Split Precision.init_context (#18734)

src/lightning/fabric/CHANGELOG.md

@@ -49,7 +49,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added support for meta-device initialization with `Fabric.init_module(empty_init=True)` in FSDP ([#18122](https://github.com/Lightning-AI/lightning/pull/18122))
 
 
-- Added `lightning.fabric.plugins.Precision.init_context()` and `lightning.fabric.strategies.Strategy.module_init_context()` context managers to control model and tensor instantiation ([#17462](https://github.com/Lightning-AI/lightning/pull/17462))
+- Added `lightning.fabric.plugins.Precision.module_init_context()` and `lightning.fabric.strategies.Strategy.module_init_context()` context managers to control model and tensor instantiation ([#17462](https://github.com/Lightning-AI/lightning/pull/17462))
 
 
 - `lightning.fabric.strategies.Strategy.tensor_init_context()` context manager to instantiate tensors efficiently directly on device and dtype ([#17607](https://github.com/Lightning-AI/lightning/pull/17607))

src/lightning/fabric/plugins/precision/bitsandbytes.py

@@ -116,7 +116,10 @@ def convert_module(self, module: torch.nn.Module) -> torch.nn.Module:
                 m.compute_type_is_set = False
         return module
 
-    def init_context(self) -> ContextManager:
+    def tensor_init_context(self) -> ContextManager:
+        return _DtypeContextManager(self.dtype)
+
+    def module_init_context(self) -> ContextManager:
         if self.ignore_modules:
             # cannot patch the Linear class if the user wants to skip some submodules
             raise RuntimeError(
@@ -125,7 +128,7 @@ def init_context(self) -> ContextManager:
                 " may initialize the layers on-device, defeating the purpose of quantization. You can remove"
                 " `ignore_modules` or remove the `init_module` context manager."
             )
-        dtype_ctx = _DtypeContextManager(self.dtype)
+        dtype_ctx = self.tensor_init_context()
         # TODO: this could also support replacing `Embedding` and `Conv1D`
         context_manager = _ClassReplacementContextManager({"torch.nn.Linear": self._linear_cls})
         stack = ExitStack()

src/lightning/fabric/plugins/precision/deepspeed.py

@@ -66,11 +66,14 @@ def convert_module(self, module: Module) -> Module:
             return module.to(dtype=self._desired_dtype)
         return module
 
-    def init_context(self) -> ContextManager:
+    def tensor_init_context(self) -> ContextManager:
         if "true" not in self.precision:
             return nullcontext()
         return _DtypeContextManager(self._desired_dtype)
 
+    def module_init_context(self) -> ContextManager:
+        return self.tensor_init_context()
+
     def convert_input(self, data: Any) -> Any:
         return apply_to_collection(data, function=_convert_fp_tensor, dtype=Tensor, dst_type=self._desired_dtype)
 

src/lightning/fabric/plugins/precision/double.py

@@ -30,11 +30,14 @@ class DoublePrecision(Precision):
     def convert_module(self, module: Module) -> Module:
         return module.double()
 
-    def init_context(self) -> ContextManager:
+    def tensor_init_context(self) -> ContextManager:
         return _DtypeContextManager(torch.double)
 
+    def module_init_context(self) -> ContextManager:
+        return self.tensor_init_context()
+
     def forward_context(self) -> ContextManager:
-        return _DtypeContextManager(torch.double)
+        return self.tensor_init_context()
 
     def convert_input(self, data: Any) -> Any:
         return apply_to_collection(data, function=_convert_fp_tensor, dtype=Tensor, dst_type=torch.double)

src/lightning/fabric/plugins/precision/fsdp.py

@@ -103,13 +103,16 @@ def mixed_precision_config(self) -> "TorchMixedPrecision":
             buffer_dtype=buffer_dtype,
         )
 
