RAPIDS cuML

Common errors

• RuntimeError: CUDA out of memory. Tried to allocate X GiB

  cause Attempting to allocate more GPU memory than is available on the device.
  fix Reduce the size of the dataset, decrease batch sizes in iterative algorithms, or use Dask with cuML for out-of-core processing if your dataset exceeds single-GPU memory. Consider using a GPU with more VRAM.

• AttributeError: 'numpy.ndarray' object has no attribute 'get_nrows'

  cause cuML estimators primarily expect `cudf.DataFrame` or `cupy.ndarray` objects as input, not `numpy.ndarray`.
  fix Convert your input data from NumPy arrays to cuDF DataFrames or CuPy arrays before passing them to cuML estimators, e.g., `X_gdf = cudf.DataFrame(X_numpy)`.

• No module named 'cudf'

  cause The `cudf` library, which provides GPU-accelerated DataFrames and is essential for most cuML workflows, is not installed in the environment.
  fix Install `cudf` explicitly with the correct CUDA version, e.g., `pip install cudf-cu12` for CUDA 12, or follow the full RAPIDS installation guide for your environment.

• ImportError: libcudart.so.12: cannot open shared object file: No such file or directory

  cause The CUDA Toolkit (specifically, the CUDA Runtime library) for the expected CUDA version (e.g., v12 for `libcuml-cu12`) is either not installed or its library path is not correctly configured in `LD_LIBRARY_PATH`.
  fix Ensure that the appropriate CUDA Toolkit is installed for your system and that its `lib` directory is included in your `LD_LIBRARY_PATH` environment variable. Using `conda` to manage RAPIDS installations often handles this automatically.

Warnings

• breaking Sparse input validation now raises `TypeError` if sparse input is not supported by the algorithm.
  Fix: Ensure that sparse input data is only provided to cuML algorithms that explicitly support it, or convert sparse data to a dense format before passing it to unsupported algorithms.
• breaking `check_is_fitted` validation and `feature_names_in_` support have been added, requiring estimators to conform more strictly to scikit-learn's API.
  Fix: If using custom or wrapped estimators, ensure they implement `_validate_data` and set `feature_names_in_` after fitting, consistent with scikit-learn's guidelines.
• breaking The `handle` object has been deprecated from public APIs, affecting low-level GPU resource management.
  Fix: Review existing code for direct usage of `handle` objects and refactor to use higher-level cuML APIs. cuML now manages GPU resources internally for most use cases.
• breaking `output_type=None` in estimator `__init__` will no longer implicitly coerce to a global `output_type` setting.
  Fix: Explicitly specify `output_type` in the constructor for cuML estimators if you rely on a particular output type (e.g., `output_type='cudf'`).
• breaking `dask` is now an optional dependency. Dask-related features will require `dask` and `distributed` to be installed explicitly.
  Fix: If using cuML with Dask, ensure you install `dask` and `distributed` separately (e.g., `pip install dask distributed`).
• deprecated The `convert_to_*` methods (e.g., `convert_to_cudf`) in `cuml.ensemble` have been deprecated in favor of `as_*` methods (e.g., `as_cudf`).
  Fix: Update method calls to use the `as_*` prefix (e.g., `model.as_cudf(X)` instead of `model.convert_to_cudf(X)`).

Install

• pip install libcuml-cu12 Install core cuML for CUDA 12
• pip install cudf-cu12 # Recommended for data handling Install cuDF (GPU DataFrames)
• pip install dask distributed # Recommended for Dask integration Install Dask for distributed computing

Imports

• KMeans

  from cuml.cluster import KMeans

• RandomForestClassifier

  from cuml.ensemble import RandomForestClassifier

• LinearRegression

  from cuml.linear_model import LinearRegression

Quickstart

This example demonstrates how to perform k-means clustering using cuML. It generates synthetic data with scikit-learn, converts it to a cuDF DataFrame for GPU processing, and then fits a KMeans model to find clusters. It requires `cudf` and `scikit-learn`.

import cuml
import cudf
from sklearn.datasets import make_blobs

# Generate synthetic data on CPU
X, _ = make_blobs(n_samples=1000, n_features=10, centers=5, random_state=42)

# Convert to cuDF DataFrame for GPU processing
X_gdf = cudf.DataFrame(X)

# Initialize and fit a cuML KMeans model
kmeans = cuml.cluster.KMeans(n_clusters=5, random_state=42)
kmeans.fit(X_gdf)

# Predict cluster labels
labels = kmeans.predict(X_gdf)

print("Cluster labels (first 5):\n", labels.head())
print("Cluster centers (first 5 rows):\n", kmeans.cluster_centers_.head())