lightning 2.6.0.dev20251019__tar.gz → 2.6.0.dev20251102__tar.gz

{lightning-2.6.0.dev20251019/src/lightning.egg-info → lightning-2.6.0.dev20251102}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lightning
-Version: 2.6.0.dev20251019
+Version: 2.6.0.dev20251102
 Summary: The Deep Learning framework to train, deploy, and ship AI products Lightning fast.
 Home-page: https://github.com/Lightning-AI/lightning
 Download-URL: https://github.com/Lightning-AI/lightning
@@ -48,6 +48,7 @@ Provides-Extra: fabric-test
 Requires-Dist: click==8.1.8; python_version < "3.11" and extra == "fabric-test"
 Requires-Dist: click==8.3.0; python_version > "3.10" and extra == "fabric-test"
 Requires-Dist: coverage==7.10.7; extra == "fabric-test"
+Requires-Dist: huggingface-hub; extra == "fabric-test"
 Requires-Dist: numpy<2.0,>=1.21.0; extra == "fabric-test"
 Requires-Dist: pytest-cov==6.3.0; extra == "fabric-test"
 Requires-Dist: pytest-random-order==1.2.0; extra == "fabric-test"
@@ -75,6 +76,7 @@ Provides-Extra: pytorch-test
 Requires-Dist: cloudpickle<4.0,>=1.3; extra == "pytorch-test"
 Requires-Dist: coverage==7.10.7; extra == "pytorch-test"
 Requires-Dist: fastapi; extra == "pytorch-test"
+Requires-Dist: huggingface-hub; extra == "pytorch-test"
 Requires-Dist: numpy<2.0,>1.20.0; extra == "pytorch-test"
 Requires-Dist: onnx<2.0,>1.12.0; extra == "pytorch-test"
 Requires-Dist: onnxruntime<2.0,>=1.12.0; extra == "pytorch-test"
@@ -103,6 +105,7 @@ Requires-Dist: click==8.1.8; python_version < "3.11" and extra == "fabric-dev"
 Requires-Dist: click==8.3.0; python_version > "3.10" and extra == "fabric-dev"
 Requires-Dist: coverage==7.10.7; extra == "fabric-dev"
 Requires-Dist: deepspeed<0.17.0,>=0.15.0; (platform_system != "Windows" and platform_system != "Darwin") and extra == "fabric-dev"
+Requires-Dist: huggingface-hub; extra == "fabric-dev"
 Requires-Dist: hydra-core<2.0,>=1.2.0; extra == "fabric-dev"
 Requires-Dist: numpy<2.0,>=1.21.0; extra == "fabric-dev"
 Requires-Dist: pytest-cov==6.3.0; extra == "fabric-dev"
@@ -133,6 +136,7 @@ Requires-Dist: cloudpickle<4.0,>=1.3; extra == "pytorch-dev"
 Requires-Dist: coverage==7.10.7; extra == "pytorch-dev"
 Requires-Dist: deepspeed<0.17.0,>=0.15.0; (platform_system != "Windows" and platform_system != "Darwin") and extra == "pytorch-dev"
 Requires-Dist: fastapi; extra == "pytorch-dev"
+Requires-Dist: huggingface-hub; extra == "pytorch-dev"
 Requires-Dist: hydra-core<2.0,>=1.2.0; extra == "pytorch-dev"
 Requires-Dist: ipython[all]<11.0,>=8.0.0; extra == "pytorch-dev"
 Requires-Dist: jsonargparse[jsonnet,signatures]<5.0,>=4.39.0; extra == "pytorch-dev"
@@ -183,6 +187,7 @@ Requires-Dist: click==8.3.0; python_version > "3.10" and extra == "test"
 Requires-Dist: cloudpickle<4.0,>=1.3; extra == "test"
 Requires-Dist: coverage==7.10.7; extra == "test"
 Requires-Dist: fastapi; extra == "test"
+Requires-Dist: huggingface-hub; extra == "test"
 Requires-Dist: numpy<2.0,>1.20.0; extra == "test"
 Requires-Dist: numpy<2.0,>=1.21.0; extra == "test"
 Requires-Dist: onnx<2.0,>1.12.0; extra == "test"
@@ -222,6 +227,7 @@ Requires-Dist: cloudpickle<4.0,>=1.3; extra == "dev"
 Requires-Dist: coverage==7.10.7; extra == "dev"
 Requires-Dist: deepspeed<0.17.0,>=0.15.0; (platform_system != "Windows" and platform_system != "Darwin") and extra == "dev"
 Requires-Dist: fastapi; extra == "dev"
+Requires-Dist: huggingface-hub; extra == "dev"
 Requires-Dist: hydra-core<2.0,>=1.2.0; extra == "dev"
 Requires-Dist: ipython[all]<11.0,>=8.0.0; extra == "dev"
 Requires-Dist: jsonargparse[jsonnet,signatures]<5.0,>=4.39.0; extra == "dev"
@@ -286,7 +292,7 @@ ______________________________________________________________________
   <a href="#lightning-fabric-expert-control">Fabric</a> •
   <a href="https://lightning.ai/?utm_source=ptl_readme&utm_medium=referral&utm_campaign=ptl_readme">Lightning Cloud</a> •   
   <a href="#community">Community</a> •
-  <a href="https://pytorch-lightning.readthedocs.io/en/2.6.0.dev20251019">Docs</a>
+  <a href="https://pytorch-lightning.readthedocs.io/en/2.6.0.dev20251102">Docs</a>
 </p>
 
