Remove AcceleratorConnector.num_nodes (#12107)

Co-authored-by: Aki Nitta <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Author: 3 people

CHANGELOG.md

@@ -601,6 +601,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Removed the `AcceleratorConnector.device_type` property ([#12081](https://github.com/PyTorchLightning/pytorch-lightning/pull/12081))
 
 
+- Removed `AcceleratorConnector.num_nodes` ([#12107](https://github.com/PyTorchLightning/pytorch-lightning/pull/12107))
+
+
 - Removed `AcceleratorConnector.has_ipu` property ([#12111](https://github.com/PyTorchLightning/pytorch-lightning/pull/12111))
 
 

pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -768,10 +768,6 @@ def _lazy_init_strategy(self) -> None:
     def parallel_devices(self) -> List[Union[torch.device, int]]:
         return self._parallel_devices
 
-    @property
-    def num_nodes(self) -> int:
-        return self._num_nodes_flag
-
     @property
     def num_processes(self) -> int:
         return self.devices if self.devices is not None else 1

pytorch_lightning/trainer/trainer.py

@@ -2018,7 +2018,7 @@ def should_rank_save_checkpoint(self) -> bool:
 
     @property
     def num_nodes(self) -> int:
-        return self._accelerator_connector.num_nodes
+        return getattr(self.strategy, "num_nodes", 1)
 
     @property
     def num_processes(self) -> int:

tests/trainer/test_trainer.py

@@ -2059,7 +2059,6 @@ def training_step(self, batch, batch_idx):
         ({"strategy": None}, SingleDeviceStrategy, "single_device", CPUAccelerator, 0),
         ({"strategy": "dp"}, DDPStrategy, "ddp", CPUAccelerator, 0),
         ({"strategy": "ddp"}, DDPStrategy, "ddp", CPUAccelerator, 0),
-        ({"strategy": "ddp", "num_processes": 2}, DDPStrategy, "ddp", CPUAccelerator, 0),
         ({"strategy": "ddp", "num_nodes": 2}, DDPStrategy, "ddp", CPUAccelerator, 0),
         ({"strategy": "ddp2"}, DDPStrategy, "ddp", CPUAccelerator, 0),
         ({"strategy": None, "gpus": 1}, SingleDeviceStrategy, "single_device", GPUAccelerator, 1),
@@ -2097,6 +2096,23 @@ def training_step(self, batch, batch_idx):
             2,
         ),
         ({"strategy": DDPShardedStrategy(), "gpus": 2}, DDPShardedStrategy, "ddp_sharded", GPUAccelerator, 2),
+        ({"strategy": "ddp2", "gpus": 2, "num_nodes": 2}, DDP2Strategy, "ddp2", GPUAccelerator, 2),
+        ({"strategy": "ddp_spawn", "gpus": 2, "num_nodes": 2}, DDPSpawnStrategy, "ddp_spawn", GPUAccelerator, 2),
+        (
+            {"strategy": "ddp_fully_sharded", "gpus": 1, "num_nodes": 2},
+            DDPFullyShardedStrategy,
+            "ddp_fully_sharded",
+            GPUAccelerator,
+            1,
+        ),
+        ({"strategy": "ddp_sharded", "gpus": 2, "num_nodes": 2}, DDPShardedStrategy, "ddp_sharded", GPUAccelerator, 2),
+        (
+            {"strategy": "ddp_sharded_spawn", "gpus": 2, "num_nodes": 2},
+            DDPSpawnShardedStrategy,
+            "ddp_sharded_spawn",
+            GPUAccelerator,
+            2,
+        ),
     ],
 )
 def test_trainer_config_strategy(monkeypatch, trainer_kwargs, strategy_cls, strategy_name, accelerator_cls, num_gpus):
@@ -2110,6 +2126,7 @@ def test_trainer_config_strategy(monkeypatch, trainer_kwargs, strategy_cls, stra
     assert strategy_cls.strategy_name == strategy_name
     assert isinstance(trainer.accelerator, accelerator_cls)
     assert trainer.num_gpus == num_gpus
+    assert trainer.num_nodes == trainer_kwargs.get("num_nodes", 1)
 
 
 @pytest.mark.parametrize(