Awwscale implements V-JEPA2-AC inference on Ray Serve A100s

An engineering implementation demonstrates how moving V-JEPA2-AC world model inference from an on-board GPU to an 8-GPU Ray Serve cluster reduces action planning time from 81.8 seconds to 2.86 seconds per step. The design uses Ray Core and Ray Serve to distribute candidate rollouts across multiple A100 GPUs while keeping model weights resident. Further optimizations, including KV caching and non-blocking control, can reduce execution time even further. Open article →