 <!-- DO NOT ADD CONDA DOWNLOADS... README CHANGES MUST BE APPROVED BY EDEN OR WILL -->
@@ -828,7 +834,7 @@ Lightning is rigorously tested across multiple CPUs, GPUs and TPUs and against m
 
 |       System / PyTorch ver.        | 1.13                                                                                                                                                                                                                            | 2.0                                                                                                                                                                                                                             |                                                                                                               2.1                                                                                                               |
 | :--------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
-|        Linux py3.9 \[GPUs\]        |  |  | [![Build Status](https://dev.azure.com/Lightning-AI/lightning/_apis/build/status%2Fpytorch-lightning%20%28GPUs%29?branchName=refs%2Ftags%2F2.6.0.dev20251019)](https://dev.azure.com/Lightning-AI/lightning/_build/latest?definitionId=24&branchName=master) |
+|        Linux py3.9 \[GPUs\]        |  |  | [![Build Status](https://dev.azure.com/Lightning-AI/lightning/_apis/build/status%2Fpytorch-lightning%20%28GPUs%29?branchName=refs%2Ftags%2F2.6.0.dev20251102)](https://dev.azure.com/Lightning-AI/lightning/_build/latest?definitionId=24&branchName=master) |
 |  Linux (multiple Python versions)  | [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                                 | [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                                 |                 [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                 |
 |   OSX (multiple Python versions)   | [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                                 | [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                                 |                 [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                 |
 | Windows (multiple Python versions) | [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                                 | [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                                 |                 [![Test PyTorch](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml/badge.svg)](https://github.com/Lightning-AI/lightning/actions/workflows/ci-tests-pytorch.yml)                 |

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/pyproject.toml

@@ -91,7 +91,6 @@ ignore = [
 ]
 "src/**" = [
     "S101",  # todo: Use of `assert` detected
-    "S105", "S106", "S107",  # todo: Possible hardcoded password: ...
     "S113",  # todo: Probable use of requests call without timeout
     "S301",  # todo: `pickle` and modules that wrap it can be unsafe when used to deserialize untrusted data, possible security issue
     "S324",  # todo: Probable use of insecure hash functions in `hashlib`
@@ -106,7 +105,6 @@ ignore = [
 ]
 "tests/**" = [
     "S101",  # Use of `assert` detected
-    "S105", "S106",  # todo: Possible hardcoded password: ...
     "S301",  # `pickle` and modules that wrap it can be unsafe when used to deserialize untrusted data, possible security issue
     "S113",  # todo: Probable use of requests call without timeout
     "S311",  # todo: Standard pseudo-random generators are not suitable for cryptographic purposes

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/requirements/fabric/examples.txt

@@ -1,5 +1,5 @@
 # NOTE: the upper bound for the package version is only set for CI stability, and it is dropped while installing this package
 #  in case you want to preserve/enforce restrictions on the latest compatible version, add "strict" as an in-line comment
 
-torchvision >=0.16.0, <0.24.0
+torchvision >=0.16.0, <0.25.0
 torchmetrics >=0.10.0, <1.9.0

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/requirements/fabric/test.txt

@@ -9,3 +9,4 @@ pytest-random-order ==1.2.0
 click ==8.1.8; python_version < "3.11"
 click ==8.3.0; python_version > "3.10"
 tensorboardX >=2.6, <2.7.0  # todo: relax it back to `>=2.2` after fixing tests
+huggingface-hub

