Add cuDNN discovery and NCCL header support

  - cuDNN: added dynamic-library loading and header discovery.
  - NCCL: dynamic-library loading already existed; added the missing header discovery.

Author: DEKHTIARJonathan

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/descriptor_catalog.py

@@ -411,6 +411,17 @@ class DescriptorSpec:
         dependencies=("nvshmem_host",),
         requires_rtld_deepbind=True,
     ),
+    DescriptorSpec(
+        name="cudnn",
+        packaged_with="other",
+        linux_sonames=("libcudnn.so.9",),
+        windows_dlls=("cudnn64_9.dll",),
+        supported_windows_arch=("x64",),
+        site_packages_linux=("nvidia/cudnn/lib",),
+        site_packages_windows=WindowsSearchDirs.x64_only("nvidia/cudnn/bin"),
+        anchor_rel_dirs_windows=WindowsSearchDirs.x64_only("bin/x64", "bin"),
+        requires_add_dll_directory=True,
+    ),
     DescriptorSpec(
         name="cusolverMp",
         packaged_with="other",

cuda_pathfinder/cuda/pathfinder/_headers/header_descriptor_catalog.py

@@ -150,6 +150,14 @@ class HeaderDescriptorSpec:
     # -----------------------------------------------------------------------
     # Third-party / separately packaged headers
     # -----------------------------------------------------------------------
+    HeaderDescriptorSpec(
+        name="cudnn",
+        packaged_with="other",
+        header_basename="cudnn.h",
+        site_packages_dirs=("nvidia/cudnn/include",),
+        conda_targets_layout=False,
+        use_ctk_root_canary=False,
+    ),
     HeaderDescriptorSpec(
         name="cusolverMp",
         packaged_with="other",
@@ -244,6 +252,15 @@ class HeaderDescriptorSpec:
         conda_targets_layout=False,
         use_ctk_root_canary=False,
     ),
+    HeaderDescriptorSpec(
+        name="nccl",
+        packaged_with="other",
+        header_basename="nccl.h",
+        site_packages_dirs=("nvidia/nccl/include",),
+        available_on_windows=False,
+        conda_targets_layout=False,
+        use_ctk_root_canary=False,
+    ),
     HeaderDescriptorSpec(
         name="nvshmem",
         packaged_with="other",

cuda_pathfinder/pyproject.toml

@@ -24,6 +24,7 @@ cu12 = [
     "cuquantum-cu12; sys_platform != 'win32'",
     "cutensor-cu12",
     "nvidia-cublasmp-cu12; sys_platform != 'win32'",
+    "nvidia-cudnn-cu12",
     "nvidia-cudss-cu12",
     "nvidia-cufftmp-cu12; sys_platform != 'win32'",
     "nvidia-cusolvermp-cu12; sys_platform != 'win32'",
@@ -39,6 +40,7 @@ cu13 = [
     "cutensor-cu13",
     "nvidia-cublasmp-cu13; sys_platform != 'win32'",
     "nvidia-cudla; platform_system == 'Linux' and platform_machine == 'aarch64'",
+    "nvidia-cudnn-cu13",
     "nvidia-cudss-cu13",
     "nvidia-cufftmp-cu13; sys_platform != 'win32'",
     "nvidia-cusolvermp-cu13; sys_platform != 'win32'",

cuda_pathfinder/tests/test_descriptor_catalog.py

@@ -115,6 +115,20 @@ def test_cusparselt_windows_metadata_matches_wheel_layouts():
     )
 
 
+@pytest.mark.agent_authored(model="gpt-5")
+def test_cudnn_metadata_matches_wheel_layouts():
+    spec = _CATALOG_BY_NAME["cudnn"]
+    assert spec.packaged_with == "other"
+    assert spec.linux_sonames == ("libcudnn.so.9",)
+    assert spec.windows_dlls == ("cudnn64_9.dll",)
+    assert spec.supported_windows_arch == ("x64",)
+    assert spec.site_packages_linux == ("nvidia/cudnn/lib",)
+    assert spec.site_packages_windows == WindowsSearchDirs.x64_only("nvidia/cudnn/bin")
+    assert spec.anchor_rel_dirs_windows == WindowsSearchDirs.x64_only("bin/x64", "bin")
+    assert spec.dependencies == ()
+    assert spec.requires_add_dll_directory
+
+
 @pytest.mark.parametrize("spec", DESCRIPTOR_CATALOG, ids=lambda s: s.name)
 def test_ctk_root_canary_anchors_reference_known_ctk_libs(spec: DescriptorSpec):
     for anchor in spec.ctk_root_canary_anchor_libnames:

