Here's a comprehensive comparison of memory-mapped model weights versus copying all weights into private buffers on macOS:

## Memory-Mapped Weights (mmap)

### How It Works
```swift
// Example: Memory-mapping a model file
let fileHandle = FileHandle(forReadingAtPath: modelPath)
