diff --git a/.github/workflows/ci-nightly.yml b/.github/workflows/ci-nightly.yml
new file mode 100644
index 00000000000..72778445a91
--- /dev/null
+++ b/.github/workflows/ci-nightly.yml
@@ -0,0 +1,214 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# Nightly CI pipeline that tests optional dependencies (PyTorch, numba-cuda)
+# against the latest cuda-python wheels built on main.
+#
+# This workflow does NOT build wheels — it downloads them from the latest
+# successful CI run on main and runs integration tests with optional deps.
+
+name: "CI: Nightly optional-deps"
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.event_name }}
+  cancel-in-progress: true
+
+on:
+  push:
+    branches:
+      - "main"
+      - "pull-request/[0-9]+"
+  schedule:
+    # 2 AM UTC daily, after the midnight main CI build finishes
+    - cron: "0 2 * * *"
+  workflow_dispatch:
+    inputs:
+      run-id:
+        description: >
+          Override the CI run ID to download artifacts from.
+          Leave empty to auto-detect the latest successful main run.
+        type: string
+        default: ''
+
+jobs:
+  find-wheels:
+    runs-on: ubuntu-latest
+    outputs:
+      RUN_ID: ${{ steps.find.outputs.run_id }}
+      HEAD_SHA: ${{ steps.find.outputs.head_sha }}
+      CUDA_BUILD_VER: ${{ steps.get-vars.outputs.cuda_build_ver }}
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          fetch-depth: 1
+
+      - name: Get CUDA build versions
+        id: get-vars
+        run: |
+          cuda_build_ver=$(yq '.cuda.build.version' ci/versions.yml)
+          echo "cuda_build_ver=$cuda_build_ver" >> $GITHUB_OUTPUT
+
+      - name: Find latest successful CI run on main
+        id: find
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          if [[ -n "${{ inputs.run-id }}" ]]; then
+            RUN_ID="${{ inputs.run-id }}"
+            echo "Using manually specified run ID: $RUN_ID"
+          else
+            RUN_ID=$(gh run list \
+              -b main \
+              -L 1 \
+              -w "CI" \
+              -s success \
+              -R "${{ github.repository }}" \
+              --json databaseId \
+              | jq -r '.[0].databaseId')
+
+            if [[ -z "$RUN_ID" || "$RUN_ID" == "null" ]]; then
+              echo "::error::No successful CI run found on main"
+              exit 1
+            fi
+            echo "Using latest successful CI run: $RUN_ID"
+          fi
+
+          # Resolve the head SHA from the CI run — artifact names embed this.
+          HEAD_SHA=$(gh run view "$RUN_ID" \
+            -R "${{ github.repository }}" \
+            --json headSha \
+            | jq -r '.headSha')
+
+          echo "run_id=$RUN_ID" >> $GITHUB_OUTPUT
+          echo "head_sha=$HEAD_SHA" >> $GITHUB_OUTPUT
+          echo "Source commit: $HEAD_SHA"
+
+  # ── PyTorch interop tests ──
+
+  test-pytorch-linux:
+    name: "Nightly PyTorch (linux-64)"
+    if: ${{ github.repository_owner == 'nvidia' }}
+    needs: find-wheels
+    permissions:
+      contents: read
+      actions: read
+    secrets: inherit
+    uses: ./.github/workflows/test-wheel-linux.yml
+    with:
+      build-type: nightly
+      host-platform: linux-64
+      build-ctk-ver: ${{ needs.find-wheels.outputs.CUDA_BUILD_VER }}
+      run-id: ${{ needs.find-wheels.outputs.RUN_ID }}
+      sha: ${{ needs.find-wheels.outputs.HEAD_SHA }}
+      test-mode: nightly-pytorch
+      matrix_filter: 'map(select(.MODE == "nightly-pytorch"))'
+
+  test-pytorch-windows:
+    name: "Nightly PyTorch (win-64)"
+    if: ${{ github.repository_owner == 'nvidia' }}
+    needs: find-wheels
+    permissions:
+      contents: read
+      actions: read
+    secrets: inherit
+    uses: ./.github/workflows/test-wheel-windows.yml
+    with:
+      build-type: nightly
+      host-platform: win-64
+      build-ctk-ver: ${{ needs.find-wheels.outputs.CUDA_BUILD_VER }}
+      run-id: ${{ needs.find-wheels.outputs.RUN_ID }}
+      sha: ${{ needs.find-wheels.outputs.HEAD_SHA }}
+      test-mode: nightly-pytorch
+      matrix_filter: 'map(select(.MODE == "nightly-pytorch"))'
+
+  # ── numba-cuda tests ──
+
+  test-numba-cuda-linux-64:
