diff --git a/.github/scripts/build-cuda.sh b/.github/scripts/build-cuda.sh
index 8985327f2..672ab1121 100644
--- a/.github/scripts/build-cuda.sh
+++ b/.github/scripts/build-cuda.sh
@@ -11,14 +11,14 @@ if [[ -v cuda_targets ]]; then
 elif [ "${build_arch}" = "aarch64" ]; then
     build_capability="75;80;90"
 
-    # CUDA 12.8: Add sm100
-    [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;90;100"
+    # CUDA 12.8+: Add sm100/sm120
+    [[ "${cuda_version}" == 12.8.* || "${cuda_version}" == 12.9.* ]] && build_capability="75;80;90;100;120"
 else
     # By default, target Maxwell through Hopper.
     build_capability="50;52;60;61;70;75;80;86;89;90"
 
-    # CUDA 12.8: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
-    [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;86;89;90;100;120"
+    # CUDA 12.8+: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
+    [[ "${cuda_version}" == 12.8.* || "${cuda_version}" == 12.9.* ]] && build_capability="75;80;86;89;90;100;120"
 fi
 
 [[ "${build_os}" = windows-* ]] && python3 -m pip install ninja
diff --git a/.github/workflows/python-package.yml b/.github/workflows/python-package.yml
index d3deb26ee..2ca19e32f 100644
--- a/.github/workflows/python-package.yml
+++ b/.github/workflows/python-package.yml
@@ -72,16 +72,17 @@ jobs:
           - os: windows-latest
             arch: x86_64
         cuda_version:
-          ["11.8.0", "12.0.1", "12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1", "12.6.3", "12.8.1"]
+          ["11.8.0", "12.0.1", "12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1", "12.6.3", "12.8.1", "12.9.1"]
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
         # Windows: We install Cuda on the agent (slow)
-      - uses: Jimver/cuda-toolkit@v0.2.22
+      - uses: Jimver/cuda-toolkit@c35baa1a18fd1fc9dcf47c5bd839bf30559c0bc3 # v0.2.24
         if: startsWith(matrix.os, 'windows')
         id: cuda-toolkit
         with:
-          cuda: ${{ matrix.cuda_version }}
+          # Temporary: Use CUDA 12.9.0 for Windows until 12.9.1 is supported with this action.
+          cuda: ${{ matrix.cuda_version == '12.9.1' && '12.9.0' || matrix.cuda_version }}
           method: "network"
           sub-packages: '["nvcc","cudart","cusparse","cublas","thrust","nvrtc_dev","cublas_dev","cusparse_dev"]'
           linux-local-args: '["--toolkit"]'
diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index c6423b1f8..483ddc6ec 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -49,8 +49,8 @@ jobs:
   build-cuda:
     strategy:
       matrix:
-        cuda_version: ["11.8.0", "12.6.3", "12.8.1"]
-        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025]
+        cuda_version: ["11.8.0", "12.6.3", "12.8.1", "12.9.1"]
+        os: [ubuntu-22.04, ubuntu-22.04-arm]
         include:
           - os: ubuntu-22.04
             arch: x86_64
@@ -58,13 +58,14 @@ jobs:
             arch: aarch64
           - os: windows-2025
             arch: x86_64
+            cuda_version: "11.8.0"
     runs-on: ${{ matrix.os }}
 
     steps:
       - uses: actions/checkout@v4
 
       - name: Install CUDA Toolkit
-        uses: Jimver/cuda-toolkit@v0.2.23
+        uses: Jimver/cuda-toolkit@c35baa1a18fd1fc9dcf47c5bd839bf30559c0bc3 # v0.2.24
         if: startsWith(matrix.os, 'windows')
         id: cuda-toolkit
         with:
@@ -231,7 +232,7 @@ jobs:
         os: [ubuntu-22.04, windows-2025]
         arch: [x86_64]
         gpu: [T4, L40S]
-        cuda_version: ["11.8.0", "12.6.3", "12.8.1"]
+        cuda_version: ["11.8.0", "12.6.3", "12.8.1", "12.9.1"]
         include:
           - cuda_version: "11.8.0"
             torch_version: "2.2.2"
@@ -242,6 +243,9 @@ jobs:
           - cuda_version: "12.8.1"
             torch_version: "2.7.1"
             pypi_index: "https://download.pytorch.org/whl/cu128"
+          - cuda_version: "12.9.1"
+            torch_version: "2.8.0"
+            pypi_index: "https://download.pytorch.org/whl/nightly/cu129"
 
 
           # Linux L40S runners
@@ -274,12 +278,14 @@ jobs:
             gpu: T4
             runner: CUDA-Windows-x64
             cuda_version: "11.8.0"
-            torch_version: "2.7.1"
+            torch_version: "2.7.1"    # Note: this is the last PyTorch release supporting CUDA 11.8.
             pypi_index: "https://download.pytorch.org/whl/cu118"
 
         exclude:
           # Our current T4 Windows runner has a driver too old (471.11)
           # and cannot support CUDA 12+. Skip for now.
+          - os: windows-2025
+            cuda_version: "12.9.1"
           - os: windows-2025
             cuda_version: "12.8.1"
           - os: windows-2025
@@ -311,7 +317,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip install torch==${{ matrix.torch_version }} --index-url ${{ matrix.pypi_index }}
+          pip install --pre torch~=${{ matrix.torch_version }}.dev0 --index-url ${{ matrix.pypi_index }}
           pip install -e ".[test]"
           pip install pytest-cov