Can we do tensor-parallel or pipline-parallel infra with sam3？

Can we use tensor parallelism (TP) or pipeline parallelism (PP) techniques for SAM3 inference? Currently, SAM3's GPU memory consumption is very high during inference. For a 200M video, the memory usage reaches 50GB during propagation, and a 3000-frame video causes OOM (Out of Memory) on A100 80GB. Is multi-GPU parallelism feasible? How can we implement it?