libcuvs-cu12

26.4.0 verified Mon Apr 27 auth: no python

cuVS C++ shared library for GPU-accelerated vector search (includes CAGRA, IVF-PQ, HNSW). Version 26.4.0. Released as part of RAPIDS monthly releases.

pip install libcuvs-cu12

Common errors

error ModuleNotFoundError: No module named 'libcuvs_cu12' ↓

cause Importing the C++ backend directly instead of the Python wrapper.

fix

Install and import the cuvs package: pip install cuvs ; import cuvs

error ImportError: libcuvs_c.so: cannot open shared object file ↓

cause Missing libcuvs-cu12 installation or broken environment.

fix

Reinstall libcuvs-cu12: pip install --upgrade libcuvs-cu12

Warnings

breaking HNSW GPU hierarchy is now default as of v26.04.00. Existing HNSW parameters may need adjustment. ↓

fix Review HNSW build parameters; refer to migration guide.

breaking Static linking of libcudart is now default (v26.04.00). May conflict with existing CUDA runtime libraries. ↓

fix Set CUDA dynamic linking if needed; see RAPIDS docs.

breaking CAGRA C API enums changed to stable values (v25.12.00). Old enum values no longer work. ↓

fix Update enum usage to match new C API header.

deprecated Java bindings: `destroyIndex()` replaced by `close()` (v25.10.00). ↓

fix Use `close()` method instead of `destroyIndex()`.

gotcha libcuvs-cu12 is a C++ library; do not import directly in Python. Instead, use the `cuvs` Python package. ↓

fix Install `cuvs` (pip install cuvs) which depends on libcuvs-cu12.

Imports

cuvs
wrong
```
import libcuvs_cu12
```
correct
```
import cuvs
```
libcuvs-cu12 is a C++ backend; the Python API is in the `cuvs` package.
CagraIndex
wrong
```
from libcuvs_cu12 import CagraIndex
```
correct
```
from cuvs.neighbors import cagra
```
CAGRA index is accessed via the `cuvs` Python package.

Quickstart

Build a CAGRA index and search using the `cuvs` Python package. libcuvs-cu12 is the C++ backend installed automatically.

import cupy as cp
import cuvs
from cuvs.neighbors import cagra

rs = cuvs.raft.RaftResourceManager()
data = cp.random.random((1000, 128))
index = cagra.build(rs, data)
queries = cp.random.random((10, 128))
distances, neighbors = cagra.search(index, queries, k=10)