Skip to contents

Changelog

Source: NEWS.md

cuda.ml 0.4.0

This is a breaking release targeting one backend: CUDA Toolkit 13.2.2, RAPIDS cuML and nvForest 26.06, and Treelite 4.7.0.

  • Added Linux x86_64 backends targeting glibc 2.28 or newer and hosted as GitHub Release assets. The CRAN package remains a portable R installer and loader; it contains neither compiled code nor the approximately 1.6 GiB CUDA and RAPIDS runtime.

  • Added explicit cuda_ml_install() runtime preparation. Model operations no longer download libraries implicitly; they direct users to run the installer when the managed runtime is absent. cuda_ml_runtime_audit() performs a full content audit, and cuda_ml_cache_clean() removes cuda.ml cache generations.

  • Added cuda_ml_install(source = TRUE) for host builds without Docker or a prebuilt cuda.ml backend. By default it bootstraps exact locked CUDA, RAPIDS, Treelite, CMake, and Ninja build inputs, requiring only GNU C++ 14 or newer on a supported Linux host. dependencies = "host" uses explicit native build inputs and makes no downloads. Managed builds detect CUDA-visible GPU architectures by default and otherwise use the portable package list. architectures = "native" requires detection, while architectures = "portable" forces the package list. Source backend selection persists across R sessions.

  • Package loading and cuda_ml_backend_info() are silent and side-effect free. They do not inspect the GPU, create a cache, contact the network, or load the native backend. CRAN checks use an explicit network-free stub backend.

  • Parsnip is now optional. Loading cuda.ml does not load parsnip, ggplot2, or S7; cuda.ml registers its engines when parsnip is loaded.

  • Replaced the removed cuML FIL interface with current nvForest loading, prediction, model inspection, leaf-ID, per-tree, and persistence APIs. Random-forest inference and persistence now use the same nvForest backend.

  • Corrected random-forest arguments: mtry now controls sampled predictors, sample_fraction controls sampled rows, trees defaults to 100, and seed is forwarded to cuML. Classification uses predict(..., type = "class") and predict(..., type = "prob").

  • Added parsnip engines for linear_reg(), logistic_reg(), and multinom_reg() using the customary penalty and mixture arguments. Normalization is no longer emulated inside regularized linear models; use a recipe preprocessing step when scaling is required.

  • Removed interfaces that the supported backend cannot implement: FIL, random projection, KNN IVFSQ, cuML log-level controls, has_cuML(), and the split cuML version queries. cuda_ml_backend_info() is the single backend metadata interface.

  • Model persistence now stores explicit versioned state through cuda_ml_serialize() and bundle::bundle(). The package version is recorded as provenance and does not by itself prevent restoration. Schema 1 compatibility is defined by the model ABI: linear and logistic-regression states require no backend identity match; PCA, SVC, one-vs-rest SVC, SVR, and UMAP states require the recorded RAPIDS version; and random-forest and nvForest states require the recorded Treelite version. Unknown schemas or ABIs, missing payloads, and missing or incompatible required backend fields fail without implicit migration or native-pointer fallback behavior.

  • Added GPU-less fat-binary compilation for compute capabilities 7.5, 8.0, 8.6, 8.9, 9.0, 10.0, and 12.0, with PTX forward compatibility from 12.0.

  • Organized the function reference and added guides for getting started, runtime installation and management, tidymodels, model persistence, and nvForest inference and deployment.

  • Removed native compatibility branches for historical cuML releases and the implicit bootstrap fallbacks for Python, pip, CMake, wheel layouts, CUDA architectures, and missing local toolchains.