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/requirements/pytorch/examples.txt

@@ -2,6 +2,6 @@
 #  in case you want to preserve/enforce restrictions on the latest compatible version, add "strict" as an in-line comment
 
 requests <2.33.0
-torchvision >=0.16.0, <0.24.0
+torchvision >=0.16.0, <0.25.0
 ipython[all] >=8.0.0, <10.0.0
 torchmetrics >=0.10.0, <1.9.0

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/requirements/pytorch/extra.txt

@@ -5,7 +5,7 @@
 matplotlib>3.1, <3.11.0
 omegaconf >=2.2.3, <2.4.0
 hydra-core >=1.2.0, <1.4.0
-jsonargparse[signatures,jsonnet] >=4.39.0, <4.42.0
+jsonargparse[signatures,jsonnet] >=4.39.0, <4.43.0
 rich >=12.3.0, <14.2.0
 tensorboardX >=2.2, <2.7.0  # min version is set by torch.onnx missing attribute
 bitsandbytes >=0.45.2,<0.47.0; platform_system != "Darwin"

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/requirements/pytorch/test.txt

@@ -13,7 +13,7 @@ numpy >1.20.0, <1.27.0
 onnx >1.12.0, <1.20.0
 onnxruntime >=1.12.0, <1.24.0
 onnxscript >= 0.1.0, < 0.5.0
-psutil <7.1.1 # for `DeviceStatsMonitor`
+psutil <7.1.2 # for `DeviceStatsMonitor`
 pandas >2.0, <2.4.0  # needed in benchmarks
 fastapi  # for `ServableModuleValidator`  # not setting version as re-defined in App
 uvicorn  # for `ServableModuleValidator`  # not setting version as re-defined in App
@@ -21,3 +21,4 @@ uvicorn  # for `ServableModuleValidator`  # not setting version as re-defined in
 tensorboard >=2.11, <2.21.0  # for `TensorBoardLogger`
 
 torch-tensorrt; platform_system == "Linux" and python_version >= "3.12"
+huggingface-hub

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/fabric/CHANGELOG.md

@@ -27,7 +27,7 @@ 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))
 
 
 ---

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/fabric/plugins/environments/lightning.py

@@ -13,11 +13,11 @@
 # limitations under the License.
 
 import os
-import socket
 
 from typing_extensions import override
 
 from lightning.fabric.plugins.environments.cluster_environment import ClusterEnvironment
+from lightning.fabric.utilities.port_manager import get_port_manager
 from lightning.fabric.utilities.rank_zero import rank_zero_only
 
 
@@ -104,6 +104,13 @@ class LightningEnvironment(ClusterEnvironment):
         if "WORLD_SIZE" in os.environ:
             del os.environ["WORLD_SIZE"]
 
+        if self._main_port != -1:
+            get_port_manager().release_port(self._main_port)
+            self._main_port = -1
+
+        os.environ.pop("MASTER_PORT", None)
+        os.environ.pop("MASTER_ADDR", None)
+
 
 def find_free_network_port() -> int:
     """Finds a free port on localhost.
@@ -111,9 +118,24 @@ def find_free_network_port() -> int:
     It is useful in single-node training when we don't want to connect to a real main node but have to set the
     `MASTER_PORT` environment variable.
 
+    The allocated port is reserved and won't be returned by subsequent calls until it's explicitly released.
+
+    Returns:
+        A port number that is reserved and free at the time of allocation
+
     """
-    s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
-    s.bind(("", 0))
-    port = s.getsockname()[1]
-    s.close()
-    return port
+    # If an external launcher already specified a MASTER_PORT (for example, torch.distributed.spawn or
+    # multiprocessing helpers), reserve it through the port manager so no other test reuses the same number.
+    if "MASTER_PORT" in os.environ:
+        master_port_str = os.environ["MASTER_PORT"]
+        try:
+            existing_port = int(master_port_str)
+        except ValueError:
+            pass
+        else:
+            port_manager = get_port_manager()
+            if port_manager.reserve_existing_port(existing_port):
+                return existing_port
+
+    port_manager = get_port_manager()
+    return port_manager.allocate_port()

