Trajectory Releases a Concurrent Multi-LoRA Training Stack for Continual Learning, Reporting a 2.81× Experiment-Throughput Gain
2.81×. That's the experiment-throughput gain Trajectory just unlocked with concurrent multi-LoRA training—and they open-sourced it.
Why it matters
Trajectory's concurrent multi-LoRA stack removes a major bottleneck in RL training pipelines, enabling teams to run more experiments faster without sacrificing performance. This is infrastructure-layer tooling that directly impacts iteration speed for AI labs and enterprises scaling reinforcement learning workflows.
The key facts
7 to know2.81× experiment-throughput gain over single-tenant baseline
No reward regression reported
Maps each RL experiment to dedicated LoRA adapter on always-hot engine
Collaboration: Trajectory + UC Berkeley Sky Lab + Anyscale
Code open-sourced: NovaSky-AI/SkyRL
Addresses continual learning and multi-experiment concurrency
Published May 31, 2026
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Trajectory, working with UC Berkeley Sky Lab and Anyscale, built a concurrent multi-LoRA training stack for continual learning. It maps each RL experiment to a dedicated LoRA adapter on an always-hot engine, reporting a 2.81× end-to-end experiment-throughput gain over a single-tenant baseline with…