Minimize idle accelerators: Native RL job interleaving with co-operative time-slicing in llm-d

A new solution, co-operative time-slicing through the llm-d project, minimizes idle accelerators in large language models (LLMs) by interleaving independent RL jobs onto shared hardware. This increases accelerator duty cycles from 40% to 70% without impacting model convergence or accuracy, improving price-performance and lowering TCO. The solution targets synchronous and asynchronous RL workloads, eliminating wasted compute time. The llm-d project aims to eliminate accelerator idle time for various workloads, including inference, agentic, and RL. The solution is now available through the llm-d project, with a detailed technical description and future roadmap to be discussed.