chore: handle the variants

Author: cpcloud

cuda_pathfinder/cuda/pathfinder/_binaries/find_nvidia_binary_utility.py

@@ -84,19 +84,20 @@ def find_nvidia_binary_utility(utility_name: str) -> str | None:
 
     # 1. Search in site-packages (NVIDIA wheels)
     candidate_dirs = supported_nvidia_binaries.SITE_PACKAGES_BINDIRS.get(utility_name, ())
-    dirs = [bindir for sub_dir in candidate_dirs if (bindir := find_sub_dirs_all_sitepackages(sub_dir.split("/")))]
+    dirs = [bin_dir for sub_dir in candidate_dirs if (bin_dir := find_sub_dirs_all_sitepackages(sub_dir.split(os.sep)))]
 
     # 2. Search in Conda environment
-    conda_prefix = os.environ.get("CONDA_PREFIX")
-    if conda_prefix is not None:
+    if (conda_prefix := os.environ.get("CONDA_PREFIX")) is not None:
         if IS_WINDOWS:
-            dirs.append(os.path.join(conda_prefix, "Library"))
-        dirs.append(conda_prefix)
+            dirs.append(os.path.join(conda_prefix, "Library", "bin"))
+        dirs.append(os.path.join(conda_prefix, "bin"))
 
     # 3. Search in CUDA Toolkit (CUDA_HOME/CUDA_PATH)
-    cuda_home = get_cuda_home_or_path()
-    if cuda_home is not None:
-        dirs.append(cuda_home)
+    if (cuda_home := get_cuda_home_or_path()) is not None:
+        if IS_WINDOWS:
+            dirs.extend(os.path.join(cuda_home, "bin", "x64"))
+            dirs.append(os.path.join(cuda_home, "bin", "x86_64"))
+        dirs.append(os.path.join(cuda_home, "bin"))
 
     normalized_name = _normalize_utility_name(utility_name)
     return shutil.which(normalized_name, path=os.pathsep.join(dirs))