-    def init_context(self) -> ContextManager:
+    def tensor_init_context(self) -> ContextManager:
+        return _DtypeContextManager(self._desired_input_dtype)
+
+    def module_init_context(self) -> ContextManager:
         return _DtypeContextManager(self.mixed_precision_config.param_dtype or torch.float32)
 
     def forward_context(self) -> ContextManager:
         if "mixed" in self.precision:
             return torch.autocast("cuda", dtype=(torch.bfloat16 if self.precision == "bf16-mixed" else torch.float16))
-        return _DtypeContextManager(self._desired_input_dtype)
+        return self.tensor_init_context()
 
     def convert_input(self, data: Any) -> Any:
         return apply_to_collection(data, function=_convert_fp_tensor, dtype=Tensor, dst_type=self._desired_input_dtype)

src/lightning/fabric/plugins/precision/half.py

@@ -39,11 +39,14 @@ def __init__(self, precision: Literal["bf16-true", "16-true"] = "16-true") -> No
     def convert_module(self, module: Module) -> Module:
         return module.to(dtype=self._desired_input_dtype)
 
-    def init_context(self) -> ContextManager:
+    def tensor_init_context(self) -> ContextManager:
         return _DtypeContextManager(self._desired_input_dtype)
 
+    def module_init_context(self) -> ContextManager:
+        return self.tensor_init_context()
+
     def forward_context(self) -> ContextManager:
-        return _DtypeContextManager(self._desired_input_dtype)
+        return self.tensor_init_context()
 
     def convert_input(self, data: Any) -> Any:
         return apply_to_collection(data, function=_convert_fp_tensor, dtype=Tensor, dst_type=self._desired_input_dtype)

src/lightning/fabric/plugins/precision/precision.py

@@ -53,7 +53,11 @@ def convert_module(self, module: Module) -> Module:
         """
         return module
 
-    def init_context(self) -> ContextManager:
+    def tensor_init_context(self) -> ContextManager:
+        """Controls how tensors get created (device, dtype)."""
+        return nullcontext()
+
+    def module_init_context(self) -> ContextManager:
         """Instantiate module parameters or tensors in the precision type this plugin handles.
 
         This is optional and depends on the precision limitations during optimization.

src/lightning/fabric/plugins/precision/transformer_engine.py

@@ -93,8 +93,11 @@ def convert_module(self, module: torch.nn.Module) -> torch.nn.Module:
         module = module.to(dtype=self.dtype)
         return module
 
-    def init_context(self) -> ContextManager:
-        dtype_ctx = _DtypeContextManager(self.dtype)
+    def tensor_init_context(self) -> ContextManager:
+        return _DtypeContextManager(self.dtype)
+
+    def module_init_context(self) -> ContextManager:
+        dtype_ctx = self.tensor_init_context()
         stack = ExitStack()
         if self.replace_layers:
             import transformer_engine.pytorch as te

src/lightning/fabric/strategies/fsdp.py

@@ -329,16 +329,17 @@ def module_to_device(self, module: Module) -> None:
         pass
 
     def module_init_context(self, empty_init: Optional[bool] = None) -> ContextManager:
-        precision_init_ctx = self.precision.init_context()
+        precision_init_ctx = self.precision.module_init_context()
         module_sharded_ctx = self.module_sharded_context()
+        empty_ctx = _EmptyInit(enabled=bool(empty_init))
         stack = ExitStack()
         if _TORCH_GREATER_EQUAL_2_1 and empty_init:
             # Materialization happens in `setup`. When modules get wrapped by FSDP, the sequence of operations is:
             # 1) materialize module 2) call `reset_parameters()` 3) shard the module.
             # These operations are applied to each submodule 'bottom up' in the module hierarchy.
             stack.enter_context(torch.device("meta"))
         elif _TORCH_GREATER_EQUAL_1_13:
-            stack.enter_context(_EmptyInit(enabled=bool(empty_init)))
+            stack.enter_context(empty_ctx)
         stack.enter_context(precision_init_ctx)
         stack.enter_context(module_sharded_ctx)
         return stack

src/lightning/fabric/strategies/strategy.py

@@ -120,7 +120,7 @@ def process_dataloader(self, dataloader: DataLoader) -> DataLoader:
 
