RAFT Dask

26.4.0 verified Mon Apr 27 auth: no python

RAFT (Reusable Accelerated Functions & Tools) provides GPU-accelerated primitives for machine learning, linear algebra, and spatial computations. The Dask integration enables distributed GPU workflows via Dask. Current version: 26.4.0 (stable). RAPIDS releases every 3 months (quarterly). CUDA 12.x only.

pip install raft-dask-cu12

Common errors

error ModuleNotFoundError: No module named 'raft' ↓

cause Attempting to import 'raft' which is the C++ library, not the Python package. Python packages are named with prefixes like 'raft_dask'.

fix

Use 'import raft_dask' or install 'raft-dask-cu12' (not 'raft').

error ImportError: cannot import name 'DaskLinalg' from 'raft.dask.linalg' ↓

cause Old import path; module structure changed in v25.12.

fix

Use 'from raft_dask.linalg import DaskLinalg'.

error RuntimeError: Dask cluster not running or UCX not initialized ↓

cause Dask cluster proxy started without a running cluster or UCX transport not set up.

fix

Start a Dask scheduler/worker (e.g., 'dask-scheduler' & 'dask-worker') and install ucx-py for GPU communication.

Warnings

breaking Requires CUDA 12.2+ as of v25.12.00; CUDA 11 support dropped entirely. ↓

fix Upgrade to CUDA 12.2 or later, or use raft-dask-cu12=25.10.00 for CUDA 11.

breaking Import paths changed: most top-level imports now come from 'raft_dask' subpackages. Old 'raft.dask.*' patterns removed in v25.12. ↓

fix Replace 'import raft.dask' with 'import raft_dask' and adjust imports accordingly.

gotcha DaskClusterProxy assumes UCX-Py is installed and configured for GPU communication; missing UCX leads to runtime errors. ↓

fix Install ucx-py and configure UCX transport: pip install ucx-py.

Install

conda install -c rapidsai raft-dask

Imports

DaskClusterProxy
```
from raft_dask import DaskClusterProxy
```
Import from raft_dask package (note underscore vs hyphen).

DaskLinalg

wrong

from raft.dask import ...

correct

from raft_dask.linalg import DaskLinalg

Old import path deprecated in v25.12.

Quickstart

Initialize a Dask cluster with RAFT acceleration. Requires CUDA 12.x and a running cluster backend (e.g., UCX).

from raft_dask.cluster import DaskClusterProxy
cluster = DaskClusterProxy()
print(cluster)
# Use with Dask client
import dask.distributed as dd
client = dd.Client(cluster)
print(client)