Add CUDA 13 support and add metapackages (#320)

## Summary
This PR adds support for CUDA 13 (preserving full support for CUDA 12).

## Details

* Bump CUDA-Q commit to one that supports CUDA 13
(NVIDIA/cuda-quantum#3502)
* Produce *2 sets* of wheels, bringing our total number of *regular*
wheels to 4:
  * cudaq-qec-cu12
  * cudaq-qec-cu13
  * cudaq-solvers-cu12
  * cudaq-solvers-cu13
* Additionally, produce meta-packages (cudaq-qec and cudaq-solvers)
which will automatically choose the appropriate subpackage. We already
have cudaq-qec and cudaq-solvers wheels; these 0.5.0 meta-packages will
just replace the older 0.4.0 packages at the time of release.
* These meta-packages will contain the same optional dependencies as the
child packages. That is - `pip install
cudaq-qec[tensor-network-decoder]` will use the metapackage to
automatically transform that into `pip install
cudaq-qec-cuXY[tensor-network-decoder]`, where `XY` is the appropriate
CUDA version. This also applies to `cudaq-solvers[gqe]`.
* Matrix the build and test portion of the CI to be running both 12.6
and 13.0.

## Todo (in a separate PR)

* Once internal pipelines are updated, update the "release" workflows to
incorporate private artifacts from both cu12 and cu13 images.
* Consider removing the `torch` dependency, and instead update the docs
to have users manually install the version that they want (pending
discussion).

---------

Signed-off-by: Ben Howe <[email protected]>

Author: bmhowe23

.cudaq_version

@@ -1,6 +1,6 @@
 {
   "cudaq": {
     "repository": "NVIDIA/cuda-quantum",
-    "ref": "7781a8c856b84d5d7714a66ce6a26e4dc804830e"
+    "ref": "43f539c7e45a0a47c56d43be237ffc673b23222b"
   }
 }

.github/actions/get-cudaq-build/build_cudaq.sh

@@ -13,9 +13,13 @@ LAUNCHER=${2:-""}
 CC=${3:-"gcc"}
 CXX=${4:-"g++"}
 
+# Set CUDA suffix to, for example, "cu12" or "cu13", based on the cuquantum
+# package name.
+CUDA_SUFFIX=$(pip list | grep cuquantum-cu | head -1 | cut -d' ' -f1 | cut -d'-' -f2)
+
 LLVM_INSTALL_PREFIX=/usr/local/llvm
-CUQUANTUM_INSTALL_PREFIX="$(pip show cuquantum-cu12 | grep "Location:" | cut -d " " -f 2)/cuquantum"
-CUTENSOR_INSTALL_PREFIX="$(pip show cutensor-cu12 | grep "Location:" | cut -d " " -f 2)/cutensor"
+CUQUANTUM_INSTALL_PREFIX="$(pip show cuquantum-${CUDA_SUFFIX} | grep "Location:" | cut -d " " -f 2)/cuquantum"
+CUTENSOR_INSTALL_PREFIX="$(pip show cutensor-${CUDA_SUFFIX} | grep "Location:" | cut -d " " -f 2)/cutensor"
 
 cd cudaq
 

.github/actions/get-cudaq-wheels/action.yaml

@@ -52,7 +52,7 @@ runs:
           .cudaq_version
       run: |
         hash=${{ hashFiles(format('{0}', env.to_hash)) }}
-        echo "main=cudaq-wheels-${{ inputs.platform }}-${{ inputs.ref }}-$hash" >> $GITHUB_OUTPUT
+        echo "main=cudaq-wheels-${{ inputs.platform }}-cu${{ inputs.cuda_version }}-${{ inputs.ref }}-$hash" >> $GITHUB_OUTPUT
         if [[ -n "${{ inputs.pr-number }}" ]]; then
           echo "pr=-pr${{ inputs.pr-number }}" >> $GITHUB_OUTPUT
         fi

.github/workflows/all_libs.yaml

@@ -11,7 +11,7 @@ jobs:
       fail-fast: false
       matrix:
         platform: ['amd64', 'arm64']
-        cuda_version: ['12.6']
+        cuda_version: ['12.6', '13.0']
     runs-on: ${{ startsWith(github.repository, 'NVIDIA/cudaqx') && format('linux-{0}-cpu8', matrix.platform) || 'ubuntu-latest' }}
     container: ghcr.io/nvidia/cuda-quantum-devdeps:ext-${{ matrix.platform }}-cu${{ matrix.cuda_version }}-gcc11-main
     permissions:

