TRL enables co-located vLLM in GRPO training for unified GPU usage

AI Impact Summary

TRL now supports co-locating vLLM with the training process for GRPO, enabling shared GPUs and eliminating REST API communication between training and inference. This reduces idle time during generation, increases training throughput, and lowers hardware costs for online learning workloads. It uses the external_launcher backend to run vLLM inline within the training job, preserving tensor/data parallelism and torchrun scalability. Operators should tune vllm_gpu_memory_utilization to fit model size and avoid OOM or underutilization.