sklam · Mar 19, 2024 · Mar 10, 2024 · Mar 10, 2024 · Mar 10, 2024 · Mar 10, 2024
diff --git a/buildscripts/condarecipe.local/meta.yaml b/buildscripts/condarecipe.local/meta.yaml
@@ -54,8 +54,9 @@ requirements:
     # Exclude 0.3.20 too
     # https://github.com/numba/numba/issues/8096
     - libopenblas >=0.3.18, !=0.3.20     # [arm64]
-    # CUDA 10.2 or later is required for CUDA support
-    - cudatoolkit >=10.2
+    # CUDA 11.2 or later is required for CUDA support
+    - cuda-version >=11.2
+    - cudatoolkit >=11.2
     # scipy 1.0 or later
     - scipy >=1.0
     # CUDA Python 11.6 or later

diff --git a/docs/source/cuda/minor_version_compatibility.rst b/docs/source/cuda/minor_version_compatibility.rst
@@ -5,39 +5,61 @@ CUDA Minor Version Compatibility
 
 CUDA `Minor Version Compatibility
 <https://docs.nvidia.com/deploy/cuda-compatibility/index.html#minor-version-compatibility>`_
-(MVC) enables the use of a newer CUDA toolkit version than the CUDA version
-supported by the driver, provided that the toolkit and driver both have the same
-major version. For example, use of CUDA toolkit 11.5 with CUDA driver 450 (CUDA
+(MVC) enables the use of a newer CUDA Toolkit version than the CUDA version
+supported by the driver, provided that the Toolkit and driver both have the same
+major version. For example, use of CUDA Toolkit 11.5 with CUDA driver 450 (CUDA
 version 11.0) is supported through MVC.
 
+Numba supports MVC for CUDA 12 on Linux using the external ``pynvjitlink``
+package.
+
 Numba supports MVC for CUDA 11 on Linux using the external ``cubinlinker`` and
 ``ptxcompiler`` packages, subject to the following limitations:
 
 - Linking of archives is unsupported.
 - Cooperative Groups are unsupported, because they require an archive to be
   linked.
 
-MVC is not yet supported on CUDA 12 - this will be added in a future release of
-Numba.
-
 MVC is not supported on Windows.
 
 
-Enabling MVC Support
---------------------
+Installation
+------------
+
+CUDA 12
+~~~~~~~
+
+To use MVC support, the ``pynvjitlink`` package must be installed. To install
+using conda, use:
+
+.. code:: bash
+
+   conda install -c rapidsai pynvjitlink
+
+To install with pip, use the NVIDIA package index:
+
+.. code:: bash
+
+   pip install --extra-index-url https://pypi.nvidia.com pynvjitlink-cu12
+
+CUDA 12
+~~~~~~~
 
 To use MVC support, the ``cubinlinker`` and ``ptxcompiler`` compiler packages
 must be installed from the appropriate channels. To install using conda, use:
 
 .. code:: bash
 
-   conda install rapidsai::cubinlinker conda-forge::ptxcompiler
+   conda install -c rapidsai -c conda-forge cubinlinker ptxcompiler
 
 To install with pip, use the NVIDIA package index:
 
 .. code:: bash
 
-   pip install ptxcompiler-cu11 cubinlinker-cu11 --extra-index-url=https://pypi.nvidia.com
+   pip install --extra-index-url https://pypi.nvidia.com ptxcompiler-cu11 cubinlinker-cu11
+
+Enabling MVC Support
+--------------------
 
 MVC support is enabled by setting the environment variable:
 
@@ -62,6 +84,7 @@ Further information about Minor Version Compatibility may be found in:
 
 - The `CUDA Compatibility Guide
   <https://docs.nvidia.com/deploy/cuda-compatibility/index.html>`_.
+- The `README for pynvjitlink
+  <https://github.com/rapidsai/pynvjitlink/blob/main/README.md>`_.
 - The `README for ptxcompiler
   <https://github.com/rapidsai/ptxcompiler/blob/main/README.md>`_.
-
diff --git a/docs/source/cuda/overview.rst b/docs/source/cuda/overview.rst
@@ -60,9 +60,15 @@ Presently 11.2 is the minimum required toolkit version. An NVIDIA driver
 sufficient for the toolkit version is also required (see also
 :ref:`minor-version-compatibility`).
 
-If you are using Conda, you can install the CUDA toolkit with::
+Conda users can install the CUDA Toolkit into a conda environment.
 
-   $ conda install cudatoolkit
+For CUDA 12, ``cuda-nvcc`` and ``cuda-nvrtc`` are required::
+
+    $ conda install -c conda-forge cuda-nvcc cuda-nvrtc "cuda-version>=12.0"
+
+For CUDA 11, ``cudatoolkit`` is required::
+
+    $ conda install -c conda-forge cudatoolkit "cuda-version>=11.2,<12.0"
 
 If you are not using Conda or if you want to use a different version of CUDA
 toolkit, the following describes how Numba searches for a CUDA toolkit
@@ -82,7 +88,7 @@ release.
 