     def tensor_init_context(self) -> ContextManager:
         """Controls how tensors get created (device, dtype)."""
-        precision_init_ctx = self.precision.init_context()
+        precision_init_ctx = self.precision.tensor_init_context()
         stack = ExitStack()
         if _TORCH_GREATER_EQUAL_2_0:
             stack.enter_context(self.root_device)

src/lightning/fabric/strategies/xla_fsdp.py

@@ -194,7 +194,7 @@ def module_to_device(self, module: Module) -> None:
         pass
 
     def module_init_context(self, empty_init: Optional[bool] = None) -> ContextManager:
-        precision_init_ctx = self.precision.init_context()
+        precision_init_ctx = self.precision.module_init_context()
         module_sharded_ctx = self.module_sharded_context()
         stack = ExitStack()
         if _TORCH_GREATER_EQUAL_1_13:

src/lightning/pytorch/CHANGELOG.md

@@ -83,7 +83,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added support for meta-device initialization with `Trainer.init_module(empty_init=True)` in FSDP ([#18385](https://github.com/Lightning-AI/lightning/pull/18385))
 
 
-- Added `lightning.pytorch.plugins.PrecisionPlugin.init_context()` and `lightning.pytorch.strategies.Strategy.tensor_init_context()` context managers to control model and tensor instantiation ([#18004](https://github.com/Lightning-AI/lightning/pull/18004))
+- Added `lightning.pytorch.plugins.PrecisionPlugin.module_init_context()` and `lightning.pytorch.strategies.Strategy.tensor_init_context()` context managers to control model and tensor instantiation ([#18004](https://github.com/Lightning-AI/lightning/pull/18004))
 
 
 - Automatically call `xla_model.mark_step()` before saving checkpoints with XLA ([#17882](https://github.com/Lightning-AI/lightning/pull/17882))

src/lightning/pytorch/plugins/precision/deepspeed.py

@@ -11,8 +11,8 @@
 # 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.
-from contextlib import contextmanager
-from typing import TYPE_CHECKING, Any, Callable, Generator, Optional, Union
+from contextlib import nullcontext
+from typing import TYPE_CHECKING, Any, Callable, ContextManager, Optional, Union
 
 import torch
 from lightning_utilities import apply_to_collection
@@ -23,7 +23,7 @@
 
 import lightning.pytorch as pl
 from lightning.fabric.plugins.precision.deepspeed import _PRECISION_INPUT
-from lightning.fabric.plugins.precision.utils import _convert_fp_tensor
+from lightning.fabric.plugins.precision.utils import _convert_fp_tensor, _DtypeContextManager
 from lightning.fabric.utilities.types import Steppable
 from lightning.pytorch.plugins.precision.precision_plugin import PrecisionPlugin
 from lightning.pytorch.utilities import GradClipAlgorithmType
@@ -77,18 +77,13 @@ def convert_module(self, module: Module) -> Module:
     def convert_input(self, data: Any) -> Any:
         return apply_to_collection(data, function=_convert_fp_tensor, dtype=Tensor, dst_type=self._desired_dtype)
 
-    @contextmanager
-    def init_context(self) -> Generator[None, None, None]:
+    def tensor_init_context(self) -> ContextManager:
         if "true" not in self.precision:
-            yield
-            return
-
-        default_dtype = torch.get_default_dtype()
-        torch.set_default_dtype(self._desired_dtype)
-        try:
-            yield
-        finally:
-            torch.set_default_dtype(default_dtype)
+            return nullcontext()
+        return _DtypeContextManager(self._desired_dtype)
+
+    def module_init_context(self) -> ContextManager:
+        return self.tensor_init_context()
 
     def backward(  # type: ignore[override]
         self,

src/lightning/pytorch/plugins/precision/double.py

@@ -12,15 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from contextlib import contextmanager
-from typing import Any, Generator, Literal
+from typing import Any, ContextManager, Generator, Literal
 
