data-infra/cube-studio

CubeStudio – A Chinese‑made, cloud‑native AI platform

What it is – CubeStudio is an open‑source, MIT‑licensed AI/MLops platform built for enterprises that need a private, end‑to‑end solution covering the whole data‑to‑model lifecycle. It bundles data ingestion, annotation, notebook development, pipeline orchestration, distributed training (including large‑model fine‑tuning), inference serving, resource‑marketplace, and monitoring, all on top of Kubernetes.

Key capabilities

Area	What you get
User & access	SSO (LDAP, OIDC, remote‑user), multi‑tenant project groups, RBAC with audit logs, per‑project resource quotas.
Compute scheduling	Multi‑cluster, multi‑resource‑group GPU/CPU/NPU/vGPU scheduling; supports NVIDIA, AMD, Ascend, Cambricon, Hygon, Moore‑Thread, Metax, etc.; RDMA (IB/ RoCE); token‑based compute marketplace; fine‑grained limits per user, project, or task; built‑in billing & usage dashboards.
Storage	Web UI to attach NFS, CephFS, OSS, S3/MinIO, GlusterFS, etc.; per‑user PVC mounted automatically.
Data & annotation	SQLLab for interactive queries (MySQL, PostgreSQL, ClickHouse, Hive, Presto, etc.); dataset manager for images/audio/text; full‑stack annotation UI supporting image, video, audio, text and multimodal labeling; automated labeling via built‑in models (object detection, OCR, speech‑to‑text, NER, summarization, etc.).
Online development	Browser‑based JupyterLab / VS‑Code IDEs (Chinese UI) with many Python versions, R, Julia, Spark, etc.; custom image builder; notebook GPU/CPU/vGPU allocation; SSH remote linking; TensorBoard integration.
Pipeline & training	Drag‑and‑drop workflow editor; supports single‑task and DAG pipelines; distributed training on PyTorch, TensorFlow, MindSpore, DeepSpeed, Horovod, Ray, Spark, Paddle, etc.; automatic hyper‑parameter search (NNI/auto‑search); feature‑engineering operators (join, scaling, encoding, PCA, sampling, etc.).
Large‑model support	Multi‑node, multi‑GPU inference with vLLM, Ollama, Xinference; training/fine‑tuning (LoRA, SFT, RLHF) for models like Llama‑3, Qwen‑2, DeepSeek, ChatGLM‑4; model quantization, pruning, evaluation (OpenCompass); model marketplace (AIHub) with 400+ pre‑trained vision, audio, NLP, multimodal models.
Serving	Zero‑code model deployment; versioned model registry; GPU/CPU/vGPU autoscaling; traffic routing, rate‑limiting, JWT auth, health checks; supports ONNX, TensorRT, TorchServe, etc.; can expose OpenAI‑compatible endpoints.
Monitoring & alerts	Cluster‑wide CPU/GPU/Memory/Network/IO dashboards; pod‑level metrics; inference QPS/latency; IB traffic; NPU monitoring; alerts via WeChat, DingTalk, Feishu, webhook.
Infrastructure	Runs on any Kubernetes distribution (Rancher, KubeSphere, K3s, TKEStack, etc.); supports ARM64 and x86, edge clusters, and private‑cloud deployments; integrates with existing data‑center hardware and Chinese‑made accelerators.

Why it matters – Many Chinese enterprises require on‑prem or private‑cloud AI platforms that can run on domestic chips for security and regulatory reasons. CubeStudio provides a single pane of glass for data scientists, ML engineers, and ops teams, eliminating the “glue code” that normally ties together separate notebooks, schedulers, and serving stacks.

Getting started – The repository links to a Wiki with deployment guides, and the platform can be installed via Helm charts onto a Kubernetes cluster. The UI includes a visual pipeline builder, making it approachable for users without deep DevOps expertise.

---

All details above are taken directly from the project's README.