+    name: "Nightly numba-cuda (linux-64)"
+    if: ${{ github.repository_owner == 'nvidia' }}
+    needs: find-wheels
+    permissions:
+      contents: read
+      actions: read
+    secrets: inherit
+    uses: ./.github/workflows/test-wheel-linux.yml
+    with:
+      build-type: nightly
+      host-platform: linux-64
+      build-ctk-ver: ${{ needs.find-wheels.outputs.CUDA_BUILD_VER }}
+      run-id: ${{ needs.find-wheels.outputs.RUN_ID }}
+      sha: ${{ needs.find-wheels.outputs.HEAD_SHA }}
+      test-mode: nightly-numba-cuda
+      matrix_filter: 'map(select(.MODE == "nightly-numba-cuda"))'
+
+  test-numba-cuda-linux-aarch64:
+    name: "Nightly numba-cuda (linux-aarch64)"
+    if: ${{ github.repository_owner == 'nvidia' }}
+    needs: find-wheels
+    permissions:
+      contents: read
+      actions: read
+    secrets: inherit
+    uses: ./.github/workflows/test-wheel-linux.yml
+    with:
+      build-type: nightly
+      host-platform: linux-aarch64
+      build-ctk-ver: ${{ needs.find-wheels.outputs.CUDA_BUILD_VER }}
+      run-id: ${{ needs.find-wheels.outputs.RUN_ID }}
+      sha: ${{ needs.find-wheels.outputs.HEAD_SHA }}
+      test-mode: nightly-numba-cuda
+      matrix_filter: 'map(select(.MODE == "nightly-numba-cuda"))'
+
+  test-numba-cuda-windows:
+    name: "Nightly numba-cuda (win-64)"
+    if: ${{ github.repository_owner == 'nvidia' }}
+    needs: find-wheels
+    permissions:
+      contents: read
+      actions: read
+    secrets: inherit
+    uses: ./.github/workflows/test-wheel-windows.yml
+    with:
+      build-type: nightly
+      host-platform: win-64
+      build-ctk-ver: ${{ needs.find-wheels.outputs.CUDA_BUILD_VER }}
+      run-id: ${{ needs.find-wheels.outputs.RUN_ID }}
+      sha: ${{ needs.find-wheels.outputs.HEAD_SHA }}
+      test-mode: nightly-numba-cuda
+      matrix_filter: 'map(select(.MODE == "nightly-numba-cuda"))'
+
+  # ── Status check ──
+
+  checks:
+    name: Nightly check status
+    if: always()
+    runs-on: ubuntu-latest
+    needs:
+      - test-pytorch-linux
+      - test-pytorch-windows
+      - test-numba-cuda-linux-64
+      - test-numba-cuda-linux-aarch64
+      - test-numba-cuda-windows
+    steps:
+      - name: Exit
+        run: |
+          # If any dependency was cancelled or failed, that's a failure.
+          #
+          # See ci.yml for the full rationale on why we must use always()
+          # and explicitly check each result rather than relying on the
+          # default behaviour.
+          if ${{ needs.test-pytorch-linux.result == 'cancelled' ||
+                 needs.test-pytorch-linux.result == 'failure' ||
+                 needs.test-pytorch-windows.result == 'cancelled' ||
+                 needs.test-pytorch-windows.result == 'failure' ||
+                 needs.test-numba-cuda-linux-64.result == 'cancelled' ||
+                 needs.test-numba-cuda-linux-64.result == 'failure' ||
+                 needs.test-numba-cuda-linux-aarch64.result == 'cancelled' ||
+                 needs.test-numba-cuda-linux-aarch64.result == 'failure' ||
+                 needs.test-numba-cuda-windows.result == 'cancelled' ||
+                 needs.test-numba-cuda-windows.result == 'failure' }}; then
+            exit 1
+          fi
+          exit 0
diff --git a/.github/workflows/test-wheel-linux.yml b/.github/workflows/test-wheel-linux.yml
index 35c5e6c3734..e1a36bc086b 100644
--- a/.github/workflows/test-wheel-linux.yml
+++ b/.github/workflows/test-wheel-linux.yml
@@ -29,6 +29,24 @@ on:
       skip-bindings-test:
         type: boolean
         default: false
+      run-id:
+        description: >
+          Workflow run ID to download artifacts from.
+          Defaults to the current run when empty.
+        type: string
+        default: ''
+      test-mode:
+        description: >
+          Test mode: 'standard' (default), 'nightly-pytorch', or
+          'nightly-numba-cuda'.
+        type: string
+        default: 'standard'
+      sha:
+        description: >
+          Commit SHA used to construct artifact names.
+          Defaults to github.sha (current run) when empty.
+        type: string
+        default: ''
 
 defaults:
   run:
@@ -119,7 +137,7 @@ jobs:
           HOST_PLATFORM: ${{ inputs.host-platform }}
           LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
           PY_VER: ${{ matrix.PY_VER }}
-          SHA: ${{ github.sha }}
+          SHA: ${{ inputs.sha || github.sha }}
           SKIP_BINDINGS_TEST_OVERRIDE: ${{ inputs.skip-bindings-test && '1' || '0' }}
         run: ./ci/tools/env-vars test
 
@@ -128,6 +146,8 @@ jobs:
         with:
           name: cuda-pathfinder-wheel
           path: ./cuda_pathfinder
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Download cuda-python build artifacts
         if: ${{ env.USE_BACKPORT_BINDINGS == '0' }}
@@ -135,6 +155,8 @@ jobs:
         with:
           name: cuda-python-wheel
           path: .
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Download cuda.bindings build artifacts
         if: ${{ env.USE_BACKPORT_BINDINGS == '0' }}
@@ -142,6 +164,8 @@ jobs:
         with:
           name: ${{ env.CUDA_BINDINGS_ARTIFACT_NAME }}
           path: ${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Download cuda-python & cuda.bindings build artifacts from the prior branch
         if: ${{ env.USE_BACKPORT_BINDINGS == '1' }}
@@ -181,7 +205,7 @@ jobs:
       - name: Display structure of downloaded cuda-python artifacts
         run: |
           pwd
-          ls -lahR .
+          ls -lah cuda_python*.whl cuda_pathfinder/
 
       - name: Display structure of downloaded cuda.bindings artifacts
         run: |
@@ -194,6 +218,8 @@ jobs:
         with:
           name: ${{ env.CUDA_BINDINGS_ARTIFACT_NAME }}-tests
           path: ${{ env.CUDA_BINDINGS_CYTHON_TESTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Display structure of downloaded cuda.bindings Cython tests
         if: ${{ env.SKIP_CYTHON_TEST == '0' }}
@@ -206,6 +232,8 @@ jobs:
         with:
           name: ${{ env.CUDA_CORE_ARTIFACT_NAME }}
           path: ${{ env.CUDA_CORE_ARTIFACTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Display structure of downloaded cuda.core build artifacts
         run: |
@@ -218,6 +246,8 @@ jobs:
         with:
           name: ${{ env.CUDA_CORE_ARTIFACT_NAME }}-tests
           path: ${{ env.CUDA_CORE_CYTHON_TESTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Display structure of downloaded cuda.core Cython tests
         if: ${{ env.SKIP_CYTHON_TEST == '0' }}
@@ -256,7 +286,9 @@ jobs:
       - name: Set up test repetition on nightly runs
         run: echo "PYTEST_ADDOPTS=\"--count=${{ inputs.nruns }}\"" >> "$GITHUB_ENV"
 
+      # ── Standard test steps (skipped for nightly modes) ──
       - name: Run cuda.pathfinder tests with see_what_works
+        if: ${{ inputs.test-mode == 'standard' }}
         env:
           CUDA_PATHFINDER_TEST_LOAD_NVIDIA_DYNAMIC_LIB_STRICTNESS: see_what_works
           CUDA_PATHFINDER_TEST_FIND_NVIDIA_HEADERS_STRICTNESS: see_what_works
@@ -264,14 +296,14 @@ jobs:
         run: run-tests pathfinder
 
       - name: Run cuda.bindings tests
-        if: ${{ env.SKIP_CUDA_BINDINGS_TEST == '0' }}
+        if: ${{ inputs.test-mode == 'standard' && env.SKIP_CUDA_BINDINGS_TEST == '0' }}
         env:
           CUDA_VER: ${{ matrix.CUDA_VER }}
           LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
         run: run-tests bindings
 
