Source Layout Reference

This page is a concise reference for where each major subsystem lives.

Public package targets

MereRunCore

Shared inference and model-management code.

Main areas:

• ACEStep/: music generation
• CodeGen/: code-generation support
• Embeddings/: embedding support
• Flux2Klein/: Klein image family
• LTX/: video generation
• LightOnOCR/: OCR
• LoRA/: LoRA support
• MLX/: MLX utilities
• Gemma4/, LFM2/, MeBot/, Psi/, Q35/: text model families
• QwenImageEdit/: image editing
• Support/: model paths, manifests, resolver, config helpers
• Training/: training-specific support retained in the package
• VLM/: vision-language model support
• ZImageI2L/, ZImageTurbo/: image-family support

Linux CLI compatibility should enter through reusable core surfaces here, not through app bundle code. Media-tool discovery belongs behind typed APIs and should honor MERERUN_FFMPEG, MERERUN_FFPROBE, and then PATH.

AudioCore

Shared audio types and utilities.

AudioCodecs

Audio codecs and conversion helpers.

Keep Linux audio/video probing fixture-sized: use ffmpeg and ffprobe stubs or tiny generated files in tests, not real model checkpoints.

AudioSTT

Speech-to-text backends.

• Parakeet/
• Qwen3ASR/

AudioTTS

Text-to-speech backends.

• Qwen3TTS/
• TTS/

MereRunCLI

The public executable target.

• Commands/
• Support/

This is the product Linux compatibility work should exercise.

MereRunApp

The optional SwiftUI studio target. It stays macOS-only and should not become a Linux compatibility dependency.

Tests

Tests/MereRunCoreTests

Use this for:

• runtime behavior
• model resolution
• subsystem-level integration

Tests/MereRunCLITests

Use this for:

• command parsing
• public CLI UX contracts
• help and command-tree expectations

Operational files

• Package.swift: public package definition
• scripts/check.sh: main validation gate
• scripts/e2e_smoke.sh: installed-model smoke runner
• vendor/llama.xcframework: vendored native dependency for code and API paths
• vendor/mlx-swift_Cmlx.bundle: macOS MLX shader resources; hosted Linux CI should use CPU MLX-sized fixtures, while Linux arm64 package validation needs the CUDA path on real arm64 CUDA hardware