Fine-Tune Llama 2 Models with Ray and DeepSpeed

In this repo, you will find an example demonstrating how to fine-tune LLMs with Ray on OpenShift AI, using HF Transformers, Accelerate, PEFT (LoRA), and DeepSpeed, for Llama 2 models.

It adapts the Fine-tuning Llama-2 series models with Deepspeed, Accelerate, and Ray Train TorchTrainer¹ example from the Ray project, so it runs using the Distributed Workloads stack, on OpenShift AI.

Overview of the Ray Dashboard during the fine-tuning:

Example of a visualization comparing experiments with different combinations of context length and batch size:

---

1. https://github.com/ray-project/ray/tree/master/doc/source/templates/04_finetuning_llms_with_deepspeed ↩