GPU installation

CTBoost reports CUDA support at runtime:

import ctboost
print(ctboost.build_info())

Train on CUDA by setting task_type="GPU":

from ctboost import CTBoostClassifier

model = CTBoostClassifier(task_type="GPU", iterations=500)
model.fit(X_train, y_train)

Starting with CTBoost 0.1.54, releases use unified Linux x86-64 and Windows AMD64 wheels for CPython 3.10 through 3.14: the ordinary command below installs one wheel that can train on CPU and use CUDA when an NVIDIA device is available.

python -m pip install -U ctboost

Released artifacts are authoritative

Check the GitHub release notes for the authoritative Python, operating-system, CUDA-runtime, architecture, and minimum-driver matrix. macOS and ARM wheels are CPU-only. A CUDA-enabled build still works for CPU training on a machine without an NVIDIA GPU. The bundled CUDA runtime remains subject to the NVIDIA CUDA Toolkit license included in each CUDA-enabled wheel.

The 0.1.54 through 0.1.56 unified wheels bundle the CUDA 12.8 runtime. GPU use requires an NVIDIA driver compatible with CUDA 12.x (at least 525.60.13 on Linux or 528.33 on Windows) but does not require a local CUDA toolkit. Released CUDA wheels target compute capability 6.0 or newer, with native code through current architectures and forward-compatible PTX. Linux aarch64, macOS, and CPython 3.8/3.9 artifacts are CPU-only.

Older releases that published CUDA wheels only as GitHub assets retain the legacy ctboost-install-gpu command. It verifies the GitHub-provided SHA-256 digest before asking pip to replace the installed wheel.