Remove the deprecated auto_select_gpus Trainer argument (#16184)

Co-authored-by: Carlos Mocholí <[email protected]>

Author: 2 people

src/pytorch_lightning/CHANGELOG.md

@@ -50,6 +50,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Removed the deprecated `resume_from_checkpoint` Trainer argument ([#16167](https://github.com/Lightning-AI/lightning/pull/16167))
 
+- Removed the deprecated automatic GPU selection ([#16184](https://github.com/Lightning-AI/lightning/pull/16184))
+  * Removed the `Trainer(auto_select_gpus=...)` argument
+  * Removed the `pytorch_lightning.tuner.auto_gpu_select.{pick_single_gpu,pick_multiple_gpus}` functions
+
 
 ### Fixed
 

src/pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -75,10 +75,9 @@
     TPUSpawnStrategy,
 )
 from pytorch_lightning.strategies.ddp_spawn import _DDP_FORK_ALIASES
-from pytorch_lightning.tuner.auto_gpu_select import pick_multiple_gpus
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.imports import _IPU_AVAILABLE
-from pytorch_lightning.utilities.rank_zero import rank_zero_deprecation, rank_zero_info, rank_zero_warn
+from pytorch_lightning.utilities.rank_zero import rank_zero_info, rank_zero_warn
 
 log = logging.getLogger(__name__)
 
@@ -101,7 +100,6 @@ def __init__(
         benchmark: Optional[bool] = None,
         replace_sampler_ddp: bool = True,
         deterministic: Optional[Union[bool, _LITERAL_WARN]] = False,
-        auto_select_gpus: Optional[bool] = None,  # TODO: Remove in v2.0.0
     ) -> None:
         """The AcceleratorConnector parses several Trainer arguments and instantiates the Strategy including other
         components such as the Accelerator and Precision plugins.
@@ -163,7 +161,6 @@ def __init__(
         self._parallel_devices: List[Union[int, torch.device, str]] = []
         self._layer_sync: Optional[LayerSync] = NativeSyncBatchNorm() if sync_batchnorm else None
         self.checkpoint_io: Optional[CheckpointIO] = None
-        self._auto_select_gpus: Optional[bool] = auto_select_gpus  # TODO: Remove in v2.0.0
 
         self._check_config_and_set_final_flags(
             strategy=strategy,
@@ -442,7 +439,6 @@ def _set_parallel_devices_and_init_accelerator(self) -> None:
             )
 
         self._set_devices_flag_if_auto_passed()
-        self._set_devices_flag_if_auto_select_gpus_passed()
         self._devices_flag = accelerator_cls.parse_devices(self._devices_flag)
         if not self._parallel_devices:
             self._parallel_devices = accelerator_cls.get_parallel_devices(self._devices_flag)
@@ -451,24 +447,6 @@ def _set_devices_flag_if_auto_passed(self) -> None:
         if self._devices_flag == "auto" or self._devices_flag is None:
             self._devices_flag = self.accelerator.auto_device_count()
 
-    def _set_devices_flag_if_auto_select_gpus_passed(self) -> None:
-        if self._auto_select_gpus is not None:
-            rank_zero_deprecation(
-                "The Trainer argument `auto_select_gpus` has been deprecated in v1.9.0 and will be removed in v2.0.0."
-                " Please use the function `pytorch_lightning.accelerators.find_usable_cuda_devices` instead."
-            )
-        if (
-            self._auto_select_gpus
-            and isinstance(self._devices_flag, int)
-            and isinstance(self.accelerator, CUDAAccelerator)
-        ):
-            self._devices_flag = pick_multiple_gpus(
-                self._devices_flag,
-                # we already show a deprecation message when user sets Trainer(auto_select_gpus=...)
-                _show_deprecation=False,
-            )
-            log.info(f"Auto select gpus: {self._devices_flag}")
-
     def _choose_and_init_cluster_environment(self) -> ClusterEnvironment:
         if isinstance(self._cluster_environment_flag, ClusterEnvironment):
             return self._cluster_environment_flag

