README.md

Examples using the Fine Tuning Framework.

This repo uses the llm-research-fine-tuning and llm-research-orchestration repos to run fine tuning jobs, locally or in the cloud.

Setup

Requirements

pip install -r requirements.txt
pip install -r requirements-dev.txt

Llama 2

Make sure you've signed the T&Cs to be able to access the llama-2 weights on Huggingface.

Environment

You'll also need a .env file in the folder from where you'll run the jobs.

# For reporting experiments to W&B
WANDB_API_KEY=
WANDB_USERNAME=

# For loading/saving data/models to Huggingface
HF_TOKEN=

# For downloading data to from S3
S3_ENDPOINT=s3.amazonaws.com
S3_ACCESS_KEY_ID=
S3_SECRET_ACCESS_KEY=
S3_REGION=us-east-2
S3_SECURE=true

Data

Set up data for training or inference using the instructions in the llm-research-data project.

Running in colab.

Please checkout notebooks/Launch_SFT_jobs_in_the_Notebook.ipynb.

Running on Coreweave: Launching a Fine-Tuning Job

Kubernetes

Currently we are only supporting Coreweave.

• Coreweave - We assume that you have downloaded the kubeconfig file and can see your pods using kubectl get pods.

Update the Config for the Job

Update k8s/configs/.yaml as needed for your job.

Launch the Job

python k8s/train.py launch rparundekar/fine_tune_research:a327fd0 ./k8s/configs/simple_text.yaml

You'll see the name of the job. And instructions to see the logs and delete the job. NOTE: The container is created in the fine-tuning project.

The hash above should be the 7-hex hash in the llm-research-fine-tuning project's main branch.

If launching with a distributed config

python k8s/train.py launch rparundekar/fine_tune_research:fbea00f distributed_default.yaml -d fsdp_single_worker.yaml

Deleting a Job

Use the same program to delete the job so that all resources are deleted.

python k8s/train.py delete <job-name>

Batch Inference with Pretrained LLM

python k8s/infer.py launch rparundekar/fine_tune_research:fbea00f open_instruct_batch_inference.yaml

Serving a Trained Model

We'll use the Huggingface TGI library to serve a model. The program will read details of the output model from the training config and serve in Kubernetes. Then, you can port forward locally and run the PoC

python k8s/serve.py launch mmlu_peft.yaml

Running the PoC

The PoC assumes that the model server is running at http://127.0.0.1:8080/.

streamlit run poc/poc.py