Swift-zenz-CoreML-APP

🇯🇵 Swiftで Skyline23/zenz-coreml の Core ML アーティファクトをビルド時に取り込み、実機で推論性能をベンチマークするためのサンプルリポジトリです。
🇰🇷 Swift에서 Skyline23/zenz-coreml Hugging Face 리포지토리의 Core ML 아티팩트를 빌드 시점에 받아와 번들에 넣고, 실기기에서 추론 성능을 벤치마크하기 위한 샘플 리포지토리입니다.
🇺🇸 Sample repository for pulling the Core ML artifacts from Skyline23/zenz-coreml during the build and benchmarking them on real iOS devices.

Artifact Source

• Hugging Face model repo: Skyline23/zenz-coreml
• Manifest: hf_manifest.json
• Stateless FP16: Artifacts/stateless/zenz-stateless-fp16.mlpackage
• Stateless 8-bit: Artifacts/stateless/zenz-stateless-8bit.mlpackage
• Stateful: Artifacts/stateful/zenz-stateful-fp16.mlpackage

Build-Time Bootstrap

The app no longer fetches model artifacts at runtime.

• The Xcode build phase hydrates Resources/Artifacts, Resources/tokenizer, and Resources/hf_manifest.json before compilation.
• If those files already exist, the bootstrap skips the network step.
• If the network is unavailable or the download fails, the build still continues.
• At runtime the app reads only from bundled Resources; if the artifacts are missing, the app reports that instead of trying to download them itself.
• Round 1 benchmark numbers remain valid as legacy bundled-model results; new benchmark rounds measure the single-stateful pipeline staged during the build.

ベンチマーク (Core ML greedy decoding) / 벤치마크 (Core ML greedy decoding) / Benchmarks (Core ML greedy decoding)

Detailed benchmark material is organized as:

• Legacy Round 1 results: iPhone 12 details
• Legacy Round 1 results: iPhone Air details
• New HF-backed benchmark plan: Round 2 single-stateful plan