NVVM bindings not working on Windows + CUDA conda packages

[leofang]

As discussed offline, the search logic here

cuda-python/cuda_bindings/cuda/bindings/_internal/nvjitlink_windows.pyx

Lines 57 to 87 in 19563d5

	# First check if the DLL has been loaded by 3rd parties
	try:
	handle = win32api.GetModuleHandle(dll_name)
	except:
	pass
	else:
	break
	# Next, check if DLLs are installed via pip
	for sp in get_site_packages():
	mod_path = os.path.join(sp, "nvidia", "nvJitLink", "bin")
	if not os.path.isdir(mod_path):
	continue
	os.add_dll_directory(mod_path)
	try:
	handle = win32api.LoadLibraryEx(
	# Note: LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR needs an abs path...
	os.path.join(mod_path, dll_name),
	0, LOAD_LIBRARY_SEARCH_DEFAULT_DIRS | LOAD_LIBRARY_SEARCH_DLL_LOAD_DIR)
	except:
	pass
	else:
	break
	# Finally, try default search
	try:
	handle = win32api.LoadLibrary(dll_name)
	except:
	pass
	else:
	break

does not look for conda installations , which has the following path:

%CONDA_PREFIX%\Library\nvvm\lib\x64\nvvm.lib

One way to fix it is to patch to the internal template and regenerate the above file.

[leofang]

(FWIW this is the reason why we want to have a general solution #451 🙂)

[leofang]

To make the situation worse, CUDA 11 and CUDA 12 conda packages have different layouts:
https://github.com/conda-forge/cudatoolkit-feedstock/blob/0f01f6edfe22e77ac068daa46505711a1481272e/recipe/meta.yaml#L681-L683
(see docs in https://github.com/conda-forge/cuda-feedstock/tree/main/recipe for more info on the "new" layout)