Hardware Starter Packs

Hardware Starter Packs

Section: AI & Inference · Model-Manifests-and-Governance · Home · FFmpeg-Integration

Hardware Starter Packs map machine capabilities to ONNX model sets and execution providers. The app profiles hardware, recommends presets, and shows the first-run dialog and Model Manager. Never show Optimize/readiness banners before downloads succeed.

Architecture and Coordination

The Hardware Starter Pack system is orchestrated by the StarterPackCoordinator, which works in tandem with the HardwareProfilerService. This architecture ensures that model selection is not guessed from GPU names alone but is instead based on structured performance telemetry.

Logic Flow and Provisioning

The system operates through a sequence of discovery, profiling, and provisioning. The HardwareProfilerService benchmarks representative workloads (VAD, ASR, Translation, TTS) to establish baseline timings. Based on these benchmarks, the StarterPackCoordinator identifies which "Starter Pack" (set of models and providers) is most appropriate for the current machine.

flowchart TD
    Start[First Startup Gate] --> Discovery[Hardware Discovery]
    Discovery --> Profiler[Hardware Profiler Service]
    Profiler --> Benchmarks[Run Stage Benchmarks]
    Benchmarks --> Evidence[Generate Evidence JSON]
    Evidence --> Coordinator[StarterPackCoordinator]
    Coordinator --> Recommendation[Recommend Quality Preset]
    Recommendation --> ModelManager[Model Manager Provisioning]
    ModelManager --> Ready[Honest Readiness State]

Loading

The flow ensures that users are not presented with "Optimize" or readiness banners until the necessary models are downloaded and verified against their specific hardware.

Hardware Profiling and Presets

The system uses a structured approach to hardware evaluation. The HardwareProfilerService generates evidence stored locally to inform the application's runtime planning.

Performance Tiers

Benchmarks are run against four critical stages:

• VAD (Voice Activity Detection): Quick inference test.
• ASR (Automatic Speech Recognition): Evaluation of transcription throughput.
• Translation: Matrix smoke tests for language pairs.
• TTS (Text-to-Speech): Evaluation of synthesis latency.

Component	Description	Relevance
Evidence JSON	Benchmarking results stored in %LOCALAPPDATA%\Trackdub\hardware-profiler\	Used for driver/fingerprint invalidation and preset selection.
Quality Presets	Threshold-based configurations (CPU, DirectML, TensorRT RTX).	Maps hardware capabilities to specific execution provider paths.
IRuntimePlanningPreferences	User-facing overrides and service-layer preferences.	Allows manual intervention if the automated recommendation is insufficient.

Execution Provider Strategy

Hardware Starter Packs specifically target platform-appropriate execution providers (EPs). Trackdub follows a strict policy regarding Windows acceleration, prioritizing native integration layers over ad-hoc installs.

Windows Runtime Routing

On Windows, the system prioritizes Windows ML as the integration surface for ONNX Runtime. This allows for catalog EP registration and explicit device selection.

1. TensorRT RTX: Standalone ONNX Runtime EP ABI plugin for NVIDIA RTX hardware.
2. Windows ML Catalog EPs: Used for MIGraphX (AMD), OpenVINO (Intel), or QNN.
3. DirectML: Maintained as the legacy GPU fallback for hardware not supported by newer catalog EPs.
4. CPU Fallback: Always available but marked with performance warnings for heavy synthesis tasks.

graph TD
    subgraph Windows_Inference
    A[Runtime Planner] --> B{Hardware Check}
    B -->|NVIDIA RTX| C[TensorRT RTX Plugin]
    B -->|AMD/Intel| D[Windows ML Catalog EP]
    B -->|Legacy GPU| E[DirectML Fallback]
    B -->|No GPU| F[CPU Execution]
    end
    
    C -.-> G[Starter Pack: RTX Optimized]
    D -.-> H[Starter Pack: Vendor Specific]
    E -.-> I[Starter Pack: Standard GPU]
    F -.-> J[Starter Pack: Baseline CPU]

Loading

Model Governance and Integrity

Starter Packs must adhere to the product's "Commercial-Safe" and "Honest Readiness" mandates. No model can be provisioned by a starter pack without a corresponding manifest entry.

Manifest Requirements

The bundled-models.manifest.json acts as the authoritative source for starter pack provisioning. Every pack must verify:

• License Metadata: Only commercial-safe models are included in shipping starter packs.
• Integrity: SHA-256 checksums are verified post-download.
• Expected Runtime: Models are tagged with their intended execution provider (e.g., windows-ml|onnxruntime-directml).

Starter Packs are explicitly prohibited from downloading end-user runtime dependencies like Python, Conda, or CUDA Toolkits during the setup process. They must rely on bundled native binaries or validated plugin DLL bundles.

Summary

Hardware Starter Packs represent the orchestration of Trackdub's hardware-aware configuration. By utilizing the HardwareProfilerService to generate measurable evidence, the StarterPackCoordinator can transition the application from an unconfigured state to a "Ready" state with validated models and execution providers. This system ensures that the application remains high-performance on Windows, macOS, and Linux while maintaining strict adherence to model governance and commercial safety policies.