faiss-cuda

GPU-enabled FAISS wheels published to PyPI, maintained by Taylor Geospatial.

pip install faiss-cuda

Built against CUDA 12.8. Your only host prerequisite is an NVIDIA driver (R525+). No system CUDA toolkit required — the CUDA runtime libraries install from PyPI (nvidia-cuda-runtime-cu12, nvidia-cublas-cu12) and are preloaded by a small loader shipped in the wheel.

What's in the wheel

• faiss Python package, GPU-enabled, built against FAISS upstream v1.14.3.
• CUDA arch 70;80;86;89;90 — V100, A100, A10/A30/RTX-30, RTX-40, H100/H200. T4 (sm_75), B100/B200 (sm_100) and RTX-50 (sm_120) return in the next .post release once PyPI grants a larger file limit (pypi/support#11444).
• Works on GPU-less machines too: the wheel contains the full CPU implementation of FAISS and loading the CUDA runtime does not require a driver.
• manylinux_2_28_x86_64, CPython 3.11–3.14. libstdc++ is statically linked so no GCC 11+ runtime is required on the host — works on RHEL 8 / glibc 2.28 out of the box.
• CUDA libs (libcudart, libcublas, libcublasLt) are excluded from the wheel — they install transitively from PyPI as nvidia-* packages, which keeps the wheel small and avoids ABI clashes if you also have PyTorch installed.

To pin the runtime CUDA libs to the same minor as the build (reproducibility):

pip install 'faiss-cuda[fix-cuda]'

Verify

import faiss, numpy as np
res = faiss.StandardGpuResources()
xb = np.random.random((10000, 64)).astype("float32")
xq = np.random.random((5,     64)).astype("float32")
index = faiss.GpuIndexFlatL2(res, 64)
index.add(xb)
D, I = index.search(xq, 4)
print("GPU OK", I.shape)

CUDA 13

A CUDA 13 variant (packages/cuda13, arch 75;80;86;89;90;100;103;120) is maintained in this repo but not published to PyPI — CUDA 12.8 already covers V100 through Blackwell B200/RTX-50, and CUDA 13 dropped Volta/V100 support entirely. Build it from source if you need sm_103 (GB300) or a pure cu13 dependency stack.

Deprecated: faiss-cuda-cu128

Before 1.14.3, faiss-cuda shipped CUDA 13 and faiss-cuda-cu128 carried the CUDA 12.8 build. faiss-cuda is now the CUDA 12.8 build; faiss-cuda-cu128 remains on PyPI only as an empty shim depending on faiss-cuda>=1.14.3 so old pins keep resolving.

Build from source

Wheels are built natively on TGI RAILS (no container). The build script uses the cluster CUDA toolkit and system gcc via modules, and runs auditwheel repair to certify the result as manylinux_2_28_x86_64. ccache persists builds across SLURM jobs.

Prereqs on the build host: CUDA 12.8 toolkit (13.x for the cuda13 variant), RHEL 8 / glibc 2.28 baseline, gcc 11+, uv, openblas-devel.

scripts/submit_matrix.sh --account <PROJECT>   # all (variant, python) pairs
ls wheelhouse/

Or directly on a node:

scripts/rails_build.sh cuda12         # one package
scripts/rails_build.sh                # both

Releasing

1. Build wheels on RAILS: scripts/submit_matrix.sh --account <PROJECT>.
2. Build the cu128 shim: python -m build --wheel packages/cu128-shim -o wheelhouse/.
3. Tag and create a GitHub release, attaching every .whl from wheelhouse/:

   gh release create v1.14.3 wheelhouse/*.whl --notes "..."
   

4. The release.yml workflow fires on release: published, downloads the wheel assets, and runs uv publish with OIDC trusted publishing into the pypi environment. Both faiss-cuda and faiss-cuda-cu128 need trusted-publisher configs on PyPI pointing at release.yml + environment pypi.

Versioning

Package version follows upstream FAISS: <faiss_version> on a fresh upstream bump, then <faiss_version>.postN for packaging-only changes (e.g. FAISS 1.14.3 → 1.14.3, next packaging fix → 1.14.3.post0). The faiss/ git submodule is the source of truth for the upstream version.

License

MIT. See LICENSE.

The wheels redistribute compiled binaries built from FAISS (Meta, MIT-licensed); the LICENSE file includes Meta's copyright and license text alongside ours.