Skip to main content
Siverse is a one-stop AI training and inference platform based on a GPU cluster, designed for algorithm R&D and model engineering teams that require large-scale computing power. The platform supports the complete AI development lifecycle: obtaining computing resources on demand, preparing data and runtime environments, interactive development, submitting distributed training jobs, managing models, and publishing models as online inference services. You can also directly deploy existing models or custom container applications without starting from the training stage.

​
Functional Architecture

Siverse is built on a containerized GPU cluster and organizes workloads, assets, resources, and operations and maintenance capabilities around user workflows. Siverse product architecture The platform is responsible for resource scheduling, environment provisioning, status recording, and runtime monitoring; code, data, model adaptation, and business result validation must still be handled by the user. For example, the platform can restart a training job, but the training program must save and load checkpoints itself in order to resume completed training progress.

​
Core Features

​
Resource and Asset Preparation

  • Resources: Manage instance resources, storage resources, resource pools, quotas, orders, and resource usage statistics.
  • Images: Maintain container images used for development, training, and inference, standardizing framework versions, system dependencies, Python packages, and service startup environments.
  • Dataset: Provides data that can be selected and mounted for reading in development environments or training jobs.

​
Development and Training

  • Develop: Provides JupyterLab and VSCS for code development, data exploration, dependency validation, and small-scale debugging.
  • Training: Submit and manage training jobs such as PyTorchJob and RayJob; view status, events, logs, and monitoring; and configure scheduling and fault tolerance.

​
Models and Inference

  • Models: Register built-in and custom models, and manage model versions, sources, and reference information.
  • Inference Service: Publish models or custom container applications as online services, supporting LLM inference service and general service.

​
Operations and Governance

  • Alerts: Subscribe to platform events and metric changes, and send them through notification channels.
  • Config Management: Manage tenant-level configuration, including project groups, tags, task environment variables, and monitoring dashboards.

​
Developer Integration

  • Developer Integration: Access platform capabilities through the Python SDK or MCP for scripting, automation, or AI tool scenarios.

​
Next Steps