Efficient training, fine-tuning and inference of large-scale ML models

• Constantine Dovrolis (The Cyprus Institute)

This presentation reviews the main algorithmic and systems-level ideas for making large-scale generative AI more efficient across the full model lifecycle: pre-training, fine-tuning, and inference. It introduces the computational challenges created by scaling laws, model size, memory footprint, and long-context generation, then surveys model-centric approaches such as quantization, pruning, low-rank approximation, knowledge distillation, mixed-precision training, sparse initialization, and parameter-efficient fine-tuning methods including adapters, prompt tuning, and LoRA-style techniques. The presentation also covers inference-time acceleration, including speculative decoding, KV-cache optimization, efficient attention variants, mixture-of-experts architectures, and long-context methods. The talk is intended for audiences interested in deploying, adapting, or studying large-scale ML models under realistic compute and cost constraints.

Fine-Tuning Transformers for Medical Reasoning with LoRA and Hugging Face Trainer

• Roman Dolgopolyi (GRNET)

This talk demonstrates a practical, end-to-end notebook for fine-tuning a reasoning-capable transformer model for medical question answering. Participants will learn how to use Hugging Face Datasets and Trainer to handle the training workflow, from loading and cleaning data to tokenization, checkpointing, evaluation, and inference. The session demonstrates parameter-efficient fine-tuning with LoRA, showing how a 3B-class Mistral reasoning model can be adapted on a single 16 GB GPU by training only small adapter weights instead of the full model. The notebook combines MedReason and medical-o1 reasoning datasets into a unified question, chain-of-thought, and answer format, then trains and evaluates the model on a small demo subset. By the end, attendees will understand the key engineering choices behind efficient LLM fine-tuning and see a side-by-side comparison of base and fine-tuned model behavior on medical reasoning tasks, including practical notes on GPU setup, mixed precision, and resource cleanup for reproducible classroom demos.

Introduction to Fully Sharded Data Parallel (FSDP)

• Nikos Bakas (GRNET)
• Roman Dolgopolyi (GRNET)

The Pharos training platform

• Margarita Markoulatou (GRNET)

In this talk, we will briefly navigate you through our recently launched Pharos Training Platform. You may navigate various talks and hands-on material on Machine Learning, Deep Learning, NLP, LLMs, Computer Vision, Advanced GenAI, HPC, AI Ethics, AI4Health, AI4Sustainability, and AI4LanguageCulture: https://www.pharos-aifactory.eu/pharos-training-platform/