       - name: Run cuda.bindings benchmarks (smoke test)
-        if: ${{ env.SKIP_CUDA_BINDINGS_TEST == '0' }}
+        if: ${{ inputs.test-mode == 'standard' && env.SKIP_CUDA_BINDINGS_TEST == '0' }}
         run: |
           pip install pyperf
           pushd benchmarks/cuda_bindings
@@ -279,12 +311,14 @@ jobs:
           popd
 
       - name: Run cuda.core tests
+        if: ${{ inputs.test-mode == 'standard' }}
         env:
           CUDA_VER: ${{ matrix.CUDA_VER }}
           LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
         run: run-tests core
 
       - name: Ensure cuda-python installable
+        if: ${{ inputs.test-mode == 'standard' }}
         run: |
           if [[ "${{ matrix.LOCAL_CTK }}" == 1 ]]; then
             pip install --only-binary=:all: cuda_python*.whl
@@ -293,6 +327,7 @@ jobs:
           fi
 
       - name: Install cuda.pathfinder extra wheels for testing
+        if: ${{ inputs.test-mode == 'standard' }}
         run: |
           set -euo pipefail
           pushd cuda_pathfinder
@@ -301,8 +336,38 @@ jobs:
           popd
 
       - name: Run cuda.pathfinder tests with all_must_work
+        if: ${{ inputs.test-mode == 'standard' }}
         env:
           CUDA_PATHFINDER_TEST_LOAD_NVIDIA_DYNAMIC_LIB_STRICTNESS: all_must_work
           CUDA_PATHFINDER_TEST_FIND_NVIDIA_HEADERS_STRICTNESS: all_must_work
           CUDA_PATHFINDER_TEST_FIND_NVIDIA_BITCODE_LIB_STRICTNESS: all_must_work
         run: run-tests pathfinder
+
+      # ── Nightly: install wheels + optional dep together ──
+      - name: Install cuda-python wheels + PyTorch
+        if: ${{ inputs.test-mode == 'nightly-pytorch' }}
+        env:
+          CUDA_VER: ${{ matrix.CUDA_VER }}
+          LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
+          TORCH_VER: ${{ matrix.TORCH_VER }}
+          TORCH_CUDA: ${{ matrix.TORCH_CUDA }}
+        run: run-tests nightly-pytorch
+
+      - name: Install cuda-python wheels + numba-cuda
+        if: ${{ inputs.test-mode == 'nightly-numba-cuda' }}
+        env:
+          CUDA_VER: ${{ matrix.CUDA_VER }}
+          LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
+        run: run-tests nightly-numba-cuda
+
+      # ── Nightly: run tests ──
+      - name: Run PyTorch interop tests
+        if: ${{ inputs.test-mode == 'nightly-pytorch' }}
+        run: |
+          pushd cuda_core
+          pytest -rxXs -v --durations=0 tests/test_utils.py tests/example_tests/
+          popd
+
+      - name: Run numba-cuda tests
+        if: ${{ inputs.test-mode == 'nightly-numba-cuda' }}
+        run: python -m numba.runtests numba.cuda.tests
diff --git a/.github/workflows/test-wheel-windows.yml b/.github/workflows/test-wheel-windows.yml
index 765823c6bfc..eefc9273594 100644
--- a/.github/workflows/test-wheel-windows.yml
+++ b/.github/workflows/test-wheel-windows.yml
@@ -29,6 +29,24 @@ on:
       skip-bindings-test:
         type: boolean
         default: false
+      run-id:
+        description: >
+          Workflow run ID to download artifacts from.
+          Defaults to the current run when empty.
+        type: string
+        default: ''
+      test-mode:
+        description: >
+          Test mode: 'standard' (default), 'nightly-pytorch', or
+          'nightly-numba-cuda'.
+        type: string
+        default: 'standard'
+      sha:
+        description: >
+          Commit SHA used to construct artifact names.
+          Defaults to github.sha (current run) when empty.
+        type: string
+        default: ''
 
