Merge branch 'master' into feat/to_tensorrt

GdoongMathew · web-flow · commit 8489571b3316 · 2025-07-13T15:21:58.000+08:00
diff --git a/.azure/gpu-tests-fabric.yml b/.azure/gpu-tests-fabric.yml
@@ -130,7 +130,7 @@ jobs:
       - bash: |
           set -e
           extra=$(python -c "print({'lightning': 'fabric-'}.get('$(PACKAGE_NAME)', ''))")
-          pip install -e ".[${extra}dev]" -U --extra-index-url="${TORCH_URL}"
+          pip install -e ".[${extra}dev]" -U --upgrade-strategy=eager --extra-index-url="${TORCH_URL}"
         displayName: "Install package & dependencies"
 
       - bash: |
diff --git a/.azure/gpu-tests-pytorch.yml b/.azure/gpu-tests-pytorch.yml
@@ -134,7 +134,7 @@ jobs:
       - bash: |
           set -e
           extra=$(python -c "print({'lightning': 'pytorch-'}.get('$(PACKAGE_NAME)', ''))")
-          pip install -e ".[${extra}dev]" -U --extra-index-url="${TORCH_URL}"
+          pip install -e ".[${extra}dev]" -U --upgrade-strategy=eager --extra-index-url="${TORCH_URL}"
         displayName: "Install package & dependencies"
 
       - bash: pip uninstall -y lightning
diff --git a/.github/CONTRIBUTING.md b/.github/CONTRIBUTING.md
@@ -109,6 +109,36 @@ ______________________________________________________________________
 
 ## Guidelines
 
+### Development environment
+
+To set up a local development environment, we recommend using `uv`, which can be installed following their [instructions](https://docs.astral.sh/uv/getting-started/installation/).
+
+Once `uv` has been installed, begin by cloning the repository:
+
+```bash
+git clone https://github.com/Lightning-AI/lightning.git
+cd lightning
+```
+
+Once in root level of the repository, create a new virtual environment and install the project dependencies.
+
+```bash
+uv venv
+# uv venv --python 3.11 # use this instead if you need a specific python version
+
+source .venv/bin/activate # command may differ based on your shell
+uv pip install ".[dev, examples]"
+```
+
+Once the dependencies have been installed, install pre-commit and set up the git hook scripts:
+
+```bash
+uv pip install pre-commit
+pre-commit install
+```
+
+If you would like more information regarding the uv commands, please refer to uv's documentation for more information on their [pip interface](https://docs.astral.sh/uv/pip/).
+
 ### Developments scripts
 
 To build the documentation locally, simply execute the following commands from project root (only for Unix):
diff --git a/.github/workflows/ci-tests-fabric.yml b/.github/workflows/ci-tests-fabric.yml
@@ -138,7 +138,8 @@ jobs:
       - name: Install package & dependencies
         timeout-minutes: 20
         run: |
-          pip install -e ".[${EXTRA_PREFIX}test,${EXTRA_PREFIX}strategies]" -U --prefer-binary \
+          pip install -e ".[${EXTRA_PREFIX}test,${EXTRA_PREFIX}strategies]" \
+            -U --upgrade-strategy=eager --prefer-binary \
             --extra-index-url="${TORCH_URL}" --find-links="${PYPI_CACHE_DIR}"
           pip list
       - name: Dump handy wheels
diff --git a/.github/workflows/ci-tests-pytorch.yml b/.github/workflows/ci-tests-pytorch.yml
@@ -136,7 +136,8 @@ jobs:
       - name: Install package & dependencies
         timeout-minutes: 20
         run: |
-          pip install ".[${EXTRA_PREFIX}extra,${EXTRA_PREFIX}test,${EXTRA_PREFIX}strategies]" -U --prefer-binary \
+          pip install ".[${EXTRA_PREFIX}extra,${EXTRA_PREFIX}test,${EXTRA_PREFIX}strategies]" \
+            -U --upgrade-strategy=eager --prefer-binary \
             -r requirements/_integrations/accelerators.txt \
             --extra-index-url="${TORCH_URL}" --find-links="${PYPI_CACHE_DIR}"
           pip list
diff --git a/.github/workflows/probot-check-group.yml b/.github/workflows/probot-check-group.yml
@@ -14,7 +14,7 @@ jobs:
     if: github.event.pull_request.draft == false
     timeout-minutes: 61 # in case something is wrong with the internal timeout
     steps:
-      - uses: Lightning-AI/probot@v5.4
+      - uses: Lightning-AI/probot@v5.5
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
diff --git a/docs/source-pytorch/advanced/transfer_learning.rst b/docs/source-pytorch/advanced/transfer_learning.rst
@@ -32,7 +32,7 @@ Let's use the `AutoEncoder` as a feature extractor in a separate model.
     class CIFAR10Classifier(LightningModule):
         def __init__(self):
             # init the pretrained LightningModule
-            self.feature_extractor = AutoEncoder.load_from_checkpoint(PATH)
+            self.feature_extractor = AutoEncoder.load_from_checkpoint(PATH).encoder
             self.feature_extractor.freeze()
 
