Prepare Patch Release 2.6.5 (#21745)

* Use fs.pipe() for S3/GCS checkpoint uploads in _atomic_save (#21595)

* Use fs.pipe() for S3/GCS checkpoint uploads in _atomic_save

Replace fs.open() + f.write() with fs.pipe() for S3 and GCS
filesystems, enabling parallel multipart uploads. This gives
4-5x throughput improvement for checkpoints >= 500 MB.

Azure is excluded because adlfs stages blocks sequentially,
making fs.pipe() slower than f.write().

Fixes #21499

Signed-off-by: c-pozzi <corina.rios@gmail.com>

* Use isinstance check for Azure exclusion in _atomic_save

Use module_available + isinstance instead of string comparison
to detect AzureBlobFileSystem. Fix test mocking to patch
module_available so the isinstance check works without adlfs
installed.

Signed-off-by: c-pozzi <corina.rios@gmail.com>

* Fix ruff SIM117: combine nested with statements in tests

Signed-off-by: c-pozzi <corina.rios@gmail.com>

---------

Signed-off-by: c-pozzi <corina.rios@gmail.com>

* Update version and changelog

---------

Signed-off-by: c-pozzi <corina.rios@gmail.com>
Co-authored-by: c-pozzi <37026778+c-pozzi@users.noreply.github.com>

Author: justusschock

src/lightning/fabric/CHANGELOG.md

@@ -24,6 +24,12 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 --
 
+## [2.6.5] - 2026-05-27
+
+### Changed
+
+- Use fs.pipe() for S3/GCS checkpoint uploads in _atomic_save ([#21595](https://github.com/Lightning-AI/pytorch-lightning/pull/21595)
+
 ## [2.6.2] - 2026-03-19
 
 ### Fixed

src/lightning/fabric/utilities/cloud_io.py

@@ -100,8 +100,21 @@ def _atomic_save(checkpoint: dict[str, Any], filepath: _PATH) -> None:
     try:
         # We use a transaction here to avoid file corruption if the save gets interrupted
         fs, urlpath = fsspec.core.url_to_fs(str(filepath))
-        with fs.transaction, fs.open(urlpath, "wb") as f:
-            f.write(bytesbuffer.getvalue())
+        with fs.transaction:
+            is_azure = False
+            if module_available("adlfs"):
+                from adlfs import AzureBlobFileSystem
+
+                is_azure = isinstance(fs, AzureBlobFileSystem)
+
+            if _is_object_storage(fs) and not is_azure:
+                # Use fs.pipe() for S3/GCS where it triggers parallel multipart uploads,
+                # giving 4-5x throughput improvement for checkpoints >= 500 MB.
+                # Azure is excluded because adlfs stages blocks sequentially, making pipe() slower.
+                fs.pipe(urlpath, bytesbuffer.getvalue())
+            else:
+                with fs.open(urlpath, "wb") as f:
+                    f.write(bytesbuffer.getvalue())
     except PermissionError as e:
         if isinstance(e.__context__, OSError) and getattr(e.__context__, "errno", None) == errno.EXDEV:
             raise RuntimeError(

src/version.info

@@ -1 +1 @@
-2.6.4
+2.6.5

tests/tests_fabric/utilities/test_cloud_io.py

@@ -12,12 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import os
+from unittest import mock
 
 import fsspec
+import torch
 from fsspec.implementations.local import LocalFileSystem
 from fsspec.spec import AbstractFileSystem
 
-from lightning.fabric.utilities.cloud_io import _is_dir, get_filesystem
+from lightning.fabric.utilities.cloud_io import _atomic_save, _is_dir, get_filesystem
 
 
 def test_get_filesystem_custom_filesystem():
@@ -90,3 +92,61 @@ def isfile(self, path):
     assert _is_dir(get_filesystem(s3_directory), s3_directory, strict=True)
     assert not _is_dir(get_filesystem(s3_directory), s3_file)
     assert not _is_dir(get_filesystem(s3_directory), s3_file, strict=True)
+
+
+def test_atomic_save_uses_pipe_for_s3(tmp_path):
+    """Test that _atomic_save uses fs.pipe() for S3 filesystems."""
+    checkpoint = {"key": torch.tensor([1, 2, 3])}
+    filepath = "s3://bucket/checkpoint.ckpt"
+
+    mock_fs = mock.MagicMock()
+    mock_fs.__class__.__name__ = "S3FileSystem"
+
+    with (
+        mock.patch("lightning.fabric.utilities.cloud_io._is_object_storage", return_value=True),
+        mock.patch("fsspec.core.url_to_fs", return_value=(mock_fs, "bucket/checkpoint.ckpt")),
+    ):
+        _atomic_save(checkpoint, filepath)
+
+    mock_fs.pipe.assert_called_once()
+    mock_fs.open.assert_not_called()
+
+
+def test_atomic_save_uses_write_for_azure(tmp_path):
+    """Test that _atomic_save uses f.write() for Azure filesystems."""
+    import sys
+    import types
+
+    checkpoint = {"key": torch.tensor([1, 2, 3])}
+    filepath = "azure://container/checkpoint.ckpt"
+
+    # Create a fake adlfs module so isinstance check works
+    AzureBlobFileSystem = type("AzureBlobFileSystem", (), {})
+    fake_adlfs = types.ModuleType("adlfs")
+    fake_adlfs.AzureBlobFileSystem = AzureBlobFileSystem
+
+    mock_fs = mock.MagicMock()
+    mock_fs.__class__ = AzureBlobFileSystem
+
+    with (
+        mock.patch.dict(sys.modules, {"adlfs": fake_adlfs}),
+        mock.patch("lightning.fabric.utilities.cloud_io.module_available", return_value=True),
+        mock.patch("lightning.fabric.utilities.cloud_io._is_object_storage", return_value=True),
+        mock.patch("fsspec.core.url_to_fs", return_value=(mock_fs, "container/checkpoint.ckpt")),
+    ):
+        _atomic_save(checkpoint, filepath)
+
+    mock_fs.pipe.assert_not_called()
+    mock_fs.open.assert_called_once()
+
+
+def test_atomic_save_uses_write_for_local(tmp_path):
+    """Test that _atomic_save uses f.write() for local filesystems."""
+    checkpoint = {"key": torch.tensor([1, 2, 3])}
+    filepath = tmp_path / "checkpoint.ckpt"
+
+    _atomic_save(checkpoint, filepath)
+
+    assert filepath.exists()
+    loaded = torch.load(filepath, weights_only=True)
+    torch.testing.assert_close(loaded["key"], checkpoint["key"])