Add inference_mode flag to Trainer

docs/source-pytorch/common/trainer.rst

@@ -1525,6 +1525,39 @@ Whether to enable or disable the model summarization. Defaults to True.
 
     trainer = Trainer(enable_model_summary=True, callbacks=[ModelSummary(max_depth=-1)])
 
+
+eval_inference_mode
+^^^^^^^^^^^^^^^^^^^
+
+Whether to use :meth:`~torch.inference_mode` or :meth:`~torch.no_grad` mode during evaluation
+
+.. testcode::
+
+    # default used by the Trainer
+    trainer = Trainer(eval_inference_mode=True)
+
+    # enables no_grad mode
+    trainer = Trainer(eval_inference_mode=False)
+
+
+With :meth:`~torch.inference_mode` disabled, you can enable the grad of your model layers if required during validat/test/predict.
+
+.. code-block:: python
+
+    class LitModel(LightningModule):
+        def validation_step(self, batch, batch_idx):
+            preds = self.layer1(batch)
+
+            with torch.enable_grad():
+                grad_preds = preds.requires_grad_()
+                preds2 = self.layer2(batch)
+
+
+    model = LitModel()
+    trainer = Trainer(eval_inference_mode=False)
+    trainer.validate(model)
+
+
 -----
 
 Trainer class API

src/pytorch_lightning/CHANGELOG.md

@@ -67,7 +67,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added a more descriptive error message when attempting to fork processes with pre-initialized CUDA context ([#14709](https://github.com/Lightning-AI/lightning/issues/14709))
 
 
-- Added support for custom parameters in subclasses of `SaveConfigCallback` ([#14998](https://github.com/Lightning-AI/lightning/pull/14998)
+- Added support for custom parameters in subclasses of `SaveConfigCallback` ([#14998](https://github.com/Lightning-AI/lightning/pull/14998))
+
+
+- Added `eval_inference_mode` flag to Trainer to let users enable/disable inference mode during evaluation ([#15034](https://github.com/Lightning-AI/lightning/pull/15034))
 
 
 ### Changed

src/pytorch_lightning/trainer/trainer.py

@@ -162,6 +162,7 @@ def __init__(
         amp_level: Optional[str] = None,
         move_metrics_to_cpu: bool = False,
         multiple_trainloader_mode: str = "max_size_cycle",
+        eval_inference_mode: bool = True,
     ) -> None:
         r"""
         Customize every aspect of training via flags.
@@ -388,6 +389,9 @@ def __init__(
                 and smaller datasets reload when running out of their data. In 'min_size' mode, all the datasets
                 reload when reaching the minimum length of datasets.
                 Default: ``"max_size_cycle"``.
+
+            eval_inference_mode: Control whether to use inference mode or no grad mode during
+                evaluation (validate/test/predict).
         """
         super().__init__()
         Trainer._log_api_event("init")
@@ -487,6 +491,8 @@ def __init__(
         )
         self.track_grad_norm: float = float(track_grad_norm)
 
+        self._eval_inference_mode: str = eval_inference_mode
+
         self._detect_anomaly: bool = detect_anomaly
         self._setup_on_init()
 
@@ -1169,7 +1175,9 @@ def _run_evaluate(self) -> _EVALUATE_OUTPUT:
         # reset trainer on this loop and all child loops in case user connected a custom loop
         self._evaluation_loop.trainer = self
 
-        with self.profiler.profile(f"run_{self.state.stage}_evaluation"), _evaluation_context(self.accelerator):
+        with self.profiler.profile(f"run_{self.state.stage}_evaluation"), _evaluation_context(
+            self.accelerator, self._eval_inference_mode
+        ):
             eval_loop_results = self._evaluation_loop.run()
 
         # remove the tensors from the eval results
@@ -1185,7 +1193,7 @@ def _run_predict(self) -> Optional[_PREDICT_OUTPUT]:
         self.reset_predict_dataloader(self.lightning_module)
         # reset trainer on this loop and all child loops in case user connected a custom loop
         self.predict_loop.trainer = self
-        with _evaluation_context(self.accelerator):
+        with _evaluation_context(self.accelerator, self._eval_inference_mode):
             return self.predict_loop.run()
 
     def _run_sanity_check(self) -> None:
@@ -2228,15 +2236,20 @@ def configure_optimizers(self):
 
 
 @contextmanager
-def _evaluation_context(accelerator: Accelerator) -> Generator:
+def _evaluation_context(accelerator: Accelerator, eval_inference_mode: bool = False) -> Generator:
     # inference mode is not supported with gloo backend (#9431),
     # and HPU & TPU accelerators.
+
     context_manager_class = (
         torch.inference_mode
         if not (dist.is_available() and dist.is_initialized() and dist.get_backend() == "gloo")
         and not isinstance(accelerator, HPUAccelerator)
         and not isinstance(accelerator, TPUAccelerator)
         else torch.no_grad
     )
+
+    if not eval_inference_mode:
+        context_manager_class = torch.no_grad
+
     with context_manager_class():
         yield

tests/tests_pytorch/trainer/flags/test_eval_inference_mode.py

@@ -0,0 +1,39 @@
+# Copyright The PyTorch Lightning 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.
+
+import torch
+
+from pytorch_lightning import Trainer
+from pytorch_lightning.demos.boring_classes import BoringModel
+
+
+def test_eval_inference_mode():
+    """Testing overwriting trainer arguments."""
+
+    class BoringModelNoGrad(BoringModel):
+        def on_test_epoch_start(self) -> None:
+            assert not torch.is_grad_enabled()
+            assert not torch.is_inference_mode_enabled()
+            return super().on_test_epoch_start()
+
+    class BoringModelForInferenceMode(BoringModel):
+        def on_test_epoch_start(self) -> None:
+            assert not torch.is_grad_enabled()
+            assert torch.is_inference_mode_enabled()
+            return super().on_test_epoch_start()
+
+    trainer = Trainer(logger=False, eval_inference_mode=False, fast_dev_run=True)
+    trainer.test(BoringModelNoGrad())
+    trainer = Trainer(logger=False, eval_inference_mode=True, fast_dev_run=True)
+    trainer.test(BoringModelForInferenceMode())