 jobs:
   compute-matrix:
@@ -113,7 +131,7 @@ jobs:
           HOST_PLATFORM: ${{ inputs.host-platform }}
           LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
           PY_VER: ${{ matrix.PY_VER }}
-          SHA: ${{ github.sha }}
+          SHA: ${{ inputs.sha || github.sha }}
           SKIP_BINDINGS_TEST_OVERRIDE: ${{ inputs.skip-bindings-test && '1' || '0' }}
         shell: bash --noprofile --norc -xeuo pipefail {0}
         run: ./ci/tools/env-vars test
@@ -123,6 +141,8 @@ jobs:
         with:
           name: cuda-pathfinder-wheel
           path: ./cuda_pathfinder
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Download cuda-python build artifacts
         if: ${{ env.USE_BACKPORT_BINDINGS == '0' }}
@@ -130,6 +150,8 @@ jobs:
         with:
           name: cuda-python-wheel
           path: .
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Download cuda.bindings build artifacts
         if: ${{ env.USE_BACKPORT_BINDINGS == '0' }}
@@ -137,6 +159,8 @@ jobs:
         with:
           name: ${{ env.CUDA_BINDINGS_ARTIFACT_NAME }}
           path: ${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Download cuda-python & cuda.bindings build artifacts from the prior branch
         if: ${{ env.USE_BACKPORT_BINDINGS == '1' }}
@@ -167,7 +191,8 @@ jobs:
       - name: Display structure of downloaded cuda-python artifacts
         run: |
           Get-Location
-          Get-ChildItem -Recurse -Force | Select-Object Mode, LastWriteTime, Length, FullName
+          Get-ChildItem cuda_python*.whl | Select-Object Mode, LastWriteTime, Length, FullName
+          Get-ChildItem cuda_pathfinder/ | Select-Object Mode, LastWriteTime, Length, FullName
 
       - name: Display structure of downloaded cuda.bindings artifacts
         run: |
@@ -180,6 +205,8 @@ jobs:
         with:
           name: ${{ env.CUDA_BINDINGS_ARTIFACT_NAME }}-tests
           path: ${{ env.CUDA_BINDINGS_CYTHON_TESTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Display structure of downloaded cuda.bindings Cython tests
         if: ${{ env.SKIP_CYTHON_TEST == '0' }}
@@ -192,6 +219,8 @@ jobs:
         with:
           name: ${{ env.CUDA_CORE_ARTIFACT_NAME }}
           path: ${{ env.CUDA_CORE_ARTIFACTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Display structure of downloaded cuda.core build artifacts
         run: |
@@ -204,6 +233,8 @@ jobs:
         with:
           name: ${{ env.CUDA_CORE_ARTIFACT_NAME }}-tests
           path: ${{ env.CUDA_CORE_CYTHON_TESTS_DIR }}
+          run-id: ${{ inputs.run-id || github.run_id }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Display structure of downloaded cuda.core Cython tests
         if: ${{ env.SKIP_CYTHON_TEST == '0' }}
@@ -237,7 +268,9 @@ jobs:
         shell: bash --noprofile --norc -xeuo pipefail {0}
         run: echo "PYTEST_ADDOPTS=\"--count=${{ inputs.nruns }}\"" >> "$GITHUB_ENV"
 
+      # ── Standard test steps (skipped for nightly modes) ──
       - name: Run cuda.pathfinder tests with see_what_works
+        if: ${{ inputs.test-mode == 'standard' }}
         env:
           CUDA_PATHFINDER_TEST_LOAD_NVIDIA_DYNAMIC_LIB_STRICTNESS: see_what_works
           CUDA_PATHFINDER_TEST_FIND_NVIDIA_HEADERS_STRICTNESS: see_what_works
@@ -246,7 +279,7 @@ jobs:
         run: run-tests pathfinder
 
       - name: Run cuda.bindings tests
-        if: ${{ env.SKIP_CUDA_BINDINGS_TEST == '0' }}
+        if: ${{ inputs.test-mode == 'standard' && env.SKIP_CUDA_BINDINGS_TEST == '0' }}
         env:
           CUDA_VER: ${{ matrix.CUDA_VER }}
           LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
@@ -254,6 +287,7 @@ jobs:
         run: run-tests bindings
 
       - name: Run cuda.core tests
+        if: ${{ inputs.test-mode == 'standard' }}
         env:
           CUDA_VER: ${{ matrix.CUDA_VER }}
           LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
@@ -261,6 +295,7 @@ jobs:
         run: run-tests core
 
