Repolode
Submit
data-infra

data-infra/cube-studio

Cloud-native MLOps platform on Kubernetes: notebooks, drag-and-drop pipelines, distributed training and LLM serving.

View on GitHubWebsite
Stars
2,564
Forks
214
Open issues
28
Last commit
10 days ago

What is it?

CubeStudio is a cloud-native, Kubernetes-based platform that covers the whole machine-learning lifecycle in one place: data ingestion, labeling, browser-based development, training, model management and inference deployment. Its stated goal is to stop algorithm engineers from spending their time on requesting machines, preparing environments, pulling data and hand-rolling deployment scripts.

At the center is a drag-and-drop pipeline editor. The platform ships layered operators for data synchronization from Hugging Face and ModelScope, feature engineering, classical ML (sklearn, XGBoost, LightGBM and more), TF/PyTorch distributed training, and large-model fine-tuning. It supports multi-machine multi-GPU training, NNI hyperparameter search, TensorBoard, task-level input/output visualization, and zero-code publishing of models as inference services. Online IDEs include JupyterLab, VSCode and RStudio, with full-card, virtual and shared GPU modes, plus SSH remote development.

For infrastructure, CubeStudio isolates tenants, projects, resource groups and multiple Kubernetes clusters, and manages NVIDIA plus domestic accelerators such as Ascend, Hygon DCU, Cambricon MLU, MetaX, Moore Threads and Kunlunxin, with IB/RoCE RDMA and vGPU virtualization. It can use Dameng, Kingbase or OceanBase as its metadata database, adapts to ARM64 CPUs, and supports fully offline private deployment.

Key features

  • 01Drag-and-drop pipeline editor for building DAGs, with scheduled runs, backfill, concurrency limits, timeouts and instance dependencies.
  • 02Online IDEs (JupyterLab, VSCode, RStudio) with multi-language kernels, TensorBoard, Git integration and SSH remote access.
  • 03Data labeling for image, text, audio and multimodal data, plus automatic labeling driven by vision models and LLMs.
  • 04Distributed training for TensorFlow and PyTorch, integrating DeepSpeed, Megatron, Colossal-AI, Horovod, MPI, MindSpore and PaddlePaddle.
  • 05NNI-based hyperparameter search on single machines or clusters, with a graphical view of each trial's results.
  • 06Zero-code inference publishing for ML, TF, PyTorch, TensorRT, ONNX and LLM models, with canary release, traffic splitting, rate limiting, autoscaling and rollback.
  • 07Large-model support: vLLM/Ollama/MindIE/Xinference serving, LLaMA-Factory SFT, reward-model and RL training, plus OpenCompass evaluation and pruning.
  • 08Compute scheduling across clusters and resource groups with vGPU virtualization, shared/exclusive card modes, binpack strategy and RDMA.
  • 09AIHub model market with 400+ pretrained models offering one-click deployment, one-click Notebook development and one-click fine-tuning pipelines.
  • 10Monitoring and alerting via Prometheus/Grafana dashboards covering GPU, pod and vGPU load, QPS, throughput and token metrics, pushed to email, WeCom, DingTalk or Feishu.

Use cases

  • Organizations that need an on-premises AI platform with multi-tenant isolation and their own compute pool.
  • Teams running training or inference on domestic accelerators such as Ascend, Hygon, Cambricon or MetaX for compliance reasons.
  • Algorithm teams that want a standardized path from experiment to deployed model instead of ad-hoc scripts.
  • Teams building private knowledge bases, chatbots or AIGC applications on top of self-hosted LLMs.
  • Groups that need image, text and audio annotation combined with automated pre-labeling.
  • Platform operators offering compute rental, metering and token relay to internal or external users.

Things to know

  • Runs on Kubernetes, so you need a cluster before you can use it.
  • The README says MIT, but the listed license is unverified, so check the LICENSE file before commercial use.

Topics

#mlops#kubernetes#llmops#model-serving#distributed-training#pipeline#gpu-scheduling#vgpu#data-labeling#finetuning

FAQ

What is cube-studio?
Cloud-native MLOps platform on Kubernetes: notebooks, drag-and-drop pipelines, distributed training and LLM serving. CubeStudio is a cloud-native, Kubernetes-based platform that covers the whole machine-learning lifecycle in one place: data ingestion, labeling, browser-based development, training, model management and inference deployment. Its stated goal is to stop algorithm engineers from spending their time on requesting machines, preparing environments, pulling data and hand-rolling deployment scripts.
Is cube-studio free to use? What license does it have?
cube-studio is open source under the Other license. Check the license terms before using it in commercial products.
Is cube-studio still maintained?
Actively maintained. The last commit to cube-studio was on Oct 12, 2026.
What can I use cube-studio for?
Organizations that need an on-premises AI platform with multi-tenant isolation and their own compute pool. Teams running training or inference on domestic accelerators such as Ascend, Hygon, Cambricon or MetaX for compliance reasons. Algorithm teams that want a standardized path from experiment to deployed model instead of ad-hoc scripts. Teams building private knowledge bases, chatbots or AIGC applications on top of self-hosted LLMs. Groups that need image, text and audio annotation combined with automated pre-labeling. Platform operators offering compute rental, metering and token relay to internal or external users.
Where can I get cube-studio?
The source code and installation instructions are on GitHub: https://github.com/data-infra/cube-studio

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects