remove legacy plugins (#5950)

* remove legacy plugins

* imports

* formatting

* fix docs references

* fix cluster environment inheritance

Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>

.yapfignore

@@ -1,5 +1 @@
 .git/*
-
-
-# TODO
-pytorch_lightning/plugins/legacy/*

docs/source/advanced/multi_gpu.rst

@@ -580,9 +580,9 @@ Below are the possible configurations we support.
 
 Implement Your Own Distributed (DDP) training
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-If you need your own way to init PyTorch DDP you can override :meth:`pytorch_lightning.plugins.legacy.ddp_plugin.DDPPlugin.init_ddp_connection`.
+If you need your own way to init PyTorch DDP you can override :meth:`pytorch_lightning.plugins.training_type.ddp.DDPPlugin.init_ddp_connection`.
 
-If you also need to use your own DDP implementation, override :meth:`pytorch_lightning.plugins.legacy.ddp_plugin.DDPPlugin.configure_ddp`.
+If you also need to use your own DDP implementation, override :meth:`pytorch_lightning.plugins.training_type.ddp.DDPPlugin.configure_ddp`.
 
 
 ----------
@@ -679,7 +679,7 @@ In addition, we use Gradient Checkpointing to reduce GPU memory requirements fur
 
 Reference: https://arxiv.org/abs/1811.06965
 
-.. note:: DDPSequentialPlugin is currently supported only for Pytorch 1.6.
+.. note:: RPCSequentialPlugin is currently supported only for Pytorch 1.6.
 
 To get started, install FairScale using the command below. We install a specific branch which contains PyTorch related fixes for Sequential Parallelism.
 
@@ -692,7 +692,7 @@ This should be kept within the ``sequential_module`` variable within your ``Ligh
 
 .. code-block:: python
 
-    from pytorch_lightning.plugins.legacy.ddp_sequential_plugin import DDPSequentialPlugin
+    from pytorch_lightning.plugins.training_type.rpc_sequential import RPCSequentialPlugin
     from pytorch_lightning import LightningModule
 
     class MyModel(LightningModule):
@@ -702,7 +702,7 @@ This should be kept within the ``sequential_module`` variable within your ``Ligh
 
     # Split my module across 4 gpus, one layer each
     model = MyModel()
-    plugin = DDPSequentialPlugin(balance=[1, 1, 1, 1])
+    plugin = RPCSequentialPlugin(balance=[1, 1, 1, 1])
     trainer = Trainer(accelerator='ddp', gpus=4, plugins=[plugin])
     trainer.fit(model)
 

pl_examples/basic_examples/conv_sequential_example.py

@@ -18,7 +18,7 @@
 to balance across your GPUs.
 
 To run:
-python conv_model_sequential_example.py --accelerator ddp --gpus 4 --max_epochs 1  --batch_size 256 --use_ddp_sequential
+python conv_model_sequential_example.py --accelerator ddp --gpus 4 --max_epochs 1  --batch_size 256 --use_rpc_sequential
 """
 import math
 from argparse import ArgumentParser
@@ -32,7 +32,7 @@
 from pl_examples import cli_lightning_logo
 from pytorch_lightning import Trainer
 from pytorch_lightning.metrics.functional import accuracy
-from pytorch_lightning.plugins.legacy.ddp_sequential_plugin import DDPSequentialPlugin
+from pytorch_lightning.plugins import RPCSequentialPlugin
 from pytorch_lightning.utilities import _BOLTS_AVAILABLE, _FAIRSCALE_PIPE_AVAILABLE
 
 if _BOLTS_AVAILABLE:
@@ -201,7 +201,7 @@ def instantiate_datamodule(args):
 if __name__ == "__main__":
     cli_lightning_logo()
     parser = ArgumentParser(description="Pipe Example")
-    parser.add_argument("--use_ddp_sequential", action="store_true")
+    parser.add_argument("--use_rpc_sequential", action="store_true")
     parser = Trainer.add_argparse_args(parser)
     parser = pl_bolts.datamodules.CIFAR10DataModule.add_argparse_args(parser)
     args = parser.parse_args()
@@ -212,8 +212,8 @@ def instantiate_datamodule(args):
     cifar10_dm = instantiate_datamodule(args)
 
     plugins = None
-    if args.use_ddp_sequential:
-        plugins = DDPSequentialPlugin()
+    if args.use_rpc_sequential:
+        plugins = RPCSequentialPlugin()
 
     model = LitResnet(batch_size=args.batch_size, manual_optimization=not args.automatic_optimization)
 
@@ -223,4 +223,4 @@ def instantiate_datamodule(args):
 
     if trainer.accelerator_backend.rpc_enabled:
         # Called at the end of trainer to ensure all processes are killed
-        trainer.accelerator_backend.ddp_plugin.exit_rpc_process()
+        trainer.training_type_plugin.exit_rpc_process()

pytorch_lightning/plugins/environments/cluster_environment.py

@@ -12,10 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from pytorch_lightning.plugins.legacy.plugin import LightningPlugin
 
-
-class ClusterEnvironment(LightningPlugin):
+class ClusterEnvironment:
 
     def __init__(self):
         self._world_size = None