             # the autoencoder outputs a 100-dim representation and CIFAR-10 has 10 classes
diff --git a/docs/source-pytorch/common/precision_intermediate.rst b/docs/source-pytorch/common/precision_intermediate.rst
@@ -165,7 +165,7 @@ Under the hood, we use `transformer_engine.pytorch.fp8_autocast <https://docs.nv
 Quantization via Bitsandbytes
 *****************************
 
-`bitsandbytes <https://github.com/TimDettmers/bitsandbytes>`__ (BNB) is a library that supports quantizing :class:`torch.nn.Linear` weights.
+`bitsandbytes <https://github.com/bitsandbytes-foundation/bitsandbytes>`__ (BNB) is a library that supports quantizing :class:`torch.nn.Linear` weights.
 
 Both 4-bit (`paper reference <https://arxiv.org/abs/2305.14314v1>`__) and 8-bit (`paper reference <https://arxiv.org/abs/2110.02861>`__) quantization is supported.
 Specifically, we support the following modes:
diff --git a/docs/source-pytorch/glossary/index.rst b/docs/source-pytorch/glossary/index.rst
@@ -209,7 +209,7 @@ Glossary
 
 .. displayitem::
    :header: LightningModule
-   :description: A base class organizug your neural network module
+   :description: A base class organizing your neural network module
    :col_css: col-md-12
    :button_link: ../common/lightning_module.html
    :height: 100
diff --git a/requirements/ci.txt b/requirements/ci.txt
@@ -1,6 +1,6 @@
 setuptools <80.9.1
 wheel <0.46.0
-awscli >=1.30.0, <1.41.0
+awscli >=1.30.0, <1.42.0
 twine ==6.1.0
 importlib-metadata <9.0.0
 wget
diff --git a/requirements/fabric/strategies.txt b/requirements/fabric/strategies.txt
@@ -6,5 +6,4 @@
 # note: is a bug around 0.10 with `MPS_Accelerator must implement all abstract methods`
 #  shall be resolved by https://github.com/microsoft/DeepSpeed/issues/4372
 deepspeed >=0.9.3, <=0.9.3; platform_system != "Windows" and platform_system != "Darwin"  # strict
-# skip bitsandbytes==0.46, due to ValueError: too many values to unpack (expected 2)
-bitsandbytes >=0.45.2,!=0.46,<0.47.0; platform_system != "Darwin"
+bitsandbytes >=0.45.2,<0.47.0; platform_system != "Darwin"
diff --git a/requirements/fabric/test.txt b/requirements/fabric/test.txt
@@ -1,4 +1,4 @@
-coverage ==7.9.1
+coverage ==7.9.2
 numpy >=1.17.2, <1.27.0
 pytest ==8.4.1
 pytest-cov ==6.2.1
diff --git a/requirements/pytorch/base.txt b/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 >=2.1.0, <2.8.0
+torch >=2.1.0, <=2.8.0
 tqdm >=4.57.0, <4.68.0
 PyYAML >5.4, <6.1.0
 fsspec[http] >=2022.5.0, <2025.6.0
diff --git a/requirements/pytorch/extra.txt b/requirements/pytorch/extra.txt
@@ -8,5 +8,4 @@ hydra-core >=1.2.0, <1.4.0
 jsonargparse[signatures,jsonnet] >=4.39.0, <4.41.0
 rich >=12.3.0, <14.1.0
 tensorboardX >=2.2, <2.7.0  # min version is set by torch.onnx missing attribute
-# skip bitsandbytes==0.46, due to ValueError: too many values to unpack (expected 2)
-bitsandbytes >=0.45.2,!=0.46,<0.47.0; platform_system != "Darwin"
+bitsandbytes >=0.45.2,<0.47.0; platform_system != "Darwin"
diff --git a/requirements/pytorch/test.txt b/requirements/pytorch/test.txt
@@ -1,4 +1,4 @@
-coverage ==7.9.1
+coverage ==7.9.2
 pytest ==8.4.1
 pytest-cov ==6.2.1
 pytest-timeout ==2.4.0
diff --git a/src/lightning/fabric/plugins/precision/bitsandbytes.py b/src/lightning/fabric/plugins/precision/bitsandbytes.py
@@ -256,10 +256,12 @@ def quantize(
             if int8params.has_fp16_weights:
                 int8params.data = B
             else:
-                if hasattr(bnb.functional, "double_quant"):
+                # bitsandbytes >= 0.45 supports an improved API
+                if hasattr(bnb.functional, "int8_vectorwise_quant"):
+                    CB, SCB, _ = bnb.functional.int8_vectorwise_quant(B)
+                else:  # old method is deprecated in 0.45, removed in 0.46+.
                     CB, _, SCB, _, _ = bnb.functional.double_quant(B)
-                else:  # for bitsandbytes versions ≥0.46
-                    CB, SCB = bnb.functional.int8_double_quant(B)
+
                 int8params.data = CB
                 setattr(int8params, "CB", CB)
                 setattr(int8params, "SCB", SCB)
diff --git a/src/lightning/pytorch/callbacks/model_checkpoint.py b/src/lightning/pytorch/callbacks/model_checkpoint.py
@@ -133,9 +133,15 @@ class ModelCheckpoint(Checkpoint):
             will only save checkpoints at epochs 0 < E <= N
             where both values for ``every_n_epochs`` and ``check_val_every_n_epoch`` evenly divide E.
         save_on_train_epoch_end: Whether to run checkpointing at the end of the training epoch.
-            If this is ``False``, then the check runs at the end of the validation.
+            If ``True``, checkpoints are saved at the end of every training epoch.
+            If ``False``, checkpoints are saved at the end of validation.
+            If ``None`` (default), checkpointing behavior is determined based on training configuration.
+            If ``check_val_every_n_epoch != 1``, checkpointing will not be performed at the end of
+            every training epoch. If there are no validation batches of data, checkpointing will occur at the
+            end of the training epoch. If there is a non-default number of validation runs per training epoch
+            (``val_check_interval != 1``), checkpointing is performed after validation.
         enable_version_counter: Whether to append a version to the existing file name.
-            If this is ``False``, then the checkpoint files will be overwritten.
+            If ``False``, then the checkpoint files will be overwritten.
 
