JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models
JoyNexus is pitched as shared infrastructure for post-training VLA models without giving each tenant a dedicated GPU stack.
The paper says the system separates training, inference, and environment services behind APIs, with shared resident base models and tenant-specific slots. It supports supervised fine-tuning, reinforcement learning, rollout, and evaluation while keeping each tenant’s modules, optimizers, records, and policy versions isolated. Its scheduling uses global training and inference queues, plus group batching to share backbone passes across compatible heterogeneous VLA data. In simulations and an embodied scenario pipeline, the authors report lower aggregate GPU time and better utilization than isolated single-tenant execution. HF Daily Papers' note
The paper says the system separates training, inference, and environment services behind APIs, with shared resident base models and tenant-specific slots. It supports supervised fine-tuning, reinforcement learning, rollout, and evaluation while keeping each tenant’s modules, optimizers, records, and policy versions isolated. Its scheduling uses global training and inference queues, plus group batching to share backbone passes across compatible heterogeneous VLA data. In simulations and an embodied scenario pipeline, the authors report lower aggregate GPU time and better utilization than isolated single-tenant execution. HF Daily Papers' note
score 4