ci: add CUDA 13.2 build and nightly test support

.github/workflows/python-package.yml

@@ -74,12 +74,12 @@ jobs:
           - os: windows-2025
             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", "12.9.1", "13.0.2"]
+          ["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", "13.0.2", "13.2.0"]
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
         # Windows: We install Cuda on the agent (slow)
-      - uses: Jimver/cuda-toolkit@6008063726ffe3309d1b22e413d9e88fed91a2f2 # v0.2.29
+      - uses: Jimver/cuda-toolkit@3d45d157f327c09c04b50ee6ccdea2d9d017ec76 # v0.2.35
         if: startsWith(matrix.os, 'windows')
         id: cuda-toolkit
         with:

.github/workflows/test-runner.yml

@@ -32,6 +32,10 @@ on:
         type: string
         default: ""
         description: "CPU architecture for testing: icelake, cascadelake (default: platform default runner)"
+      torch_nightly:
+        type: boolean
+        default: false
+        description: "Install nightly torch (--pre, no version pin) instead of a stable release"
 
 env:
   BNB_SKIP_CMAKE: 1
@@ -144,7 +148,7 @@ jobs:
       # Windows + CUDA: Install CUDA Toolkit
       - name: Install CUDA Toolkit
         if: inputs.backend == 'cuda' && inputs.platform == 'windows'
-        uses: Jimver/cuda-toolkit@6008063726ffe3309d1b22e413d9e88fed91a2f2 # v0.2.29
+        uses: Jimver/cuda-toolkit@3d45d157f327c09c04b50ee6ccdea2d9d017ec76 # v0.2.35
         with:
           cuda: ${{ inputs.cuda_version }}
           method: "network"
@@ -208,9 +212,14 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip install torch==${{ inputs.torch_version }} --index-url ${{ inputs.pypi_index }}
+          if [[ "${{ inputs.torch_nightly }}" == "true" ]]; then
+            pip install --pre torch --index-url ${{ inputs.pypi_index }}
+          else
+            pip install torch==${{ inputs.torch_version }} --index-url ${{ inputs.pypi_index }}
+          fi
           pip install -e ".[test]" -v
           pip install pytest-cov
+        shell: bash
 
       # Windows: Downgrade NumPy for torch<2.4.1 compatibility
       # See: https://github.com/pytorch/pytorch/issues/131668

.github/workflows/tests-nightly.yml

@@ -57,7 +57,7 @@ jobs:
         # Linux x64 cross-product
         platform: [linux-x64]
         gpu_type: [T4, A10, L40S]
-        cuda_version: ["11.8.0", "12.6.3", "12.8.1", "13.0.2"]
+        cuda_version: ["11.8.0", "12.6.3", "12.8.1", "13.0.2", "13.2.0"]
 
         include:
           # Map CUDA version to torch version and PyPI index
@@ -73,6 +73,10 @@ jobs:
           - cuda_version: "13.0.2"
             torch_version: "2.11.0"
             pypi_index: "https://download.pytorch.org/whl/cu130"
+          - cuda_version: "13.2.0"
+            torch_version: "2.12.0"
+            torch_nightly: true
+            pypi_index: "https://download.pytorch.org/whl/nightly/cu132"
 
           # Windows CUDA Tests - T4 GPU (CUDA 11.8 only, multiple torch versions)
           - platform: windows
@@ -98,4 +102,5 @@ jobs:
       cuda_version: ${{ matrix.cuda_version }}
       gpu_type: ${{ matrix.gpu_type }}
       torch_version: ${{ matrix.torch_version }}
+      torch_nightly: ${{ matrix.torch_nightly || false }}
       pypi_index: ${{ matrix.pypi_index }}

CMakeLists.txt

@@ -194,6 +194,11 @@ if(BUILD_CUDA)
         string(APPEND CMAKE_CUDA_FLAGS " --compress-mode=size")
     endif()
 
+    # CUDA 13.2+ CCCL headers require MSVC's standard conforming preprocessor.
+    if(MSVC AND CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL "13.2")
+        string(APPEND CMAKE_CUDA_FLAGS " -Xcompiler=/Zc:preprocessor")
+    endif()
+
     if(PTXAS_VERBOSE)
         string(APPEND CMAKE_CUDA_FLAGS " -Xptxas=-v")
     endif()