swift-lm：Apple Silicon 上的原生 LLM 推理框架

swift-lm is a high-performance LLM inference framework built with Swift and Metal for Apple Silicon. It supports direct loading of Hugging Face models without model-specific Swift type definitions, enabling efficient native inference on macOS/iOS for edge computing and local AI applications.

As LLMs gain traction in mobile/edge scenarios, Apple Silicon developers need native solutions. Traditional Python/CUDA approaches are not ideal for macOS/iOS. swift-lm addresses this gap as a Swift package using Metal compute shaders, supporting direct Hugging Face model loading without custom Swift types.

swift-lm uses a layered architecture: LMIR (backend-agnostic intermediate representation), LMArchitecture (DSL + validation), ModelDeclarations (predefined models), MetalCompiler (protocol-driven backend), and SwiftLM (user API). LMIR abstracts compute logic (e.g., AttentionAttributes), MetalCompiler uses protocol scheduling, and DispatchOptimizer offers pluggable strategies (NoOptimizer:242, Standard:179, Aggressive:144).

Usage example: Load Hugging Face repo (e.g., LiquidAI/LFM2.5-1.2B-Instruct) with ModelBundleLoader and generate text. Key highlights: 1. Zero model-specific code (supports standard config/safetensors/tokenizer). 2. Dynamic context awareness (adjusts based on KernelContext). 3. Extensible components (3 steps: OperationAttributes → ModelComponent → MetalKernelFragment).

swift-lm is suitable for: macOS/iOS local AI apps (pure Swift, Xcode-friendly), edge inference (uses Neural Engine/GPU), privacy apps (local data processing), and quick prototyping (Hugging Face integration).

swift-lm provides a modern, scalable framework for Apple's ecosystem with layered/protocol-driven design. It paves the way for supporting more hardware (Neural Engine, TPU) and helps developers build performant native AI apps on macOS/iOS.