[Feature] Support torch ZeroRedundancyOptimizer

docs/en/api/optim.rst

@@ -23,6 +23,7 @@ Optimizer
     OptimWrapper
     OptimWrapperDict
     DefaultOptimWrapperConstructor
+    ZeroRedundancyOptimizer
 
 .. autosummary::
    :toctree: generated

docs/zh_cn/api/optim.rst

@@ -23,6 +23,7 @@ Optimizer
     OptimWrapper
     OptimWrapperDict
     DefaultOptimWrapperConstructor
+    ZeroRedundancyOptimizer
 
 .. autosummary::
    :toctree: generated

mmengine/hooks/checkpoint_hook.py

@@ -6,7 +6,7 @@
 from pathlib import Path
 from typing import Callable, Dict, List, Optional, Sequence, Union
 
-from mmengine.dist import master_only
+from mmengine.dist import is_main_process
 from mmengine.fileio import FileClient, get_file_backend
 from mmengine.registry import HOOKS
 from mmengine.utils import is_list_of, is_seq_of
@@ -309,7 +309,6 @@ def _get_metric_score(self, metrics, key_indicator):
 
         return eval_res[key_indicator]
 
-    @master_only
     def _save_checkpoint(self, runner) -> None:
         """Save the current checkpoint and delete outdated checkpoint.
 
@@ -331,6 +330,11 @@ def _save_checkpoint(self, runner) -> None:
             backend_args=self.backend_args,
             **self.args)
 
+        # Model parallel-like training should involve pulling sharded states
+        # from all ranks, but skip the following procedure.
+        if not is_main_process():
+            return
+
         runner.message_hub.update_info(
             'last_ckpt',
             self.file_backend.join_path(self.out_dir, ckpt_filename))
@@ -357,7 +361,6 @@ def _save_checkpoint(self, runner) -> None:
         with open(save_file, 'w') as f:
             f.write(filepath)
 
-    @master_only
     def _save_best_checkpoint(self, runner, metrics) -> None:
         """Save the current checkpoint and delete outdated checkpoint.
 

mmengine/optim/optimizer/__init__.py

@@ -5,9 +5,10 @@
 from .default_constructor import DefaultOptimWrapperConstructor
 from .optimizer_wrapper import OptimWrapper
 from .optimizer_wrapper_dict import OptimWrapperDict
+from .zero_optimizer import ZeroRedundancyOptimizer
 
 __all__ = [
     'OPTIM_WRAPPER_CONSTRUCTORS', 'OPTIMIZERS',
     'DefaultOptimWrapperConstructor', 'build_optim_wrapper', 'OptimWrapper',
-    'AmpOptimWrapper', 'OptimWrapperDict'
+    'AmpOptimWrapper', 'OptimWrapperDict', 'ZeroRedundancyOptimizer'
 ]

