[bugfix] Prevent a DDP failure using copy (#9239)

CHANGELOG.md

@@ -273,9 +273,13 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed not setting a default value for `max_epochs` if `max_time` was specified on the `Trainer` constructor ([#9072](https://github.com/PyTorchLightning/pytorch-lightning/pull/9072))
 
+
 - Fixed the CometLogger, no longer modifies the metrics in place. Instead creates a copy of metrics before performing any operations ([#9150](https://github.com/PyTorchLightning/pytorch-lightning/pull/9150))
 
 
+- Fixed `DDP` "CUDA error: initialization error" due to a `copy` instead of `deepcopy` on `ResultCollection` ([#9239](https://github.com/PyTorchLightning/pytorch-lightning/pull/9239))
+
+
 ## [1.4.3] - 2021-08-17
 
 - Fixed plateau scheduler stepping on incomplete epoch ([#8861](https://github.com/PyTorchLightning/pytorch-lightning/pull/8861))

pytorch_lightning/loops/batch/training_batch_loop.py

@@ -11,7 +11,7 @@
 # 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 copy import copy
+from copy import deepcopy
 from functools import partial
 from typing import Any, Callable, Dict, List, Optional, Tuple
 
@@ -142,12 +142,12 @@ def advance(self, batch, batch_idx):
 
                 result = self._run_optimization(batch_idx, split_batch, opt_idx, optimizer)
                 if result:
-                    self.batch_outputs[opt_idx].append(copy(result.result_collection))
+                    self.batch_outputs[opt_idx].append(deepcopy(result.result_collection))
         else:
             # in manual optimization, there is no looping over optimizers
             result = self._run_optimization(batch_idx, split_batch)
             if result:
-                self.batch_outputs[0].append(copy(result.result_collection))
+                self.batch_outputs[0].append(deepcopy(result.result_collection))
 
     def teardown(self) -> None:
         # release memory

pytorch_lightning/trainer/connectors/logger_connector/result.py

@@ -17,6 +17,7 @@
 from typing import Any, Callable, Dict, List, Mapping, Optional, Tuple, Union
 
 import torch
+from torch.functional import Tensor
 from torchmetrics import Metric
 
 from pytorch_lightning.core.mixins import DeviceDtypeModuleMixin
@@ -435,8 +436,12 @@ def log(
     ) -> None:
         """See :meth:`~pytorch_lightning.core.lightning.LightningModule.log`"""
         # no metrics should be logged with graphs
-        if not enable_graph and isinstance(value, torch.Tensor):
-            value = value.detach()
+        if not enable_graph:
+
+            def detach_fn(tensor: Tensor) -> Tensor:
+                return tensor.detach()
+
+            value = apply_to_collection(value, Tensor, detach_fn)
 
         # move metrics to cpu on TPU.
         if isinstance(value, torch.Tensor) and value.device.type == "xla":