Optimizing Continuous RL @ Fireworks: Rollout, Reward, Update, Repeat!

​Zoom link: https://us02web.zoom.us/j/82308186562

​Talk #0: Introductions and Meetup Updates by Chris Fregly and Antje Barth

​Talk #1: Rollout, Reward, Update, Repeat: How Fireworks Optimizes Continuous Post-Training by Sinan Ozdemir @ Fireworks.ai

​The RL loop is three simple steps: rollout, reward, weight update. What makes it work continuously in production is all in the data plumbing and in how you design your learning environment.

​We'll walk through a post-training RL session on Kimi K3 using Fireworks' serverless API, and talk about the choices of algorithms, loss functions, LoRA rank, hyperparameters, reward design, and how they all interact with how a model learns (or doesn't).

​Zoom link: https://us02web.zoom.us/j/82308186562

​Related Links

​Github Repo: /cfregly/ai-performance-engineering

​O'Reilly Book: https://www.amazon.com/Systems-Performance-Engineering-Optimizing-Algorithms/dp/B0F47689K8/

​YouTube: /AIPerformanceEngineering

​DeepLearning.ai: https://bit.ly/gllm