[bug] Fix Pytorch profiler with emit_nvtx (#6260)

* resolve bug

* update changelog

* Update tests/trainer/test_trainer.py

* Update pytorch_lightning/profiler/profilers.py

Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>

* resolve comments

* resolve flake8

Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>

CHANGELOG.md

@@ -83,6 +83,12 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed DP reduction with collection ([#6324](https://github.com/PyTorchLightning/pytorch-lightning/pull/6324))
 
 
+- Fixed PyTorch Profiler with `emit_nvtx` ([#6260](https://github.com/PyTorchLightning/pytorch-lightning/pull/6260))
+
+
+- Fixed `trainer.test` from `best_path` hangs after calling `trainer.fit`  ([#6272](https://github.com/PyTorchLightning/pytorch-lightning/pull/6272))
+
+
 ## [1.2.2] - 2021-03-02
 
 ### Added

pytorch_lightning/profiler/profilers.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Profiler to check if there are any bottlenecks in your code."""
-
 import cProfile
 import io
 import logging

pytorch_lightning/profiler/pytorch.py

@@ -205,7 +205,7 @@ def start(self, action_name: str) -> None:
 
     def _start(self, action_name: str) -> None:
         if self.emit_nvtx:
-            self._create_profiler(action_name, torch.cuda.profiler.profile, enter=False)
+            self._parent_profiler = self._create_profiler(action_name, torch.cuda.profiler.profile, enter=True)
             self._create_profiler(action_name, torch.autograd.profiler.emit_nvtx)
         else:
             self._create_profiler(action_name, torch.autograd.profiler.profile)
@@ -215,15 +215,24 @@ def _create_profiler(self, action_name, profiler, enter=True):
         profiler_args = {k: v for k, v in vars(self).items() if k in init_args}
         pr = profiler(**profiler_args)
         if enter:
-            pr = pr.__enter__()
+            out_pr = pr.__enter__()
+            if out_pr is not None:
+                pr = out_pr
         self.profiler = pr
+        return self.profiler
 
     def _stop(self, action_name: str) -> None:
         if self.profiler is None:
             return
 
         self.profiler.__exit__(exc_type=None, exc_val=None, exc_tb=None)
 
+        if isinstance(self.profiler, torch.autograd.profiler.emit_nvtx):
+            # when running ``emit_nvtx``, PyTorch requires 2 context manager.
+            # The parent_profiler is being closed too.
+            self._parent_profiler.__exit__(None, None, None)
+            return
+
         function_events = self.profiler.function_events
         self.profiler = None
         for name in self.running_stack:

tests/special_tests.sh

@@ -32,3 +32,4 @@ python ${DEFAULTS} tests/trainer/test_trainer.py::test_pytorch_profiler_trainer_
 python ${DEFAULTS} tests/models/test_hooks.py::test_transfer_batch_hook_ddp
 python ${DEFAULTS} tests/trainer/test_data_loading.py::test_replace_distrubuted_sampler_custom_dataloader_custom_batch_sampler
 python ${DEFAULTS} tests/trainer/optimization/test_manual_optimization.py::test_step_with_optimizer_closure_with_different_frequencies_ddp_with_toggle_model
+nvprof --profile-from-start off -o trace_name.prof -- python ${DEFAULTS} tests/trainer/test_trainer.py::test_pytorch_profiler_nested_emit_nvtx

tests/trainer/test_trainer.py

@@ -1554,6 +1554,22 @@ def test_pytorch_profiler_nested(tmpdir):
         assert pa[n] == expected_[n]
 
 
+@RunIf(min_gpus=1, special=True)
+def test_pytorch_profiler_nested_emit_nvtx(tmpdir):
+    """
+    This test check emit_nvtx is correctly supported
+    """
+    profiler = PyTorchProfiler(use_cuda=True, emit_nvtx=True)
+
+    model = BoringModel()
+    trainer = Trainer(
+        fast_dev_run=True,
+        profiler=profiler,
+        gpus=1,
+    )
+    trainer.fit(model)
+
+
 @pytest.mark.parametrize(
     ["limit_train_batches", "global_step", "num_training_batches", "current_epoch", "should_train"],
     [(0.2, 0, 0, 0, False), (0.5, 10, 2, 4, True)],