lightning-2.6.0.dev20251102/src/lightning/fabric/utilities/port_manager.py

@@ -0,0 +1,233 @@
+# Copyright The Lightning AI team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Port allocation manager to prevent race conditions in distributed training."""
+
+import atexit
+import logging
+import socket
+import threading
+from collections import deque
+from collections.abc import Iterator
+from contextlib import contextmanager
+from typing import Optional
+
+log = logging.getLogger(__name__)
+
+# Size of the recently released ports queue
+# This prevents immediate reuse of ports that were just released
+# Set to 1024 to balance memory usage vs TIME_WAIT protection
+_RECENTLY_RELEASED_PORTS_MAXLEN = 1024
+
+
+class PortManager:
+    """Thread-safe port manager to prevent EADDRINUSE errors.
+
+    This manager maintains a global registry of allocated ports to ensure that multiple concurrent tests don't try to
+    use the same port. While this doesn't completely eliminate the race condition with external processes, it prevents
+    internal collisions within the test suite.
+
+    """
+
+    def __init__(self) -> None:
+        self._lock = threading.Lock()
+        self._allocated_ports: set[int] = set()
+        # Recently released ports are kept in a queue to avoid immediate reuse
+        self._recently_released: deque[int] = deque(maxlen=_RECENTLY_RELEASED_PORTS_MAXLEN)
+        # Register cleanup to release all ports on exit
+        atexit.register(self.release_all)
+
+    def allocate_port(self, preferred_port: Optional[int] = None, max_attempts: int = 1000) -> int:
+        """Allocate a free port, ensuring it's not already reserved.
+
+        Args:
+            preferred_port: If provided, try to allocate this specific port first
+            max_attempts: Maximum number of attempts to find a free port
+
+        Returns:
+            An allocated port number
+
+        Raises:
+            RuntimeError: If unable to find a free port after max_attempts
+
+        """
+        with self._lock:
+            # If a preferred port is specified and available, use it
+            if (
+                preferred_port is not None
+                and preferred_port not in self._allocated_ports
+                and preferred_port not in self._recently_released
+                and self._is_port_free(preferred_port)
+            ):
+                self._allocated_ports.add(preferred_port)
+                return preferred_port
+
+            # Let the OS choose a free port, but verify it's not in our tracking structures
+            # The OS naturally avoids ports in TIME_WAIT (without SO_REUSEADDR)
+            for attempt in range(max_attempts):
+                port = self._find_free_port()
+
+                # Skip if already allocated by us or recently released
+                # This prevents race conditions within our process
+                if port not in self._allocated_ports and port not in self._recently_released:
+                    self._allocated_ports.add(port)
+
+                    # Log diagnostics if queue utilization is high
+                    queue_count = len(self._recently_released)
+                    if queue_count > _RECENTLY_RELEASED_PORTS_MAXLEN * 0.8:  # >80% full
+                        log.warning(
+                            f"Port queue utilization high: {queue_count}/{_RECENTLY_RELEASED_PORTS_MAXLEN} "
+                            f"({queue_count / _RECENTLY_RELEASED_PORTS_MAXLEN * 100:.1f}% full). "
+                            f"Allocated port {port}. Active allocations: {len(self._allocated_ports)}"
+                        )
+
+                    return port
+
+            # Provide detailed diagnostics to understand allocation failures
+            allocated_count = len(self._allocated_ports)
+            queue_count = len(self._recently_released)
+            queue_capacity = _RECENTLY_RELEASED_PORTS_MAXLEN
+            queue_utilization = (queue_count / queue_capacity * 100) if queue_capacity > 0 else 0
+
+            raise RuntimeError(
+                f"Failed to allocate a free port after {max_attempts} attempts. "
+                f"Diagnostics: allocated={allocated_count}, "
+                f"recently_released={queue_count}/{queue_capacity} ({queue_utilization:.1f}% full). "
+                f"If queue is near capacity, consider increasing _RECENTLY_RELEASED_PORTS_MAXLEN."
+            )
+
+    def release_port(self, port: int) -> None:
+        """Release a previously allocated port.
+
+        Args:
+            port: Port number to release
+
+        """
+        with self._lock:
+            if port in self._allocated_ports:
+                self._allocated_ports.remove(port)
+                # Add to the back of the queue; oldest will be evicted when queue is full
+                self._recently_released.append(port)
+
+    def release_all(self) -> None:
+        """Release all allocated ports."""
+        with self._lock:
+            self._allocated_ports.clear()
+            self._recently_released.clear()
+
+    def reserve_existing_port(self, port: int) -> bool:
+        """Reserve a port that was allocated externally.
+
+        Args:
+            port: The externally assigned port to reserve.
+
+        Returns:
+            True if the port was reserved (or already reserved), False if the port value is invalid.
+
+        """
+        if port <= 0 or port > 65535:
+            return False
+
+        with self._lock:
+            if port in self._allocated_ports:
+                return True
+
+            # Remove from recently released queue if present (we're explicitly reserving it)
+            if port in self._recently_released:
+                # Create a new deque without this port
+                self._recently_released = deque(
+                    (p for p in self._recently_released if p != port), maxlen=_RECENTLY_RELEASED_PORTS_MAXLEN
+                )
+
+            self._allocated_ports.add(port)
+            return True
+
+    @contextmanager
+    def allocated_port(self, preferred_port: Optional[int] = None) -> Iterator[int]:
+        """Context manager for automatic port cleanup.
+
+        Usage:
+            with manager.allocated_port() as port:
+                # Use port here
+                pass
+            # Port automatically released
+
+        Args:
+            preferred_port: Optional preferred port number
+
+        Yields:
+            Allocated port number
+
+        """
+        port = self.allocate_port(preferred_port=preferred_port)
+        try:
+            yield port
+        finally:
+            self.release_port(port)
+
+    @staticmethod
+    def _find_free_port() -> int:
+        """Find a free port using OS allocation.
+
+        Returns:
+            A port number that was free at the time of checking
+
+        """
+        s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        # Don't use SO_REUSEADDR - we need to match the behavior of TCPStore
+        # which binds without it, so ports in TIME_WAIT will be rejected
+        s.bind(("", 0))
+        port = s.getsockname()[1]
+        s.close()
+        return port
+
+    @staticmethod
+    def _is_port_free(port: int) -> bool:
+        """Check if a specific port is available.
+
+        Args:
+            port: Port number to check
+
+        Returns:
+            True if the port is free, False otherwise
+
+        """
+        try:
+            s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+            # Don't use SO_REUSEADDR - we need to match the behavior of TCPStore
+            # which binds without it, so ports in TIME_WAIT will be rejected
+            s.bind(("", port))
+            s.close()
+            return True
+        except OSError:
+            return False
+
+
+# Global singleton instance
+_port_manager: Optional[PortManager] = None
+_port_manager_lock = threading.Lock()
+
+
+def get_port_manager() -> PortManager:
+    """Get or create the global port manager instance.
+
+    Returns:
+        The global PortManager singleton
+
+    """
+    global _port_manager
+    if _port_manager is None:
+        with _port_manager_lock:
+            if _port_manager is None:
+                _port_manager = PortManager()
+    return _port_manager

