dfalbel · Apr 6, 2026 · Apr 6, 2026
diff --git a/.github/workflows/test-install.yaml b/.github/workflows/test-install.yaml
@@ -12,7 +12,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        package: [cuda12.6, cuda12.8, cuda13.2]
+        package: [cuda12.6, cuda12.8, cuda12.9, cuda13.2]
 
     name: ${{ matrix.package }}
 

diff --git a/cuda12.9/DESCRIPTION b/cuda12.9/DESCRIPTION
@@ -0,0 +1,15 @@
+Package: cuda12.9
+Title: CUDA 12.9 Toolkit Libraries
+Version: 1.0.0
+Authors@R:
+    person("Daniel", "Falbel", , "daniel@posit.co", role = c("aut", "cre"))
+Description: Provides CUDA 12.9 native libraries (runtime, cuBLAS, cuDNN, cuFFT,
+    cuSOLVER, cuSPARSE, CUPTI, NVRTC, NvJitLink, NCCL, NVSHMEM, and NVCC) by
+    downloading pre-built binaries from PyPI at install time. Intended to be used
+    as a dependency by packages that require CUDA 12.9.
+License: MIT + file LICENSE
+URL: https://github.com/mlverse/cudatoolkit
+BugReports: https://github.com/mlverse/cudatoolkit/issues
+Encoding: UTF-8
+Roxygen: list(markdown = TRUE)
+RoxygenNote: 7.3.2
diff --git a/cuda12.9/LICENSE b/cuda12.9/LICENSE
@@ -0,0 +1,2 @@
+YEAR: 2026
+COPYRIGHT HOLDER: Daniel Falbel
diff --git a/cuda12.9/NAMESPACE b/cuda12.9/NAMESPACE
@@ -0,0 +1,7 @@
+# Generated by roxygen2: do not edit by hand
+
+export(all_lib_paths)
+export(bin_path)
+export(cuda_path)
+export(include_path)
+export(lib_path)
diff --git a/cuda12.9/R/paths.R b/cuda12.9/R/paths.R
@@ -0,0 +1,70 @@
+# Read component -> wheel_subdir mapping from installed components.tsv
+.read_components <- function() {
+  tsv <- system.file("components.tsv", package = packageName(), mustWork = TRUE)
+  df <- read.delim(tsv, stringsAsFactors = FALSE)
+  setNames(df$wheel_subdir, df$component)
+}
+
+# Lazily cached component map
+.component_subdir_env <- new.env(parent = emptyenv())
+
+.component_subdir <- function(component) {
+  if (is.null(.component_subdir_env$map)) {
+    .component_subdir_env$map <- .read_components()
+  }
+  subdir <- .component_subdir_env$map[[component]]
+  if (is.null(subdir)) {
+    stop(sprintf("Unknown component: '%s'. Available: %s",
+                 component, paste(names(.component_subdir_env$map), collapse = ", ")))
+  }
+  subdir
+}
+
+#' Path to a CUDA component installation
+#'
+#' @param component Component name (e.g., "runtime", "cublas", "cudnn").
+#' @return A character string with the path to the installed component files.
+#' @export
+cuda_path <- function(component) {
+  system.file(file.path("nvidia", .component_subdir(component)),
+              package = packageName(), mustWork = TRUE)
+}
+
+#' Path to the shared library directory
+#'
+#' All CUDA component libraries are installed into a single shared directory.
+#'
+#' @return A character string with the path to the lib directory.
+#' @export
+lib_path <- function() {
+  system.file("lib", package = packageName(), mustWork = TRUE)
+}
+
+#' Path to a CUDA component's headers
+#'
+#' @param component Component name (e.g., "runtime", "cublas", "cudnn").
+#' @return A character string with the path to the include directory.
+#' @export
+include_path <- function(component) {
+  file.path(cuda_path(component), "include")
+}
+
+#' Path to a CUDA component's binaries
+#'
+#' @param component Component name (e.g., "nvcc").
+#' @return A character string with the path to the bin directory.
