echo-2

The distributed AI lab Gradient today released the Echo-2 distributed reinforcement learning framework, aimed at breaking the barriers of training efficiency in AI research. By achieving a complete decoupling of Learner and Actor at the architectural level, Echo-2 has reduced the post-training cost of a 30B model from $4,500 to $425. This results in over 10 times the research throughput under the same budget.The framework utilizes compute-storage separation technology for asynchronous training (Async RL), offloading massive sampling computation to unstable GPU instances and Parallax-based heterogeneous GPUs. Coupled with breakthroughs in bounded staleness, instance fault-tolerant scheduling, and the self-developed Lattica communication protocol, it significantly enhances training efficiency while ensuring model accuracy. Along with the framework release, Gradient will soon launch the RLaaS platform Logits, promoting the shift in AI research from a "capital accumulation" paradigm to an "efficiency iteration" paradigm. Logits is now open for reservations for students and researchers worldwide.

Distributed AI Lab Gradient has released the Echo-2 distributed reinforcement learning framework, aiming to break the barriers of training efficiency in AI research by decoupling the Learner and Actor at the architectural level to reduce the post-training costs of large models.Official data shows that this framework can reduce the post-training cost of a 30B model from $4500 to $425. Echo-2 utilizes separation of storage and computation technology for asynchronous training (Async RL), supporting the offloading of sampling computation to unstable GPU instances and Parallax-based heterogeneous GPUs. Additionally, Gradient plans to launch the RLaaS (Reinforcement Learning as a Service) platform Logits, which is currently open for reservations for students and researchers.