 import torch
 import torch.nn as nn
 from lightning_utilities.core.apply_func import apply_to_collection
 from torch import Tensor
 
 import lightning.pytorch as pl
-from lightning.fabric.plugins.precision.utils import _convert_fp_tensor
+from lightning.fabric.plugins.precision.utils import _convert_fp_tensor, _DtypeContextManager
 from lightning.fabric.utilities.device_dtype_mixin import _DeviceDtypeModuleMixin
 from lightning.pytorch.plugins.precision.precision_plugin import PrecisionPlugin
 from lightning.pytorch.utilities.rank_zero import rank_zero_deprecation
@@ -34,19 +34,11 @@ class DoublePrecisionPlugin(PrecisionPlugin):
     def convert_module(self, module: nn.Module) -> nn.Module:
         return module.double()
 
-    @contextmanager
-    def init_context(self) -> Generator[None, None, None]:
-        """A context manager to change the default tensor type when initializing module parameters or tensors.
-
-        See: :func:`torch.set_default_dtype`
+    def tensor_init_context(self) -> ContextManager:
+        return _DtypeContextManager(torch.float64)
 
-        """
-        default_dtype = torch.get_default_dtype()
-        torch.set_default_dtype(torch.float64)
-        try:
-            yield
-        finally:
-            torch.set_default_dtype(default_dtype)
+    def module_init_context(self) -> ContextManager:
+        return self.tensor_init_context()
 
     @contextmanager
     def forward_context(self) -> Generator[None, None, None]:
@@ -55,12 +47,8 @@ def forward_context(self) -> Generator[None, None, None]:
         See: :func:`torch.set_default_dtype`
 
         """
-        default_dtype = torch.get_default_dtype()
-        torch.set_default_dtype(torch.float64)
-        try:
+        with self.tensor_init_context():
             yield
-        finally:
-            torch.set_default_dtype(default_dtype)
 
     def convert_input(self, data: Any) -> Any:
         return apply_to_collection(data, function=_convert_fp_tensor, dtype=Tensor, dst_type=torch.double)

src/lightning/pytorch/plugins/precision/fsdp.py

@@ -112,7 +112,10 @@ def mixed_precision_config(self) -> "TorchMixedPrecision":
             buffer_dtype=buffer_dtype,
         )
 
-    def init_context(self) -> ContextManager:
+    def tensor_init_context(self) -> ContextManager:
+        return _DtypeContextManager(self._desired_input_dtype)
+
+    def module_init_context(self) -> ContextManager:
         return _DtypeContextManager(self.mixed_precision_config.param_dtype or torch.float32)
 
     def forward_context(self) -> ContextManager:

src/lightning/pytorch/plugins/precision/half.py

@@ -12,14 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from contextlib import contextmanager
-from typing import Any, Generator, Literal
+from typing import Any, ContextManager, Generator, Literal
 
 import torch
 from lightning_utilities import apply_to_collection
 from torch import Tensor
 from torch.nn import Module
 
-from lightning.fabric.plugins.precision.utils import _convert_fp_tensor
+from lightning.fabric.plugins.precision.utils import _convert_fp_tensor, _DtypeContextManager
 from lightning.pytorch.plugins.precision.precision_plugin import PrecisionPlugin
 
 
@@ -40,19 +40,11 @@ def __init__(self, precision: Literal["bf16-true", "16-true"] = "16-true") -> No
     def convert_module(self, module: Module) -> Module:
         return module.to(dtype=self._desired_input_dtype)
 
-    @contextmanager
-    def init_context(self) -> Generator[None, None, None]:
-        """A context manager to change the default tensor type when initializing module parameters or tensors.
-
-        See: :func:`torch.set_default_dtype`
+    def tensor_init_context(self) -> ContextManager:
+        return _DtypeContextManager(self._desired_input_dtype)
 
-        """
-        default_dtype = torch.get_default_dtype()
-        torch.set_default_dtype(self._desired_input_dtype)
-        try:
-            yield
-        finally:
-            torch.set_default_dtype(default_dtype)
+    def module_init_context(self) -> ContextManager:
+        return self.tensor_init_context()
 
