Merge branch 'master' into fix/fsdp-mixed-precision

Author: justusschock

.github/CONTRIBUTING.md

@@ -212,6 +212,8 @@ We welcome any useful contribution! For your convenience here's a recommended wo
    - [Test README](https://github.com/Lightning-AI/pytorch-lightning/blob/master/tests/README.md)
    - [CI/CD README](https://github.com/Lightning-AI/pytorch-lightning/tree/master/.github/workflows#readme)
 
+1. Once you have a PR opened (and thereby a PR number), please update the respective changelog for [fabric](https://github.com/Lightning-AI/pytorch-lightning/blob/master/src/lightning/fabric/CHANGELOG.md) or [pytorch](https://github.com/Lightning-AI/pytorch-lightning/blob/master/src/lightning/pytorch/CHANGELOG.md) subpackage depending on where you made your changes.
+
 1. When you feel ready for integrating your work, mark your PR "Ready for review".
 
    - Your code should be readable and follow the project's design principles.

.github/workflows/ci-tests-fabric.yml

@@ -40,23 +40,20 @@ jobs:
       matrix:
         os: [macOS-14, ubuntu-22.04, windows-2022]
         config:
-          # only run PyTorch latest
+          # Test unified "lightning" package with PyTorch 2.1-2.5
           - { pkg-name: "lightning", python-version: "3.10", pytorch-version: "2.1" }
           - { pkg-name: "lightning", python-version: "3.11", pytorch-version: "2.2.2" }
           - { pkg-name: "lightning", python-version: "3.11", pytorch-version: "2.3" }
           - { pkg-name: "lightning", python-version: "3.12.7", pytorch-version: "2.4.1" }
           - { pkg-name: "lightning", python-version: "3.12.7", pytorch-version: "2.5.1" }
 
-          # only run PyTorch latest with Python latest, use Fabric scope to limit dependency issues
+          # Test "fabric" package with PyTorch 2.6-2.9
           - { pkg-name: "fabric", python-version: "3.12.7", pytorch-version: "2.6" }
-
-          # "fabric" installs the standalone package
-          - { pkg-name: "fabric", python-version: "3.10", pytorch-version: "2.7" }
-
-          # adding recently cut Torch 2.7 - FUTURE
+          - { pkg-name: "fabric", python-version: "3.12", pytorch-version: "2.7" }
           - { pkg-name: "fabric", python-version: "3.12", pytorch-version: "2.8" }
+          - { pkg-name: "fabric", python-version: "3.12", pytorch-version: "2.9" }
 
-          # "oldest" versions tests, only on minimum Python
+          # Test minimum supported versions (oldest)
           - { pkg-name: "fabric", pytorch-version: "2.1", requires: "oldest" }
     timeout-minutes: 25 # because of building grpcio on Mac
     env:

.github/workflows/ci-tests-pytorch.yml

@@ -44,23 +44,20 @@ jobs:
       matrix:
         os: [macOS-14, ubuntu-22.04, windows-2022]
         config:
-          # only run PyTorch latest
+          # Test unified "lightning" package with PyTorch 2.1-2.5
           - { pkg-name: "lightning", python-version: "3.10", pytorch-version: "2.1" }
           - { pkg-name: "lightning", python-version: "3.11", pytorch-version: "2.2.2" }
           - { pkg-name: "lightning", python-version: "3.11", pytorch-version: "2.3" }
           - { pkg-name: "lightning", python-version: "3.12.7", pytorch-version: "2.4.1" }
           - { pkg-name: "lightning", python-version: "3.12.7", pytorch-version: "2.5.1" }
 
-          # only run PyTorch latest with Python latest, use PyTorch scope to limit dependency issues
+          # Test "pytorch" package with PyTorch 2.6-2.9
           - { pkg-name: "pytorch", python-version: "3.12.7", pytorch-version: "2.6" }
-
-          # "pytorch" installs the standalone package
-          - { pkg-name: "pytorch", python-version: "3.10", pytorch-version: "2.7" }
-
-          # adding recently cut Torch 2.7 - FUTURE
+          - { pkg-name: "pytorch", python-version: "3.12", pytorch-version: "2.7" }
           - { pkg-name: "pytorch", python-version: "3.12", pytorch-version: "2.8" }
+          - { pkg-name: "pytorch", python-version: "3.12", pytorch-version: "2.9" }
 
