Metadata-Version: 2.4
Name: faiss-gpu-cu12-cuvs
Version: 1.14.0
Summary: FAISS 1.14.0 with GPU + NVIDIA cuVS support — unofficial pip wheel (Linux x86_64, Ampere/SM86+)
Home-page: https://github.com/Gabrieliam42/faiss-gpu-cu12-cuvs
Author: Matthijs Douze, Jeff Johnson, Herve Jegou, Lucas Hosseini
Author-email: faiss@meta.com
License: MIT
Keywords: search nearest neighbors faiss gpu cuda cuvs
Description-Content-Type: text/plain
Provides-Dist: faiss
Requires-Dist: numpy
Requires-Dist: packaging
Requires-Dist: libcuvs-cu12==25.10.0
Requires-Dist: librmm-cu12==25.10.0
Requires-Dist: libraft-cu12==25.10.0
Requires-Dist: rapids-logger
Requires-Dist: nvidia-nvjitlink-cu12>=12.9

FAISS 1.14.0 built from source with full GPU and NVIDIA cuVS support, and AVX2 CPU fallback.

This fills a gap that does not exist elsewhere:
- faiss-gpu-cu12 on PyPI (including 1.14.0): GPU only, no cuVS
- faiss-gpu-cuvs official: conda only, not on PyPI

| Property | Value |
|---|---|
| FAISS version | 1.14.0 |
| Python | 3.12 |
| CUDA | 12.x |
| GPU architecture | Ampere (SM86 native; SM89/90 via PTX JIT) |
| cuVS | Enabled (FAISS_ENABLE_CUVS=ON) |
| CPU SIMD | AVX2 |
| BLAS | OpenBLAS |
| Platform | Linux x86_64 |

GPU Architecture Compatibility:
- SM 86 (native): RTX 3080, RTX 3090, RTX 3090 Ti, A40
- SM 89, 90 (PTX JIT): RTX 4000 series, H100
- SM 80 and below (will NOT work): A100, RTX 2000 series, V100 and older

Installation:

    # 1. Install RAPIDS dependencies (hosted on https://pypi.nvidia.com)
    pip install libcuvs-cu12==25.10.0 librmm-cu12==25.10.0 libraft-cu12==25.10.0 \
        rapids-logger "nvidia-nvjitlink-cu12>=12.9" --extra-index-url https://pypi.nvidia.com

    # 2. Install system dependency
    sudo apt-get install -y libopenblas-dev

    # 3. Install this package
    pip install faiss-gpu-cu12-cuvs --extra-index-url https://pypi.nvidia.com

Note: Conflicts with faiss-cpu and faiss-gpu-cu12. Do not install alongside them.
Note: RAPIDS dependencies (libcuvs-cu12, librmm-cu12, libraft-cu12, rapids-logger,
      nvidia-nvjitlink-cu12) are hosted at https://pypi.nvidia.com and must be
      available via that index when installing.
