Fix wrong behavior of DDPStrategy option with simple GAN training using DDP

examples/pytorch/domain_templates/generative_adversarial_net.py

@@ -29,13 +29,22 @@
 from lightning.pytorch import cli_lightning_logo
 from lightning.pytorch.core import LightningModule
 from lightning.pytorch.demos.mnist_datamodule import MNISTDataModule
+from lightning.pytorch.strategies.ddp import MultiModelDDPStrategy
 from lightning.pytorch.trainer import Trainer
 from lightning.pytorch.utilities.imports import _TORCHVISION_AVAILABLE
 
 if _TORCHVISION_AVAILABLE:
     import torchvision
 
 
+def _block(in_feat: int, out_feat: int, normalize: bool = True) -> list:
+    layers = [nn.Linear(in_feat, out_feat)]
+    if normalize:
+        layers.append(nn.BatchNorm1d(out_feat, 0.8))
+    layers.append(nn.LeakyReLU(0.2, inplace=True))
+    return layers
+
+
 class Generator(nn.Module):
     """
     >>> Generator(img_shape=(1, 8, 8))  # doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
@@ -47,19 +56,11 @@ class Generator(nn.Module):
     def __init__(self, latent_dim: int = 100, img_shape: tuple = (1, 28, 28)):
         super().__init__()
         self.img_shape = img_shape
-
-        def block(in_feat, out_feat, normalize=True):
-            layers = [nn.Linear(in_feat, out_feat)]
-            if normalize:
-                layers.append(nn.BatchNorm1d(out_feat, 0.8))
-            layers.append(nn.LeakyReLU(0.2, inplace=True))
-            return layers
-
         self.model = nn.Sequential(
-            *block(latent_dim, 128, normalize=False),
-            *block(128, 256),
-            *block(256, 512),
-            *block(512, 1024),
+            *_block(latent_dim, 128, normalize=False),
+            *_block(128, 256),
+            *_block(256, 512),
+            *_block(512, 1024),
             nn.Linear(1024, int(math.prod(img_shape))),
             nn.Tanh(),
         )
@@ -209,10 +210,18 @@ def main(args: Namespace) -> None:
     # ------------------------
     # 2 INIT TRAINER
     # ------------------------
-    # If use distributed training  PyTorch recommends to use DistributedDataParallel.
-    # See: https://pytorch.org/docs/stable/nn.html#torch.nn.DataParallel
     dm = MNISTDataModule()
-    trainer = Trainer(accelerator="gpu", devices=1)
+
+    if args.use_ddp:
+        # `MultiModelDDPStrategy` is critical for multi-gpu GAN training
+        # There are two ways to run training codes with existed `DDPStrategy`:
+        # 1) Activate `find_unused_parameters` option
+        # 2) change from self.manual_backward(loss) to loss.backward()
+        # Neither of them is desirable.
+        trainer = Trainer(accelerator="gpu", devices=-1, strategy=MultiModelDDPStrategy())
+    else:
+        # If you want to run on a single GPU, you can use the default strategy.
+        trainer = Trainer(accelerator="gpu", devices=1)
 
     # ------------------------
     # 3 START TRAINING
@@ -229,6 +238,8 @@ def main(args: Namespace) -> None:
     parser.add_argument("--b1", type=float, default=0.5, help="adam: decay of first order momentum of gradient")
     parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of second order momentum of gradient")
     parser.add_argument("--latent_dim", type=int, default=100, help="dimensionality of the latent space")
+    parser.add_argument("--use_ddp", action="store_true", help="distributed strategy to use")
+
     args = parser.parse_args()
 
     main(args)

src/lightning/pytorch/strategies/__init__.py

@@ -15,7 +15,7 @@
 
 from lightning.fabric.strategies.registry import _StrategyRegistry
 from lightning.fabric.utilities.registry import _register_classes
-from lightning.pytorch.strategies.ddp import DDPStrategy
+from lightning.pytorch.strategies.ddp import DDPStrategy, MultiModelDDPStrategy
 from lightning.pytorch.strategies.deepspeed import DeepSpeedStrategy
 from lightning.pytorch.strategies.fsdp import FSDPStrategy
 from lightning.pytorch.strategies.model_parallel import ModelParallelStrategy
@@ -30,6 +30,7 @@
 