-          # "oldest" versions tests, only on minimum Python
+          # Test minimum supported versions (oldest)
           - { pkg-name: "pytorch", pytorch-version: "2.1", requires: "oldest" }
     timeout-minutes: 50
     env:

src/lightning/fabric/CHANGELOG.md

@@ -25,7 +25,11 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
--
+- Fixed `EADDRINUSE` errors in distributed tests with port manager and retry logic ([#21309](https://github.com/Lightning-AI/pytorch-lightning/pull/21309))
+
+
+- Learning rate scheduler is stepped at the end of epoch when `on_train_batch_start` returns -1 ([#21296](https://github.com/Lightning-AI/pytorch-lightning/issues/21296)).
+
 
 
 ---

src/lightning/pytorch/CHANGELOG.md

@@ -25,6 +25,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added support for variable batch size in `ThroughputMonitor` ([#20236](https://github.com/Lightning-AI/pytorch-lightning/pull/20236))
 
 
+- Added `EMAWeightAveraging` callback that wraps Lightning's `WeightAveraging` class ([#21260](https://github.com/Lightning-AI/pytorch-lightning/pull/21260))
+
+
 ### Changed
 
 - Expose `weights_only` argument for `Trainer.{fit,validate,test,predict}` and let `torch` handle default value ([#21072](https://github.com/Lightning-AI/pytorch-lightning/pull/21072))

src/lightning/pytorch/callbacks/__init__.py

@@ -32,7 +32,7 @@
 from lightning.pytorch.callbacks.stochastic_weight_avg import StochasticWeightAveraging
 from lightning.pytorch.callbacks.throughput_monitor import ThroughputMonitor
 from lightning.pytorch.callbacks.timer import Timer
-from lightning.pytorch.callbacks.weight_averaging import WeightAveraging
+from lightning.pytorch.callbacks.weight_averaging import EMAWeightAveraging, WeightAveraging
 
 __all__ = [
     "BackboneFinetuning",
@@ -59,5 +59,6 @@
     "ThroughputMonitor",
     "Timer",
     "TQDMProgressBar",
+    "EMAWeightAveraging",
     "WeightAveraging",
 ]

src/lightning/pytorch/callbacks/weight_averaging.py

@@ -21,7 +21,7 @@
 from typing import Any, Optional, Union
 
 import torch
-from torch.optim.swa_utils import AveragedModel
+from torch.optim.swa_utils import AveragedModel, get_ema_avg_fn
 from typing_extensions import override
 
 import lightning.pytorch as pl
@@ -361,3 +361,55 @@ def _copy_average_to_current(self, pl_module: "pl.LightningModule") -> None:
         current_params = itertools.chain(pl_module.parameters(), pl_module.buffers())
         for average_param, current_param in zip(average_params, current_params):
             current_param.data.copy_(average_param.data)
+
+
+class EMAWeightAveraging(WeightAveraging):
+    """Exponential Moving Average (EMA) Weight Averaging callback."""
+
+    def __init__(
+        self,
+        device: Optional[Union[torch.device, str, int]] = None,
+        use_buffers: bool = True,
+        decay: float = 0.999,
+        update_every_n_steps: int = 1,
+        update_starting_at_step: Optional[int] = None,
+        update_starting_at_epoch: Optional[int] = None,
+        **kwargs: Any,
+    ):
+        super().__init__(
+            device=device,
+            use_buffers=use_buffers,
+            **kwargs,
+            avg_fn=get_ema_avg_fn(decay=decay),
+        )
+
+        self.update_every_n_steps = update_every_n_steps
+        self.update_starting_at_step = update_starting_at_step
+        self.update_starting_at_epoch = update_starting_at_epoch
+
+    def should_update(self, step_idx: Optional[int] = None, epoch_idx: Optional[int] = None) -> bool:
+        """Decide when to update the model weights.
+
+        Args:
+            step_idx: The current step index.
+            epoch_idx: The current epoch index.
+        Returns:
+            bool: True if the model weights should be updated, False otherwise.
+
+        """
+        if step_idx is not None:
+            # Check step-based conditions only if we have a valid step_idx
+            meets_step_requirement = self.update_starting_at_step is None or step_idx >= self.update_starting_at_step
+            meets_step_frequency = self.update_every_n_steps > 0 and step_idx % self.update_every_n_steps == 0
+            if meets_step_requirement and meets_step_frequency:
+                return True
+
+        if epoch_idx is not None:
+            # Check epoch-based condition only if we specify one
+            meets_epoch_requirement = (
+                self.update_starting_at_epoch is not None and epoch_idx >= self.update_starting_at_epoch
+            )
+            if meets_epoch_requirement:
+                return True
+
+        return False

src/lightning/pytorch/core/hooks.py

@@ -69,6 +69,7 @@ def on_train_batch_start(self, batch: Any, batch_idx: int) -> Optional[int]:
         """Called in the training loop before anything happens for that batch.
 