+#' @export
+bin_path <- function(component) {
+  file.path(cuda_path(component), "bin")
+}
+
+#' Path to the shared library directory
+#'
+#' Alias for \code{\link{lib_path}}. All CUDA component libraries are in a
+#' single directory, so this simply returns that path.
+#'
+#' @return A character vector of length one with the library path.
+#' @export
+all_lib_paths <- function() {
+  lib_path()
+}
diff --git a/cuda12.9/configure b/cuda12.9/configure
@@ -0,0 +1,222 @@
+#!/bin/sh
+
+# Downloads CUDA toolkit components from PyPI during package installation.
+# Component names and pinned versions are read from inst/components.tsv
+# (generated at build time by CI).
+#
+# Environment variables:
+#   CUDA_COMPONENTS  Comma-separated list of components to install.
+#                    Default: all components. Example: "runtime,cublas,cudnn"
+#                    Use @version to override a component's pinned version:
+#                    "runtime,cublas,cudnn@9.10.0.56"
+#   CUDA_PLATFORM    Override platform detection (linux-x64, linux-arm64, windows-x64).
+
+detect_platform() {
+  if [ -n "$CUDA_PLATFORM" ]; then
+    case "$CUDA_PLATFORM" in
+      linux-x64)    echo "x86_64" ;;
+      linux-arm64)  echo "aarch64" ;;
+      windows-x64)  echo "win_amd64" ;;
+      *)
+        echo "ERROR: Unknown CUDA_PLATFORM: $CUDA_PLATFORM" >&2
+        exit 1
+        ;;
+    esac
+    return
+  fi
+
+  os=$(uname -s)
+  arch=$(uname -m)
+
+  case "$os" in
+    Linux)
+      case "$arch" in
+        x86_64)  echo "x86_64" ;;
+        aarch64) echo "aarch64" ;;
+        *)
+          echo "ERROR: Unsupported Linux architecture: $arch" >&2
+          exit 1
+          ;;
+      esac
+      ;;
+    MINGW*|MSYS*|CYGWIN*|Windows_NT)
+      echo "win_amd64"
+      ;;
+    *)
+      echo "ERROR: Unsupported OS: $os" >&2
+      exit 1
+      ;;
+  esac
+}
+
+is_windows() {
+  case "$PLATFORM" in
+    win_amd64) return 0 ;;
+    *) return 1 ;;
+  esac
+}
+
+download_file() {
+  url="$1"; dest="$2"
+  if command -v curl > /dev/null 2>&1; then
+    curl -sL -o "$dest" "$url"
+  elif command -v wget > /dev/null 2>&1; then
+    wget -q -O "$dest" "$url"
+  else
+    echo "ERROR: Neither curl nor wget found." >&2
+    exit 1
+  fi
+}
+
+if ! command -v unzip > /dev/null 2>&1; then
+  echo "ERROR: unzip not found. Please install unzip." >&2
+  exit 1
+fi
+
+PLATFORM=$(detect_platform) || exit 1
+TMPDIR=$(mktemp -d)
+trap "rm -rf '$TMPDIR'" EXIT
+
+# Parse CUDA_COMPONENTS filter and version overrides
+SELECTED_COMPONENTS="${CUDA_COMPONENTS:-}"
+
+is_selected() {
+  comp="$1"
+  # If no filter set, install all
+  [ -z "$SELECTED_COMPONENTS" ] && return 0
+  # Check if component (with or without @version) is in the comma-separated list
+  echo ",$SELECTED_COMPONENTS," | grep -q ",${comp}," || echo ",$SELECTED_COMPONENTS," | grep -q ",${comp}@"
+}
+
+# Get version override for a component, or empty if not overridden
+get_version_override() {
+  comp="$1"
+  [ -z "$SELECTED_COMPONENTS" ] && return
+  override=$(echo ",$SELECTED_COMPONENTS," | grep -o ",${comp}@[^,]*," | sed "s/,//g;s/${comp}@//")
+  echo "$override"
+}
+
+echo "=== CUDA Toolkit ==="
+echo "* Platform: ${PLATFORM}"
+if [ -n "$SELECTED_COMPONENTS" ]; then
+  echo "* Components: ${SELECTED_COMPONENTS}"
+fi
+echo ""
+
+# Shared lib directory for all components
+mkdir -p inst/lib
+
+# Read components.tsv (skip header) and install each component
+tail -n +2 inst/components.tsv | while IFS="	" read -r component pypi_package version wheel_subdir extract linux_only; do
+  [ -z "$component" ] && continue
+
+  # Skip components not in the selection
+  if ! is_selected "$component"; then
+    echo "* Skipping ${component} (not selected)"
+    continue
+  fi
+
+  # Skip linux-only components on Windows
+  if [ "$linux_only" = "yes" ] && is_windows; then
+    echo "* Skipping ${component} (Linux-only)"
+    continue
+  fi
+
+  # Check for version override
+  ver_override=$(get_version_override "$component")
+  if [ -n "$ver_override" ]; then
+    version="$ver_override"
+    echo "* Installing ${component} ${version} (overridden)..."
