advantage of this merging process is to streamline inference operations, as it eliminates the need to dynamically incorporate the LoRA weights during runtime, which can improve inference speed.

For example, after finetuning produced a checkpoint folder out/lora/step-002000, merge it as follows:

litgpt merge_lora --checkpoint_dir "out/lora/step-002000"

The command above creates a full lit_model.pth checkpoint file.