Description: This advanced level course will cover how to fine-tune a large language model with custom data on HPC resources. Users will be shown examples of the following:

• Full-parameter fine-tuning with unsupervised learning on custom data using PyTorch Fully Sharded Data Parallel
• Low-rank adaptation fine-tuning with unsupervised learning on custom data using Parameter-Efficient Fine-Tuning (PEFT) and PyTorch Distributed Data Parallel (DDP)
• Low-rank adaptation with supervised learning to instruction-tune a pretrained model using open-source data.

The examples will use a single initial pretrained model to add understanding of DoD-related concepts and instruction-following capabilities. The course will cover how to format data for unsupervised and supervised fine-tuning, how to leverage multiple GPUs across multiple nodes to accelerate training, hardware constraints on model selection for fine-tuning, and hyperparameter selection.

Presenter(s): Dr. Mathew Boyer, GDIT / PET Location: Webinar Date & Time: October 30, 2024, 2:00p - 3:30p ET

Controlled by: DoD HPCMP Controlled by: PET Program CUI Category: OPSEC Limited Dissemination Control: FEDCON POC: Mr. Ronald Hedgepeth, pet@hpc.mil

---

CUI