@@ -6,6 +6,9 @@ cuda/_version.py export-subst
 # we do not own any headers checked in, don't touch them
 *.h binary
 *.hpp binary
+# Exception: headers we own (cuda_core C++ implementation)
+cuda_core/cuda/core/_cpp/*.h -binary text diff
+cuda_core/cuda/core/_cpp/*.hpp -binary text diff
 # git should not convert line endings in PNG files
 *.png binary
 *.svg binary
 
@@ -9,14 +9,25 @@ on:
     types: [closed, labeled]
     branches:
       - main
+  workflow_dispatch:
+    inputs:
+      backport-branch:
+        description: "Branch to backport commits onto"
+        required: false
+        type: string
+      pull-request:
+        description: "PR to backport"
+        required: true
+        type: number
+
 
 permissions:
   contents: write       # so it can comment
   pull-requests: write  # so it can create pull requests
 
 jobs:
-  backport:
-    name: Backport pull request
+  backport-from-pr:
+    name: Backport directly from a pull request
     if: ${{ github.repository_owner == 'nvidia' &&
             github.event.pull_request.merged == true &&
             contains( github.event.pull_request.labels.*.name, 'to-be-backported')
@@ -38,3 +49,28 @@ jobs:
           copy_labels_pattern: true
           copy_requested_reviewers: true
           target_branches: ${{ env.OLD_BRANCH }}
+  backport-to-branch:
+    name: Backport a specific PR against a specific branch
+    if: github.repository_owner == 'nvidia' && github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8  # v6.0.1
+
+      - name: Load branch from environment name
+        if: inputs.backport-branch == null
+        run: |
+          BACKPORT_BRANCH=$(yq '.backport_branch' ci/versions.yml)
+          echo "BACKPORT_BRANCH=${BRANCH}" >> $GITHUB_ENV
+
+      - name: Load branch name from input
+        if: inputs.backport-branch != null
+        run: echo "BACKPORT_BRANCH=${{ inputs.backport-branch }}" >> $GITHUB_ENV
+
+      - name: Create backport pull requests
+        uses: korthout/backport-action@c656f5d5851037b2b38fb5db2691a03fa229e3b2  # v4.0.1
+        with:
+          copy_assignees: true
+          copy_labels_pattern: true
+          copy_requested_reviewers: true
+          target_branches: ${{ env.BACKPORT_BRANCH }}
+          source_pr_number: ${{ inputs.pull-request }}
@@ -11,6 +11,7 @@ __pycache__/
 .pytest_cache/
 .benchmarks/
 *.cpp
+!*_impl.cpp
 !cuda_bindings/cuda/bindings/_lib/param_packer.cpp
 !cuda_bindings/cuda/bindings/_bindings/loader.cpp
 cache_driver
 
@@ -150,15 +150,21 @@ def merge_wheels(wheels: List[Path], output_dir: Path, show_wheel_contents: bool
             "__init__.py",
             "_version.py",
             "_include",
+            "_cpp",  # Headers for Cython development
             "cu12",
             "cu13",
         )
+        # _resource_handles is shared (not CUDA-version-specific) and must stay
+        # at top level. It's imported early in __init__.py before versioned code.
+        items_to_keep_prefix = ("_resource_handles",)
         all_items = os.scandir(base_wheel / base_dir)
         removed_count = 0
         for f in all_items:
             f_abspath = f.path
             if f.name in items_to_keep:
                 continue
+            if any(f.name.startswith(prefix) for prefix in items_to_keep_prefix):
+                continue
             if f.is_dir():
                 print(f"  Removing directory: {f.name}", file=sys.stderr)
                 shutil.rmtree(f_abspath)
 
@@ -50,6 +50,7 @@ cdef class _HelperCUpointer_attribute:
     # Return values
     cdef driver.CUcontext _ctx
     cdef unsigned int _uint
+    cdef int _int
     cdef driver.CUdeviceptr _devptr
     cdef void** _void
     cdef driver.CUDA_POINTER_ATTRIBUTE_P2P_TOKENS _token
 
@@ -247,12 +247,14 @@ cdef class _HelperCUpointer_attribute:
             else:
                 self._cptr = <void*><void_ptr>init_value.getPtr()
         elif self._attr in ({{if 'CU_POINTER_ATTRIBUTE_MEMORY_TYPE'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_MEMORY_TYPE,{{endif}}
-                            {{if 'CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_ACCESS_FLAGS'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ACCESS_FLAGS,{{endif}}):
             self._uint = init_value
             self._cptr = <void*>&self._uint
+        elif self._attr in ({{if 'CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,{{endif}}):
+            self._int = init_value
+            self._cptr = <void*>&self._int
         elif self._attr in ({{if 'CU_POINTER_ATTRIBUTE_DEVICE_POINTER'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_POINTER,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_RANGE_START_ADDR'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_RANGE_START_ADDR,{{endif}}):
             if self._is_getter: