research Spotlight Cost-efficient DiT reinforcement learning with spot GPUs, dynamic seed exploration, and elastic sequence parallelism. TokenScale Proactive autoscaling for disaggregated LLM serving, driven by token-level workload dynamics. Liquid Multidimensional resource allocation for LLM inference through dynamic parallelism and autoscaling. open source ROLL Diffusion RL End-to-end Qwen-Image post-training with FlowGRPO, DiffNFT, FSDP2, and vLLM-Omni, released in ROLL v0.4.0.