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/fabric/utilities/testing/_runif.py

@@ -113,7 +113,7 @@ def _runif_reasons(
             reasons.append("Standalone execution")
         kwargs["standalone"] = True
 
-    if deepspeed and not (_DEEPSPEED_AVAILABLE and not _TORCH_GREATER_EQUAL_2_4):
+    if deepspeed and not (_DEEPSPEED_AVAILABLE and _TORCH_GREATER_EQUAL_2_4):
         reasons.append("Deepspeed")
 
     if dynamo:

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/pytorch/CHANGELOG.md

@@ -58,6 +58,13 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed `LightningCLI` loading of hyperparameters from `ckpt_path` failing for subclass model mode ([#21246](https://github.com/Lightning-AI/pytorch-lightning/pull/21246))
 
 
+- Fixed how `ThroughputMonitor` calculated training time ([#21291](https://github.com/Lightning-AI/pytorch-lightning/pull/21291))
+
+
+- Fixed synchronization of gradients in manual optimization with `DDPStrategy(static_graph=True)` ([#21251](https://github.com/Lightning-AI/pytorch-lightning/pull/21251))
+
+
+
 ---
 
 ## [2.5.5] - 2025-09-05

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/pytorch/callbacks/lr_monitor.py

@@ -241,7 +241,7 @@ class LearningRateMonitor(Callback):
         self.lrs[name].append(lr)
         return {name: lr}
 
-    def _remap_keys(self, names: list[list[str]], token: str = "/pg1") -> None:
+    def _remap_keys(self, names: list[list[str]], token: str = "/pg1") -> None:  # noqa: S107
         """This function is used the remap the keys if param groups for a given optimizer increased."""
         for group_new_names in names:
             for new_name in group_new_names:

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/pytorch/callbacks/throughput_monitor.py

@@ -109,7 +109,9 @@ class ThroughputMonitor(Callback):
         stage = trainer.state.stage
         assert stage is not None
 
-        if stage not in self._samples:
+        reset_needed = trainer.state.fn == TrainerFn.FITTING or stage not in self._samples
+
+        if reset_needed:
             self._throughputs[stage].reset()
             self._lengths[stage] = 0
             self._samples[stage] = 0
@@ -202,10 +204,17 @@ class ThroughputMonitor(Callback):
     def on_validation_end(self, trainer: "Trainer", *_: Any) -> None:
         if trainer.sanity_checking or trainer.state.fn != TrainerFn.FITTING:
             return
+
+        train_times = self._throughputs[RunningStage.TRAINING]._time
+        val_times = self._throughputs[RunningStage.VALIDATING]._time
+
+        train_elapsed = train_times[-1] if train_times else 0.0
+        val_elapsed = val_times[-1] if val_times else 0.0
+
         # add the validation time to the training time before continuing to avoid sinking the training throughput
-        training_finished = self._t0s[RunningStage.TRAINING] + sum(self._throughputs[RunningStage.TRAINING]._time)
+        training_finished = self._t0s[RunningStage.TRAINING] + train_elapsed
         time_between_train_and_val = self._t0s[RunningStage.VALIDATING] - training_finished
-        val_time = sum(self._throughputs[RunningStage.VALIDATING]._time)
+        val_time = val_elapsed
         self._t0s[RunningStage.TRAINING] += time_between_train_and_val + val_time
 
     @override

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/pytorch/cli.py

@@ -66,6 +66,13 @@ ModuleType = TypeVar("ModuleType")
 
 
 class ReduceLROnPlateau(torch.optim.lr_scheduler.ReduceLROnPlateau):
+    """Custom ReduceLROnPlateau scheduler that extends PyTorch's ReduceLROnPlateau.
+
+    This class adds a `monitor` attribute to the standard PyTorch ReduceLROnPlateau to specify which metric should be
+    tracked for learning rate adjustment.
+
+    """
+
     def __init__(self, optimizer: Optimizer, monitor: str, *args: Any, **kwargs: Any) -> None:
         super().__init__(optimizer, *args, **kwargs)
         self.monitor = monitor

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/pytorch/profilers/profiler.py

@@ -77,7 +77,10 @@ class Profiler(ABC):
             log.info(*args, **kwargs)
 
     def _prepare_filename(
-        self, action_name: Optional[str] = None, extension: str = ".txt", split_token: str = "-"
+        self,
+        action_name: Optional[str] = None,
+        extension: str = ".txt",
+        split_token: str = "-",  # noqa: S107
     ) -> str:
         args = []
         if self._stage is not None:

{lightning-2.6.0.dev20251019 → lightning-2.6.0.dev20251102}/src/lightning/pytorch/strategies/ddp.py

@@ -103,6 +103,7 @@ class DDPStrategy(ParallelStrategy):
         self._process_group_backend: Optional[str] = process_group_backend
         self._timeout: Optional[timedelta] = timeout
         self._start_method = start_method
+        self._pl_static_graph_delay_done = False
 
     @property
     def is_distributed(self) -> bool:  # pragma: no-cover
@@ -319,6 +320,27 @@ class DDPStrategy(ParallelStrategy):
         if not self.lightning_module.automatic_optimization:
             prepare_for_backward(self.model, closure_loss)
 
+    @override
+    def post_backward(self, closure_loss: Tensor) -> None:
+        # Only for first static-graph iteration with manual optimization
+        model = self.model
+        lm = self.lightning_module
+        if not isinstance(model, DistributedDataParallel):
+            return
+        if lm is None or lm.automatic_optimization:
+            return
+        if not getattr(model, "static_graph", False):
+            return
+        if self._pl_static_graph_delay_done:
+            return
+
+        # Call DDP's own first-iter static-graph flush.
+        # This is what actually launches the bucket all-reduces.
+        reducer = model.reducer
+        reducer._delay_all_reduce()
+
+        self._pl_static_graph_delay_done = True
+
     @override
     def model_to_device(self) -> None:
         log.debug(f"{self.__class__.__name__}: moving model to device [{self.root_device}]...")

lightning-2.6.0.dev20251102/src/lightning/version.info

@@ -0,0 +1 @@
+2.6.0.dev20251102