deprecate Trainer.root_gpu and remove AcceleratorConnector.root_gpu

Author: DuYicong515

CHANGELOG.md

@@ -518,6 +518,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Deprecated `ParallelPlugin.torch_distributed_backend` in favor of `DDPStrategy.process_group_backend` property ([#11745](https://github.com/PyTorchLightning/pytorch-lightning/pull/11745))
 
 
+- Deprecated `Trainer.root_gpu` in favor of `Trainer.strategy.root_device.index` when GPU is used. ([#12262](https://github.com/PyTorchLightning/pytorch-lightning/pull/11978))
+
+
 ### Removed
 
 - Removed deprecated parameter `method` in `pytorch_lightning.utilities.model_helpers.is_overridden` ([#10507](https://github.com/PyTorchLightning/pytorch-lightning/pull/10507))
@@ -705,6 +708,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Removed public attribute `sync_batchnorm` from strategies ([#11754](https://github.com/PyTorchLightning/pytorch-lightning/pull/11754))
 
 
+- Removed `AcceleratorConnector.root_gpu` property ([#12262](https://github.com/PyTorchLightning/pytorch-lightning/pull/12262))
+
+
 ### Fixed
 
 - Fixed an issue where `ModelCheckpoint` could delete older checkpoints when `dirpath` has changed during resumed training ([#12045](https://github.com/PyTorchLightning/pytorch-lightning/pull/12045))

pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -788,14 +788,6 @@ def parallel_devices(self) -> List[Union[torch.device, int]]:
     def num_processes(self) -> int:
         return self.devices if self.devices is not None else 1
 
-    @property
-    def root_gpu(self) -> Optional[int]:
-        return (
-            self.strategy.root_device.index
-            if not isinstance(self.accelerator, (IPUAccelerator, TPUAccelerator))
-            else None
-        )
-
     @property
     def devices(self) -> int:
         if isinstance(self.strategy, SingleDeviceStrategy):

pytorch_lightning/trainer/trainer.py

@@ -2016,7 +2016,11 @@ def num_processes(self) -> int:
 
     @property
     def root_gpu(self) -> Optional[int]:
-        return self._accelerator_connector.root_gpu
+        rank_zero_deprecation(
+            "`Trainer.root_gpu` is deprecated in v1.6 and will be removed in v1.8. Please use "
+            "`Trainer.strategy.root_device.index if isinstance(Trainer.accelerator, GPUAccelerator) else None` instead."
+        )
+        return self.strategy.root_device.index if isinstance(self.accelerator, GPUAccelerator) else None
 
     @property
     def tpu_cores(self) -> int:

tests/deprecated_api/test_remove_1-8.py

@@ -878,3 +878,41 @@ def all_gather(self, tensor):
         match="ParallelStrategy.torch_distributed_backend was deprecated" " in v1.6 and will be removed in v1.8."
     ):
         strategy.torch_distributed_backend
+
+
+@pytest.mark.parametrize(
+    ["gpus", "expected_root_gpu", "strategy"],
+    [
+        pytest.param(None, None, "ddp", id="None is None"),
+        pytest.param(0, None, "ddp", id="O gpus, expect gpu root device to be None."),
+        pytest.param(1, 0, "ddp", id="1 gpu, expect gpu root device to be 0."),
+        pytest.param(-1, 0, "ddp", id="-1 - use all gpus, expect gpu root device to be 0."),
+        pytest.param("-1", 0, "ddp", id="'-1' - use all gpus, expect gpu root device to be 0."),
+        pytest.param(3, 0, "ddp", id="3 gpus, expect gpu root device to be 0.(backend:ddp)"),
+    ],
+)
+def test_root_gpu_property(monkeypatch, gpus, expected_root_gpu, strategy):
+    monkeypatch.setattr(torch.cuda, "is_available", lambda: True)
+    monkeypatch.setattr(torch.cuda, "device_count", lambda: 16)
+    with pytest.deprecated_call(
+        match="`Trainer.root_gpu` is deprecated in v1.6 and will be removed in v1.8. Please use "
+        r"`Trainer.strategy.root_device.index if isinstance\(Trainer.accelerator, GPUAccelerator\) else None` instead."
+    ):
+        assert Trainer(gpus=gpus, strategy=strategy).root_gpu == expected_root_gpu
+
+
+@pytest.mark.parametrize(
+    ["gpus", "expected_root_gpu", "strategy"],
+    [
+        pytest.param(None, None, None, id="None is None"),
+        pytest.param(None, None, "ddp", id="None is None"),
+        pytest.param(0, None, "ddp", id="None is None"),
+    ],
+)
+def test_root_gpu_property_0_passing(monkeypatch, gpus, expected_root_gpu, strategy):
+    monkeypatch.setattr(torch.cuda, "device_count", lambda: 0)
+    with pytest.deprecated_call(
+        match="`Trainer.root_gpu` is deprecated in v1.6 and will be removed in v1.8. Please use "
+        r"`Trainer.strategy.root_device.index if isinstance\(Trainer.accelerator, GPUAccelerator\) else None` instead."
+    ):
+        assert Trainer(gpus=gpus, strategy=strategy).root_gpu == expected_root_gpu

