Set find unused parameters to True by default to fix breaking compatibility

CHANGELOG.md

@@ -44,6 +44,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Changed `setup()` and `teardown()` stage argument to take any of `{fit,validate,test,predict}` ([#6386](https://github.com/PyTorchLightning/pytorch-lightning/pull/6386))
 
 
+- Changed the default of `find_unused_parameters` back to `True` in DDP and DDP Spawn ([#6438](https://github.com/PyTorchLightning/pytorch-lightning/pull/6438))
+
+
 ### Deprecated
 
 

docs/source/benchmarking/performance.rst

@@ -94,6 +94,21 @@ DP performs three GPU transfers for EVERY batch:
 
 Whereas DDP only performs 1 transfer to sync gradients. Because of this, DDP is MUCH faster than DP.
 
+When using DDP set find_unused_parameters=False
+-----------------------------------------------
+
+By default we have enabled find unused parameters to True. This is for compatibility issues that have arisen in the past (see the `discussion <https://github.com/PyTorchLightning/pytorch-lightning/discussions/6219>`_ for more information).
+This by default comes with a performance hit, and can be disabled in most cases.
+
+.. code-block:: python
+
+    from pytorch_lightning.plugins import DDPPlugin
+
+    trainer = pl.Trainer(
+        gpus=2,
+        plugins=DDPPlugin(find_unused_parameters=False),
+    )
+
 ----------
 
 16-bit precision

pytorch_lightning/plugins/training_type/ddp.py

@@ -175,6 +175,13 @@ def set_world_ranks(self):
         self.world_size = self.num_nodes * self.num_processes
 
     def pre_configure_ddp(self):
+        # if unset, default `find_unused_parameters` `True`
+        # Many models require setting this parameter to True, as there are corner cases
+        # when not all parameter backward hooks are fired by the autograd engine even if require_grad is set to True.
+        # This flag does come with a performance hit, so it is suggested to disable in cases where it is possible.
+        self._ddp_kwargs["find_unused_parameters"] = self._ddp_kwargs.get(
+            "find_unused_parameters", True
+        )
         # todo: PyTorch 1.7.0 DDP introduces ``self.reducer._rebuild_buckets()`` breaking manual_optimization
         if _TORCH_GREATER_EQUAL_1_7 and not self.lightning_module.automatic_optimization and not self._ddp_kwargs.get(
             "find_unused_parameters", False

pytorch_lightning/plugins/training_type/ddp_spawn.py

@@ -168,6 +168,13 @@ def post_dispatch(self):
         self.__recover_child_process_weights(best_path, last_path)
 
     def pre_configure_ddp(self):
+        # if unset, default `find_unused_parameters` `True`
+        # Many models require setting this parameter to True, as there are corner cases
+        # when not all parameter backward hooks are fired by the autograd engine even if require_grad is set to True.
+        # This flag does come with a performance hit, so it is suggested to disable in cases where it is possible.
+        self._ddp_kwargs["find_unused_parameters"] = self._ddp_kwargs.get(
+            "find_unused_parameters", True
+        )
         # todo: PyTorch 1.7.0 DDP introduces ``self.reducer._rebuild_buckets()`` breaking manual_optimization
         if _TORCH_GREATER_EQUAL_1_7 and not self.lightning_module.automatic_optimization and not self._ddp_kwargs.get(
             "find_unused_parameters", False