Use PurePosixPath for cross-platform path handling (#15113)

Replace string split("/") with PurePosixPath().name for extracting
filenames from cloud storage paths, URLs, and HuggingFace repo names.

While these paths use forward slashes by convention (not affected by
the host OS), using PurePosixPath is more explicit, self-documenting,
and follows Python best practices for path manipulation.

Files modified:
- nemo/core/connectors/save_restore_connector.py (cloud storage paths)
- nemo/core/classes/common.py (cloud URLs and HuggingFace repo names)

Note: tarutils.py and nemo_file.py were NOT modified as they handle
tar archive internal paths (which mandate POSIX format) and dictionary
keys (not filesystem paths) respectively.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <[email protected]>
Signed-off-by: yurekami <[email protected]>

Author: yurekami

nemo/core/classes/common.py

@@ -28,7 +28,7 @@
 from dataclasses import dataclass, field
 from enum import Enum
 from functools import total_ordering
-from pathlib import Path
+from pathlib import Path, PurePosixPath
 from typing import Any, Dict, List, Optional, Tuple, Union
 
 import hydra
@@ -934,7 +934,8 @@ def _get_ngc_pretrained_model_info(cls, model_name: str, refresh_cache: bool = F
                 f"Model {model_name} was not found. Check cls.list_available_models()\n"
                 f"for the list of all available models."
             )
-        filename = location_in_the_cloud.split("/")[-1]
+        # Use PurePosixPath for cloud URLs which always use forward slashes
+        filename = PurePosixPath(location_in_the_cloud).name
         url = location_in_the_cloud.replace(filename, "")
         cache_dir = Path.joinpath(resolve_cache_dir(), f'{filename[:-5]}')
         # If either description and location in the cloud changes, this will force re-download
@@ -972,7 +973,8 @@ def _get_hf_hub_pretrained_model_info(cls, model_name: str, refresh_cache: bool
             -   The path to the NeMo model (.nemo file) in some cached directory (managed by HF Hub).
         """
         # Resolve the model name without origin for filename
-        resolved_model_filename = model_name.split("/")[-1] + '.nemo'
+        # Use PurePosixPath since HuggingFace repo names use forward slashes (e.g., "nvidia/model-name")
+        resolved_model_filename = PurePosixPath(model_name).name + '.nemo'
 
         # Try to take from cache first - if not fallback to options below
         if not refresh_cache:

nemo/core/connectors/save_restore_connector.py

@@ -21,6 +21,7 @@
 import time
 import uuid
 from contextlib import contextmanager
+from pathlib import PurePosixPath
 from typing import Callable, Generator, Optional, Set, Union
 
 import torch
@@ -610,7 +611,8 @@ def _make_nemo_file_from_folder(filename, source_dir):
     @staticmethod
     def _make_nemo_file_from_folder_with_multistorageclient(filename, source_dir):
         msc = import_multistorageclient()
-        filename_with_extension = filename.split("/")[-1]  # get the filename and extension
+        # Use PurePosixPath for cloud storage paths which always use forward slashes
+        filename_with_extension = PurePosixPath(filename).name
         with tempfile.TemporaryDirectory() as tmpdir:
             tar_file = os.path.join(tmpdir, filename_with_extension)
             with tarfile.open(tar_file, "w:") as tar:
@@ -726,7 +728,8 @@ def _unpack_nemo_file_with_multistorageclient(
             raise FileNotFoundError(f"{path2file} does not exist")
 
         with tempfile.TemporaryDirectory() as tmpdir:
-            filename_with_extension = path2file.split("/")[-1]  # get the filename with extension
+            # Use PurePosixPath for cloud storage paths which always use forward slashes
+            filename_with_extension = PurePosixPath(path2file).name
             downloaded_file_path = os.path.join(tmpdir, filename_with_extension)
             start_time = time.time()
             msc.download_file(path2file, downloaded_file_path)