mineru-llama-cpp 0.1.3

mineru-llama-cpp 0.1.3 upgrades the native Python binding to the CPython 3.10 stable ABI and ships six wheels reused across standard, GIL-enabled Python 3.10–3.14.

• Linux x86_64/aarch64 and Windows AMD64/ARM64: CPU and optional Vulkan.
• macOS 14+ arm64: CPU and Metal; macOS 14+ x86_64: CPU.
• Backend selection supports auto, cpu, vulkan, and metal. Automatic selection prefers discrete GPUs; missing optional GPU runtimes preserve CPU fallback. CUDA and SYCL are removed.
• Pins llama.cpp to 86a283532072722c5f3363d37d59a874d09fa99b, preserves the Python Engine API, and fixes incomplete UTF-8 tails at generation length limits.
• Linux requires glibc 2.28+; Linux x86_64 retains the x86_64-v3 instruction requirement. Vulkan drivers and loaders come from the system.

Install from PyPI with python -m pip install mineru-llama-cpp==0.1.3, or download the wheel for your platform from this release.

Release wheels are built and tested on all six platforms, with strict stable-ABI, architecture, dependency and deployment-target audits. Hardware-specific GPU inference evidence and historical model/output compatibility checks are documented in docs/abi3-backends-validation.md; CPU-only CI does not establish GPU inference performance.

Local validation of the installed 0.1.3 macOS arm64 wheel: 199 tests passed, strict CPython 3.10 ABI auditing passed, all 11 Mach-O artifacts passed macOS 14 deployment/dependency checks, and a fresh process outside the checkout loaded the packaged CPU backend with external GPU SDK paths removed.

Final six-platform build and aggregate audit: CI 38080665745, commit 5189a17. This commit fixes publishing-test collection after the v0.1.3 tag; package and native sources are unchanged. All 71 publishing checks also passed in a clean environment without the native package. The attached wheels match the PyPI upload from this build, and SHA256SUMS records their hashes.

v0.1.2-rc1 (Windows backend DLL fix, pre-release verification)

Pre-release build of 0.1.2 for verifying the Windows backend-DLL fix (make_cpu_buft_list: no CPU backend found). Identical bits to what will be published to PyPI as 0.1.2. Install the wheel matching your Python version directly — no PATH workaround needed.

Benchmark bundle

Temp release for 11-platform two_step_extract benchmark.