 __all__ = [
     "DDPStrategy",
+    "MultiModelDDPStrategy",
     "DeepSpeedStrategy",
     "FSDPStrategy",
     "ModelParallelStrategy",

src/lightning/pytorch/strategies/ddp.py

@@ -214,7 +214,7 @@ def set_world_ranks(self) -> None:
         rank_zero_only.rank = utils_rank_zero_only.rank = self.global_rank
 
     def _register_ddp_hooks(self) -> None:
-        log.debug(f"{self.__class__.__name__}: registering ddp hooks")
+        log.debug(f"{self.__class__.__name__}: registering DDP hooks")
         # currently, DDP communication hooks only work with NCCL backend and SPSD (single process single device) mode
         # https://github.com/pytorch/pytorch/blob/v1.8.0/torch/nn/parallel/distributed.py#L1080-L1084
         if self.root_device.type == "cuda":
@@ -419,6 +419,47 @@ def teardown(self) -> None:
         super().teardown()
 
 
+class MultiModelDDPStrategy(DDPStrategy):
+    """Specific strategy for training on multiple models with multiple optimizers (e.g. GAN training).
+
+    This strategy wraps each individual child module in :class:`~torch.nn.parallel.distributed.DistributedDataParallel`
+    module. Ensures manual backward only updates parameters of the targeted child module, preventing cross-references
+    between modules' parameters.
+
+    """
+
+    @override
+    def _setup_model(self, model: Module) -> DistributedDataParallel:
+        device_ids = self.determine_ddp_device_ids()
+        log.debug(f"setting up DDP model with device ids: {device_ids}, kwargs: {self._ddp_kwargs}")
+        # https://pytorch.org/docs/stable/notes/cuda.html#id5
+        ctx = torch.cuda.stream(torch.cuda.Stream()) if device_ids is not None else nullcontext()
+        with ctx:
+            for name, module in model.named_children():
+                if isinstance(module, Module):
+                    ddp_module = DistributedDataParallel(module, device_ids=device_ids, **self._ddp_kwargs)
+                    setattr(model, name, ddp_module)
+            return model
+
+    @override
+    def _register_ddp_hooks(self) -> None:
+        log.debug(f"{self.__class__.__name__}: registering DDP hooks")
+        # currently, DDP communication hooks only work with NCCL backend and SPSD (single process single device) mode
+        # https://github.com/pytorch/pytorch/blob/v1.8.0/torch/nn/parallel/distributed.py#L1080-L1084
+        if self.root_device.type != "cuda":
+            return
+        assert isinstance(self.model, Module)
+
+        for name, module in self.model.named_children():
+            assert isinstance(module, DistributedDataParallel)
+            _register_ddp_comm_hook(
+                model=module,
+                ddp_comm_state=self._ddp_comm_state,
+                ddp_comm_hook=self._ddp_comm_hook,
+                ddp_comm_wrapper=self._ddp_comm_wrapper,
+            )
+
+
 class _DDPForwardRedirection(_ForwardRedirection):
     @override
     def on_after_inner_forward(self, wrapper_module: Module, original_module: "pl.LightningModule") -> None:

