SkyRL has turned large-model cold-start into a resident memory process, boosting throughput directly by ×3—this idea is too dirty.

CryptoWorld News, OneMillion_AI announced that the open-source SkyRL concurrent training stack has achieved hour-level self-evolution of large models. The overall experimental throughput has increased by 2.81 times, and the throughput within a single node's absolute time has increased by approximately 3.25 times. This architecture maintains a resident shared model base in GPU memory, reducing the cold start overhead of repeatedly loading large models, aiming to help developers perform real-time production data training of large models at extremely low GPU memory costs.