         If you return -1 here, you will skip training for the rest of the current epoch.
+        Learning rate scheduler will still be stepped at the end of epoch.
 
         Args:
             batch: The batched data as it is returned by the training DataLoader.

src/lightning/pytorch/loops/training_epoch_loop.py

@@ -325,30 +325,33 @@ def advance(self, data_fetcher: _DataFetcher) -> None:
         trainer._logger_connector.on_batch_start(batch)
 
         batch_output: _BATCH_OUTPUTS_TYPE = None  # for mypy
+        should_skip_rest_of_epoch = False
+
         if batch is None and not using_dataloader_iter:
             self._warning_cache.warn("train_dataloader yielded None. If this was on purpose, ignore this warning...")
         else:
             # hook
             call._call_callback_hooks(trainer, "on_train_batch_start", batch, batch_idx)
             response = call._call_lightning_module_hook(trainer, "on_train_batch_start", batch, batch_idx)
             call._call_strategy_hook(trainer, "on_train_batch_start", batch, batch_idx)
-            if response == -1:
-                self.batch_progress.increment_processed()
-                raise StopIteration
-
-            self.batch_progress.increment_started()
-
-            kwargs = (
-                self._build_kwargs(OrderedDict(), batch, batch_idx)
-                if not using_dataloader_iter
-                else OrderedDict(any=dataloader_iter)
-            )
-            with trainer.profiler.profile("run_training_batch"):
-                if trainer.lightning_module.automatic_optimization:
-                    # in automatic optimization, there can only be one optimizer
-                    batch_output = self.automatic_optimization.run(trainer.optimizers[0], batch_idx, kwargs)
-                else:
-                    batch_output = self.manual_optimization.run(kwargs)
+            should_skip_rest_of_epoch = response == -1
+            # Signal this is the last batch for the current epoch
+            if should_skip_rest_of_epoch:
+                self.batch_progress.increment_by(0, is_last_batch=True)
+            else:
+                self.batch_progress.increment_started()
+
+                kwargs = (
+                    self._build_kwargs(OrderedDict(), batch, batch_idx)
+                    if not using_dataloader_iter
+                    else OrderedDict(any=dataloader_iter)
+                )
+                with trainer.profiler.profile("run_training_batch"):
+                    if trainer.lightning_module.automatic_optimization:
+                        # in automatic optimization, there can only be one optimizer
+                        batch_output = self.automatic_optimization.run(trainer.optimizers[0], batch_idx, kwargs)
+                    else:
+                        batch_output = self.manual_optimization.run(kwargs)
 
         self.batch_progress.increment_processed()
 
@@ -358,6 +361,10 @@ def advance(self, data_fetcher: _DataFetcher) -> None:
         if self._num_ready_batches_reached():
             self.update_lr_schedulers("epoch", update_plateau_schedulers=False)
 
+        if should_skip_rest_of_epoch:
+            # Only raise StopIteration now so that the training epoch loop can finish
+            raise StopIteration
+
         if using_dataloader_iter:
             # update the hook kwargs now that the step method might have consumed the iterator
             batch = data_fetcher._batch

src/lightning/pytorch/trainer/trainer.py

@@ -563,14 +563,16 @@ def fit(
                 recommend using ``weights_only=True``. For more information, please refer to the
                 `PyTorch Developer Notes on Serialization Semantics <https://docs.pytorch.org/docs/main/notes/serialization.html#id3>`_.
 
+        For more information about multiple dataloaders, see this :ref:`section <multiple-dataloaders>`.
+
+        :rtype: :py:obj:`None`
+
         Raises:
             TypeError:
                 If ``model`` is not :class:`~lightning.pytorch.core.LightningModule` for torch version less than
                 2.0.0 and if ``model`` is not :class:`~lightning.pytorch.core.LightningModule` or
                 :class:`torch._dynamo.OptimizedModule` for torch versions greater than or equal to 2.0.0 .
 
-        For more information about multiple dataloaders, see this :ref:`section <multiple-dataloaders>`.
-
         """
         model = _maybe_unwrap_optimized(model)
         self.strategy._lightning_module = model