Remove calls to profile model_forward (#12032)

Co-authored-by: ananthsub <[email protected]>

Author: daniellepintz

CHANGELOG.md

@@ -567,6 +567,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Removed `log_text` and `log_image` from the `LightningLoggerBase` API ([#11857](https://github.com/PyTorchLightning/pytorch-lightning/pull/11857))
 
 
+- Removed calls to `profile("model_forward")` in favor of profiling `training_step` ([#12032](https://github.com/PyTorchLightning/pytorch-lightning/pull/12032))
+
 ### Fixed
 
 - Fixed an issue where `HorovodStrategy.teardown()` did not complete gracefully if an exception was thrown during callback setup [#11752](https://github.com/PyTorchLightning/pytorch-lightning/pull/11752)

docs/source/advanced/profiler.rst

@@ -19,7 +19,6 @@ PyTorch Lightning supports profiling standard actions in the training loop out o
 - on_train_epoch_start
 - on_train_epoch_end
 - on_train_batch_start
-- model_forward
 - model_backward
 - on_after_backward
 - optimizer_step
@@ -66,7 +65,6 @@ The profiler's results will be printed at the completion of a training ``trainer
     |  run_training_epoch                              |  6.1558         	|  6.1558         |
     |  run_training_batch                              |  0.0022506      	|  0.015754       |
     |  [LightningModule]BoringModel.optimizer_step     |  0.0017477      	|  0.012234       |
-    |  model_forward                                   |  0.00055868     	|  0.0039108      |
     |  [LightningModule]BoringModel.val_dataloader     |  0.00024388     	|  0.00024388     |
     |  on_train_batch_start                            |  0.00014637     	|  0.0010246      |
     |  [LightningModule]BoringModel.teardown           |  2.15e-06       	|  2.15e-06       |

pytorch_lightning/loops/optimization/manual_loop.py

@@ -107,30 +107,28 @@ def advance(self, batch: Any, batch_idx: int) -> None:  # type: ignore[override]
         assert self.trainer is not None
         lightning_module = self.trainer.lightning_module
 
-        with self.trainer.profiler.profile("model_forward"):
+        step_kwargs = _build_training_step_kwargs(
+            lightning_module, self.trainer.optimizers, batch, batch_idx, opt_idx=None, hiddens=self._hiddens
+        )
 
-            step_kwargs = _build_training_step_kwargs(
-                lightning_module, self.trainer.optimizers, batch, batch_idx, opt_idx=None, hiddens=self._hiddens
-            )
-
-            # manually capture logged metrics
-            training_step_output = self.trainer._call_strategy_hook("training_step", *step_kwargs.values())
-            self.trainer.strategy.post_training_step()
+        # manually capture logged metrics
+        training_step_output = self.trainer._call_strategy_hook("training_step", *step_kwargs.values())
+        self.trainer.strategy.post_training_step()
 
-            del step_kwargs
+        del step_kwargs
 
-            model_output = self.trainer._call_lightning_module_hook("training_step_end", training_step_output)
-            strategy_output = self.trainer._call_strategy_hook("training_step_end", training_step_output)
-            training_step_output = strategy_output if model_output is None else model_output
-            self._hiddens = _extract_hiddens(training_step_output, lightning_module.truncated_bptt_steps)
+        model_output = self.trainer._call_lightning_module_hook("training_step_end", training_step_output)
+        strategy_output = self.trainer._call_strategy_hook("training_step_end", training_step_output)
+        training_step_output = strategy_output if model_output is None else model_output
+        self._hiddens = _extract_hiddens(training_step_output, lightning_module.truncated_bptt_steps)
 
-            result = self.output_result_cls.from_training_step_output(training_step_output)
+        result = self.output_result_cls.from_training_step_output(training_step_output)
 
-            if self.trainer.move_metrics_to_cpu:
-                # hiddens and the training step output are not moved as they are not considered "metrics"
-                # the user might need them on the correct device for an operation in `training_epoch_end`
-                assert self.trainer._results is not None
-                self.trainer._results.cpu()
+        if self.trainer.move_metrics_to_cpu:
+            # hiddens and the training step output are not moved as they are not considered "metrics"
+            # the user might need them on the correct device for an operation in `training_epoch_end`
+            assert self.trainer._results is not None
+            self.trainer._results.cpu()
 
         self._done = True
         self._output = result.asdict()

pytorch_lightning/loops/optimization/optimizer_loop.py

@@ -414,32 +414,30 @@ def _training_step(self, split_batch: Any, batch_idx: int, opt_idx: int) -> Clos
         # give the PL module a result for logging
         lightning_module = self.trainer.lightning_module
 
-        with self.trainer.profiler.profile("model_forward"):
-
-            step_kwargs = _build_training_step_kwargs(
-                lightning_module, self.trainer.optimizers, split_batch, batch_idx, opt_idx, self._hiddens
-            )
+        step_kwargs = _build_training_step_kwargs(
+            lightning_module, self.trainer.optimizers, split_batch, batch_idx, opt_idx, self._hiddens
+        )
 
-            # manually capture logged metrics
-            training_step_output = self.trainer._call_strategy_hook("training_step", *step_kwargs.values())
-            self.trainer.strategy.post_training_step()
+        # manually capture logged metrics
+        training_step_output = self.trainer._call_strategy_hook("training_step", *step_kwargs.values())
+        self.trainer.strategy.post_training_step()
 
-            model_output = self.trainer._call_lightning_module_hook("training_step_end", training_step_output)
-            strategy_output = self.trainer._call_strategy_hook("training_step_end", training_step_output)
-            training_step_output = strategy_output if model_output is None else model_output
+        model_output = self.trainer._call_lightning_module_hook("training_step_end", training_step_output)
+        strategy_output = self.trainer._call_strategy_hook("training_step_end", training_step_output)
+        training_step_output = strategy_output if model_output is None else model_output
 
-            self._hiddens = _extract_hiddens(training_step_output, lightning_module.truncated_bptt_steps)
+        self._hiddens = _extract_hiddens(training_step_output, lightning_module.truncated_bptt_steps)
 
-            result = self.output_result_cls.from_training_step_output(
-                training_step_output, self.trainer.accumulate_grad_batches
-            )
+        result = self.output_result_cls.from_training_step_output(
+            training_step_output, self.trainer.accumulate_grad_batches
+        )
 
-            if self.trainer._terminate_on_nan:
-                check_finite_loss(result.closure_loss)
+        if self.trainer._terminate_on_nan:
+            check_finite_loss(result.closure_loss)
 
-            if self.trainer.move_metrics_to_cpu:
-                # hiddens and the training step output are not moved as they are not considered "metrics"
-                assert self.trainer._results is not None
-                self.trainer._results.cpu()
+        if self.trainer.move_metrics_to_cpu:
+            # hiddens and the training step output are not moved as they are not considered "metrics"
+            assert self.trainer._results is not None
+            self.trainer._results.cpu()
 
         return result