build(deps): bump torch from 2.5.1 to 2.6.0 & torchvision from <0.21.0,>=0.16.0 to >=0.16.0,<0.22.0 in /requirements (#20728)

* build(deps): bump torch from 2.5.1 to 2.6.0 in /requirements

Bumps [torch](https://github.com/pytorch/pytorch) from 2.5.1 to 2.6.0.
- [Release notes](https://github.com/pytorch/pytorch/releases)
- [Changelog](https://github.com/pytorch/pytorch/blob/main/RELEASE.md)
- [Commits](pytorch/pytorch@v2.5.1...v2.6.0)

---
updated-dependencies:
- dependency-name: torch
  dependency-version: 2.6.0
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <[email protected]>

* build(deps): update torchvision requirement from <0.21.0,>=0.16.0 to >=0.16.0,<0.22.0 in /requirements (#20736)

build(deps): update torchvision requirement in /requirements

Updates the requirements on [torchvision](https://github.com/pytorch/vision) to permit the latest version.
- [Release notes](https://github.com/pytorch/vision/releases)
- [Commits](pytorch/vision@v0.16.0...v0.21.0)

---
updated-dependencies:
- dependency-name: torchvision
  dependency-version: 0.21.0
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <[email protected]>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* # type: ignore[arg-type]
* type: ignore[arg-type]

---------

Signed-off-by: dependabot[bot] <[email protected]>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Jirka B <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Author: 3 people

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 >=2.1.0, <2.6.0
+torch >=2.1.0, <2.7.0
 fsspec[http] >=2022.5.0, <2024.4.0
 packaging >=20.0, <=25.0
 typing-extensions >=4.4.0, <4.11.0

requirements/fabric/examples.txt

@@ -1,6 +1,6 @@
 # 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, <0.21.0
+torchvision >=0.16.0, <0.22.0
 torchmetrics >=0.10.0, <1.8.0
 lightning-utilities >=0.8.0, <0.15.0

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 >=2.1.0, <2.6.0
+torch >=2.1.0, <2.7.0
 tqdm >=4.57.0, <4.67.0
 PyYAML >=5.4, <6.1.0
 fsspec[http] >=2022.5.0, <2024.4.0

requirements/pytorch/examples.txt

@@ -2,7 +2,7 @@
 #  in case you want to preserve/enforce restrictions on the latest compatible version, add "strict" as an in-line comment
 
 requests <2.32.0
-torchvision >=0.16.0, <0.21.0
+torchvision >=0.16.0, <0.22.0
 ipython[all] <8.19.0
 torchmetrics >=0.10.0, <1.8.0
 lightning-utilities >=0.8.0, <0.15.0

requirements/typing.txt

@@ -1,5 +1,5 @@
 mypy==1.11.0
-torch==2.5.1
+torch==2.6.0
 
 types-Markdown
 types-PyYAML

src/lightning/fabric/plugins/collectives/torch_collective.py

@@ -50,7 +50,7 @@ def world_size(self) -> int:
 
     @override
     def broadcast(self, tensor: Tensor, src: int) -> Tensor:
-        dist.broadcast(tensor, src, group=self.group)
+        dist.broadcast(tensor, src, group=self.group)  # type: ignore[arg-type]
         return tensor
 
     @override
@@ -62,7 +62,7 @@ def all_reduce(self, tensor: Tensor, op: Union[str, ReduceOp, RedOpType] = "sum"
     @override
     def reduce(self, tensor: Tensor, dst: int, op: Union[str, ReduceOp, RedOpType] = "sum") -> Tensor:
         op = self._convert_to_native_op(op)
-        dist.reduce(tensor, dst, op=op, group=self.group)
+        dist.reduce(tensor, dst, op=op, group=self.group)  # type: ignore[arg-type]
         return tensor
 
     @override
@@ -72,12 +72,12 @@ def all_gather(self, tensor_list: list[Tensor], tensor: Tensor) -> list[Tensor]:
 
     @override
     def gather(self, tensor: Tensor, gather_list: list[Tensor], dst: int = 0) -> list[Tensor]:
-        dist.gather(tensor, gather_list, dst, group=self.group)
+        dist.gather(tensor, gather_list, dst, group=self.group)  # type: ignore[arg-type]
         return gather_list
 
