address comments

Author: four4fish

pytorch_lightning/accelerators/gpu.py

@@ -82,7 +82,8 @@ def auto_device_count() -> int:
 
     @staticmethod
     def is_available() -> bool:
-        return torch.cuda.device_count() > 0
+        print(torch.cuda.is_available() and torch.cuda.device_count() > 0)
+        return torch.cuda.is_available() and torch.cuda.device_count() > 0
 
 
 def get_nvidia_gpu_stats(device: _DEVICE) -> dict[str, float]:

pytorch_lightning/strategies/tpu_spawn.py

@@ -52,7 +52,7 @@
 class TPUSpawnStrategy(DDPSpawnStrategy):
     """Strategy for training multiple TPU devices using the :func:`torch.multiprocessing.spawn` method."""
 
-    strategy_name = "tpu_spawn_strategy"
+    strategy_name = "tpu_spawn"
 
     def __init__(
         self,

pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -451,6 +451,8 @@ def _set_parallel_devices_and_init_accelerator(self) -> None:
         elif self._accelerator_flag == "gpu":
             self.accelerator = GPUAccelerator()
             self._set_devices_flag_if_auto_passed()
+            # TODO add device availablity check for all devices, not only GPU
+            self._check_device_availability()
             if isinstance(self._devices_flag, int) or isinstance(self._devices_flag, str):
                 self._devices_flag = int(self._devices_flag)
                 self._parallel_devices = (
@@ -459,7 +461,7 @@ def _set_parallel_devices_and_init_accelerator(self) -> None:
                     else []
                 )
             else:
-                self._parallel_devices = [torch.device("cuda", i) for i in self._devices_flag]
+                self._parallel_devices = [torch.device("cuda", i) for i in self._devices_flag]  # type: ignore
 
         elif self._accelerator_flag == "cpu":
             self.accelerator = CPUAccelerator()
@@ -479,6 +481,12 @@ def _set_devices_flag_if_auto_passed(self) -> None:
         if self._devices_flag == "auto" or not self._devices_flag:
             self._devices_flag = self.accelerator.auto_device_count()
 
+    def _check_device_availability(self) -> None:
+        if not self.accelerator.is_available():
+            raise MisconfigurationException(
+                f"You requested {self._accelerator_flag}, " f"but {self._accelerator_flag} is not available"
+            )
+
     def _choose_and_init_cluster_environment(self) -> ClusterEnvironment:
         if isinstance(self._cluster_environment_flag, ClusterEnvironment):
             return self._cluster_environment_flag
@@ -514,7 +522,7 @@ def _choose_strategy(self) -> Union[Strategy, str]:
             return DDPStrategy.strategy_name
         if len(self._parallel_devices) <= 1:
             device = (
-                device_parser.determine_root_gpu_device(self._parallel_devices)
+                device_parser.determine_root_gpu_device(self._parallel_devices)  # type: ignore
                 if self._accelerator_flag == "gpu"
                 else "cpu"
             )

pytorch_lightning/utilities/device_parser.py

@@ -19,6 +19,7 @@
 from pytorch_lightning.tuner.auto_gpu_select import pick_multiple_gpus
 from pytorch_lightning.utilities import _TPU_AVAILABLE
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
+from pytorch_lightning.utilities.types import _DEVICE
 
 
 def determine_root_gpu_device(gpus: List[_DEVICE]) -> Optional[_DEVICE]:
@@ -164,7 +165,7 @@ def _sanitize_gpu_ids(gpus: List[int]) -> List[int]:
     for gpu in gpus:
         if gpu not in all_available_gpus:
             raise MisconfigurationException(
-                f"You requested GPUs: {gpus}\n But your machine only has: {all_available_gpus}"
+                f"You requested gpu: {gpus}\n But your machine only has: {all_available_gpus}"
             )
     return gpus
 

tests/accelerators/test_accelerator_connector.py

@@ -453,12 +453,11 @@ def test_accelerator_cpu(mack_gpu_avalible):
     assert trainer._device_type == "cpu"
     assert isinstance(trainer.accelerator, CPUAccelerator)
 
-    with pytest.raises(MisconfigurationException):
+    with pytest.raises(MisconfigurationException, match="You requested gpu"):
         trainer = Trainer(gpus=1)
-    # with pytest.raises(MisconfigurationException):
-    #     trainer = Trainer(accelerator="gpu")
-
-    with pytest.raises(MisconfigurationException, match="You requested GPUs:"):
+    with pytest.raises(MisconfigurationException, match="You requested gpu, but gpu is not available"):
+        trainer = Trainer(accelerator="gpu")
+    with pytest.raises(MisconfigurationException, match="You requested gpu:"):
         trainer = Trainer(accelerator="cpu", gpus=1)
 
 
@@ -470,9 +469,6 @@ def test_accelerator_gpu():
     assert trainer._device_type == "gpu"
     assert isinstance(trainer.accelerator, GPUAccelerator)
 
-    # with pytest.raises(
-    #     MisconfigurationException, match="You passed `accelerator='gpu'`, but you didn't pass `gpus` to `Trainer`"
-    # ):
     trainer = Trainer(accelerator="gpu")
 
     trainer = Trainer(accelerator="auto", gpus=1)

tests/utilities/test_cli.py

@@ -582,8 +582,8 @@ def on_fit_start(self):
     (
         # dict(strategy="ddp_spawn")
         # dict(strategy="ddp")
-        # !! old accl_conn will choose singleDeviceStrategy for both strategy=ddp/ddp_spawn
-        # this test never worked with DDPSpawnStrategy
+        # the previous accl_conn will choose singleDeviceStrategy for both strategy=ddp/ddp_spawn
+        # TODO revisit this test as it never worked with DDP or DDPSpawn
         dict(strategy="single_device"),
         pytest.param({"tpu_cores": 1}, marks=RunIf(tpu=True)),
     ),