OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism

In three linesOxyGen proposes unified KV cache management for Vision-Language-Action (VLA) model inference under multi-task parallelism. Implemented on π₀.₅, the system achieves 3.7× speedup on RTX 4090 and Jetson AGX Thor, delivering 200+ tokens/s and 70 Hz simultaneously without quality degradation.

Summary generated by Claude — human-verified