Merge pull request #462 from cupy/copilot/fix-build-scripts-cupycuda13

Enable cuTENSOR preload support for CUDA 13.x binary packages

Author: kmaehashi

.pfnci/wheel-windows/build.ps1

@@ -17,8 +17,9 @@ function UninstallCuTENSOR($cuda_path) {
     echo "Uninstalling cuTENSOR installation from ${cuda_path}"
     if(Test-Path ${cuda_path}\bin\cutensor.dll) {
         Remove-Item -Force -Verbose ${cuda_path}\bin\cutensor.dll
-        Remove-Item -Force -Verbose ${cuda_path}\bin\cutensor.lib
-        Remove-Item -Force -Verbose ${cuda_path}\bin\cutensor_static.lib
+        Remove-Item -Force -Verbose ${cuda_path}\bin\cutensorMg.dll
+        Remove-Item -Force -Verbose ${cuda_path}\lib\x64\cutensor.lib
+        Remove-Item -Force -Verbose ${cuda_path}\lib\x64\cutensorMg.lib
         Remove-Item -Force -Verbose ${cuda_path}\include\cutensor.h
         Remove-Item -Recurse -Force -Verbose ${cuda_path}\include\cutensor
     } else {

builder/setup_cuda_opt_lib.py

@@ -98,7 +98,8 @@ def main() -> None:
     elif sys.platform == 'win32':
         _install_library(
             'cutensor', src_dir, dst_dir, {
-                'lib': 'bin',
+                'bin': 'bin',
+                'lib': 'lib/x64',
                 'include': 'include',
             })
     else:

dist_config.py

@@ -179,7 +179,7 @@ class _WheelLinuxConfig(TypedDict):
         'image': 'cupy/cupy-release-tools:cuda-runfile-13.0.0-centos7',
         'libs': [],
         'includes': [],
-        'preloads': ['nccl'],  # TODO(kmaehashi): Re-add cuTENSOR
+        'preloads': ['cutensor', 'nccl'],
         'verify_image': 'nvidia/cuda:{system}',
         'verify_systems': [
             # Test on all supported CUDA version variants.
@@ -273,7 +273,7 @@ class _WheelWindowsConfig(TypedDict):
         'name': 'cupy-cuda13x',
         'kind': 'cuda',
         'libs': [],
-        'preloads': [],  # TODO(kmaehashi): Re-add cuTENSOR
+        'preloads': ['cutensor'],
         'cudart_lib': 'cudart64_13',  # binary compatible between CUDA 13.x
         'check_version': lambda x: 13000 <= x < 13010,  # CUDA 13.0
     }

release-tests/pkg_wheel/test_preload.py

@@ -34,9 +34,6 @@ def test_cutensor(self):
         if cupy.cuda.runtime.runtimeGetVersion() < 10010:
             # cuTENSOR is only available for CUDA 10.1+.
             return
-        if 13000 <= cupy.cuda.runtime.runtimeGetVersion():
-            # TODO(kmaehashi): cuTENSOR is not yet available in CUDA 13+.
-            return
         preload_version = self._get_config()['cutensor']['version']
         major, minor, patchlevel = (int(x) for x in preload_version.split('.'))
         expected_version = major * 10000 + minor * 100 + patchlevel