tests/models/data/horovod/train_default_model.py

@@ -20,6 +20,7 @@
 import os
 import sys
 
+import pytest
 import torch
 
 # this is needed because Conda does not use `PYTHONPATH` env var while pip and virtualenv do
@@ -102,7 +103,12 @@ def training_epoch_end(self, outputs) -> None:
     if on_gpu:
         trainer = Trainer(gpus=1, strategy="horovod", max_epochs=1)
         # Test the root_gpu property
-        assert trainer.root_gpu == hvd.local_rank()
+        with pytest.deprecated_call(
+            match="`Trainer.root_gpu` is deprecated in v1.6 and will be removed in v1.8. Please use "
+            r"`Trainer.strategy.root_device.index if isinstance\(Trainer.accelerator, GPUAccelerator\) else None`"
+            " instead."
+        ):
+            assert trainer.root_gpu == hvd.local_rank()
 
 
 if __name__ == "__main__":

tests/models/test_gpu.py

@@ -118,33 +118,6 @@ def test_trainer_num_gpu_0(mocked_device_count_0, gpus, expected_num_gpus, strat
     assert Trainer(gpus=gpus, strategy=strategy).num_gpus == expected_num_gpus
 
 
-@pytest.mark.parametrize(
-    ["gpus", "expected_root_gpu", "strategy"],
-    [
-        pytest.param(None, None, "ddp", id="None is None"),
-        pytest.param(0, None, "ddp", id="O gpus, expect gpu root device to be None."),
-        pytest.param(1, 0, "ddp", id="1 gpu, expect gpu root device to be 0."),
-        pytest.param(-1, 0, "ddp", id="-1 - use all gpus, expect gpu root device to be 0."),
-        pytest.param("-1", 0, "ddp", id="'-1' - use all gpus, expect gpu root device to be 0."),
-        pytest.param(3, 0, "ddp", id="3 gpus, expect gpu root device to be 0.(backend:ddp)"),
-    ],
-)
-def test_root_gpu_property(mocked_device_count, gpus, expected_root_gpu, strategy):
-    assert Trainer(gpus=gpus, strategy=strategy).root_gpu == expected_root_gpu
-
-
-@pytest.mark.parametrize(
-    ["gpus", "expected_root_gpu", "strategy"],
-    [
-        pytest.param(None, None, None, id="None is None"),
-        pytest.param(None, None, "ddp", id="None is None"),
-        pytest.param(0, None, "ddp", id="None is None"),
-    ],
-)
-def test_root_gpu_property_0_passing(mocked_device_count_0, gpus, expected_root_gpu, strategy):
-    assert Trainer(gpus=gpus, strategy=strategy).root_gpu == expected_root_gpu
-
-
 # Asking for a gpu when non are available will result in a MisconfigurationException
 @pytest.mark.parametrize(
     ["gpus", "expected_root_gpu", "strategy"],