support launching Lightning ddp with traditional command

CHANGELOG.md

@@ -146,6 +146,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added `FastForwardSampler` and `CaptureIterableDataset` ([#8307](https://github.com/PyTorchLightning/pytorch-lightning/pull/8307))
 
 
+- Enabled traditional/manual launching of DDP processes through `LOCAL_RANK` and `NODE_RANK` environment variable assignments ([#7480](https://github.com/PyTorchLightning/pytorch-lightning/pull/7480))
+
+
 ### Changed
 
 

pytorch_lightning/plugins/environments/lightning_environment.py

@@ -23,12 +23,16 @@ class LightningEnvironment(ClusterEnvironment):
     """
     The default environment used by Lightning for a single node or free cluster (not managed).
 
-    The master process must be launched by the user and Lightning will spawn new
-    worker processes for distributed training, either in a single node or across multiple nodes.
+    There are two modes the Lightning environment can operate with:
+
+    1.  The user only launches the main process by :code:`python train.py ...` with no additional environment variables
+        set. Lightning will spawn new worker processes for distributed training in the current node.
+    2.  The user launches all processes manually or with utilities like :code:`torch.distributed.launch`.
+        The appropriate environment variables need to be set, and at minimum :code:`LOCAL_RANK`.
 
     If the master address and port are not provided, the default environment will choose them
     automatically. It is recommended to use this default environment for single-node distributed
-    training as it provides the most convenient way to launch the training script.
+    training as it provides a convenient way to launch the training script.
     """
 
     def __init__(self):
@@ -38,7 +42,12 @@ def __init__(self):
         self._world_size: int = 1
 
     def creates_children(self) -> bool:
-        return False
+        """
+        Returns whether the cluster creates the processes or not.
+        If at least :code:`LOCAL_RANK` is available as environment variable, Lightning assumes the user acts as the
+        process launcher/job scheduler and Lightning will not launch new processes.
+        """
+        return "LOCAL_RANK" in os.environ
 
     def master_address(self) -> str:
         return os.environ.get("MASTER_ADDR", "127.0.0.1")

pytorch_lightning/plugins/training_type/ddp.py

@@ -156,7 +156,7 @@ def _is_single_process_single_device(self) -> bool:
 
     def setup_environment(self) -> None:
         # start the other scripts
-        if not self.cluster_environment.creates_children() and os.environ.get("PL_IN_DDP_SUBPROCESS", "0") != "1":
+        if not self.cluster_environment.creates_children():
             self._call_children_scripts()
 
         # set the task idx
@@ -208,8 +208,6 @@ def _call_children_scripts(self):
         if self.parallel_devices is None:
             raise MisconfigurationException("you selected (distribute_backend = ddp) but did not set Trainer(gpus=?)")
 
-        os.environ["PL_IN_DDP_SUBPROCESS"] = "1"
-
         os.environ["WORLD_SIZE"] = f"{self.num_processes * self.num_nodes}"
 
         self.interactive_ddp_procs = []

pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -452,11 +452,6 @@ def select_training_type_plugin(self) -> TrainingTypePlugin:
             use_ddp_sharded_spawn = self._distrib_type == DistributedType.DDP_SHARDED_SPAWN
             use_ddp_fully_sharded = self._distrib_type == DistributedType.DDP_FULLY_SHARDED
 
-            # TODO: decouple from TE
-            # ddp script mode uses the same flags as TE
-            if os.environ.get("PL_IN_DDP_SUBPROCESS", False):
-                use_torchelastic_ddp = False
-
             if use_tpu_spawn:
                 ddp_plugin_cls = TPUSpawnPlugin
             elif use_ddp_sharded:

tests/plugins/environments/test_lightning_environment.py

@@ -1,6 +1,8 @@
 import os
 from unittest import mock
 
+import pytest
+
 from pytorch_lightning.plugins.environments import LightningEnvironment
 
 
@@ -37,6 +39,24 @@ def test_attributes_from_environment_variables():
     assert env.world_size() == 100
 
 
+@pytest.mark.parametrize(
+    "environ, creates_children", [
+        ({}, False),
+        ({
+            "LOCAL_RANK": "2"
+        }, True),
+        ({
+            "NODE_RANK": "1"
+        }, False),
+    ]
+)
+def test_manual_user_launch(environ, creates_children):
+    """ Test that the environment switches to manual user mode when LOCAL_RANK env variable detected. """
+    with mock.patch.dict(os.environ, environ):
+        env = LightningEnvironment()
+        assert env.creates_children() == creates_children
+
+
 @mock.patch.dict(os.environ, {
     "GROUP_RANK": "1",
 })

tests/special_tests.sh

@@ -78,6 +78,12 @@ if [ $? -eq 0 ]; then
     report+="Ran\ttests/utilities/test_warnings.py\n"
 fi
 
+# test that a user can manually launch individual processes
+args="--trainer.gpus 2 --trainer.accelerator ddp --trainer.fast_dev_run=True
+MASTER_ADDR="localhost" MASTER_PORT=1234 LOCAL_RANK=1 python pl_examples/basic_examples/simple_image_classifier.py ${args} &
+MASTER_ADDR="localhost" MASTER_PORT=1234 LOCAL_RANK=0 python pl_examples/basic_examples/simple_image_classifier.py ${args}
+report+="Ran\tmanual ddp launch test\n"
+
 # echo test report
 printf '=%.s' {1..80}
 printf "\n$report"