diff --git a/tensorflow_datasets/core/utils/py_utils.py b/tensorflow_datasets/core/utils/py_utils.py
index c50ada43b21..81594e82b69 100644
--- a/tensorflow_datasets/core/utils/py_utils.py
+++ b/tensorflow_datasets/core/utils/py_utils.py
@@ -312,7 +312,7 @@ def _tmp_file_name(
     path: epath.PathLike,
     subfolder: str | None = None,
 ) -> epath.Path:
-  """Returns the temporary file name for the given path.
+  """Returns the temporary file path dependent on the given path and subfolder.
 
   Args:
     path: The path to the file.
@@ -322,9 +322,12 @@ def _tmp_file_name(
   path = epath.Path(path)
   file_name = f'{_tmp_file_prefix()}.{path.name}'
   if subfolder:
-    return path.parent / subfolder / file_name
+    tmp_path = path.parent / subfolder / file_name
   else:
-    return path.parent / file_name
+    tmp_path = path.parent / file_name
+  # Create the parent directory if it doesn't exist.
+  tmp_path.parent.mkdir(parents=True, exist_ok=True)
+  return tmp_path
 
 
 @contextlib.contextmanager
@@ -334,7 +337,6 @@ def incomplete_file(
 ) -> Iterator[epath.Path]:
   """Writes to path atomically, by writing to temp file and renaming it."""
   tmp_path = _tmp_file_name(path, subfolder=subfolder)
-  tmp_path.parent.mkdir(exist_ok=True)
   try:
     yield tmp_path
     tmp_path.replace(path)
@@ -346,20 +348,24 @@ def incomplete_file(
 @contextlib.contextmanager
 def incomplete_files(
     path: epath.Path,
+    subfolder: str | None = None,
 ) -> Iterator[epath.Path]:
   """Writes to path atomically, by writing to temp file and renaming it."""
-  tmp_file_prefix = _tmp_file_prefix()
-  tmp_path = path.parent / f'{tmp_file_prefix}.{path.name}'
+  tmp_path = _tmp_file_name(path, subfolder=subfolder)
+  tmp_file_prefix = tmp_path.name.removesuffix(f'.{path.name}')
   try:
     yield tmp_path
     # Rename all tmp files to their final name.
-    for tmp_file in path.parent.glob(f'{tmp_file_prefix}.*'):
+    for tmp_file in tmp_path.parent.glob(f'{tmp_file_prefix}.*'):
       file_name = tmp_file.name.removeprefix(tmp_file_prefix + '.')
       tmp_file.replace(path.parent / file_name)
   finally:
     # Eventually delete the tmp_path if exception was raised
-    for tmp_file in path.parent.glob(f'{tmp_file_prefix}.*'):
-      tmp_file.unlink(missing_ok=True)
+    if subfolder:
+      tmp_path.parent.unlink(missing_ok=True)
+    else:
+      for tmp_file in tmp_path.parent.glob(f'{tmp_file_prefix}.*'):
+        tmp_file.unlink(missing_ok=True)
 
 
 def is_incomplete_file(path: epath.Path) -> bool:
diff --git a/tensorflow_datasets/core/utils/py_utils_test.py b/tensorflow_datasets/core/utils/py_utils_test.py
index 548bb4a43d2..8d774f621ed 100644
--- a/tensorflow_datasets/core/utils/py_utils_test.py
+++ b/tensorflow_datasets/core/utils/py_utils_test.py
@@ -373,15 +373,23 @@ def test_make_valid_name(name: str, expected: str):
 
 
 @pytest.mark.parametrize(
-    ['path', 'subfolder', 'expected'],
+    ['file_name', 'subfolder', 'expected_tmp_file_name'],
     [
-        ('/a/file.ext', None, '/a/foobar.file.ext'),
-        ('/a/file.ext', 'sub', '/a/sub/foobar.file.ext'),
+        ('file.ext', None, 'foobar.file.ext'),
+        ('file.ext', 'sub', 'sub/foobar.file.ext'),
     ],
 )
-def test_tmp_file_name(path, subfolder, expected):
+def test_tmp_file_name(
+    tmp_path: pathlib.Path,
+    file_name: str,
+    subfolder: str | None,
+    expected_tmp_file_name: str,
+):
   with mock.patch.object(py_utils, '_tmp_file_prefix', return_value='foobar'):
-    assert os.fspath(py_utils._tmp_file_name(path, subfolder)) == expected
+    assert (
+        py_utils._tmp_file_name(tmp_path / file_name, subfolder)
+        == tmp_path / expected_tmp_file_name
+    )
 
 
 if __name__ == '__main__':
diff --git a/tensorflow_datasets/core/writer.py b/tensorflow_datasets/core/writer.py
index 45237d32e45..66651afc538 100644
--- a/tensorflow_datasets/core/writer.py
+++ b/tensorflow_datasets/core/writer.py
@@ -572,12 +572,14 @@ def _write_final_shard(
     shard_path = self._filename_template.sharded_filepath(
         shard_index=shard_id, num_shards=len(non_empty_shard_ids)
     )
-    with utils.incomplete_files(epath.Path(shard_path)) as tmp_path:
+    with utils.incomplete_files(shard_path, subfolder="tmp") as tmp_shard_path:
       logging.info(
-          "Writing %d examples to %s.", len(example_by_key), os.fspath(tmp_path)
+          "Writing %d examples to %s.",
+          len(example_by_key),
+          os.fspath(tmp_shard_path),
       )
       record_keys = self._example_writer.write(
-          tmp_path, sorted(example_by_key.items())
+          tmp_shard_path, sorted(example_by_key.items())
       )
     self.inc_counter(name="written_shards")
     # If there are record_keys, create index files.