pathfinder: add cudla and nvcudla support

Add pathfinder support for loading ``libcudla.so.1`` from the ``nvidia-cudla``
package and probing ``libnvcudla.so`` through the existing canary subprocess
path. Use that probe in the cudla load test so hosts without the platform
runtime are skipped, while real ``libcudla.so.1`` load failures still surface
when ``libnvcudla.so`` is available.

Made-with: Cursor

Author: rwgk

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/descriptor_catalog.py

@@ -324,6 +324,12 @@ class DescriptorSpec:
         site_packages_windows=("nvidia/cu13/bin", "nvidia/cu12/bin"),
         dependencies=("nvrtc",),
     ),
+    DescriptorSpec(
+        name="cudla",
+        packaged_with="other",
+        linux_sonames=("libcudla.so.1",),
+        site_packages_linux=("nvidia/cu13/lib",),
+    ),
     DescriptorSpec(
         name="cudss",
         packaged_with="other",
@@ -386,6 +392,11 @@ class DescriptorSpec:
         linux_sonames=("libcuda.so.1",),
         windows_dlls=("nvcuda.dll",),
     ),
+    DescriptorSpec(
+        name="nvcudla",
+        packaged_with="driver",
+        linux_sonames=("libnvcudla.so",),
+    ),
     DescriptorSpec(
         name="nvml",
         packaged_with="driver",

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/load_nvidia_dynamic_lib.py

@@ -99,14 +99,18 @@ def _raise_canary_probe_child_process_error(
 
 
 @functools.cache
-def _resolve_system_loaded_abs_path_in_subprocess(libname: str) -> str | None:
+def _resolve_system_loaded_abs_path_in_subprocess(
+    libname: str,
+    *,
+    timeout: float = _CANARY_PROBE_TIMEOUT_SECONDS,
+) -> str | None:
     """Resolve a canary library's absolute path in a fresh Python subprocess."""
     try:
         result = subprocess.run(  # noqa: S603 - trusted argv: current interpreter + internal probe module
             build_dynamic_lib_subprocess_command(MODE_CANARY, libname),
             capture_output=True,
             text=True,
-            timeout=_CANARY_PROBE_TIMEOUT_SECONDS,
+            timeout=timeout,
             check=False,
             cwd=DYNAMIC_LIB_SUBPROCESS_CWD,
         )
@@ -127,6 +131,11 @@ def _resolve_system_loaded_abs_path_in_subprocess(libname: str) -> str | None:
     return None
 
 
+def _loadable_via_canary_subprocess(libname: str, *, timeout: float = _CANARY_PROBE_TIMEOUT_SECONDS) -> bool:
+    """Return True if the canary subprocess can resolve ``libname`` via system search."""
+    return _resolve_system_loaded_abs_path_in_subprocess(libname, timeout=timeout) is not None
+
+
 def _try_ctk_root_canary(ctx: SearchContext) -> str | None:
     """Try CTK-root canary fallback for descriptor-configured libraries."""
     for canary_libname in ctx.desc.ctk_root_canary_anchor_libnames:

cuda_pathfinder/pyproject.toml

@@ -36,6 +36,7 @@ cu13 = [
     "cuda-toolkit[cufile]==13.*; sys_platform != 'win32'",
     "cutensor-cu13",
     "nvidia-cublasmp-cu13; sys_platform != 'win32'",
+    "nvidia-cudla; platform_system == 'Linux' and platform_machine == 'aarch64'",
     "nvidia-cudss-cu13",
     "nvidia-cufftmp-cu13; sys_platform != 'win32'",
     "nvidia-cusolvermp-cu13; sys_platform != 'win32'",

cuda_pathfinder/tests/test_load_nvidia_dynamic_lib.py

@@ -14,7 +14,10 @@
 from cuda.pathfinder import DynamicLibNotAvailableError, DynamicLibUnknownError, load_nvidia_dynamic_lib
 from cuda.pathfinder._dynamic_libs import load_nvidia_dynamic_lib as load_nvidia_dynamic_lib_module
 from cuda.pathfinder._dynamic_libs import supported_nvidia_libs
-from cuda.pathfinder._dynamic_libs.subprocess_protocol import STATUS_NOT_FOUND, parse_dynamic_lib_subprocess_payload
+from cuda.pathfinder._dynamic_libs.subprocess_protocol import (
+    STATUS_NOT_FOUND,
+    parse_dynamic_lib_subprocess_payload,
+)
 from cuda.pathfinder._utils.platform_aware import IS_WINDOWS, quote_for_shell
 
 STRICTNESS = os.environ.get("CUDA_PATHFINDER_TEST_LOAD_NVIDIA_DYNAMIC_LIB_STRICTNESS", "see_what_works")
@@ -95,7 +98,7 @@ def test_known_but_platform_unavailable_libname_raises_dynamic_lib_not_available
 
 
 def _is_expected_load_nvidia_dynamic_lib_failure(libname):
-    if libname == "nvpl_fftw" and platform.machine().lower() != "aarch64":
+    if libname in ("cudla", "nvcudla", "nvpl_fftw") and platform.machine().lower() != "aarch64":
         return True
     dist_name_pattern = IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES.get(libname)
     if dist_name_pattern is not None:
@@ -117,6 +120,10 @@ def raise_child_process_failed():
         raise RuntimeError(build_child_process_failed_for_libname_message(libname, result))
 
     if result.returncode != 0:
+        if libname == "cudla" and not load_nvidia_dynamic_lib_module._loadable_via_canary_subprocess(
+            "nvcudla", timeout=timeout
+        ):
+            pytest.skip("libnvcudla.so is not loadable via canary subprocess on this host.")
         raise_child_process_failed()
     assert not result.stderr
     payload = parse_dynamic_lib_subprocess_payload(