JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

2026-07-17

AI Summary

This work addresses the inefficiencies of existing post-training services for vision-language-action (VLA) models, which typically suffer from resource monopolization, high costs, low utilization, and burdensome infrastructure adaptation requirements for users. To overcome these limitations, we propose a service-oriented, multi-tenant VLA post-training platform that decouples training, inference, and environment services, enabling concurrent task submission by multiple tenants while ensuring module and data isolation. The platform features a shared-resource scheduling architecture, a global task queue manager, and a prefix-compatible batching mechanism for heterogeneous data, complemented by both high- and low-level APIs to support flexible algorithmic composition. Experimental results demonstrate that, compared to single-tenant exclusive deployment, our approach significantly reduces total GPU time while improving resource utilization and training efficiency.

Abstract

The post-training of Vision-Language-Action (VLA) models is essential due to the diversity of simulators, robot embodiments, and task objectives. Existing compute services, whether offered as direct accelerator rental or batch-workload submission, typically allocate an exclusive set of GPU and CPU resources to a single tenant. While this paradigm maximizes client flexibility, it burdens users with infrastructure adaptation, and the fixed card-hour accounting model renders short or bursty workloads both expensive for tenants and inefficient for the service provider. To address these challenges, we present JoyNexus, a unified service for multi-tenant VLA supervised fine-tuning, reinforcement learning, and evaluation. JoyNexus decouples the Training Model Service, Inference Model Service, and Environment Service, each accessed through APIs and backed by resident shared base models with tenant-specific slots. Tenants can directly invoke high-level semantic APIs for training, rollout, and evaluation, or compose custom algorithms using lower-level APIs and their assigned endpoints. Multiple tenants submit workloads concurrently; their action modules, optimizers, rollout records, and policy versions remain isolated, and the service is scheduled by the global Training Queue and Inference Queue. To further improve multi-tenant training efficiency, JoyNexus introduces group batching for heterogeneous VLA data schemas that share a compatible model-facing prefix, enabling a single shared backbone forward pass over grouped samples. Finally, we evaluate JoyNexus through workload simulation and a group-batching pipeline in a realistic embodied scenario. Results show that, compared with isolated single-tenant execution, JoyNexus reduces aggregate GPU time and improves service utilization via cross-tenant scheduling on shared resources.

Problem

• multi-tenant
• VLA models
• post-training
• resource efficiency
• service utilization

Innovation

• multi-tenant
• group batching
• service-oriented
• VLA models
• shared backbone