+  else
+    echo "* Installing ${component} ${version}..."
+  fi
+
+  # Fetch PyPI simple index for this package
+  INDEX_FILE="${TMPDIR}/index_${pypi_package}.html"
+  download_file "https://pypi.org/simple/${pypi_package}/" "$INDEX_FILE"
+
+  # Find matching wheel URL
+  WHEEL_URL=$(grep -o "href=\"[^\"]*${pypi_package}-${version}[^\"]*${PLATFORM}[^\"]*\"" "$INDEX_FILE" | head -1 | sed 's/href="//;s/"//')
+
+  if [ -z "$WHEEL_URL" ]; then
+    echo "  ERROR: Could not find wheel for ${pypi_package}==${version} on ${PLATFORM}" >&2
+    exit 1
+  fi
+
+  WHEEL_URL=$(echo "$WHEEL_URL" | sed 's/#.*//')
+  WHEEL_NAME=$(basename "$WHEEL_URL")
+  WHEEL_PATH="${TMPDIR}/${WHEEL_NAME}"
+
+  echo "  Downloading ${WHEEL_NAME}..."
+  download_file "$WHEEL_URL" "$WHEEL_PATH"
+
+  if [ ! -f "$WHEEL_PATH" ]; then
+    echo "  ERROR: Download failed." >&2
+    exit 1
+  fi
+
+  # Extract requested directories from the wheel
+  for dir in $(echo "$extract" | tr ',' ' '); do
+    unzip -q -o "$WHEEL_PATH" "nvidia/${wheel_subdir}/${dir}/*" -d inst/
+  done
+
+  # Move shared libraries into the unified lib directory
+  if [ -d "inst/nvidia/${wheel_subdir}/lib" ]; then
+    mv inst/nvidia/${wheel_subdir}/lib/* inst/lib/ 2>/dev/null || true
+    rmdir inst/nvidia/${wheel_subdir}/lib 2>/dev/null || true
+  fi
+
+  # Ensure binaries are executable (for nvcc/ptxas)
+  if echo "$extract" | grep -q "bin"; then
+    chmod +x inst/nvidia/${wheel_subdir}/bin/* 2>/dev/null || true
+  fi
+
+  rm -f "$WHEEL_PATH"
+  echo "  Done."
+done
+
+# Patch RUNPATH on Linux so each .so resolves dependencies from its own
+# directory first, preventing LD_LIBRARY_PATH from pulling in wrong versions
+# (e.g. conda's older cuDNN).
