Best Practices

This section collects recommended configurations and operational paths for common business scenarios on the Rise Computing Platform, helping users avoid common pitfalls.

End-to-End Operation Guides

The following three guides walk through the complete workflow from environment preparation to production deployment (adapted from the admin manual for version 2.2):

• Inference Service Go-Live: Image upload → model download → deployment template → model registry → start inference → web/API access → operations troubleshooting.
• Container Instance Development Environment: Dev image tagging → create instance → web terminal / VS Code → save image → instance management.
• Distributed Training Jobs: Training image/data/scheduling policy/queue preparation → create job (with torchrun parameter examples) → monitor training → job management.

Security Configuration

• 2FA Security Configuration: Admin setup for SMS/email/MFA two-factor authentication methods, protected API endpoints, and login 2FA; user-side authenticator binding, verification code login, and operation verification.

Billing & Gateway

• AI LLM Gateway Billing Options: Using the platform as an enterprise AI LLM gateway to provide unified access to self-hosted inference instances and external AI channels with differentiated billing and cost allocation across departments. Covers dual billing dimensions (user payable / channel cost), four billing strategies, multi-level scope and price priority, cost reports, and deployment recommendations.

Scenario 1: Provisioning Compute for a New Team

1. Platform administrator creates or onboards clusters under Platform Management > Platform & Members.
2. Creates a new workspace and assigns projects under AI Computing Management > Workspaces & Projects.
3. Configures resource quota and resource pool bindings for the workspace.
4. Invites members to join the workspace/project and assigns roles (e.g., project administrator, developer).
5. Members can then create instances and jobs under AI Computing Development after logging in.

Scenario 2: One-Click Inference Deployment from Model Marketplace

1. Navigate to LLM Services > Model Marketplace and filter for the desired model (by engine, chipset, scenario, pricing).
2. Click "One-Click Deploy," select the target project, spec, and instance count.
3. After deployment completes, check the service status under AI Computing Development > Compute Services > Inference Service.
4. Copy the invocation URL and API Key to call from Playground or local code.

Scenario 3: Training Job Troubleshooting

1. After job submission, monitor the status in the list; if it stays in Pending for a long time, check the detail page Events / Scheduling Events.
2. Common causes:
   • Insufficient resource pool quota → contact the administrator to expand or adjust quota.
   • Image pull failure → verify the image address, credentials, and network.
   • Incorrect start command → check the Logs tab to fix the start command.
3. Once the job is running, enable TensorBoard to monitor training curves.

Scenario 4: Data Preparation Workflow

1. Create a dataset and upload files under AI Computing Development > Data Management > Datasets.
2. If data quality is poor, create a Data Cleaning job (based on Data-Juicer).
3. If sample expansion is needed, create a Data Augmentation job (LLM-based generative augmentation).
4. After cleaning/augmentation completes, a new dataset version is produced and can be referenced directly in fine-tuning jobs.

Scenario 5: API Channel Cost Reconciliation

1. Platform administrator connects external channels under AI Computing Management > AI Gateway > API Channels.
2. View channel overview and call records under Metering & Billing > API Channels.
3. Compare upstream bills with the platform-recorded cost_microcents to identify discrepancies.
4. To adjust procurement prices, maintain channel-level pricing under Channel Cost Strategy.

General Recommendations

• Naming conventions: Use semantic names for instances/jobs/services to facilitate later search and cost attribution.
• Regular cleanup: Release unused instances, services, and jobs promptly to avoid resource waste.
• Recycle bin safety net: Recover accidentally deleted resources from the recycle bin; only permanently delete after confirming they are no longer needed.
• Quota alerts: Configure resource allocation/utilization alerts in Alert Center to detect bottlenecks early.