Add pathfinder.locate_nvidia_header_directory support for mathdx, cutlass, cute (#1816)

* Add nvidia-libmathdx-cu13 in cuda_pathfinder/pyproject.toml

* Remove cufftMp, mathdx from _is_expected_load_nvidia_dynamic_lib_failure

cufftMp should have been removed in PR #1315 (cuda-bindings v13.1.0 release), but this was overlooked.

* Add HeaderDescriptorSpec for mathdx

* Add HeaderDescriptorSpec entries for cute, cutlass

* Restore mathdx in IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES

The cu13 Windows CI jobs do not run `pip install --group test-cu13`,
so nvidia-libmathdx is not installed when the all_must_work pathfinder
test run executes. Without the distribution-presence check the test
fails unconditionally. Restore the entry so the test treats a missing
mathdx wheel as an expected failure under all_must_work strictness.

Made-with: Cursor

* Skip all_must_work pathfinder tests for free-threaded Python on Windows (#1820)

The nvidia-cutlass wheel contains paths that exceed the 260-char
Windows MAX_PATH limit when installed under the longer
x64-freethreaded site-packages prefix. Skip the pip --group install
and the all_must_work test run for free-threaded builds until the CI
runners enable LongPathsEnabled.

Made-with: Cursor

* Fix mathdx site_packages_windows path for cu13

The nvidia-libmathdx-cu13 wheel installs mathdx64_0.dll under
nvidia/cu13/bin, not nvidia/cu13/bin/x86_64. The x86_64 subdirectory
is used by cuda-toolkit CTK packages, but nvidia-libmathdx is packaged
separately and follows the flat layout (matching nvidia-cudss and the
cu12 mathdx wheel).

Made-with: Cursor

Author: rwgk

.github/workflows/test-wheel-windows.yml

@@ -263,6 +263,7 @@ jobs:
           }
 
       - name: Install cuda.pathfinder extra wheels for testing
+        if: ${{ !endsWith(matrix.PY_VER, 't') }} # see issue #1820
         shell: bash --noprofile --norc -xeuo pipefail {0}
         run: |
           pushd cuda_pathfinder
@@ -271,6 +272,7 @@ jobs:
           popd
 
       - name: Run cuda.pathfinder tests with all_must_work
+        if: ${{ !endsWith(matrix.PY_VER, 't') }} # see issue #1820
         env:
           CUDA_PATHFINDER_TEST_LOAD_NVIDIA_DYNAMIC_LIB_STRICTNESS: all_must_work
           CUDA_PATHFINDER_TEST_FIND_NVIDIA_HEADERS_STRICTNESS: all_must_work

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/descriptor_catalog.py

@@ -314,7 +314,7 @@ class DescriptorSpec:
         linux_sonames=("libmathdx.so.0",),
         windows_dlls=("mathdx64_0.dll",),
         site_packages_linux=("nvidia/cu13/lib", "nvidia/cu12/lib"),
-        site_packages_windows=("nvidia/cu13/bin/x86_64", "nvidia/cu12/bin"),
+        site_packages_windows=("nvidia/cu13/bin", "nvidia/cu12/bin"),
         dependencies=("nvrtc",),
     ),
     DescriptorSpec(

cuda_pathfinder/cuda/pathfinder/_headers/header_descriptor_catalog.py

@@ -145,6 +145,14 @@ class HeaderDescriptorSpec:
         conda_targets_layout=False,
         use_ctk_root_canary=False,
     ),
+    HeaderDescriptorSpec(
+        name="cute",
+        packaged_with="other",
+        header_basename="cute/tensor.hpp",
+        site_packages_dirs=("cutlass_library/source/include",),
+        conda_targets_layout=False,
+        use_ctk_root_canary=False,
+    ),
     HeaderDescriptorSpec(
         name="cutensor",
         packaged_with="other",
@@ -153,6 +161,22 @@ class HeaderDescriptorSpec:
         conda_targets_layout=False,
         use_ctk_root_canary=False,
     ),
+    HeaderDescriptorSpec(
+        name="cutlass",
+        packaged_with="other",
+        header_basename="cutlass/cutlass.h",
+        site_packages_dirs=("cutlass_library/source/include",),
+        conda_targets_layout=False,
+        use_ctk_root_canary=False,
+    ),
+    HeaderDescriptorSpec(
+        name="mathdx",
+        packaged_with="other",
+        header_basename="libmathdx.h",
+        site_packages_dirs=("nvidia/cu13/include", "nvidia/cu12/include"),
+        conda_targets_layout=False,
+        use_ctk_root_canary=False,
+    ),
     HeaderDescriptorSpec(
         name="nvshmem",
         packaged_with="other",

cuda_pathfinder/pyproject.toml

@@ -38,10 +38,12 @@ cu13 = [
     "nvidia-cudss-cu13",
     "nvidia-cufftmp-cu13; sys_platform != 'win32'",
     "nvidia-cusparselt-cu13",
+    "nvidia-libmathdx-cu13",
     "nvidia-nccl-cu13; sys_platform != 'win32'",
     "nvidia-nvshmem-cu13; sys_platform != 'win32'",
 ]
 host = [
+    "nvidia-cutlass",
     "nvpl-fft; platform_system == 'Linux' and platform_machine == 'aarch64'",
 ]
 

cuda_pathfinder/tests/test_find_nvidia_headers.py

@@ -41,7 +41,10 @@
 
 NON_CTK_IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES = {
     "cusparseLt": r"^nvidia-cusparselt-.*$",
+    "cute": r"^nvidia-cutlass$",
     "cutensor": r"^cutensor-.*$",
+    "cutlass": r"^nvidia-cutlass$",
+    "mathdx": r"^nvidia-libmathdx-.*$",
     "nvshmem": r"^nvidia-nvshmem-.*$",
 }
 

cuda_pathfinder/tests/test_load_nvidia_dynamic_lib.py

@@ -90,7 +90,6 @@ def test_known_but_platform_unavailable_libname_raises_dynamic_lib_not_available
 
 
 IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES = {
-    "cufftMp": r"^nvidia-cufftmp-.*$",
     "mathdx": r"^nvidia-libmathdx-.*$",
 }