cuda_pathfinder/tests/test_find_nvidia_headers.py

@@ -27,6 +27,7 @@
 from cuda.pathfinder._dynamic_libs.load_nvidia_dynamic_lib import (
     _resolve_system_loaded_abs_path_in_subprocess,
 )
+from cuda.pathfinder._headers.header_descriptor import HEADER_DESCRIPTORS
 from cuda.pathfinder._headers.supported_nvidia_headers import (
     SUPPORTED_HEADERS_CTK,
     SUPPORTED_HEADERS_CTK_ALL,
@@ -43,6 +44,7 @@
 
 NON_CTK_IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES = {
     "cudensitymat": r"^cudensitymat-.*$",
+    "cudnn": r"^nvidia-cudnn-.*$",
     "cupauliprop": r"^cupauliprop-.*$",
     "cusolverMp": r"^nvidia-cusolvermp-.*$",
     "cusparseLt": r"^nvidia-cusparselt-.*$",
@@ -53,6 +55,7 @@
     "custatevec": r"^custatevec-.*$",
     "cutlass": r"^nvidia-cutlass$",
     "mathdx": r"^nvidia-libmathdx-.*$",
+    "nccl": r"^nvidia-nccl-.*$",
     "nvshmem": r"^nvidia-nvshmem-.*$",
 }
 
@@ -78,6 +81,25 @@ def test_non_ctk_importlib_metadata_distributions_names():
     assert sorted(NON_CTK_IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES) == sorted(SUPPORTED_HEADERS_NON_CTK_ALL)
 
 
+@pytest.mark.agent_authored(model="gpt-5")
+def test_cudnn_and_nccl_header_metadata_matches_wheel_layouts():
+    cudnn = HEADER_DESCRIPTORS["cudnn"]
+    assert cudnn.header_basename == "cudnn.h"
+    assert cudnn.site_packages_dirs == ("nvidia/cudnn/include",)
+    assert cudnn.available_on_linux
+    assert cudnn.available_on_windows
+    assert not cudnn.conda_targets_layout
+    assert not cudnn.use_ctk_root_canary
+
+    nccl = HEADER_DESCRIPTORS["nccl"]
+    assert nccl.header_basename == "nccl.h"
+    assert nccl.site_packages_dirs == ("nvidia/nccl/include",)
+    assert nccl.available_on_linux
+    assert not nccl.available_on_windows
+    assert not nccl.conda_targets_layout
+    assert not nccl.use_ctk_root_canary
+
+
 @functools.cache
 def have_distribution_for(libname: str) -> bool:
     pattern = re.compile(NON_CTK_IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES[libname])

toolshed/conda_create_for_pathfinder_testing.ps1

@@ -19,6 +19,7 @@ conda activate "pathfinder_testing_cu$CudaMajorMinorPatch"
 # Keep this list aligned with the Windows-installable subset of
 # cuda_pathfinder/pyproject.toml.
 $cpkgs = @(
+    "cudnn",
     "cusparselt-dev",
     "cutensor",
     "cutlass",

toolshed/conda_create_for_pathfinder_testing.sh

@@ -26,6 +26,7 @@ set -u
 # cuda_pathfinder/pyproject.toml.
 cpkgs=(
     "cuquantum"
+    "cudnn"
     "cusparselt-dev"
     "cutensor"
     "cutlass"
@@ -34,6 +35,7 @@ cpkgs=(
     "libcufftmp-dev"
     "libcusolvermp-dev"
     "libmathdx-dev"
+    "nccl"
     "libnvshmem3"
     "libnvshmem-dev"
 )