       - name: Ensure cuda-python installable
+        if: ${{ inputs.test-mode == 'standard' }}
         run: |
           if ('${{ matrix.LOCAL_CTK }}' -eq '1') {
             pip install --only-binary=:all: (Get-ChildItem -Filter cuda_python*.whl).FullName
@@ -269,6 +304,7 @@ jobs:
           }
 
       - name: Install cuda.pathfinder extra wheels for testing
+        if: ${{ inputs.test-mode == 'standard' }}
         shell: bash --noprofile --norc -xeuo pipefail {0}
         run: |
           pushd cuda_pathfinder
@@ -277,9 +313,50 @@ jobs:
           popd
 
       - name: Run cuda.pathfinder tests with all_must_work
+        if: ${{ inputs.test-mode == 'standard' }}
         env:
           CUDA_PATHFINDER_TEST_LOAD_NVIDIA_DYNAMIC_LIB_STRICTNESS: all_must_work
           CUDA_PATHFINDER_TEST_FIND_NVIDIA_HEADERS_STRICTNESS: all_must_work
           CUDA_PATHFINDER_TEST_FIND_NVIDIA_BITCODE_LIB_STRICTNESS: all_must_work
         shell: bash --noprofile --norc -xeuo pipefail {0}
         run: run-tests pathfinder
+
+      # ── Nightly: install wheels + optional dep together ──
+      - name: Install Visual C++ Redistributable (required by PyTorch on Windows)
+        if: ${{ inputs.test-mode == 'nightly-pytorch' }}
+        run: |
+          Invoke-WebRequest -Uri "https://aka.ms/vs/17/release/vc_redist.x64.exe" -OutFile "vc_redist.x64.exe"
+          Start-Process -FilePath ".\vc_redist.x64.exe" -ArgumentList "/install", "/quiet", "/norestart" -Wait
+          Remove-Item "vc_redist.x64.exe"
+
+      - name: Install cuda-python wheels + PyTorch
+        if: ${{ inputs.test-mode == 'nightly-pytorch' }}
+        env:
+          CUDA_VER: ${{ matrix.CUDA_VER }}
+          LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
+          TORCH_VER: ${{ matrix.TORCH_VER }}
+          TORCH_CUDA: ${{ matrix.TORCH_CUDA }}
+        shell: bash --noprofile --norc -xeuo pipefail {0}
+        run: run-tests nightly-pytorch
+
+      - name: Install cuda-python wheels + numba-cuda
+        if: ${{ inputs.test-mode == 'nightly-numba-cuda' }}
+        env:
+          CUDA_VER: ${{ matrix.CUDA_VER }}
+          LOCAL_CTK: ${{ matrix.LOCAL_CTK }}
+        shell: bash --noprofile --norc -xeuo pipefail {0}
+        run: run-tests nightly-numba-cuda
+
+      # ── Nightly: run tests ──
+      - name: Run PyTorch interop tests
+        if: ${{ inputs.test-mode == 'nightly-pytorch' }}
+        shell: bash --noprofile --norc -xeuo pipefail {0}
+        run: |
+          pushd cuda_core
+          pytest -rxXs -v --durations=0 tests/test_utils.py tests/example_tests/
+          popd
+
+      - name: Run numba-cuda tests
+        if: ${{ inputs.test-mode == 'nightly-numba-cuda' }}
+        shell: bash --noprofile --norc -xeuo pipefail {0}
+        run: python -m numba.runtests numba.cuda.tests
diff --git a/ci/test-matrix.yml b/ci/test-matrix.yml
index a402e3e4cf7..19931c3943a 100644
--- a/ci/test-matrix.yml
+++ b/ci/test-matrix.yml
@@ -62,7 +62,17 @@ linux:
     - { ARCH: 'amd64', PY_VER: '3.14t', CUDA_VER: '13.2.1', LOCAL_CTK: '1', GPU: 'h100',       GPU_COUNT: '2', DRIVER: 'latest' }
     - { ARCH: 'amd64', PY_VER: '3.11',  CUDA_VER: '12.9.1', LOCAL_CTK: '0', GPU: 't4',         GPU_COUNT: '1', DRIVER: 'latest', FLAVOR: 'wsl' }
     - { ARCH: 'amd64', PY_VER: '3.12',  CUDA_VER: '13.2.1', LOCAL_CTK: '0', GPU: 'rtx4090',    GPU_COUNT: '1', DRIVER: 'latest', FLAVOR: 'wsl' }
-  nightly: []
+  nightly:
+    # nightly-pytorch (amd64 only — PyTorch does not ship arm64 GPU wheels)
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '12.6.3', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', TORCH_VER: 'latest', TORCH_CUDA: 'cu126' }
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '13.0.2', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', TORCH_VER: 'latest', TORCH_CUDA: 'cu130' }
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '12.6.3', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', TORCH_VER: '2.9.1',  TORCH_CUDA: 'cu126' }
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '13.0.2', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', TORCH_VER: '2.9.1',  TORCH_CUDA: 'cu130' }
+    # nightly-numba-cuda
+    - { MODE: 'nightly-numba-cuda', ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '12.9.1', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest' }
+    - { MODE: 'nightly-numba-cuda', ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '13.2.1', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest' }
+    - { MODE: 'nightly-numba-cuda', ARCH: 'arm64', PY_VER: '3.12', CUDA_VER: '12.9.1', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest' }
+    - { MODE: 'nightly-numba-cuda', ARCH: 'arm64', PY_VER: '3.12', CUDA_VER: '13.2.1', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest' }
 
