PATCHES.md

Local patches for ZSWatch

This directory vendors fllama into the repository so iOS fixes live in source control instead of the pub cache.

Why this fork exists

The app uses both fllama and whisper_ggml_plus on iOS. Both dependencies vendor ggml/llama headers, which caused CocoaPods/Xcode header resolution conflicts during iOS device builds.

Local changes

1. src/llava.h
   • Use explicit platform-specific relative includes for iOS/macOS ggml and llama headers.
2. src/clip.cpp
   • Use explicit platform-specific relative includes for iOS/macOS ggml* and gguf headers.
3. ios/llama.cpp/include/llama.h
   • Use explicit relative includes for ggml headers.
4. ios/llama.cpp/common/log.h
   • Use explicit relative include for ggml.h.
5. ios/fllama.podspec
   • Set iOS minimum deployment target to 14.0.
   • Disable Xcode header maps for the fllama pod target to keep ggml headers isolated.
6. Force CPU-only inference at runtime.
   • src/fllama.cpp: Always set n_gpu_layers = 0 (removed simulator conditional and llama_supports_gpu_offload() checks). Metal is still compiled in but never used.
   • Deleted ios/default.metallib (pre-compiled Metal kernels — Metal shaders are now embedded in the library via GGML_METAL_EMBED_LIBRARY).

Maintenance

• Prefer rebasing these changes onto upstream fllama updates.
• If these fixes are accepted upstream, switch zswatch_app/pubspec.yaml back to the upstream dependency.