fix(build): add CUDA 13.2 support

.github/workflows/ci_linux.yml

@@ -40,6 +40,10 @@ jobs:
           - name: RockyLinux-9 / CUDA-13.0.2 / x86_64
             image: "ghcr.io/rust-gpu/rust-cuda-rockylinux9-cuda13:latest"
             runner: ubuntu-latest
+          # CUDA 13.2 entries require container images from container_images.yml
+          # to be published first. Add these back once the images exist on ghcr.io:
+          #   ghcr.io/rust-gpu/rust-cuda-ubuntu24-cuda132:latest
+          #   ghcr.io/rust-gpu/rust-cuda-rockylinux9-cuda132:latest
 
     steps:
       - name: Free up space

.github/workflows/ci_windows.yml

@@ -59,13 +59,32 @@ jobs:
                 "nvrtc_dev",
                 "nvvm",               # new subpackage in CUDA 13
               ]
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+            cuda: "13.2.0"
+            nvvm-dll-dir: "nvvm\\bin\\x64"
+            sub-packages:
+              [
+                "crt",
+                "cublas",
+                "cublas_dev",
+                "cuda_profiler_api",
+                "cudart",
+                "curand",
+                "curand_dev",
+                "nvcc",
+                "nvptxcompiler",
+                "nvrtc",
+                "nvrtc_dev",
+                "nvvm",
+              ]
 
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
 
       - name: Install CUDA
-        uses: Jimver/cuda-toolkit@v0.2.29
+        uses: Jimver/cuda-toolkit@v0.2.35
         id: cuda-toolkit
         with:
           cuda: ${{ matrix.cuda }}

.github/workflows/container_images.yml

@@ -42,6 +42,12 @@ jobs:
           - name: RockyLinux-9/CUDA-13.0.2
             image: "rust-gpu/rust-cuda-rockylinux9-cuda13"
             dockerfile: ./container/rockylinux9-cuda13/Dockerfile
+          - name: Ubuntu-24.04/CUDA-13.2.0
+            image: "rust-gpu/rust-cuda-ubuntu24-cuda132"
+            dockerfile: ./container/ubuntu24-cuda132/Dockerfile
+          - name: RockyLinux-9/CUDA-13.2.0
+            image: "rust-gpu/rust-cuda-rockylinux9-cuda132"
+            dockerfile: ./container/rockylinux9-cuda132/Dockerfile
     steps:
       - name: Free up space
         # Without this the job will likely run out of disk space.
@@ -165,6 +171,10 @@ jobs:
             image: "rust-gpu/rust-cuda-rockylinux9-cuda12"
           - name: RockyLinux-9/CUDA-13.0.2
             image: "rust-gpu/rust-cuda-rockylinux9-cuda13"
+          - name: Ubuntu-24.04/CUDA-13.2.0
+            image: "rust-gpu/rust-cuda-ubuntu24-cuda132"
+          - name: RockyLinux-9/CUDA-13.2.0
+            image: "rust-gpu/rust-cuda-rockylinux9-cuda132"
     steps:
       - name: Set artifact name
         run: |

