Trajectory Launches Open-Source SkyRL Platform, Achieves 2.81x Throughput Improvement

According to OneMillion_AI, Trajectory in collaboration with UC Berkeley's Sky Computing Lab and Anyscale recently announced the open-source SkyRL platform and Multi-LoRA Training architecture for large language model fine-tuning. The system addresses inefficiencies in traditional model optimization by maintaining a shared model base in GPU memory while treating multiple fine-tuning experiments as lightweight adapter modules. Testing showed end-to-end experiment throughput improved by 2.81 times, with single-node absolute time throughput reaching 3.25 times improvement, enabling large models to achieve hourly-level self-evolution through real-time production data. Training code is now available in the SkyRL repository.
Disclaimer: The information on this page may come from third-party sources and is for reference only. It does not represent the views or opinions of Gate and does not constitute any financial, investment, or legal advice. Virtual asset trading involves high risk. Please do not rely solely on the information on this page when making decisions. For details, see the Disclaimer.
Comment
0/400
No comments