Core Concepts

Before using the Rise AI Compute Platform console, we recommend familiarizing yourself with the following core concepts, which apply across compute allocation, task submission, service deployment, and billing analysis.

Multi-Tenant Model

Workspace
    └── Project (Namespace)
            ├── Quota
            ├── Tenant DataStorage
            ├── Image / Repository
            └── Model

• Workspace: The tenant boundary — members, quotas, and resource pool assignments are all scoped to the workspace.
• Project: Maps to a Kubernetes Namespace; it is the actual owner of tasks, services, and datasets.
• Namespace = Project name: Permission checks are enforced at the project granularity; the admin account bypasses all checks.

The overall relationship between tenants and compute resources is as follows: tenants contain multiple projects, and projects can use underlying heterogeneous compute nodes via shared pools, dedicated pools, per-card, or per-node allocation:

Compute Resource Hierarchy

Level	Purpose
Physical GPU Card	Actual GPU card on the node, managed by the HAMi device plugin.
ResourcePool	Organizes cards in shared/dedicated/oversubscription modes; can be bound to nodes or individual cards.
Flavor	Defines CPU/memory/architecture/vGPU/compute percentage/VRAM — selected directly when creating tasks.
Quota	Maximum resource limits for a workspace/project, adjustable dynamically.

Task Types

Task	Description
Container Instance (Notebook)	Interactive development environment supporting Jupyter / VS Code / CloudShell / SSH.
Training Job (TrainingJob)	Distributed training supporting PyTorch / TensorFlow / MPI / DeepSpeed.
Fine-Tuning Job (FineTuningJob)	Based on LLaMA-Factory, supporting LoRA / QLoRA / Full / PPO, etc.
Evaluation Job (EvaluationJob)	Associated with fine-tuning jobs in a 1:N relationship; supports Auto-Eval and LLM-as-Judge.
Inference Service	Online model service supporting vLLM / SGLang / MindIE / vLLM-Ascend.
Custom Service	General-purpose long-running HTTP service based on Deployment + Ingress.

Data and Models

• Dataset: Multi-version datasets supporting file upload/download/preview/edit, with ModelScope / HuggingFace Hub import.
• Model Repository: Platform model registry supporting listing/delisting, price display, and one-click deployment.
• Model-X: Large model service module including the model marketplace, Playground, API Key, and personal usage.

Gateway and Billing

• AiProvider (API Channel): External model channels supporting OpenAI / Anthropic / Zhipu protocols, etc.
• AiRoute (AI Routing): Combination of model predicates, path predicates, and upstream channels, with degradation and rate limiting.
• Token Governance: User-payable perspective, aggregated by model / workspace / api_key.
• API Channel Costs: Channel cost perspective, aggregated by provider / upstream model.