Merge remote-tracking branch 'origin/main' into explicit-graph-construction

Author: Andy-Jost

cuda_core/cuda/core/_linker.pyx

@@ -29,6 +29,7 @@ from dataclasses import dataclass
 from typing import Union
 from warnings import warn
 
+from cuda.pathfinder import optional_cuda_import
 from cuda.core._device import Device
 from cuda.core._module import ObjectCode
 from cuda.core._utils.clear_error_support import assert_type
@@ -649,23 +650,20 @@ def _decide_nvjitlink_or_driver() -> bool:
         " For best results, consider upgrading to a recent version of"
     )
 
-    try:
-        __import__("cuda.bindings.nvjitlink")  # availability check
-    except ModuleNotFoundError:
+    nvjitlink_module = optional_cuda_import(
+        "cuda.bindings.nvjitlink",
+        probe_function=lambda module: module.version(),  # probe triggers nvJitLink runtime load
+    )
+    if nvjitlink_module is None:
         warn_txt = f"cuda.bindings.nvjitlink is not available, therefore {warn_txt_common} cuda-bindings."
     else:
         from cuda.bindings._internal import nvjitlink
 
-        try:
-            if _nvjitlink_has_version_symbol(nvjitlink):
-                _use_nvjitlink_backend = True
-                return False  # Use nvjitlink
-        except RuntimeError:
-            warn_detail = "not available"
-        else:
-            warn_detail = "too old (<12.3)"
+        if _nvjitlink_has_version_symbol(nvjitlink):
+            _use_nvjitlink_backend = True
+            return False  # Use nvjitlink
         warn_txt = (
-            f"{'nvJitLink*.dll' if sys.platform == 'win32' else 'libnvJitLink.so*'} is {warn_detail}."
+            f"{'nvJitLink*.dll' if sys.platform == 'win32' else 'libnvJitLink.so*'} is too old (<12.3)."
             f" Therefore cuda.bindings.nvjitlink is not usable and {warn_txt_common} nvJitLink."
         )
 

cuda_core/cuda/core/_memoryview.pyx

@@ -1096,6 +1096,8 @@ cpdef StridedMemoryView view_as_cai(obj, stream_ptr, view=None):
     buf.exporting_obj = obj
     buf.metadata = cai_data
     buf.dl_tensor = NULL
+    # Validate shape/strides/typestr eagerly so constructor paths fail fast.
+    buf.get_layout()
     buf.ptr, buf.readonly = cai_data["data"]
     buf.is_device_accessible = True
     if buf.ptr != 0:
@@ -1138,6 +1140,8 @@ cpdef StridedMemoryView view_as_array_interface(obj, view=None):
     buf.exporting_obj = obj
     buf.metadata = data
     buf.dl_tensor = NULL
+    # Validate shape/strides/typestr eagerly so constructor paths fail fast.
+    buf.get_layout()
     buf.ptr, buf.readonly = data["data"]
     buf.is_device_accessible = False
     buf.device_id = handle_return(driver.cuCtxGetDevice())

cuda_core/cuda/core/_program.pyx

@@ -14,6 +14,7 @@ import threading
 from warnings import warn
 
 from cuda.bindings import driver, nvrtc
+from cuda.pathfinder import optional_cuda_import
 
 from libcpp.vector cimport vector
 
@@ -461,8 +462,8 @@ class ProgramOptions:
 # =============================================================================
 
 # Module-level state for NVVM lazy loading
-cdef object_nvvm_module = None
-cdef bint _nvvm_import_attempted = False
+_nvvm_module = None
+_nvvm_import_attempted = False
 
 
 def _get_nvvm_module():
@@ -484,18 +485,21 @@ def _get_nvvm_module():
                 "Please update cuda-bindings to use NVVM features."
             )
 
-        from cuda.bindings import nvvm
-        from cuda.bindings._internal.nvvm import _inspect_function_pointer
-
-        if _inspect_function_pointer("__nvvmCreateProgram") == 0:
-            raise RuntimeError("NVVM library (libnvvm) is not available in this Python environment. ")
+        nvvm = optional_cuda_import(
+            "cuda.bindings.nvvm",
+            probe_function=lambda module: module.version(),  # probe triggers libnvvm load
+        )
+        if nvvm is None:
+            raise RuntimeError(
+                "NVVM support is unavailable: cuda.bindings.nvvm is missing or libnvvm cannot be loaded."
+            )
 
         _nvvm_module = nvvm
         return _nvvm_module
 
-    except RuntimeError as e:
+    except RuntimeError:
         _nvvm_module = None
-        raise e
+        raise
 
 def _find_libdevice_path():
     """Find libdevice*.bc for NVVM compilation using cuda.pathfinder."""

cuda_core/docs/source/release/0.7.x-notes.rst

@@ -40,3 +40,6 @@ Fixes and enhancements
   linking operations to the C level and releasing the GIL during backend calls. This benefits
   workloads that create many programs or linkers, and enables concurrent compilation in
   multithreaded applications.
+- Improved optional dependency handling for NVVM and nvJitLink imports so that only genuinely
+  missing optional modules are treated as unavailable; unrelated import failures now surface
+  normally, and ``cuda.core`` now depends directly on ``cuda-pathfinder``.