Software Engineering Radio - the podcast for professional software developers Episode 549: William Falcon Optimizing Deep Learning Models
Feb 3, 2023
William Falcon, CEO of Lightning AI and creator of PyTorch Lightning, dives into optimizing deep learning models. He explains the distinctions between training and inference, emphasizing the need for fast processing in applications like self-driving cars. Falcon discusses MLOps complexities and highlights the required multidisciplinary skills for production AI. He also warns about common pitfalls for new users and the importance of using LightningModule for scalability. Plus, he shares insights on the evolving landscape of AI frameworks.
AI Snips
Chapters
Transcript
Episode notes
From Research Library To Platform
- Lightning began as William Falcon's personal research tooling, evolving from Theano to TensorFlow and later PyTorch.
- He scaled it inside NYU and Facebook research before founding a company to help teams run large cloud training.
Keep Model Logic Inside Training Step
- Put only the inner computation loop into a LightningModule training_step and let the Trainer handle hardware concerns.
- Remove .cuda calls and other accelerator logic from model code to stay hardware-agnostic.
Fabric: Control Without Losing Utilities
- Fabric lets researchers retain control of the Python training loop while getting Lightning's accelerator and utilities.
- It bridges plain PyTorch and full Lightning Trainer for advanced distributed research.