 You can install the NVIDIA bindings with::
 
-   $ conda install nvidia::cuda-python
+   $ conda install -c conda-forge cuda-python
 
 if you are using Conda, or::
 
@@ -100,23 +106,21 @@ Setting CUDA Installation Path
 
 Numba searches for a CUDA toolkit installation in the following order:
 
-1. Conda installed `cudatoolkit` package.
+1. Conda installed CUDA Toolkit packages
 2. Environment variable ``CUDA_HOME``, which points to the directory of the
-   installed CUDA toolkit (i.e. ``/home/user/cuda-10``)
+   installed CUDA toolkit (i.e. ``/home/user/cuda-12``)
 3. System-wide installation at exactly ``/usr/local/cuda`` on Linux platforms.
-   Versioned installation paths (i.e. ``/usr/local/cuda-10.0``) are intentionally
+   Versioned installation paths (i.e. ``/usr/local/cuda-12.0``) are intentionally
    ignored.  Users can use ``CUDA_HOME`` to select specific versions.
 
 In addition to the CUDA toolkit libraries, which can be installed by conda into
 an environment or installed system-wide by the `CUDA SDK installer
 <(https://developer.nvidia.com/cuda-downloads)>`_, the CUDA target in Numba
 also requires an up-to-date NVIDIA graphics driver.  Updated graphics drivers
 are also installed by the CUDA SDK installer, so there is no need to do both.
-Note that on macOS, the CUDA SDK must be installed to get the required driver,
-and the driver is only supported on macOS prior to 10.14 (Mojave).  If the
-``libcuda`` library is in a non-standard location, users can set environment
-variable ``NUMBA_CUDA_DRIVER`` to the file path (not the directory path) of the
-shared library file.
+If the ``libcuda`` library is in a non-standard location, users can set
+environment variable ``NUMBA_CUDA_DRIVER`` to the file path (not the directory
+path) of the shared library file.
 
 
 Missing CUDA Features

diff --git a/docs/source/reference/envvars.rst b/docs/source/reference/envvars.rst
@@ -489,8 +489,8 @@ GPU support
 
    The default compute capability (a string of the type ``major.minor``) to
    target when compiling to PTX using ``cuda.compile_ptx``. The default is
-   5.2, which is the lowest non-deprecated compute capability in the most
-   recent version of the CUDA toolkit supported (11.0 at present).
+   5.0, which is the lowest non-deprecated compute capability in the most
+   recent version of the CUDA toolkit supported (12.4 at present).
 
 .. envvar:: NUMBA_ENABLE_CUDASIM
 

diff --git a/docs/source/user/installing.rst b/docs/source/user/installing.rst
@@ -45,12 +45,17 @@ Note that Numba, like Anaconda, only supports PPC in 64-bit little-endian mode.
 To enable CUDA GPU support for Numba, install the latest `graphics drivers from
 NVIDIA <https://www.nvidia.com/Download/index.aspx>`_ for your platform.
 (Note that the open source Nouveau drivers shipped by default with many Linux
-distributions do not support CUDA.)  Then install the ``cudatoolkit`` package::
+distributions do not support CUDA.)  Then install the CUDA Toolkit package.
 
-    $ conda install cudatoolkit
+For CUDA 12, ``cuda-nvcc`` and ``cuda-nvrtc`` are required::
 
-You do not need to install the CUDA SDK from NVIDIA.
+    $ conda install -c conda-forge cuda-nvcc cuda-nvrtc "cuda-version>=12.0"
+
+For CUDA 11, ``cudatoolkit`` is required::
 
+    $ conda install -c conda-forge cudatoolkit "cuda-version>=11.2,<12.0"
+
+You do not need to install the CUDA SDK from NVIDIA.
 
 Installing using pip on x86/x86_64 Platforms
 --------------------------------------------

diff --git a/docs/upcoming_changes/9487.doc.rst b/docs/upcoming_changes/9487.doc.rst
@@ -0,0 +1,4 @@
+CUDA 12 conda installation documentation
+----------------------------------------
+
+Installation instructions have been added for CUDA 12 conda users.
@@ -1,8 +1,9 @@
-"""Cudatoolkit libraries lookup utilities.
+"""CUDA Toolkit libraries lookup utilities.
 
-Cudatoolkit libraries can be available via either:
+CUDA Toolkit libraries can be available via either:
 
-- the `cudatoolkit` conda package,
+- the `cuda-nvcc` and `cuda-nvrtc` conda packages for CUDA 12,
+- the `cudatoolkit` conda package for CUDA 11,
 - a user supplied location from CUDA_HOME,
 - a system wide location,
 - package-specific locations (e.g. the Debian NVIDIA packages),

@@ -330,6 +330,7 @@ def get_log(self):
     (12, 1): ((5, 0), (9, 0)),
     (12, 2): ((5, 0), (9, 0)),
     (12, 3): ((5, 0), (9, 0)),
+    (12, 4): ((5, 0), (9, 0)),
 }
 
 
@@ -416,10 +417,14 @@ def get_arch_option(major, minor):
 
 
 MISSING_LIBDEVICE_FILE_MSG = '''Missing libdevice file.
-Please ensure you have package cudatoolkit >= 11.0
-Install package by:
+Please ensure you have a CUDA Toolkit 11.2 or higher.
+For CUDA 12, ``cuda-nvcc`` and ``cuda-nvrtc`` are required:
 
-    conda install cudatoolkit
+    $ conda install -c conda-forge cuda-nvcc cuda-nvrtc "cuda-version>=12.0"
+
+For CUDA 11, ``cudatoolkit`` is required:
+
+    $ conda install -c conda-forge cudatoolkit "cuda-version>=11.2,<12.0"
 '''