Skip to content

CLAUDE.md

Latest commit

 

History

History
38 lines (30 loc) · 1.88 KB

CLAUDE.md

File metadata and controls

38 lines (30 loc) · 1.88 KB
Raw

mobius

Model conversion and optimization hub for edge AI on Apple Silicon. Converts PyTorch/ONNX models to CoreML, targeting ANE (Neural Engine).

Structure

  • models/{class}/{model}/{target}/ — self-contained conversion pipelines (each has own pyproject.toml)
  • tools/coreml-cli/ — CLI tool to profile CoreML models (device assignment, latency, compile time)
  • Documentation/ — guides (ModelConversion.md is the main one)
  • knowledge/ — curated research papers and platform docs

Key Commands

All commands run from the target directory (e.g., models/stt/parakeet-tdt-v3-0.6b/coreml/):

  • uv sync — set up environment
  • uv run python convert-coreml.py --output-dir ./build/<name> — run conversion
  • uv run python compare-models.py --audio-file <path> --coreml-dir <dir> — validate parity

Profiling (from tools/coreml-cli/):

  • uv run coreml-cli path/to/model.mlmodelc — benchmark: latency, compile time, device % across all compute unit configs
  • uv run coreml-cli model.mlmodelc --fallback — ANE optimization: show CPU fallback ops grouped by rejection reason
  • uv run coreml-cli model.mlmodelc --fallback --json — structured fallback analysis for agent parsing

Constraints

  • Trace with .CpuOnly
  • Target iOS 17+ / macOS 14+
  • Fixed input shapes only (no dynamic dimensions)
  • Use uv for all dependency management
  • Each conversion directory is self-contained with its own pyproject.toml

Style

  • 4-space indent, type hints, double-quoted strings
  • Lowercase-kebab-case for directories and files
  • Document trials, errors, and platform quirks — future agents rely on this context

Model Classes

stt, tts, vad, speaker-diarization, emb, segment-text

Workflow

See Documentation/ModelConversion.md for the full 11-step pipeline: directory setup → conversion → validation → profiling → documentation → HuggingFace upload → PR.