+get_patchelf() {
+  command -v patchelf > /dev/null 2>&1 && { echo "patchelf"; return 0; }
+
+  # Download patchelf from PyPI wheel (static binary, no Python needed)
+  PATCHELF_VERSION="0.17.2.2"
+  case "$(uname -m)" in
+    x86_64)
+      PATCHELF_URL="https://files.pythonhosted.org/packages/f2/f9/e070956e350ccdfdf059251836f757ad91ac0c01b0ba3e033ea7188d8d42/patchelf-0.17.2.2-py3-none-manylinux1_x86_64.manylinux_2_5_x86_64.musllinux_1_1_x86_64.whl"
+      ;;
+    aarch64)
+      PATCHELF_URL="https://files.pythonhosted.org/packages/56/0d/dc3ac6c6e9e9d0d3e40bee1abe95a07034f83627319e60a7dc9abdbfafee/patchelf-0.17.2.2-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.musllinux_1_1_aarch64.whl"
+      ;;
+    *) return 1 ;;
+  esac
+
+  PATCHELF_DIR="${TMPDIR}/patchelf"
+  PATCHELF_WHL="${PATCHELF_DIR}/patchelf.whl"
+  mkdir -p "$PATCHELF_DIR"
+  download_file "$PATCHELF_URL" "$PATCHELF_WHL"
+  unzip -q -o "$PATCHELF_WHL" "patchelf-${PATCHELF_VERSION}.data/scripts/patchelf" -d "$PATCHELF_DIR" || return 1
+  chmod +x "${PATCHELF_DIR}/patchelf-${PATCHELF_VERSION}.data/scripts/patchelf"
+  echo "${PATCHELF_DIR}/patchelf-${PATCHELF_VERSION}.data/scripts/patchelf"
+}
+
+if ! is_windows; then
+  PATCHELF_BIN=$(get_patchelf)
+  if [ -n "$PATCHELF_BIN" ]; then
+    echo "* Patching RPATH with \$ORIGIN (--force-rpath so it takes precedence over LD_LIBRARY_PATH)..."
+    for so in inst/lib/*.so*; do
+      [ -f "$so" ] || continue
+      "$PATCHELF_BIN" --force-rpath --add-rpath '$ORIGIN' "$so" 2>/dev/null || true
+    done
+    echo "  Done."
+  else
+    echo "* WARNING: Could not obtain patchelf, skipping RUNPATH patching."
+    echo "  Install patchelf for reliable CUDA library loading when"
+    echo "  LD_LIBRARY_PATH contains other CUDA installations."
+  fi
+fi
+
+echo ""
+echo "=== All components installed successfully ==="
diff --git a/cuda12.9/inst/components.tsv b/cuda12.9/inst/components.tsv
@@ -0,0 +1,13 @@
+component	pypi_package	version	wheel_subdir	extract	linux_only
+runtime	nvidia_cuda_runtime_cu12	12.9.79	cuda_runtime	lib,include	no
+cublas	nvidia_cublas_cu12	12.9.1.4	cublas	lib,include	no
+cupti	nvidia_cuda_cupti_cu12	12.9.79	cuda_cupti	lib,include	no
+nvrtc	nvidia_cuda_nvrtc_cu12	12.9.86	cuda_nvrtc	lib,include	no
+cufft	nvidia_cufft_cu12	11.4.1.4	cufft	lib,include	no
+cusolver	nvidia_cusolver_cu12	11.7.5.82	cusolver	lib,include	no
+cusparse	nvidia_cusparse_cu12	12.5.10.65	cusparse	lib,include	no
+nvjitlink	nvidia_nvjitlink_cu12	12.9.86	nvjitlink	lib,include	no
+nvcc	nvidia_cuda_nvcc_cu12	12.9.86	cuda_nvcc	bin,include,nvvm	no
+cudnn	nvidia_cudnn_cu12	9.20.0.48	cudnn	lib,include	no
+nccl	nvidia_nccl_cu12	2.29.7	nccl	lib,include	yes
+nvshmem	nvidia_nvshmem_cu12	3.6.5	nvshmem	lib,include	yes
diff --git a/cuda12.9/inst/cuda-version.env b/cuda12.9/inst/cuda-version.env
@@ -0,0 +1,7 @@
+CUDA_MINOR="12.9"
+CUDA_MAJOR="12"
+TOOLKIT_VERSION="12.9.1"
+PKG_VERSION="1.0.0"
+CUDNN_VERSION="9.20.0.48"
+NCCL_VERSION="2.29.7"
+NVSHMEM_VERSION="3.6.5"