Installation Guide
Version History
- Starting from v0.9.0, CSGHub will no longer support Gitea as the git backend.
- Starting from v1.1.0, the Temporal component is added as the asynchronous/scheduled task executor.
- Starting from v1.3.0, CSGHub removed Gitea from the docker-compose/helm-chart installers.
- Starting from v1.6.0, Space Builder was removed, and its functions were inherited by the runner.
- Starting from v1.7.0, Starship is integrated internally into CSGHub.
- Starting from v1.8.0, the Notification service is added.
- Starting from v1.9.0, CSGHub Helm Chart CE and EE versions are merged.
- Starting from v1.14.0, XNet storage beta testing is enabled.
- Starting from v1.16.0, Envoy-Gateway is used to replace ingress-nginx.
- Starting from v1.17.0, ArgoWorkflow, KnativeServing, and LeaderWorkSet are managed directly by the Helm Chart.
- Starting from v2.0.0, the agent-sandbox component is added.
Introduction
CSGHub is an open-source, trusted LLM asset management platform that helps users govern assets (datasets, model files, code, etc.) involved in the lifecycle of Large Language Models and their applications. Based on CSGHub, users can manage assets including model files, datasets, and code through a Web interface, Git command line, or natural language Chatbot for operations such as uploading, downloading, storing, verifying, and distributing. Meanwhile, the platform provides microservice sub-modules and standardized APIs for easy integration with existing systems.
CSGHub is committed to providing users with an asset management platform natively designed for LLMs that supports private deployment and offline operation. CSGHub offers Hugging Face-like functionality for private environments, managing LLM assets in a similar way to how OpenStack Glance manages VM images, Harbor manages container images, and Sonatype Nexus manages artifacts.
For more information about CSGHub, please refer to:
- Portal: https://github.com/OpenCSGs/csghub
- Server: https://github.com/OpenCSGs/csghub-server
- Installer:
Deployment Methods
Currently, CSGHub mainly provides the following installation methods. Users can choose the most suitable deployment plan based on their environmental configuration, deployment scale, and usage requirements. All deployment methods support private deployment to ensure data security and environmental isolation.
| Deployment Method | Infrastructure | Remarks | Target Scenarios |
|---|---|---|---|
| Docker Compose | Docker | Lightweight deployment, no complex environment configuration required. | Local debugging, functional demos, small-scale trials; ideal for quick start. |
| Helm Chart | Kubernetes | Standardized microservices deployment, supports elastic scaling. | Production environments, large-scale deployments, high-availability needs; suitable for enterprise-grade use. |
| Quick Install | K3s | One-click automated deployment, lightweight Kubernetes environment. | Single-machine deployment, no professional K8s O&M experience required, needs rapid setup. |
| Air-Gapped | K3s | Offline deployment, no public internet access required (Coming Soon). | Classified environments, government/enterprise scenarios without internet permissions; focuses on data isolation and security. |
Pre-deployment Instructions
- Version Compatibility: Please pay attention to component changes in corresponding versions (refer to the Version History above) to avoid deployment failures caused by removed or added components. It is recommended to prioritize the latest stable version for better features and compatibility.
- Environment Requirements: Different deployment methods have different basic environment requirements. Before deployment, ensure the server meets the installation conditions for the corresponding infrastructure (e.g., Docker, Kubernetes, K3s). Detailed requirements can be found in the specific documentation for each method.
- Resource Preparation: Plan server resources such as CPU, memory, and storage in advance based on the deployment scale to ensure platform stability. For large-scale deployments or scenarios involving model training and inference, it is recommended to upgrade hardware configurations and enable GPU support.
- Network Description: In non-air-gapped deployment modes, the server must have public internet access to pull the required images and dependent resources. The air-gapped mode allows distribution of resources through internal private repositories without requiring a public connection.