Skip to contents

Changelog

Source: NEWS.md

cuda.ml 0.4.0

CRAN release: 2026-08-21

This is a breaking release for users upgrading from the cuda.ml 0.3 release series. It requires R 4.1 or newer and targets CUDA Toolkit 13.2.2, RAPIDS cuML and nvForest 26.06, and Treelite 4.7.0.

Installation and runtime

  • The CRAN package is now a portable R installer and loader with no compiled code or bundled CUDA libraries. On Linux x86_64 with glibc 2.28 or newer, cuda_ml_install() explicitly prepares the pinned native backend and its approximately 1.6 GiB CUDA and RAPIDS runtime. Repeated calls reuse the prepared cache.

  • cuda_ml_install(source = TRUE) builds the backend on a supported host. Its managed mode prepares the pinned CUDA, RAPIDS, Treelite, CMake, and Ninja build inputs and requires GNU C++ 14 or newer. Use dependencies = "host" to supply all native build inputs. Managed builds detect distinct CUDA-visible architectures when possible, while architectures = "native", "portable", or an explicit CMake architecture list selects the build target policy.

  • Loading cuda.ml is silent; it does not inspect a GPU, load native code, modify the cache, or use the network. cuda_ml_backend_info() reports pinned versions and cache status without those actions. cuda_ml_runtime_audit() is the separate, explicit full content and native registration check, and cuda_ml_cache_clean() removes managed cache generations.

  • CPU nvForest inference requires a prepared cuda.ml backend. It can use the complete runtime installed by cuda_ml_install() or the separate CPU-only backend installed by cuda_ml_install(device = "cpu"). Selecting device = "cpu" when loading or restoring a model does not itself prepare a backend.

  • Prebuilt GPU backends contain real targets for compute capabilities 7.5, 8.0, 8.6, 8.9, 9.0, 10.0, and 12.0, plus PTX forward compatibility from 12.0.

Models and public APIs

  • The former FIL interface was replaced by nvForest. Use cuda_ml_nvforest_load_model() for XGBoost models, LightGBM text models, and Treelite checkpoints, standard predict() for inference, and the cuda_ml_nvforest_*() inspection, checkpoint export, and import functions for nvForest-specific operations. Random-forest fits now use nvForest for prediction and persistence as well.

  • The random-forest API changed. mtry controls predictor sampling, sample_fraction controls row sampling and defaults to 1, and the separate max_predictors_per_note_split argument was removed. When mtry is omitted, classification uses the square root of the predictor count and regression uses all predictors. trees defaults to 100. When seed is omitted, it is drawn from R’s random-number generator, so set.seed() controls the fit. The max_batch_size and n_streams defaults are now 4096 and 4. Regression split criteria are now "mse", "poisson", "gamma", and "inverse_gaussian"; "mae" was removed. Classification predictions use predict(..., type = "class") or predict(..., type = "prob").

  • Logistic and multinomial regression now use numeric penalty and mixture arguments consistent with parsnip instead of the previous penalty, C, and l1_ratio combination. The default is now unregularized; set a numeric penalty to request regularization. The iteration arguments are now max_iter and linesearch_max_iter; lbfgs_memory and penalty_normalized provide additional solver controls. cuda_ml_linear_reg() provides the corresponding parsnip-style routing across OLS, ridge, lasso, and elastic-net fits.

  • The normalize_input argument was removed from OLS, ridge, lasso, and elastic-net models. It previously requested GPU-side L2 normalization. recipes::step_normalize() is the recommended explicit preprocessing step when centering and scaling are appropriate, but it is not numerically identical to the former L2 operation.

  • cuda_ml_sgd() now fits squared-loss regression only, so its loss argument was removed, and n_iters_no_change was renamed to n_iter_no_change. Prediction methods now consistently use new_data; KNN classification uses type = "class" or type = "prob" instead of output_class_probabilities.

  • Parsnip is optional. cuda.ml registers engines when parsnip is loaded, including linear_reg(), logistic_reg(), and multinom_reg() engines that use the usual penalty and mixture arguments.

  • Random projection and KNN IVFSQ were removed and have no current replacement in the pinned upstream API. KNN continues to support brute-force, IVFFlat, and IVFPQ search. The unused use_precomputed_tables argument was removed from cuda_ml_knn_algo_ivfpq().

  • The per-call cuML_log_level arguments were removed. has_cuML(), cuML_major_version(), and cuML_minor_version() were removed in favor of fields returned by cuda_ml_backend_info().

  • cuda_ml_is_classifier() and cuda_ml_can_predict_class_probabilities() were removed. Use the documented predict() types for each model; for nvForest models, cuda_ml_nvforest_info() reports task_type and has_probability_output. The cuda_ml_serialise() and cuda_ml_unserialise() aliases were also removed; use cuda_ml_serialize() and cuda_ml_unserialize().

Model persistence

  • cuda_ml_serialize() and cuda_ml_unserialize() now provide durable model states for OLS, ridge, lasso, elastic-net, SGD, logistic and multinomial regression, PCA, binary and one-vs-rest SVC, SVR, UMAP, random forest, and nvForest models. Passing a file path writes or reads a gzip-compressed state; open connections and in-memory raw vectors are also supported. KNN and TSVD fits are not currently supported: the pinned KNN API does not expose portable approximate-index state, and the current TSVD binding does not reconstruct its native transform parameters.

  • bundle::bundle() stores the same explicit state for workflows that use the bundle package. Both interfaces support saving an artifact and restoring it in a fresh R process after the target environment prepares the required backend. nvForest models can also be exported and imported as a Treelite checkpoint plus cuda.ml metadata.

  • cuda.ml validates each saved state and its required backend before restoring the model.

Documentation

  • The function reference is reorganized and includes guides for getting started, installation and runtime management, tidymodels, model persistence, and nvForest inference and deployment.