container/rockylinux9-cuda132/Dockerfile

@@ -0,0 +1,92 @@
+FROM nvcr.io/nvidia/cuda:13.2.0-cudnn-devel-rockylinux9 AS llvm-builder
+
+RUN dnf -y install \
+    --nobest \
+    --allowerasing \
+    --setopt=install_weak_deps=False \
+    openssl-devel \
+    pkgconfig \
+    which \
+    xz \
+    zlib-devel \
+    libffi-devel \
+    ncurses-devel \
+    libxml2-devel \
+    libedit-devel \
+    python3 \
+    make \
+    cmake && \
+    dnf clean all
+
+WORKDIR /data/llvm7
+
+# Download and build LLVM 7.1.0 for all architectures.
+RUN curl -sSf -L -O https://github.com/llvm/llvm-project/releases/download/llvmorg-7.1.0/llvm-7.1.0.src.tar.xz && \
+    tar -xf llvm-7.1.0.src.tar.xz && \
+    cd llvm-7.1.0.src && \
+    mkdir build && cd build && \
+    ARCH=$(uname -m) && \
+    if [ "$ARCH" = "x86_64" ]; then \
+        TARGETS="X86;NVPTX"; \
+    else \
+        TARGETS="AArch64;NVPTX"; \
+    fi && \
+    cmake \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DLLVM_TARGETS_TO_BUILD="$TARGETS" \
+        -DLLVM_BUILD_LLVM_DYLIB=ON \
+        -DLLVM_LINK_LLVM_DYLIB=ON \
+        -DLLVM_ENABLE_ASSERTIONS=OFF \
+        -DLLVM_ENABLE_BINDINGS=OFF \
+        -DLLVM_INCLUDE_EXAMPLES=OFF \
+        -DLLVM_INCLUDE_TESTS=OFF \
+        -DLLVM_INCLUDE_BENCHMARKS=OFF \
+        -DLLVM_ENABLE_ZLIB=ON \
+        -DLLVM_ENABLE_TERMINFO=ON \
+        -DCMAKE_INSTALL_PREFIX=/opt/llvm-7 \
+        .. && \
+    make -j$(nproc) && \
+    make install && \
+    cd ../.. && \
+    rm -rf llvm-7.1.0.src* && \
+    dnf clean all
+
+FROM nvcr.io/nvidia/cuda:13.2.0-cudnn-devel-rockylinux9
+
+RUN dnf -y install \
+    --nobest \
+    --allowerasing \
+    --setopt=install_weak_deps=False \
+    clang \
+    openssl-devel \
+    fontconfig-devel \
+    libX11-devel \
+    libXcursor-devel \
+    libXi-devel \
+    libXrandr-devel \
+    libxml2-devel \
+    ncurses-devel \
+    pkgconfig \
+    which \
+    xz \
+    zlib-devel \
+    cmake && \
+    dnf clean all
+
+COPY --from=llvm-builder /opt/llvm-7 /opt/llvm-7
+RUN ln -s /opt/llvm-7/bin/llvm-config /usr/bin/llvm-config && \
+    ln -s /opt/llvm-7/bin/llvm-config /usr/bin/llvm-config-7
+
+# Get Rust (install rustup; toolchain installed from rust-toolchain.toml below)
+RUN curl -sSf -L https://sh.rustup.rs | bash -s -- -y --profile minimal --default-toolchain none
+ENV PATH="/root/.cargo/bin:${PATH}"
+
+# Setup the workspace
+WORKDIR /data/rust-cuda
+RUN --mount=type=bind,source=rust-toolchain.toml,target=/data/rust-cuda/rust-toolchain.toml \
+    rustup show
+
+# Add nvvm to LD_LIBRARY_PATH.
+ENV LD_LIBRARY_PATH="/usr/local/cuda/nvvm/lib64:${LD_LIBRARY_PATH}"
+ENV LLVM_LINK_STATIC=1
+ENV RUST_LOG=info

