Merge branch 'master' into feature/fsdp-cpu

Author: bhimrazy

.lightning/workflows/pytorch.yml

@@ -122,10 +122,11 @@ run: |
   echo "Install package"
   extra=$(python -c "print({'lightning': 'pytorch-'}.get('${PACKAGE_NAME}', ''))")
 
-  # Use find-links to prefer CUDA-specific packages from PyTorch index
+  # Prefer CUDA-specific builds from the PyTorch index.
+  # torch-tensorrt is intentionally left to PyPI: since 2.13.0 its `+cuXXX` wheels no longer
+  # bundle `libtorchtrt.so`, which silently disables the TorchScript frontend.
   uv pip install -e ".[${extra}dev,${extra}test_gpu]" --upgrade \
-    --find-links="https://download.pytorch.org/whl/${UV_TORCH_BACKEND}" \
-    --find-links="https://download.pytorch.org/whl/${UV_TORCH_BACKEND}/torch-tensorrt"
+    --find-links="https://download.pytorch.org/whl/${UV_TORCH_BACKEND}"
   uv pip list
 
   echo "Ensure only a single package is installed"

docs/source-pytorch/versioning.rst

@@ -83,6 +83,18 @@ Since the release of PyTorch `2.0`, Lightning strives to officially support the
      - ``torch``
      - ``torchmetrics``
      - Python
+   * - 2.7
+     - 2.7
+     - 2.7
+     - ≥2.6 (▼ 2.13)
+     - ≥0.7.0
+     - ≥3.10 (▼ 3.13)
+   * - 2.6
+     - 2.6
+     - 2.6
+     - ≥2.1 (▼ 2.10)
+     - ≥0.7.0
+     - ≥3.10 (▼ 3.13)
    * - 2.5
      - 2.5
      - 2.5

pyproject.toml

@@ -177,6 +177,9 @@ filterwarnings = [
     # deprecated in torch 2.12; the suggested replacement is a private API
     # TODO: migrate to torch.distributed._functional_collectives.all_gather_single for torch>=2.12
     "ignore:torch.distributed.nn.functional.all_gather is deprecated.*:FutureWarning",
+    # torch 2.13 renamed these collectives to their `_single` counterparts; DeepSpeed still calls the old names
+    "ignore:`torch.distributed.all_gather_into_tensor` is deprecated.*:FutureWarning",
+    "ignore:`torch.distributed.reduce_scatter_tensor` is deprecated.*:FutureWarning",
     # PyTorch pytree LeafSpec deprecation triggered during doctest
     "ignore:`isinstance\\(treespec, LeafSpec\\)` is deprecated.*:FutureWarning",
 ]

requirements/fabric/base.txt

@@ -1,7 +1,7 @@
 # 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
 
-torch >=2.6.0, <2.13.0
+torch >=2.6.0, <2.14.0
 fsspec[http] >=2022.5.0, <2026.4.0
 packaging >=23.0, <=26.0
 typing-extensions >4.5.0, <4.16.0

requirements/pytorch/base.txt

@@ -1,7 +1,7 @@
 # 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
 
-torch >=2.6.0, <2.13.0
+torch >=2.6.0, <2.14.0
 tqdm >=4.57.0, <4.68.0
 PyYAML >5.4, <6.1.0
 fsspec[http] >=2022.5.0, <2026.4.0

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, <4.48.0
+jsonargparse[signatures,jsonnet] >=4.39.0, <4.51.0
 rich >=12.3.0, <14.4.0
 tensorboardX >=2.2, <2.7.0  # min version is set by torch.onnx missing attribute
 bitsandbytes >=0.45.2,<0.50.0; platform_system != "Darwin"

src/lightning/fabric/CHANGELOG.md

@@ -36,6 +36,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed AccumulateGrad stream mismatch warning when using DDP with Fabric ([#21746](https://github.com/Lightning-AI/pytorch-lightning/pull/21746))
 
+- Fixed `_atomic_save` swallowing `PermissionError`, which made checkpoint saving report success without writing a file ([#21799](https://github.com/Lightning-AI/pytorch-lightning/pull/21799))
+
 ---
 
 ## [2.6.4] - 2026-05-20

src/lightning/fabric/utilities/cloud_io.py

@@ -130,6 +130,7 @@ def _atomic_save(checkpoint: dict[str, Any], filepath: _PATH) -> None:
             raise RuntimeError(
                 'Upgrade fsspec to enable cross-device local checkpoints: pip install "fsspec[http]>=2025.5.0"',
             ) from e
+        raise
 
 
 def _is_object_storage(fs: AbstractFileSystem) -> bool:

src/lightning/pytorch/CHANGELOG.md

@@ -37,6 +37,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- Fixed `LightningCLI` emitting `jsonargparse` deprecation warnings ([#21900](https://github.com/Lightning-AI/pytorch-lightning/issues/21900))
+
+- Fixed `RichProgressBar` showing a nonsensical negative epoch total (e.g. `Epoch 5/-2`) when `Trainer(max_epochs=-1)` (unlimited epochs) is used and training stops via another condition ([#21925](https://github.com/Lightning-AI/pytorch-lightning/issues/21925))
+
 - Fixed crash when calling ``self.log()`` inside a ``torch.compile``-wrapped ``LightningModule`` on PyTorch 2.12/2.13 by disabling Dynamo tracing at the ``LightningModule.log`` boundary ([#21836](https://github.com/Lightning-AI/pytorch-lightning/issues/21836))
 
 - Fixed PyTorch Lightning profiler not capturing dataloader worker initialization time ([#21771](https://github.com/Lightning-AI/pytorch-lightning/issues/21771))

src/lightning/pytorch/callbacks/progress/rich_progress.py

@@ -655,7 +655,7 @@ def on_predict_batch_end(
 
     def _get_train_description(self, current_epoch: int) -> str:
         train_description = f"Epoch {current_epoch}"
-        if self.trainer.max_epochs is not None:
+        if self.trainer.max_epochs is not None and self.trainer.max_epochs >= 0:
             train_description += f"/{self.trainer.max_epochs - 1}"
         if len(self.validation_description) > len(train_description):
             # Padding is required to avoid flickering due of uneven lengths of "Epoch X"