 windows:
   pull-request:
@@ -85,4 +95,12 @@ windows:
     - { ARCH: 'amd64', PY_VER: '3.14t', CUDA_VER: '12.9.1', LOCAL_CTK: '1', GPU: 'l4',         GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'TCC' }
     - { ARCH: 'amd64', PY_VER: '3.14t', CUDA_VER: '13.0.2', LOCAL_CTK: '0', GPU: 'a100',       GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'MCDM' }
     - { ARCH: 'amd64', PY_VER: '3.14t', CUDA_VER: '13.2.1', LOCAL_CTK: '0', GPU: 'a100',       GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'MCDM' }
-  nightly: []
+  nightly:
+    # nightly-pytorch
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '12.6.3', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'TCC', TORCH_VER: 'latest', TORCH_CUDA: 'cu126' }
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '13.0.2', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'TCC', TORCH_VER: 'latest', TORCH_CUDA: 'cu130' }
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '12.6.3', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'TCC', TORCH_VER: '2.9.1',  TORCH_CUDA: 'cu126' }
+    - { MODE: 'nightly-pytorch',    ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '13.0.2', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'TCC', TORCH_VER: '2.9.1',  TORCH_CUDA: 'cu130' }
+    # nightly-numba-cuda
+    - { MODE: 'nightly-numba-cuda', ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '12.9.1', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'TCC' }
+    - { MODE: 'nightly-numba-cuda', ARCH: 'amd64', PY_VER: '3.12', CUDA_VER: '13.2.1', LOCAL_CTK: '0', GPU: 'l4', GPU_COUNT: '1', DRIVER: 'latest', DRIVER_MODE: 'TCC' }
diff --git a/ci/tools/run-tests b/ci/tools/run-tests
index d42634a7073..a4001b7a001 100755
--- a/ci/tools/run-tests
+++ b/ci/tools/run-tests
@@ -13,19 +13,22 @@ if [[ ${#} -ne 1 ]]; then
   echo "Error: This script requires exactly 1 argument. You provided ${#}"
   exit 1
 fi
-if [[ "${1}" != "bindings" && "${1}" != "core" && "${1}" != "pathfinder" ]]; then
-  echo "Error: Invalid test module '${1}'. Must be 'bindings', 'core', or 'pathfinder'"
+if [[ "${1}" != "bindings" && "${1}" != "core" && "${1}" != "pathfinder" && "${1}" != "nightly-pytorch" && "${1}" != "nightly-numba-cuda" ]]; then
+  echo "Error: Invalid test module '${1}'. Must be 'bindings', 'core', 'pathfinder', 'nightly-pytorch', or 'nightly-numba-cuda'"
   exit 1
 fi
 
 test_module=${1}
 
-# Unconditionally install pathfinder wheel
-# (it is a direct dependency of bindings, and a transitive dependency of core)
-pushd ./cuda_pathfinder
-echo "Installing pathfinder wheel"
-pip install ./*.whl --group test
-popd
+# For standard modes, install pathfinder up front (it is a direct dependency
+# of bindings, and a transitive dependency of core).  Nightly modes install
+# all wheels together in a single pip call further below.
+if [[ "${test_module}" != "nightly-pytorch" && "${test_module}" != "nightly-numba-cuda" ]]; then
+  pushd ./cuda_pathfinder
+  echo "Installing pathfinder wheel"
+  pip install ./*.whl --group test
+  popd
+fi
 