container/ubuntu24-cuda132/Dockerfile

@@ -0,0 +1,89 @@
+FROM nvcr.io/nvidia/cuda:13.2.0-cudnn-devel-ubuntu24.04 AS llvm-builder
+
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get -qq -y install \
+    build-essential \
+    clang \
+    curl \
+    libffi-dev \
+    libedit-dev \
+    libncurses5-dev \
+    libssl-dev \
+    libtinfo-dev \
+    libxml2-dev \
+    cmake \
+    ninja-build \
+    pkg-config \
+    python3 \
+    xz-utils \
+    zlib1g-dev && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /data/llvm7
+
+# Download and build LLVM 7.1.0 for all architectures.
+RUN curl -sSf -L -O https://github.com/llvm/llvm-project/releases/download/llvmorg-7.1.0/llvm-7.1.0.src.tar.xz && \
+    tar -xf llvm-7.1.0.src.tar.xz && \
+    cd llvm-7.1.0.src && \
+    mkdir build && cd build && \
+    ARCH=$(dpkg --print-architecture) && \
+    if [ "$ARCH" = "amd64" ]; then \
+        TARGETS="X86;NVPTX"; \
+    else \
+        TARGETS="AArch64;NVPTX"; \
+    fi && \
+    cmake -G Ninja \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DLLVM_TARGETS_TO_BUILD="$TARGETS" \
+        -DLLVM_BUILD_LLVM_DYLIB=ON \
+        -DLLVM_LINK_LLVM_DYLIB=ON \
+        -DLLVM_ENABLE_ASSERTIONS=OFF \
+        -DLLVM_ENABLE_BINDINGS=OFF \
+        -DLLVM_INCLUDE_EXAMPLES=OFF \
+        -DLLVM_INCLUDE_TESTS=OFF \
+        -DLLVM_INCLUDE_BENCHMARKS=OFF \
+        -DLLVM_ENABLE_ZLIB=ON \
+        -DLLVM_ENABLE_TERMINFO=ON \
+        -DCMAKE_INSTALL_PREFIX=/opt/llvm-7 \
+        .. && \
+    ninja -j$(nproc) && \
+    ninja install && \
+    cd ../.. && \
+    rm -rf llvm-7.1.0.src*
+
+FROM nvcr.io/nvidia/cuda:13.2.0-cudnn-devel-ubuntu24.04
+
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get -qq -y install \
+    build-essential \
+    clang \
+    curl \
+    libssl-dev \
+    libtinfo-dev \
+    pkg-config \
+    xz-utils \
+    zlib1g-dev \
+    cmake \
+    libfontconfig-dev \
+    libx11-xcb-dev \
+    libxcursor-dev \
+    libxi-dev \
+    libxinerama-dev \
+    libxrandr-dev && \
+    rm -rf /var/lib/apt/lists/*
+
+COPY --from=llvm-builder /opt/llvm-7 /opt/llvm-7
+RUN ln -s /opt/llvm-7/bin/llvm-config /usr/bin/llvm-config && \
+    ln -s /opt/llvm-7/bin/llvm-config /usr/bin/llvm-config-7
+
+# Get Rust (install rustup; toolchain installed from rust-toolchain.toml below)
+RUN curl -sSf -L https://sh.rustup.rs | bash -s -- -y --profile minimal --default-toolchain none
+ENV PATH="/root/.cargo/bin:${PATH}"
+
+# Setup the workspace
+WORKDIR /data/rust-cuda
+RUN --mount=type=bind,source=rust-toolchain.toml,target=/data/rust-cuda/rust-toolchain.toml \
+    rustup show
+
+# Add nvvm to LD_LIBRARY_PATH.
+ENV LD_LIBRARY_PATH="/usr/local/cuda/nvvm/lib64:${LD_LIBRARY_PATH}"
+ENV LLVM_LINK_STATIC=1
+ENV RUST_LOG=info

crates/cust/build.rs

@@ -40,4 +40,12 @@ fn main() {
         println!("cargo::rustc-cfg=cuGraphGetEdges_v2");
         println!("cargo::rustc-cfg=cuCtxCreate_v4");
     }
+
+    // In CUDA 13.2 the `id` field in `CUmemLocation_st` was placed inside an anonymous union.
+    // Bindgen renders this as `__bindgen_anon_1: CUmemLocation_st__bindgen_ty_1` instead of a
+    // direct `id` field. This cfg gates the struct initialization syntax accordingly.
+    println!("cargo::rustc-check-cfg=cfg(cuMemLocation_anon_id)");
+    if driver_version >= 13020 {
+        println!("cargo::rustc-cfg=cuMemLocation_anon_id");
+    }
 }

crates/cust/src/memory/unified.rs

@@ -647,6 +647,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemPrefetchAsync_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemPrefetchAsync_v2))]
@@ -693,6 +696,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemPrefetchAsync_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemPrefetchAsync_v2))]
@@ -735,6 +741,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemAdvise_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemAdvise_v2))]
@@ -777,6 +786,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemAdvise_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemAdvise_v2))]
@@ -801,6 +813,9 @@ pub trait MemoryAdvise<T: DeviceCopy>: private::Sealed {
                 #[cfg(cuMemAdvise_v2)]
                 driver_sys::CUmemLocation {
                     type_: driver_sys::CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
+                    #[cfg(cuMemLocation_anon_id)]
+                    __bindgen_anon_1: driver_sys::CUmemLocation_st__bindgen_ty_1 { id },
+                    #[cfg(not(cuMemLocation_anon_id))]
                     id,
                 },
                 #[cfg(not(cuMemAdvise_v2))]