Training API now generally available | Fireworks

This post announces the general availability of the Training API and Fireworks Lab. The Training API allows users to connect custom training loops (Python) to Fireworks-managed distributed training and rollout infrastructure, offering full control over loss, data, and environment. It provides two compute options: Serverless for rapid experimentation (per-token billing, LoRA) and Dedicated for full-parameter training or sustained throughput (per-GPU-hour billing). The post highlights the importance of training and inference alignment (matching numerical formats and kernels) and performance efficiency (maximized throughput across the RL loop) for reinforcement learning.