Add base code

pytorch_lightning/overrides/fairscale.py

@@ -0,0 +1,47 @@
+# Copyright The PyTorch Lightning 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.
+from typing import Any, List, Union
+
+from fairscale.nn.data_parallel.sharded_ddp import ShardedDataParallel
+from fairscale.optim import OSS
+from torch import nn
+
+
+class LightningShardedDataParallel(ShardedDataParallel):
+    def __init__(
+            self,
+            base_model: nn.Module,
+            sharded_optimizer: Union[OSS, List[OSS]],
+            process_group: Any = None,
+            broadcast_buffers: bool = True
+    ):
+        super().__init__(
+            base_model=base_model,
+            sharded_optimizer=sharded_optimizer,
+            process_group=process_group,
+            broadcast_buffers=broadcast_buffers
+        )
+        self.module = base_model
+
+    def forward(self, *inputs, **kwargs):
+        if self.enable_broadcast_buffers:
+            self.sync_buffers()
+
+        if self.base_model.training:
+            outputs = self.base_model.training_step(*inputs, **kwargs)
+        elif self.base_model.testing:
+            outputs = self.base_model.test_step(*inputs, **kwargs)
+        else:
+            outputs = self.base_model.validation_step(*inputs, **kwargs)
+        return outputs

pytorch_lightning/plugins/sharded_native_amp_plugin.py

@@ -0,0 +1,31 @@
+# Copyright The PyTorch Lightning 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.
+from typing import cast
+
+from fairscale.optim import OSS
+from fairscale.optim.grad_scaler import ShardedGradScaler
+
+from pytorch_lightning.plugins.native_amp import NativeAMPPlugin
+
+
+class ShardedNativeAMPPlugin(NativeAMPPlugin):
+    @property
+    def scaler(self):
+        return ShardedGradScaler()
+
+    def clip_gradients(self, grad_clip_val, model, optimizer):
+        max_norm = grad_clip_val
+        norm_type = float(2.0)
+        optimizer = cast(OSS, optimizer)
+        optimizer.clip_grad_norm(max_norm, norm_type=norm_type)

pytorch_lightning/plugins/sharded_plugin.py

@@ -0,0 +1,90 @@
+# Copyright The PyTorch Lightning 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.
+from typing import List, Any, Optional
+
+from fairscale.optim import OSS
+
+from pytorch_lightning import LightningModule
+from pytorch_lightning.overrides.fairscale import LightningShardedDataParallel
+from pytorch_lightning.plugins.ddp_plugin import DDPPlugin
+from pytorch_lightning.utilities import rank_zero_only
+
+
+class DDPShardedPlugin(DDPPlugin):
+
+    def configure_ddp(
+            self, model: LightningModule, device_ids: List[int]
+    ):
+        self._wrap_optimizers(model)
+        if model.trainer.testing:  # Revert to standard DDP if using testing
+            super().configure_ddp(
+                model=model,
+                device_ids=device_ids
+            )
+        else:
+            model = LightningShardedDataParallel(model, sharded_optimizer=model.trainer.optimizers)
+        return model
+
+    def optimizer_state(self, optimizer: OSS) -> Optional[dict]:
+        optimizer.consolidate_state_dict()
+        return self._optim_state_dict(optimizer)
+
+    def on_before_forward(self, args: Any, model: LightningModule):
+        batch = args[0]
+        batch = model.transfer_batch_to_device(batch, model.trainer.root_gpu)
+        args[0] = batch
+        return args
+
+    @rank_zero_only
+    def _optim_state_dict(self, optimizer):
+        """
+        Ensure we only return the state dict from the optimizer on rank 0.
+        Other ranks do not have the complete optimizer state.
+        Args:
+            optimizer: OSS Optimizer
+        Returns:
+            State dict if rank 0 else None.
+        """
+        return optimizer.state_dict()
+
+    def _wrap_optimizers(self, model):
+        trainer = model.trainer
+        if trainer.testing is True:
+            return
+
+        self._reinit_with_fairscale_oss(trainer)
+
+    def _reinit_with_fairscale_oss(self, trainer):
+        """
+        Re-initialise optimizers to use OSS wrapper. We need to re-initialise due to
+        the parameters being sharded across distributed processes, each optimizing a partition.
+        Args:
+            trainer: trainer object to reinit optimizers.
+        """
+        optimizers = trainer.optimizers
+        lr_schedulers = trainer.lr_schedulers
+        for x, optimizer in enumerate(optimizers):
+            if not isinstance(optimizer, OSS):
+                optim_class = type(optimizer)
+                zero_optimizer = OSS(
+                    params=optimizer.param_groups,
+                    optim=optim_class,
+                    **optimizer.defaults
+                )
+                optimizers[x] = zero_optimizer
+                for scheduler in lr_schedulers:
+                    scheduler = scheduler['scheduler']
+                    if scheduler.optimizer == optimizer:
+                        scheduler.optimizer = zero_optimizer
+                del optimizer