v0.4.0

pylocalvqe v0.4.0

pylocalvqe v0.4.0 is a single Linux x86_64 wheel that contains the C++ core (flat C ABI and a header-only C++ wrapper), the CMake package configuration, and the nanobind Python extension. It includes both the CPU and the CUDA 13 backends of ggml.

Assets

• pylocalvqe-0.4.0-cp312-abi3-manylinux_2_28_x86_64.whl
• SHA256SUMS (SHA-256 of the wheel)

The same file is published on PyPI as pylocalvqe==0.4.0.

Source

• tag v0.4.0, source commit c70de0eaac360fd6fcc2bf993b647beb78ae5f8d
• LocalVQE 35b116d5eb059d552fa46fac3ce2963ed13ce153 (Apache-2.0)
• ggml c044a8eeae2591faa0950c8b5e514cbc4bbfc4ca (MIT), with the tracked ggml-gru.patch

Build

• image quay.io/pypa/manylinux_2_28_x86_64@sha256:076b19ef6e698f28649fa817c49d14443201e01da85240e8e52f93e8daa49bff
• GCC 14.2.1 (gcc-toolset-14), CMake 4.4.4, scikit-build-core 1.1.1, nanobind 3.1.0
• CUDA toolkit 13.3 (nvcc V13.3.73); real cubin for sm_86, sm_89 and sm_120a, no PTX
• GGML_CUDA=ON, GGML_OPENMP=OFF, GGML_CUDA_NO_VMM=ON, GGML_CUDA_FA=OFF
• two builds from the same tag produced byte-identical wheels

Requirements

• wheel tag cp312-abi3-manylinux_2_28_x86_64: CPython 3.12 or later, glibc 2.28 or later. The highest glibc symbol version referenced by the core and the extension is GLIBC_2.27
• numpy 1.26 or later
• CPU with SSE4.2, AVX, AVX2, FMA, F16C and BMI2. No OpenMP runtime is required
• On Linux x86_64 pip installs nvidia-cuda-runtime==13.3.29 and nvidia-cublas==13.6.0.2. cuBLAS pulls NVRTC transitively (verified with nvidia-cuda-nvrtc==13.4.92)
• The wheel does not bundle the CUDA user-space libraries or the NVIDIA driver. The host provides the driver
• Without an NVIDIA driver or a supported GPU the package still imports and processes audio with device="cpu"; device="cuda" raises LocalVQEError

GPU

• CUDA 13.3 targets Linux driver 610.43.02 (the CUDA 13.x minor compatibility floor is 580). This wheel was not verified on drivers below 610.43.02, so the supported minimum is 610.43.02
• Verified: WSL2 (Linux 6.18 x86_64) with NVIDIA RTX PRO 4000 Blackwell (sm_120) and Windows driver 616.92; the CPU and CUDA outputs of the three models (125 hops each) agreed within a maximum absolute difference of 1e-3. A native Linux driver was not verified separately

Breaking changes from v0.3

• LocalVQE(...) no longer takes library_path; the native core ships inside the wheel
• The native shared libraries are no longer separate Release assets; use the wheel

Verification

• the two builds from the tag are byte-identical (SHA-256 31c787bc495408ecd4ed97ed8e1abe0585da1b4cb782a627da078b420ab3e428)
• driverless container (Debian 12, no driver, no GPU): 68 passed, 0 skipped, 0 failed
• GPU host: 68 passed, 0 skipped, 0 failed
• Ubuntu 26.04.1 LTS without a compiler, CMake, CUDA toolkit, driver or GPU: install, CPU processing and the LocalVQEError for device="cuda" passed
• exports are limited to the public pylocalvqe_* C ABI and PyInit__core; no PTX; no bundled CUDA, driver, OpenMP or ggml libraries