tests/tests_pytorch/strategies/test_multi_model_ddp.py

@@ -0,0 +1,174 @@
+# Copyright The Lightning AI team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+
+import pytest
+import torch
+from torch.multiprocessing import ProcessRaisedException
+
+from lightning.pytorch import Trainer
+from lightning.pytorch.strategies import MultiModelDDPStrategy
+from lightning.pytorch.trainer import seed_everything
+from tests_pytorch.helpers.advanced_models import BasicGAN
+from tests_pytorch.helpers.runif import RunIf
+
+
+@RunIf(min_cuda_gpus=2, standalone=True, sklearn=True)
+def test_multi_gpu_with_multi_model_ddp_fit_only(tmp_path):
+    dm = BasicGAN.train_dataloader()
+    model = BasicGAN()
+    trainer = Trainer(
+        default_root_dir=tmp_path, max_epochs=1, accelerator="gpu", devices=-1, strategy=MultiModelDDPStrategy()
+    )
+    trainer.fit(model, datamodule=dm)
+
+
+@RunIf(min_cuda_gpus=2, standalone=True, sklearn=True)
+def test_multi_gpu_with_multi_model_ddp_predict_only(tmp_path):
+    dm = BasicGAN.train_dataloader()
+    model = BasicGAN()
+    trainer = Trainer(
+        default_root_dir=tmp_path, max_epochs=1, accelerator="gpu", devices=-1, strategy=MultiModelDDPStrategy()
+    )
+    trainer.predict(model, datamodule=dm)
+
+
+@RunIf(min_cuda_gpus=2, standalone=True, sklearn=True)
+def test_multi_gpu_multi_model_ddp_fit_predict(tmp_path):
+    seed_everything(4321)
+    dm = BasicGAN.train_dataloader()
+    model = BasicGAN()
+    trainer = Trainer(
+        default_root_dir=tmp_path, max_epochs=1, accelerator="gpu", devices=-1, strategy=MultiModelDDPStrategy()
+    )
+    trainer.fit(model, datamodule=dm)
+    trainer.predict(model, datamodule=dm)
+
+
+class UnusedParametersBasicGAN(BasicGAN):
+    def __init__(self):
+        super().__init__()
+        mnist_shape = (1, 28, 28)
+        self.intermediate_layer = torch.nn.Linear(mnist_shape[-1], mnist_shape[-1])
+
+    def training_step(self, batch, batch_idx):
+        with torch.no_grad():
+            img = self.intermediate_layer(batch[0])
+            batch[0] = img  # modify the batch to use the intermediate layer result
+        return super().training_step(batch, batch_idx)
+
+
+@RunIf(standalone=True)
+def test_find_unused_parameters_ddp_spawn_raises():
+    """Test that the DDP strategy can change PyTorch's error message so that it's more useful for Lightning users."""
+    trainer = Trainer(
+        accelerator="cpu",
+        devices=1,
+        strategy=MultiModelDDPStrategy(),
+        max_steps=2,
+        logger=False,
+    )
+    with pytest.raises(
+        ProcessRaisedException, match="It looks like your LightningModule has parameters that were not used in"
+    ):
+        trainer.fit(UnusedParametersBasicGAN())
+
+
+@RunIf(standalone=True)
+def test_find_unused_parameters_ddp_exception():
+    """Test that the DDP strategy can change PyTorch's error message so that it's more useful for Lightning users."""
+    trainer = Trainer(
+        accelerator="cpu",
+        devices=1,
+        strategy=MultiModelDDPStrategy(),
+        max_steps=2,
+        logger=False,
+    )
+    with pytest.raises(RuntimeError, match="It looks like your LightningModule has parameters that were not used in"):
+        trainer.fit(UnusedParametersBasicGAN())
+
+
+class CheckOptimizerDeviceModel(BasicGAN):
+    def configure_optimizers(self):
+        assert all(param.device.type == "cuda" for param in self.parameters())
+        super().configure_optimizers()
+
+
+@RunIf(min_cuda_gpus=1)
+def test_model_parameters_on_device_for_optimizer():
+    """Test that the strategy has moved the parameters to the device by the time the optimizer gets created."""
+    model = CheckOptimizerDeviceModel()
+    trainer = Trainer(
+        default_root_dir=os.getcwd(),
+        fast_dev_run=1,
+        accelerator="gpu",
+        devices=1,
+        strategy=MultiModelDDPStrategy(),
+    )
+    trainer.fit(model)
+
+
+class BasicGANCPU(BasicGAN):
+    def on_train_start(self) -> None:
+        # make sure that the model is on CPU when training
+        assert self.device == torch.device("cpu")
+
+
+@RunIf(skip_windows=True)
+def test_multi_model_ddp_with_cpu():
+    """Tests if device is set correctly when training for MultiModelDDPStrategy."""
+    trainer = Trainer(
+        accelerator="cpu",
+        devices=-1,
+        strategy=MultiModelDDPStrategy(),
+        fast_dev_run=True,
+    )
+    # assert strategy attributes for device setting
+    assert isinstance(trainer.strategy, MultiModelDDPStrategy)
+    assert trainer.strategy.root_device == torch.device("cpu")
+    model = BasicGANCPU()
+    trainer.fit(model)
+
+
+class BasicGANGPU(BasicGAN):
+    def on_train_start(self) -> None:
+        # make sure that the model is on GPU when training
+        assert self.device == torch.device(f"cuda:{self.trainer.strategy.local_rank}")
+        self.start_cuda_memory = torch.cuda.memory_allocated()
+
+
+@RunIf(min_cuda_gpus=2, skip_windows=True, standalone=True)
+def test_multi_model_ddp_with_gpus():
+    """Tests if device is set correctly when training and after teardown for MultiModelDDPStrategy."""
+    trainer = Trainer(
+        accelerator="gpu",
+        devices=-1,
+        strategy=MultiModelDDPStrategy(),
+        fast_dev_run=True,
+        enable_progress_bar=False,
+        enable_model_summary=False,
+    )
+    # assert strategy attributes for device setting
+    assert isinstance(trainer.strategy, MultiModelDDPStrategy)
+    local_rank = trainer.strategy.local_rank
+    assert trainer.strategy.root_device == torch.device(f"cuda:{local_rank}")
+
+    model = BasicGANGPU()
+
+    trainer.fit(model)
+
+    # assert after training, model is moved to CPU and memory is deallocated
+    assert model.device == torch.device("cpu")
+    cuda_memory = torch.cuda.memory_allocated()
+    assert cuda_memory < model.start_cuda_memory