Description

This PR consolidates the Group Multi-Model Deployment and Stacked Model Deployment features into the main branch.
The combined work enables deployment of multiple models (base + fine-tuned) on a single container instance, introduces validation and configuration updates, and includes comprehensive SDK, UI, and CLI enhancements to support both deployment and inferencing for stacked model groups.

Key Deliverables

1. Group Model Deployment

• Added multi-model deployment configuration support in AQUA SDK, UI, and CLI.
• Implemented logic to launch and manage multiple model instances in a single container via MD Groups.
• Introduced backend validation for compatible shapes, GPU allocations, and container configurations.
• Integrated multi-model evaluation and inferencing support.

2. Stacked Model Deployment

• Introduced Stacked Model Deployment for base + fine-tuned LoRA weights.
• Added dynamic merging of adapter weights in the VLLM SMC container.
• Implemented logic for versioned fine-tuned models (FT-V1 → FT-V2) migration and compatibility.
• Extended AQUA SDK/CLI to support stacked deployment lifecycle: create, edit, and infer.
• Added validator to detect fine-tuned model versions and apply correct deployment path.
• Implemented “Zero-Time Update” mechanism for fine-tuned weight management via CLI.

Testing

• Unit and integration tests added
• E2E tests covering both deployment and inferencing paths