Flash-iOS

Few tweaks for @alexintosh (macOS, iOS memory and Fanout)

Based on: https://github.com/Alexintosh/flash-moe/tree/feature/ios-app/FlashMoE-iOS

Changes from upstream

• macOS compatibility — #if os(iOS) / #if os(macOS) guards for platform-specific APIs (UIKit colors, toolbar placement, memory queries, os_proc_available_memory)
• iOS memory entitlements — extended-virtual-addressing + increased-memory-limit for running large MoE models on iPhone
• Fanout I/O — Ported chunked pread from desktop engine: split each expert read into N page-aligned chunks for parallel SSD reads. Configurable via UI (Off / 2 / 4 / 8 chunks)
• Race condition fix — Async pread uses GCD dispatch_group (not pthread pool) to eliminate generation counter conflicts
• Tiered validation fix — async_pread_wait validates each chunk against its own size (not uniform 4-bit size), fixing silent cold expert skipping
• No mmap on iOS — Expert layer files use pread-only path, saving virtual address space
• Model reload fix — Full cleanup on unload (weight file, layer cache, tracking arrays, tensor hash table) so switching models doesn't crash
• iOS storage protection — Model files marked isExcludedFromBackup to prevent iOS from purging them
• Models & Settings — Menu button to return to model list / I/O settings from chat
• Copy script — copy_model_to_iphone.sh to push models to device over USB cable with auto-detect, ETA, and per-file progress

Copy model to iPhone

Connect your iPhone via USB cable, then:

# Auto-detects connected device
./copy_model_to_iphone.sh /path/to/model-directory

# Or specify device UDID
./copy_model_to_iphone.sh /path/to/model-directory <device-udid>

The script copies all model files (config, weights, vocab, expert layers) file-by-file to the app's Documents container, with transfer speed and ETA display.