feat: add llamacpp 0.1.0 C++23 module package

[wellwei]

Summary

• add mcpplibs.llamacpp 0.1.0 as a minimal Form A descriptor backed by the released llama.cpp-m archive
• add CPU and macOS ARM64 Metal consumers that import the public llamacpp C++23 module and execute real API assertions
• supersede feat(pkg): add ggml-org.llamacpp b10069 C++23 module #132 by keeping the llama.cpp adaptation, vendored checkpoint, generated API, backend selection, and update tooling in mcpplibs/llama.cpp-m

Release identity

• package release: llama.cpp-m v0.1.0
• immutable tag archive SHA-256: 9af34d44349e520f2f9bd2eace29eb6a634ab4633a7f3bafa225f77be178ca84
• package version mapping: llama.cpp-m 0.1.0 -> llama.cpp b10069 (178a6c44937154dc4c4eff0d166f4a044c4fceba)
• exact release commit: 8876f2cec7294a938fadb50dfbd9296dff6d0beb
• package CI on that commit: all jobs passed, including real CPU inference on Linux x86_64, Linux ARM64, macOS ARM64, and Windows x86_64, plus real Metal inference on macOS ARM64
• release workflow: passed, including immutable mapping and exact-commit CI gates

The archive contains its own mcpp.toml, so the index descriptor intentionally omits the mcpp field. There is no authorized byte-identical CN mirror yet; the descriptor uses the plain upstream GitHub URL.

Use

Consumers do not download or unpack the Release manually:

mcpp add llamacpp@0.1.0

import std;
import llamacpp;

int main() {
    llama_backend_init();
    auto params = llama_model_default_params();
    llama_backend_free();
}

CPU is enabled by default. Enable Metal on macOS ARM64 in mcpp.toml:

[dependencies]
llamacpp = { version = "0.1.0", features = ["backend-metal"] }

Supported boundary

• CPU: Linux x86_64, Linux ARM64, macOS ARM64, Windows x86_64
• Metal: macOS ARM64 only, via backend-metal
• public API: import llamacpp; exposing the audited llama.cpp C API surface
• not included in 0.1.0: mtmd, CUDA, Vulkan, RPC, or other upstream backends

Update policy

llama.cpp-m does not chase every daily llama.cpp commit. Each package version pins one reviewed upstream checkpoint and keeps that mapping immutable. A new wrapper/index version is cut when an upstream change has consumer value, such as required model/API support, a relevant correctness or security fix, or a measured backend improvement.

Each update imports a new checkpoint, regenerates and audits the module API, runs cold package builds and real inference across the supported matrix, then publishes a new llama.cpp-m version. The index appends that version; it never changes the URL or SHA-256 of 0.1.0. The repository's scheduled upstream check reports drift without automatically publishing unreviewed snapshots. See the full update policy.

Validation

• mcpp 0.0.109 xpkg parse pkgs/l/llamacpp.lua: Form A, parse OK
• Lua syntax, mirror policy, atomic package name, cross-package references, and git diff --check: passed
• RED before descriptor: local checkout resolution reported no (mcpplibs, llamacpp) package
• CPU index member: downloaded the tag archive, compiled/linked the module, initialized the CPU backend, created/free'd real samplers, and printed LLAMACPP_INDEX_TEST=PASS
• Metal index member: enabled backend-metal, loaded the embedded Metal library, registered MTL0 (Apple M4), and printed LLAMACPP_METAL_INDEX_TEST=PASS
• fresh mcpp new consumer: mcpp add llamacpp@0.1.0 wrote llamacpp = "0.1.0", then cold-downloaded, built, and ran the package
• fresh Metal consumer with Qwen3.5-4B Q4_K_M: offloaded 33/33 layers, allocated a 2571.63 MiB Metal model buffer and 10.64 MiB Metal compute buffer, then decoded/sampled 32 tokens at 23.43 tokens/s on Apple M4
• affected-member scan finds both llamacpp and llamacpp-metal

No index.toml or workflow pin change is required; the existing mcpp floor remains 0.0.109.

[TPPPP72]

看起来不错

[Sunrisepeak]

感觉0.0.2上面可以加几个模板功能, 这样就可以

mcpp new xxx --template llamacpp

具体可以参考 这个 imgui 模块化的封装层里的 模板示例

• https://github.com/mcpplibs/imgui-m/tree/main/templates