src/pytorch_lightning/trainer/trainer.py

@@ -125,7 +125,6 @@ def __init__(
         gradient_clip_algorithm: Optional[str] = None,
         num_nodes: int = 1,
         devices: Optional[Union[List[int], str, int]] = None,
-        auto_select_gpus: Optional[bool] = None,  # TODO: Remove in 2.0
         enable_progress_bar: bool = True,
         overfit_batches: Union[int, float] = 0.0,
         track_grad_norm: Union[int, float, str] = -1,
@@ -368,7 +367,6 @@ def __init__(
             benchmark=benchmark,
             replace_sampler_ddp=replace_sampler_ddp,
             deterministic=deterministic,
-            auto_select_gpus=auto_select_gpus,
             precision=precision,
             plugins=plugins,
         )

src/pytorch_lightning/tuner/auto_gpu_select.py

@@ -115,7 +115,6 @@ def reset_deterministic_algorithm():
 def mock_cuda_count(monkeypatch, n: int) -> None:
     monkeypatch.setattr(lightning_fabric.accelerators.cuda, "num_cuda_devices", lambda: n)
     monkeypatch.setattr(pytorch_lightning.accelerators.cuda, "num_cuda_devices", lambda: n)
-    monkeypatch.setattr(pytorch_lightning.tuner.auto_gpu_select, "num_cuda_devices", lambda: n)
 
 
 @pytest.fixture(scope="function")

tests/tests_pytorch/conftest.py

@@ -21,7 +21,6 @@
 from torch.utils.data import DataLoader
 
 import pytorch_lightning.profiler as profiler
-from lightning_fabric.utilities.exceptions import MisconfigurationException
 from pytorch_lightning import Callback, Trainer
 from pytorch_lightning.accelerators.cpu import CPUAccelerator
 from pytorch_lightning.cli import LightningCLI
@@ -34,7 +33,6 @@
 from pytorch_lightning.strategies.bagua import LightningBaguaModule
 from pytorch_lightning.strategies.utils import on_colab_kaggle
 from pytorch_lightning.trainer.states import RunningStage, TrainerFn
-from pytorch_lightning.tuner.auto_gpu_select import pick_multiple_gpus, pick_single_gpu
 from pytorch_lightning.utilities.apply_func import (
     apply_to_collection,
     apply_to_collections,
@@ -586,23 +584,3 @@ def test_profiler_classes_deprecated_warning(cls):
         f" Use .*profilers.{cls.__name__}` class instead."
     ):
         cls()
-
-
-def test_auto_select_gpus():
-    with pytest.deprecated_call(match="The Trainer argument `auto_select_gpus` has been deprecated in v1.9.0"):
-        Trainer(auto_select_gpus=False)
-
-
-def test_pick_multiple_gpus():
-    with pytest.deprecated_call(match="The function `pick_multiple_gpus` has been deprecated in v1.9.0"), pytest.raises(
-        MisconfigurationException
-    ):
-        pick_multiple_gpus(0)
-
-
-@mock.patch("pytorch_lightning.tuner.auto_gpu_select.num_cuda_devices", return_value=0)
-def test_pick_single_gpu(_):
-    with pytest.deprecated_call(match="The function `pick_single_gpu` has been deprecated in v1.9.0"), pytest.raises(
-        RuntimeError
-    ):
-        pick_single_gpu([])

tests/tests_pytorch/deprecated_api/test_remove_2-0.py

@@ -1168,15 +1168,6 @@ def test_invalid_gradient_clip_algo(tmpdir):
         Trainer(default_root_dir=tmpdir, gradient_clip_algorithm="norm2")
 
 
-@RunIf(min_cuda_gpus=1)
-def test_invalid_gpu_choice_with_auto_select_gpus():
-    num_gpus = torch.cuda.device_count()
-    with pytest.raises(MisconfigurationException, match=r".*but your machine only has.*"), pytest.deprecated_call(
-        match="The function `pick_multiple_gpus` has been deprecated in v1.9.0"
-    ):
-        Trainer(accelerator="gpu", devices=num_gpus + 1, auto_select_gpus=True)
-
-
 @pytest.mark.parametrize("limit_val_batches", [0.0, 1, 1.0, 0.5, 5])
 def test_num_sanity_val_steps(tmpdir, limit_val_batches):
     """Test that the number of sanity check batches is clipped to `limit_val_batches`."""