\n <img src=\"https://cdn-uploads.huggingface.co/production/uploads/669912aa4ea6475a57a0cabb/lVoUKTdMia9LtcgUUlMFA.png\" width=\"60%\" alt=\"JoyNexus architecture\" />\n</div>\n\nExperimental results show that JoyNexus significantly improves GPU utilization compared to isolated single-tenant execution.\n\n<div align=\"center\">\n <img src=\"https://cdn-uploads.huggingface.co/production/uploads/669912aa4ea6475a57a0cabb/wliVySKFs4Mr_AMtAhZqD.png\" width=\"60%\" alt=\"GPU utilization comparison\" />\n</div>","html":"<p>We present JoyNexus, a service-oriented multi-tenant post-training framework for Vision-Language-Action (VLA) models. As illustrated in the following figure, JoyNexus decouples Training Model, Inference Model, and Environment Services, enabling concurrent tenant workloads with isolated action modules and shared resident base models.</p>\n<div align=\"center\">\n <img src=\"https://cdn-uploads.huggingface.co/production/uploads/669912aa4ea6475a57a0cabb/lVoUKTdMia9LtcgUUlMFA.png\" width=\"60%\" alt=\"JoyNexus architecture\">\n</div>\n\n<p>Experimental results show that JoyNexus significantly improves GPU utilization compared to isolated single-tenant execution.</p>\n<div align=\"center\">\n <img src=\"https://cdn-uploads.huggingface.co/production/uploads/669912aa4ea6475a57a0cabb/wliVySKFs4Mr_AMtAhZqD.png\" width=\"60%\" alt=\"GPU utilization comparison\">\n</div>","updatedAt":"2026-07-21T01:53:02.059Z","author":{"_id":"669912aa4ea6475a57a0cabb","avatarUrl":"/avatars/65db29620062435b1b82c41c9a73b59a.svg","fullname":"Haoran Sun","name":"Haoran231","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6570510268211365},"editors":["Haoran231"],"editorAvatarUrls":["/avatars/65db29620062435b1b82c41c9a73b59a.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.16074","authors":[{"_id":"6a5dab1e6a69ce099f4d6e78","user":{"_id":"669912aa4ea6475a57a0cabb","avatarUrl":"/avatars/65db29620062435b1b82c41c9a73b59a.svg","isPro":false,"fullname":"Haoran Sun","user":"Haoran231","type":"user","name":"Haoran231"},"name":"Haoran Sun","status":"claimed_verified","statusLastChangedAt":"2026-07-20T16:46:27.817Z","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e79","name":"Wentao Zhang","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e7a","name":"Junyang Hua","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e7b","name":"Hedan Yang","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e7c","name":"Yongjian Guo","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e7d","name":"Yifei Zhang","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e7e","name":"Xiaolong Xiang","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e7f","name":"Mingxi Luo","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e80","name":"Jing Long","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e81","name":"Chen Zhao","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e82","name":"Chen Zhou","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e83","name":"Wanting Xu","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e84","name":"Qiming Yang","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e85","name":"Hui Zhang","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e86","name":"Song Wang","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e87","name":"Xiaodong Bai","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e88","name":"Shuai Di","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e89","name":"Xu Chu","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e8a","name":"Xiaotie Deng","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e8b","name":"Yicheng Gong","hidden":false},{"_id":"6a5dab1e6a69ce099f4d6e8c","name":"Junwu Xiong","hidden":false}],"publishedAt":"2026-07-17T00:00:00.000Z","submittedOnDailyAt":"2026-07-21T00:00:00.000Z","title":"JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models","submittedOnDailyBy":{"_id":"669912aa4ea6475a57a0cabb","avatarUrl":"/avatars/65db29620062435b1b82c41c9a73b59a.svg","isPro":false,"fullname":"Haoran Sun","user":"Haoran231","type":"user","name":"Haoran231"},"summary":"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.","upvotes":3,"discussionId":"6a5dab1e6a69ce099f4d6e8d"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"669912aa4ea6475a57a0cabb","avatarUrl":"/avatars/65db29620062435b1b82c41c9a73b59a.svg","isPro":false,"fullname":"Haoran Sun","user":"Haoran231","type":"user"},{"_id":"65f929c2d0724d81ce8c771d","avatarUrl":"/avatars/e041a750cdd65b0c162554a7decddc53.svg","isPro":false,"fullname":"zhangwentao","user":"zhangwt97","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.16074.md","query":{}}">
JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models
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.
Community
We present JoyNexus, a service-oriented multi-tenant post-training framework for Vision-Language-Action (VLA) models. As illustrated in the following figure, JoyNexus decouples Training Model, Inference Model, and Environment Services, enabling concurrent tenant workloads with isolated action modules and shared resident base models.
Experimental results show that JoyNexus significantly improves GPU utilization compared to isolated single-tenant execution.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2607.16074 in a model README.md to link it from this page.
Cite arxiv.org/abs/2607.16074 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2607.16074 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.