raids-lab/crater

What it solves

Crater provides a Kubernetes-native control plane to manage shared AI computing clusters. It eliminates the complexity of using raw Kubernetes for multi-tenant GPU resource management, replacing manual YAML manifests with a unified interface for governing users, quotas, and AI workloads.

How it works

Built on top of Kubernetes and Volcano, Crater adds an operational layer that connects users, accounts, and queues to the underlying compute resources. It manages the full lifecycle of AI workloads—including LLM training, inference services, and interactive development environments—while providing a centralized system for managing datasets, models, and container images.

Who it’s for

It is designed for universities, research institutes, enterprise AI teams, and platform engineers who need to share a single GPU cluster across multiple teams, students, or projects.

Highlights

• Multi-Tenant Governance: Manages accounts, quotas, and approvals to ensure fair and accountable resource distribution.
• Policy-Aware Scheduling: Uses Volcano for queue-based admission and priority-aware execution across heterogeneous accelerators (NVIDIA and others).
• Interactive Workspaces: One-click deployment of containerized Jupyter, WebIDE, and terminals.
• Asset Management: Centralized organization of models, datasets, and images for reuse across workloads.
• AI-Assisted Operations: Provides a web console, CLI, and agent-oriented interfaces for automated and intelligent cluster management.