     @contextmanager
     def forward_context(self) -> Generator[None, None, None]:

src/lightning/pytorch/strategies/fsdp.py

@@ -360,7 +360,7 @@ def tensor_init_context(self, empty_init: Optional[bool] = None) -> Generator[No
             empty_init_context = _EmptyInit(enabled=bool(empty_init))
         else:
             empty_init_context = nullcontext()
-        with empty_init_context, self.precision_plugin.init_context():
+        with empty_init_context, self.precision_plugin.tensor_init_context():
             yield
 
     @contextmanager

src/lightning/pytorch/strategies/strategy.py

@@ -502,7 +502,7 @@ def tensor_init_context(self, empty_init: Optional[bool] = None) -> Generator[No
         """
         device_context = self.root_device if _TORCH_GREATER_EQUAL_2_0 else nullcontext()
         empty_init_context = _EmptyInit(enabled=bool(empty_init)) if _TORCH_GREATER_EQUAL_1_13 else nullcontext()
-        with empty_init_context, device_context, self.precision_plugin.init_context():
+        with empty_init_context, device_context, self.precision_plugin.tensor_init_context():
             yield
 
     @contextmanager

src/lightning/pytorch/trainer/call.py

@@ -105,7 +105,7 @@ def _call_configure_model(trainer: "pl.Trainer") -> None:
     # we don't normally check for this before calling the hook. it is done here to avoid instantiating the context
     # managers
     if is_overridden("configure_model", trainer.lightning_module):
-        with trainer.strategy.tensor_init_context(), trainer.strategy.model_sharded_context():
+        with trainer.strategy.tensor_init_context(), trainer.strategy.model_sharded_context(), trainer.precision_plugin.module_init_context():  # noqa: E501
             _call_lightning_module_hook(trainer, "configure_model")
 
 

tests/tests_fabric/plugins/precision/test_all.py

@@ -17,9 +17,9 @@ def test_default_dtype_is_restored(precision):
         precision = FSDPPrecision("16-true")
 
     contexts = (
-        (precision.init_context, precision.forward_context)
+        (precision.module_init_context, precision.forward_context)
         if not isinstance(precision, DeepSpeedPrecision)
-        else (precision.init_context,)
+        else (precision.module_init_context,)
     )
     for context in contexts:
         assert torch.get_default_dtype() is torch.float32

tests/tests_fabric/plugins/precision/test_bitsandbytes.py

@@ -46,7 +46,7 @@ def __init__(self, in_features, out_features, bias=True, *_, **__):
 
     # same logic as in `test_default_dtype_is_restored`
     assert torch.get_default_dtype() is torch.float32
-    with pytest.raises(RuntimeError, match="foo"), precision.init_context():
+    with pytest.raises(RuntimeError, match="foo"), precision.module_init_context():
         assert torch.get_default_dtype() is not torch.float32
         raise RuntimeError("foo")
     assert torch.get_default_dtype() is torch.float32
@@ -65,7 +65,7 @@ def __init__(self):
     _NF4Linear = vars(module)["_NF4Linear"]
     _NF4Linear._quantize_weight = Mock()
 
-    with precision.init_context():
+    with precision.module_init_context():
         assert torch.get_default_dtype() == torch.float16
         model = MyModule()
     assert isinstance(model.l1, _NF4Linear)

tests/tests_fabric/plugins/precision/test_deepspeed.py

@@ -79,9 +79,9 @@ def test_selected_dtype(precision, expected_dtype):
         ("16-true", torch.float16),
     ],
 )
-def test_init_context(precision, expected_dtype):
+def test_module_init_context(precision, expected_dtype):
     plugin = DeepSpeedPrecision(precision=precision)
-    with plugin.init_context():
+    with plugin.module_init_context():
         model = torch.nn.Linear(2, 2)
         assert torch.get_default_dtype() == expected_dtype
     assert model.weight.dtype == expected_dtype