refactor(qhexrt): remove model catalog registration

Closed
SH
3mo ago
+211-2094

Summary

This PR simplifies QHexRT model registration to match the current architecture:

  • Removes the engine- and SDK-owned QHexRT model catalog.
  • Removes device-specific model selection and architecture-folder URL rewriting.
  • Removes QHexRT catalog registration helpers from the Kotlin, Flutter, and React Native packages.
  • Keeps probeNpu() and isArchitectureSupported() as optional device-support checks.
  • Registers exact, caller-provided model URLs through each core SDK's existing RunAnywhere.registerModel(...) API.

Why

PR #532 followed the older QHexRT flow: seed logical catalog entries, detect the device architecture, filter models, and rewrite URLs to architecture-specific folders.

Based on the feedback on #532, current QHexRT models are published as dedicated artifacts. QHexRT should therefore not maintain a model catalog, choose a model, or modify its URL. The caller selects the model and registers its exact link through the core SDK.

This keeps model registration in the shared, backend-neutral path. QHexRT remains responsible for backend registration and optional device-capability reporting.

What changed

  • Deleted the native QHexRT catalog implementation, build wiring, and catalog tests.
  • Removed catalog policy and device-aware registration APIs from Kotlin, Flutter, and React Native QHexRT packages.
  • Updated Android, Flutter, and React Native examples to use explicit URL registration through the core SDK.
  • Removed QHexRT-specific HF authentication checks; authentication requirements now come from model metadata.
  • Updated documentation and removed tests for the retired catalog APIs.

Behavior after this PR

  • QHexRT does not seed or own a model catalog.
  • QHexRT does not select models or rewrite model URLs.
  • Applications register dedicated QHexRT model URLs through RunAnywhere.registerModel(...).
  • Applications may use probeNpu() as an optional preflight before presenting or registering NPU models.
  • Existing core download, registry, lifecycle, and compatibility behavior remains unchanged.

Breaking changes

The QHexRT catalog lookup, model eligibility, HF-auth policy, and device-aware model-registration APIs have been removed. Callers using those APIs should move to explicit model registration through the core SDK.

Validation

All PR checks pass, including:

  • Kotlin/Android SDK and example tests
  • Flutter analysis and tests
  • React Native typecheck, lint, and tests
  • Native Android, iOS, macOS, Linux, Windows, Web, and WASM builds
  • Repository validation, security, and centralization checks

Related

  • Fixes #523
  • Supersedes the catalog-based approach proposed in #532

Summary by CodeRabbit

  • New Features

    • Model registration now supports richer metadata, including download size, context length, source, and description.
    • Android, Flutter, and React Native examples can register explicit QHexRT model URLs through the core model registry.
    • QHexRT availability and architecture support can be checked before registration.

  • Breaking Changes

    • Removed QHexRT device-specific model registration and model catalog helper APIs.
    • QHexRT now focuses on NPU probing, architecture support, and backend lifecycle management.