cube-studio: a cloud-native one-stop ML platform with visual pipeline orchestration and heterogeneous compute scheduling

What it solves

CubeStudio is a cloud-native, one-stop machine learning platform designed to simplify the entire ML lifecycle. It removes the friction of managing infrastructure, environment configuration, and resource scheduling, allowing AI developers to move from data labeling and interactive development to distributed training and model deployment without manually configuring servers or Kubernetes clusters.

How it works

Built on Kubernetes, the platform orchestrates compute resources (CPU, GPU, NPU) across multiple clusters and resource groups. It provides a web-based interface for:

• Online Development: Integrated IDEs like JupyterLab and VS Code for interactive coding.
• Pipeline Orchestration: A drag-and-drop visual editor to build ML workflows (data import $\rightarrow$ preprocessing $\rightarrow$ training $\rightarrow$ evaluation $\rightarrow$ deployment).
• Resource Management: Automated scheduling of heterogeneous hardware (NVIDIA, Huawei Ascend, etc.) and support for RDMA for high-speed distributed training.
• Inference Services: A "zero-code" deployment system for serving models with support for canary releases, autoscaling, and traffic splitting.

Who it’s for

• AI Researchers and Engineers who need a scalable environment for distributed training and rapid prototyping.
• MLOps Teams looking to standardize the ML pipeline, manage multi-tenant resource quotas, and monitor GPU utilization.
• Enterprise AI Teams requiring a centralized platform that supports diverse hardware architectures (ARM, x86) and various AI accelerators.

Highlights

• Heterogeneous Compute Support: Compatible with a wide range of GPUs/NPUs (T4, V100, A100, Ascend, DCU) and ARM architectures.
• Visual Pipeline Editor: Drag-and-drop workflow orchestration with built-in operators for feature processing and model training.
• Integrated Labeling: A built-in data labeling platform supporting image, text, audio, and multimodal data, including LLM-assisted auto-labeling.
• LLM Ecosystem: Dedicated support for large model fine-tuning (via Llama-Factory), distributed inference (vLLM), and a model hub (AIHub) with 400+ pre-trained models.
• Enterprise-Grade Management: Includes SSO, RBAC, multi-tenant resource isolation, and detailed billing/metering for compute usage.