Reproduce “JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haoran Sun, Wentao Zhang, Junyang Hua, Hedan Yang, Yongjian Guo, Yifei Zhang, Xiaolong Xiang, Mingxi Luo and 13 more
JoyNexus develops a service-oriented, multi-tenant post-training infrastructure for Vision-Language-Action (VLA) models, designed to improve resource efficiency and simplify development. The system achieves a 1.39x improvement in GPU-time efficiency and nearly doubles Training Model Service utilization compared to isolated execution.
Key insights
- Decoupling training, inference, and environment services with clear API boundaries enables modularity and specialized optimization for VLA workloads.
- Sharing a memory-resident VLA base model across tenants, while isolating lightweight action modules, substantially improves resource utilization for VLA-specific fine-tuning.
- Group batching of heterogeneous inference requests amortizes the computational cost of the shared VLM backbone, significantly improving throughput for small or bursty workloads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.