@elizaos/ios-native-deps

Cross-build harness for the native dependencies that on-device iOS runtimes (Capacitor bun-runtime, full Bun engine port) need to link against:

• llama.cpp → llama.cpp/dist/LlamaCpp.xcframework (device + simulator slices, Metal embedded, baked from the milady-controlled elizaOS/llama.cpp fork — carries Q4_POLAR / QJL1_256 / TBQ4_0 / TBQ3_0 GGML types + Metal/Vulkan/CUDA kernels and DFlash spec-decode on top of stock upstream).
• sqlite-vec → static lib + headers (vector storage + KNN extension that replaces pgvector for the on-device SQLite backend; PGlite cannot run inside JSContext on iOS 16.4+ because WebAssembly is gated off).

This package replaces the pre-2026-05-13 layout at milady/native/ios-bun-port/vendor-deps/, which duplicated this tooling inside the milady workspace. Per the "milady-depends-100%-on-eliza, eliza-is-ejectable-via-submodule" architectural rule, the build harness now lives in eliza and is consumed either as a workspace symlink (MILADY_ELIZA_SOURCE=local) or as the published npm package (MILADY_ELIZA_SOURCE=packages).

Pins

See VERSIONS. Each line is <dep>=<ref> where <ref> is anything git fetch accepts (tag, branch name, or commit SHA). For llama.cpp, the default points at the elizaOS/llama.cpp fork's main branch tip; the LLAMA_CPP_REPO env var can override to upstream ggml-org/llama.cpp for an A/B parity check.

Build

# From the package dir:
bun run build:llama-cpp                 # both slices + xcframework
bun run build:llama-cpp:device          # device slice only
bun run build:llama-cpp:simulator       # simulator slice only
bun run clean:llama-cpp                 # nuke dist/ + build/

# Override source repo (parity check, fork-regression bisect, etc.):
LLAMA_CPP_REPO=https://github.com/ggml-org/llama.cpp \
  bun run build:llama-cpp

# Override iOS deployment target (default 15.0):
MILADY_IOS_MIN_VERSION=16.4 \
  bun run build:llama-cpp

# From a consumer package (workspace or installed):
bun run --filter @elizaos/ios-native-deps build:llama-cpp

The llama.cpp build produces:

llama.cpp/dist/
├── ios-arm64/                       device slice
│   ├── libllama.a                   ~80 MB; combined llama + ggml + common + shim
│   └── Headers/{llama.h, ggml.h, LlamaShim.h}
├── ios-arm64-simulator/             Apple-Silicon Mac simulator slice
│   └── (same layout)
└── LlamaCpp.xcframework/            ← consumed by Swift Pods (vendored_frameworks)

See llama.cpp/README.md for the full xcframework story including the Swift shim contract (shim/LlamaShim.h), the flag-by-flag CMake surface, and the verification recipes (nm -gU, otool -l).

Consumers

Prerequisites

• macOS host with full Xcode (Command Line Tools alone won't ship the iOS SDK or xcrun --sdk iphoneos).
• cmake >= 3.21 (brew install cmake).
• Disk: ~3 GB during build, ~250 MB after clean.
• Time: ~5–8 min on M3, ~12 min on M1.