JoyNexus turns VLA post-training into a multi-tenant service with shared GPU scheduling
Haoran Sun · hf · 2026-07-21
JoyNexus: multi-tenant post-training infrastructure for VLA models
JoyNexus proposes a unified service for supervised fine-tuning, reinforcement learning, and evaluation of Vision-Language-Action (VLA) models.
Main design points:
- Split the system into separate Training Model, Inference Model, and Environment services.
- Expose high-level semantic APIs for training, rollouts, and evaluation, while still allowing lower-level custom algorithm composition.
- Support concurrent multi-tenant workloads with isolated action modules, optimizers, rollout records, and policy versions.
- Use global training and inference queues for scheduling.
- Introduce group batching for heterogeneous VLA data schemas that share a compatible model-facing prefix, so one shared backbone forward pass can serve grouped samples.
The evaluation, including workload simulation and a realistic embodied pipeline, shows reduced aggregate GPU time and better service utilization versus isolated single-tenant execution.
More from Infra
- How to build a PostgreSQL-backed semantic search pipeline with pgvector and Ollama — KhuyenTran16 · 2026-07-21
- NeurIPS 2026 workshop calls papers on on-device intelligence — YiMaTweets · 2026-07-21
- Milled from Solid Aluminum: AI Rig Multi-GPU Case for Local Compute — dee_hw · 2026-07-21
- FutureCaribbean’s Buildathon offers $50K, H200 compute, and an NYSE pitch — HeyAmit_ · 2026-07-21
- A new series tests which data-science workflows can run on GPUs today — pandeyparul · 2026-07-21
- Former AWS operator says Bedrock margins can beat SageMaker as agentic AI lifts CPU demand — RihardJarc · 2026-07-21