.github/workflows/build_dev.yaml

@@ -25,7 +25,7 @@ jobs:
     strategy:
       matrix:
         platform: [amd64, arm64]
-        cuda_version: ['12.6']
+        cuda_version: ['12.6', '13.0']
       fail-fast: false
     runs-on: linux-${{ matrix.platform }}-cpu8
     steps:
@@ -71,17 +71,24 @@ jobs:
     - name: Build and push
       if: ${{ inputs.force_rebuild || steps.check-image.outputs.IMAGE_EXISTS == 'false' }}
       run: |
-        other_tag="latest"
+        other_tag=""
+        is_main_branch=${{ github.ref_name == 'main' }}
         is_versioned=${{ startsWith(github.ref_name, 'releases/') }}
-        if ${is_versioned}; then
+        if ${is_main_branch}; then
+          other_tag="latest"
+        elif ${is_versioned}; then
           ver=`echo ${{ github.ref_name }} | egrep -o "([0-9]{1,}\.)+[0-9]{1,}"`
           echo "Versioned release: $ver"
           other_tag=$ver
         fi
         TAGS="-t ghcr.io/nvidia/cudaqx-dev:${{ steps.get-cudaq-version-short.outputs.commit_date }}-${{ steps.get-cudaq-version-short.outputs.shortref }}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
         TAGS+=" -t ghcr.io/nvidia/cudaqx-dev:${{ steps.get-cudaq-version-short.outputs.shortref }}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
-        TAGS+=" -t ghcr.io/nvidia/cudaqx-dev:${other_tag}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
-        docker build $TAGS -f docker/build_env/cudaqx.dev.Dockerfile . --build-arg base_image=ghcr.io/nvidia/cuda-quantum-devdeps:ext-${{ matrix.platform }}-cu${{ matrix.cuda_version }}-gcc11-main
+        if [ -n "$other_tag" ]; then
+          TAGS+=" -t ghcr.io/nvidia/cudaqx-dev:${other_tag}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
+        fi
+        docker build $TAGS -f docker/build_env/cudaqx.dev.Dockerfile . \
+          --build-arg base_image=ghcr.io/nvidia/cuda-quantum-devdeps:ext-${{ matrix.platform }}-cu${{ matrix.cuda_version }}-gcc11-main \
+          --build-arg cuda_version=${{ matrix.cuda_version }}
         docker push -a ghcr.io/nvidia/cudaqx-dev
       shell: bash --noprofile --norc -euo pipefail {0}
 
@@ -92,7 +99,7 @@ jobs:
       matrix:
         python: ['3.11', '3.12', '3.13']
         platform: ['amd64', 'arm64']
-        cuda_version: ['12.6']
+        cuda_version: ['12.6', '13.0']
       fail-fast: false
     runs-on: linux-${{ matrix.platform }}-cpu8
     steps:
@@ -155,6 +162,7 @@ jobs:
           base_image=ghcr.io/nvidia/cuda-quantum-devdeps:manylinux-${{ matrix.platform }}-cu${{ matrix.cuda_version }}-gcc11-main
           release_version=0.99.99
           python_version=${{ matrix.python }}
+          cuda_version=${{ matrix.cuda_version }}
         outputs: type=local,dest=/tmp/wheels
 
     - name: Build and push dev image with prebuilt CUDA-Q wheels
@@ -168,17 +176,23 @@ jobs:
         TAGS="-t ghcr.io/nvidia/cudaqx-dev:${{ steps.get-cudaq-version-short.outputs.commit_date }}-${{ steps.get-cudaq-version-short.outputs.shortref }}-py${{ matrix.python }}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
         TAGS+=" -t ghcr.io/nvidia/cudaqx-dev:${{ steps.get-cudaq-version-short.outputs.shortref }}-py${{ matrix.python }}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
 
-        other_tag="latest"
+        other_tag=""
+        is_main_branch=${{ github.ref_name == 'main' }}
         is_versioned=${{ startsWith(github.ref_name, 'releases/') }}
-        if ${is_versioned}; then
+        if ${is_main_branch}; then
+          other_tag="latest"
+        elif ${is_versioned}; then
           ver=`echo ${{ github.ref_name }} | egrep -o "([0-9]{1,}\.)+[0-9]{1,}"`
           echo "Versioned release: $ver"
           other_tag=$ver
         fi
 
-        TAGS+=" -t ghcr.io/nvidia/cudaqx-dev:${other_tag}-py${{ matrix.python }}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
+        if [ -n "$other_tag" ]; then
+          TAGS+=" -t ghcr.io/nvidia/cudaqx-dev:${other_tag}-py${{ matrix.python }}-${{ matrix.platform }}-cu${{ matrix.cuda_version }}"
+        fi
         BUILDARGS="--build-arg base_image=ghcr.io/nvidia/cuda-quantum-devdeps:manylinux-${{ matrix.platform }}-cu${{ matrix.cuda_version }}-gcc11-main"
         BUILDARGS+=" --build-arg python_version=${{ matrix.python }}"
+        BUILDARGS+=" --build-arg cuda_version=${{ matrix.cuda_version }}"
         # For some reason, this fails on amd64 unless DOCKER_BUILDKIT=0 is set.
         # The exact error is: too many open files.
         DOCKER_BUILDKIT=0 docker build $TAGS $BUILDARGS -f docker/build_env/cudaqx.wheel.Dockerfile .