 if [[ "${test_module}" == "pathfinder" ]]; then
   pushd ./cuda_pathfinder
@@ -91,4 +94,72 @@ elif [[ "${test_module}" == "core" ]]; then
     ${SANITIZER_CMD} pytest -rxXs -v --durations=0 --randomly-dont-reorganize tests/cython
   fi
   popd
+elif [[ "${test_module}" == "nightly-pytorch" || "${test_module}" == "nightly-numba-cuda" ]]; then
+  # Nightly optional-dependency testing.
+  # Install ALL wheels (pathfinder + bindings + core) and the optional dep
+  # in a single pip call so pip resolves version constraints in one shot
+  # and avoids costly uninstall/reinstall cycles.
+
+  TEST_CUDA_MAJOR="$(cut -d '.' -f 1 <<< ${CUDA_VER})"
+  CUDA_VER_MINOR="$(cut -d '.' -f 1-2 <<< "${CUDA_VER}")"
+
+  FREE_THREADING=""
+  if python -c 'import sys; assert not sys._is_gil_enabled()' 2> /dev/null; then
+    FREE_THREADING+="-ft"
+  fi
+
+  # Resolve pathfinder wheel to absolute path before pushd.
+  # CUDA_BINDINGS_ARTIFACTS_DIR and CUDA_CORE_ARTIFACTS_DIR are already
+  # absolute (set via realpath in env-vars).
+  PATHFINDER_WHL=($(realpath ./cuda_pathfinder/*.whl))
+
+  BINDINGS_WHL=("${CUDA_BINDINGS_ARTIFACTS_DIR}"/*.whl)
+  if [[ "${LOCAL_CTK}" != 1 ]]; then
+    BINDINGS_WHL=("${BINDINGS_WHL[0]}[all]")
+  fi
+
+  CORE_WHL=("${CUDA_CORE_ARTIFACTS_DIR}"/*.whl)
+  if [[ "${LOCAL_CTK}" != 1 ]]; then
+    CORE_WHL=("${CORE_WHL[0]}[cu${TEST_CUDA_MAJOR}]")
+  fi
+
+  # pushd so --group reads test dependency groups from cuda_core/pyproject.toml.
+  # The explicit cuda-toolkit[...]==X.Y.* pin overrides the group's looser ==X.*.
+  pushd ./cuda_core
+
+  PIP_ARGS=(
+    "${PATHFINDER_WHL[@]}"
+    "${BINDINGS_WHL[@]}"
+    "${CORE_WHL[@]}"
+    --group "test-cu${TEST_CUDA_MAJOR}${FREE_THREADING}"
+  )
+
+  if [[ "${test_module}" == "nightly-pytorch" ]]; then
+    # TORCH_VER and TORCH_CUDA must be set by the caller.
+    # Use cuda-toolkit[cudart] only — torch brings its own nvcc/nvrtc/etc.
+    # This avoids version conflicts between our nvidia-* pins and torch's.
+    echo "Installing pathfinder + bindings + core + test deps + PyTorch ${TORCH_VER} (${TORCH_CUDA})"
+    PIP_ARGS+=("cuda-toolkit[cudart]==${CUDA_VER_MINOR}.*")
+    if [[ "${TORCH_VER}" == "latest" ]]; then
+      PIP_ARGS+=(torch)
+    else
+      PIP_ARGS+=("torch==${TORCH_VER}")
+    fi
+    PIP_ARGS+=(--extra-index-url "https://download.pytorch.org/whl/${TORCH_CUDA}")
+  elif [[ "${test_module}" == "nightly-numba-cuda" ]]; then
+    echo "Installing pathfinder + bindings + core + test deps + numba-cuda"
+    # numba-cuda's test-cuXX group deps (can't use --group for a wheel install):
+    PIP_ARGS+=(
+      "cuda-toolkit[curand,cublas]==${CUDA_VER_MINOR}.*"
+      "numba-cuda[cu${TEST_CUDA_MAJOR}]"
+      "cupy-cuda${TEST_CUDA_MAJOR}x"
+      psutil cffi pytest-xdist pytest-benchmark filecheck ml_dtypes statistics
+    )
+  fi
+
+  pip install "${PIP_ARGS[@]}"
+  popd
+
+  echo "Nightly install complete — installed packages:"
+  pip list
 fi