     @override
     def scatter(self, tensor: Tensor, scatter_list: list[Tensor], src: int = 0) -> Tensor:
-        dist.scatter(tensor, scatter_list, src, group=self.group)
+        dist.scatter(tensor, scatter_list, src, group=self.group)  # type: ignore[arg-type]
         return tensor
 
     @override
@@ -109,27 +109,27 @@ def all_gather_object(self, object_list: list[Any], obj: Any) -> list[Any]:
     def broadcast_object_list(
         self, object_list: list[Any], src: int, device: Optional[torch.device] = None
     ) -> list[Any]:
-        dist.broadcast_object_list(object_list, src, group=self.group, device=device)
+        dist.broadcast_object_list(object_list, src, group=self.group, device=device)  # type: ignore[arg-type]
         return object_list
 
     def gather_object(self, obj: Any, object_gather_list: list[Any], dst: int = 0) -> list[Any]:
-        dist.gather_object(obj, object_gather_list, dst, group=self.group)
+        dist.gather_object(obj, object_gather_list, dst, group=self.group)  # type: ignore[arg-type]
         return object_gather_list
 
     def scatter_object_list(
         self, scatter_object_output_list: list[Any], scatter_object_input_list: list[Any], src: int = 0
     ) -> list[Any]:
-        dist.scatter_object_list(scatter_object_output_list, scatter_object_input_list, src, group=self.group)
+        dist.scatter_object_list(scatter_object_output_list, scatter_object_input_list, src, group=self.group)  # type: ignore[arg-type]
         return scatter_object_output_list
 
     @override
     def barrier(self, device_ids: Optional[list[int]] = None) -> None:
         if self.group == dist.GroupMember.NON_GROUP_MEMBER:
             return
-        dist.barrier(group=self.group, device_ids=device_ids)
+        dist.barrier(group=self.group, device_ids=device_ids)  # type: ignore[arg-type]
 
     def monitored_barrier(self, timeout: Optional[datetime.timedelta] = None, wait_all_ranks: bool = False) -> None:
-        dist.monitored_barrier(group=self.group, timeout=timeout, wait_all_ranks=wait_all_ranks)
+        dist.monitored_barrier(group=self.group, timeout=timeout, wait_all_ranks=wait_all_ranks)  # type: ignore[arg-type]
 
     @override
     def setup(self, main_address: Optional[str] = None, main_port: Optional[str] = None, **kwargs: Any) -> Self:

src/lightning/fabric/strategies/xla_fsdp.py

@@ -295,6 +295,7 @@ def clip_gradients_norm(
     ) -> Tensor:
         """Clip gradients by norm."""
         self.precision.unscale_gradients(optimizer)
+        assert callable(module.clip_grad_norm_)
         return module.clip_grad_norm_(max_norm=max_norm, norm_type=norm_type)
 
     @override

src/lightning/fabric/utilities/init.py

@@ -67,7 +67,8 @@ def _materialize(module: Module, device: _DEVICE) -> None:
             f"Materialization requires that the `{type(module).__name__}.reset_parameters` method is implemented."
             " This method is used to initialize any children parameters or buffers in this module."
         )
-    module.reset_parameters()
+    if callable(module.reset_parameters):
+        module.reset_parameters()
 
 
 def _materialize_meta_tensors(module: Module, device: _DEVICE) -> None:

src/lightning/pytorch/callbacks/finetuning.py

@@ -133,7 +133,7 @@ def flatten_modules(modules: Union[Module, Iterable[Union[Module, Iterable]]]) -
 
         if isinstance(modules, Iterable):
             _flatten_modules = []
-            for m in modules:  # type: ignore[union-attr]
+            for m in modules:
                 _flatten_modules.extend(BaseFinetuning.flatten_modules(m))
 
             _modules = iter(_flatten_modules)

src/lightning/pytorch/callbacks/throughput_monitor.py

@@ -140,7 +140,7 @@ def _update(self, trainer: "Trainer", pl_module: "LightningModule", batch: Any,
             # this assumes that all iterations used the same batch size
             samples=iter_num * batch_size,
             lengths=None if self.length_fn is None else self._lengths[stage],
-            flops=flops_per_batch,
+            flops=flops_per_batch,  # type: ignore[arg-type]
         )
 
     def _compute(self, trainer: "Trainer", iter_num: Optional[int] = None) -> None: