⚠️ RAPIDS.jl is only supported on Julia 1.8.5+. For previous Julia versions, you have to manually upgrade to libraries from GCC 12.

RAPIDS.jl

Unofficial Julia wrapper for the RAPIDS.ai ecosystem.

The goal of this library is to provide a simple method for accessing the GPU accelerated models withing RAPIDS from Julia, and integrating the models into MLJ. This library relies on PythonCall.jl and CondaPkg.jl for efficient installations of the Python dependencies.

This wrapper could be broken up into several libraries (cuDF, cuML, cuGraph, cuSpatial), but there would be significant overlap between these libraries. Large dependencies such as cudatoolkit would be repeated.

Long term, directly wrapping libcudf, libcuml... would greatly improve this library, but I don't have time to tackle that at this moment.

CUDA/GPU requirements

More information is available here.

• CUDA 11.2+
• NVIDIA driver 470.42.01+
• Pascal architecture or better (Compute Capability >=6.0)
• Ubuntu 20.04/22.04 or CentOS 7 / Rocky Linux 8 with gcc/++ 9.0+

Installation

From the Julia General Registry:

julia> ]  # enters the pkg interface
pkg> add RAPIDS

julia> using Pkg; Pkg.add("RAPIDS")

From source:

julia> ]add https://github.com/tylerjthomas9/RAPIDS.jl

julia> using Pkg; Pkg.add(url="https://github.com/tylerjthomas9/RAPIDS.jl")

Python API

You can access the following python libraries with their standard syntax:

• cupy
• cudf
• cuml
• cugraph
• cuspatial
• cuxfilter
• dask
• dask_cuda
• dask_cudf
• numpy
• pickle

Here is an example of using LogisticRegression, make_classification via the Python API.

using RAPIDS
const make_classification = cuml.datasets.classification.make_classification

X_py, y_py = make_classification(n_samples=200, n_features=4,
                           n_informative=2, n_classes=2)
lr = cuml.LogisticRegression(max_iter=100)
lr.fit(X_py, y_py)
preds = lr.predict(X_py)

print(lr.coef_)

MLJ Interface

A MLJ interface is also available for supported models. The model hyperparameters are the same as described in the cuML docs. The only difference is that the models will always input/output numpy arrays, which will be converted back to Julia arrays (output_type="input").

using MLJBase
using RAPIDS.CuML
const make_classification = cuml.datasets.classification.make_classification

X_py, y_py = make_classification(n_samples=200, n_features=4,
                           n_informative=2, n_classes=2)
X = RAPIDS.pyconvert(Matrix{Float32}, X_py.get())
y = RAPIDS.pyconvert(Vector{Float32}, y_py.get().flatten())

lr = LogisticRegression(max_iter=100)
mach = machine(lr, X, y)
fit!(mach)
preds = predict(mach, X)

print(mach.fitresult.coef_)

MLJ Support:

• Clustering
  • KMeans
  • DBSCAN
  • AgglomerativeClustering
  • HDBSCAN
• Classification
  • LogisticRegression
  • MBSGDClassifier
  • RandomForestClassifier
  • SVC
  • LinearSVC
  • KNeighborsClassifier
• Regression
  • LinearRegression
  • Ridge
  • Lasso
  • ElasticNet
  • MBSGDRegressor
  • RandomForestRegressor
  • CD
  • SVR
  • LinearSVR
  • KNeighborsRegressor
• Dimensionality Reduction
  • PCA
  • IncrementalPCA
  • TruncatedSVD
  • UMAP
  • TSNE
  • GaussianRandomProjection
• Time Series
  • ExponentialSmoothing
  • ARIMA