[Large Models Achieve Hour-Level Self-Evolution, Open-Source SkyRL Concurrent Training Stack Brings 2.8x Efficiency Boost] According to monitoring by Beating, Trajectory, in collaboration with UC Berkeley's Sky Computing Lab and Anyscale, has announced the launch of the open-source multi-LoRA reinforcement learning training platform SkyRL, along with the release of the concurrent training architecture Multi-LoRA Training, which supports continuous learning for large models. In traditional large model fine-tuning experiments, developers must schedule GPU nodes separately for each fine-tuning task to test different training strategies, frequently loading and unloading the same massive model on servers. For instance, training a massive model with 397 billion parameters in the traditional mode requires multiple compute nodes to repeatedly load hundreds of GBs of weight files, leading to significant waste of computation time and resources. The Multi-LoRA architecture completely eliminates the dilemma of repeated cold starts. The system opts to keep an online "shared model base" resident in GPU memory and manages multiple fine-tuning experiments as lightweight "adapter modules." During the training phase, the system performs forward and backward propagation sequentially by quickly swapping in and out the states of various adapters between the CPU and GPU, avoiding the overhead of repeatedly loading the massive base model. Combined with concurrent processing during the inference phase, the overall experimental throughput for multitasking is significantly improved. Test data shows that, without compromising the performance of the large model, end-to-end experimental throughput increased by up to 2.81 times, while throughput within a single node's absolute time frame increased by approximately 3.25 times. Currently, the training code has been open-sourced in the SkyRL repository, aiming to help developers enable large models to achieve hour-level self-evolution using real-time production data at extremely low memory costs. [Original Link]