Submission files for the NeurIPS Large Language Model Efficiency Challenge 2023: 1 LLM + 1GPU + 1Day

https://llm-efficiency-challenge.github.io/index

This example demonstrates how to fine-tune and serve a Llama 2 model with llama-recipes for submission in the LLM efficiency challenge using the lit-gpt example as a template. Llama-recipes provides an easy way to fine-tune a Llama 2 model with custom datasets using efficient techniques like LoRA or Llama-adapters.

In order to use llama-recipes we need to install the following pip package:

pip install llama-recipes

To obtain access to the model weights you need to fill out this form to accept the license terms and acceptable use policy.

After access has been granted, you need to acknowledge this in your HuggingFace account for the model you want to fine-tune. In this example we will continue with the 7B parameter version available under this identifier: meta-llama/Llama-2-7b-hf

NOTE In this example the training result will be uploaded and downloaded through huggingface_hub. The authentication will be done through a token created in the settings of your HuggingFace account. Make sure to give write access to the token and set the env variables in the Dockerfiles to your token and repo:

ENV HUGGINGFACE_TOKEN="YOUR_TOKEN"
ENV HUGGINGFACE_REPO="YOUR_USERNAME/YOUR_REPO"

With llama-recipes its possible to fine-tune Llama on custom data with a single command. To fine-tune on a custom dataset we need to implement a function (get_custom_dataset) that provides the custom dataset following this example custom_dataset.py. We can then train on this dataset using this command line:

python3 -m llama_recipes.finetuning  --use_peft --peft_method lora --quantization --model_name meta-llama/Llama-2-7b --dataset custom_dataset --custom_dataset.file /workspace/custom_dataset.py --output_dir /volume/output_dir

Note The custom dataset in this example is dialog based. This is only due to the nature of the example but not a necessity of the custom dataset functionality. To see other examples of get_custom_dataset functions (btw the name of the function get_custom_dataset can be changed in the command line by using this syntax: /workspace/custom_dataset.py:get_foo_dataset) have a look at the built-in dataset in llama-recipes.

Note For a submission to the competition only the inference part (Dockerfile) will be necessary. A training docker (Dockerfile.train) will only be necessary if you need to replicate the submission in case you're within the top 3 contestants.

The inference Docker will download base and LoRA weights from huggingface_hub. For the submission it is assumed that the trained weights are uploaded to a repo on huggingface_hub and the env variables HUGGINGFACE_TOKEN and HUGGINGFACE_REPO have been updated accordingly in the Dockerfile.

To create the zip file for submission to the eval bot use the following commands:

cd neurips_llm_efficiency_challenge/sample-submissions
rm llama_recipes/Dockerfile.train
zip -r llama_recipes.zip llama_recipes

Note 1. Make sure to only zip the folder llama_recipes and do not include any other sample submission in the zipfile. 2. We delete llama_recipes/Dockerfile.train as a precaution to avoid errors if submission logic changes.

To locally build and and run the taining Docker we need to execute:

docker build -f ./Dockerfile.train -t llama_recipes_train .

docker run --gpus "device=0" --rm -ti llama_recipes_train

The inference Docker can be created and started locally with:

docker build -f ./Dockerfile -t llama_recipes_inference .

docker run --gpus "device=0" -p 8080:80 --rm -ti llama_recipes_inference

To test the inference docker we can run this query:

curl -X POST -H "Content-Type: application/json" -d '{"text": "What is the capital of france? "}' https://localhost:8080/tokenize
OR
curl -X POST -H "Content-Type: application/json" -d '{"prompt": "What is the capital of france? "}' https://localhost:8080/process