     Note:
         For extra customization, ModelCheckpoint includes the following attributes:
diff --git a/src/lightning/pytorch/trainer/connectors/accelerator_connector.py b/src/lightning/pytorch/trainer/connectors/accelerator_connector.py
@@ -453,10 +453,11 @@ def _check_strategy_and_fallback(self) -> None:
 
         if (
             strategy_flag in FSDPStrategy.get_registered_strategies() or type(self._strategy_flag) is FSDPStrategy
-        ) and self._accelerator_flag not in ("cuda", "gpu"):
+        ) and not (self._accelerator_flag in ("cuda", "gpu") or isinstance(self._accelerator_flag, CUDAAccelerator)):
             raise ValueError(
-                f"The strategy `{FSDPStrategy.strategy_name}` requires a GPU accelerator, but got:"
-                f" {self._accelerator_flag}"
+                f"The strategy `{FSDPStrategy.strategy_name}` requires a GPU accelerator, but received "
+                f"`accelerator={self._accelerator_flag!r}`. Please set `accelerator='cuda'`, `accelerator='gpu'`,"
+                " or pass a `CUDAAccelerator()` instance to use FSDP."
             )
         if strategy_flag in _DDP_FORK_ALIASES and "fork" not in torch.multiprocessing.get_all_start_methods():
             raise ValueError(
diff --git a/src/version.info b/src/version.info
@@ -1 +1 @@
-2.6.0dev
+2.6.0dev0
diff --git a/tests/tests_pytorch/trainer/connectors/test_accelerator_connector.py b/tests/tests_pytorch/trainer/connectors/test_accelerator_connector.py
@@ -582,6 +582,11 @@ class AcceleratorSubclass(CPUAccelerator):
     Trainer(accelerator=AcceleratorSubclass(), strategy=FSDPStrategySubclass())
 
 
+@RunIf(min_cuda_gpus=1)
+def test_check_fsdp_strategy_and_fallback_with_cudaaccelerator():
+    Trainer(strategy="fsdp", accelerator=CUDAAccelerator())
+
+
 @mock.patch.dict(os.environ, {}, clear=True)
 def test_unsupported_tpu_choice(xla_available, tpu_available):
     # if user didn't set strategy, _Connector will choose the SingleDeviceXLAStrategy or XLAStrategy