mmengine/optim/optimizer/zero_optimizer.py

@@ -0,0 +1,66 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+
+import torch
+from torch.distributed.rpc import is_available
+
+from mmengine.dist import is_main_process
+from mmengine.utils import digit_version
+from mmengine.utils.dl_utils import TORCH_VERSION
+
+try:
+    from torch.distributed.optim import \
+        ZeroRedundancyOptimizer as _ZeroRedundancyOptimizer
+except ImportError:
+    _ZeroRedundancyOptimizer = object
+
+from .builder import OPTIMIZERS
+
+
+@OPTIMIZERS.register_module()
+class ZeroRedundancyOptimizer(_ZeroRedundancyOptimizer):
+    """A wrapper class of :class:`ZeroRedundancyOptimizer` that gets a
+    optimizer type as string.
+
+    This class wraps an arbitrary :class:`torch.optim.Optimizer` and shards its
+    states across ranks in the group as described by ZeRO_. The local optimizer
+    instance in each rank is only responsible for updating approximately
+    ``1 / world_size`` parameters and hence only needs to keep
+    ``1 / world_size`` optimizer states. After parameters are updated locally,
+    each rank will broadcast its parameters to all other peers to keep all
+    model replicas in the same state. ``ZeroRedundancyOptimizer`` can be used
+    in conjunction with :class:`torch.nn.parallel.DistributedDataParallel` to
+    reduce per-rank peak memory consumption.
+
+    ``ZeroRedundancyOptimizer`` uses a sorted-greedy algorithm to pack a number
+    of parameters at each rank. Each parameter belongs to a single rank and is
+    not divided among ranks. The partition is arbitrary and might not match the
+    the parameter registration or usage order.
+
+    Warnings:
+        ``ZeroRedundancyOptimizer`` requires PyTorch >= 1.8.
+
+    Args:
+        params (``Iterable``): an ``Iterable`` of :class:`torch.Tensor` s
+            or :class:`dict` s giving all parameters, which will be sharded
+            across ranks.
+        optimizer_type (str): the string of the local optimizer class.
+
+    .. _ZeRO: https://arxiv.org/abs/1910.02054
+    """
+
+    def __init__(self, params, optimizer_type: str, **kwargs):
+        assert digit_version(TORCH_VERSION) >= digit_version('1.8.0'), (
+            '`torch.distributed.optim.ZeroReundancyOptimizer` is only '
+            'available when pytorch version >= 1.8.')
+        assert is_available(), 'torch.distributed.rpc is not available.'
+        optimizer_class = getattr(torch.optim, optimizer_type)
+        # TODO: Register a DDP communication hook for `overlap_with_ddp=True`.
+        # Currently only `overlap_with_ddp=False` is supported. For more
+        # details, please refer to the pytorch's official documentation.
+        super().__init__(params, optimizer_class, **kwargs)
+
+    def state_dict(self):
+        """Consolidate `state_dict`s from ranks to save the `state_dict`."""
+        self.consolidate_state_dict()
+        state_dict = super().state_dict() if is_main_process() else dict()
+        return state_dict

tests/test_optim/test_optimizer/test_optimizer.py

@@ -1,17 +1,22 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+import os
 import sys
+import unittest
 from unittest import TestCase
 from unittest.mock import MagicMock
 
 import torch
 import torch.nn as nn
+from torch.distributed.rpc import is_available
 
 from mmengine.optim import (OPTIM_WRAPPER_CONSTRUCTORS, OPTIMIZERS,
                             DefaultOptimWrapperConstructor, OptimWrapper,
                             build_optim_wrapper)
 from mmengine.optim.optimizer.builder import TORCH_OPTIMIZERS
 from mmengine.registry import build_from_cfg
-from mmengine.utils.dl_utils import mmcv_full_available
+from mmengine.testing._internal import MultiProcessTestCase
+from mmengine.utils.dl_utils import TORCH_VERSION, mmcv_full_available
+from mmengine.utils.version_utils import digit_version
 
 MMCV_FULL_AVAILABLE = mmcv_full_available()
 if not MMCV_FULL_AVAILABLE:
@@ -713,3 +718,77 @@ def test_default_optimizer_constructor_custom_key(self):
                     for setting in settings:
                         assert param_groups[i][setting] == settings[
                             setting], f'{name} {setting}'
+
+
+@unittest.skipIf(
+    (digit_version(TORCH_VERSION) < digit_version('1.8.0'))
+    or not is_available(),
+    reason='ZeRO requires pytorch>=1.8 with torch.distributed.rpc available.')
+class TestZeroOptimizer(MultiProcessTestCase):
+
+    def setUp(self) -> None:
+        super().setUp()
+        self._spawn_processes()
+
+    def _check_default_optimizer(self, optimizer, model):
+        self.assertIsInstance(optimizer.optim, torch.optim.SGD)
+        self.assertEqual(optimizer.defaults['lr'], self.base_lr)
+        self.assertEqual(optimizer.defaults['momentum'], self.momentum)
+        self.assertEqual(optimizer.defaults['weight_decay'], self.base_wd)
+        param_groups = optimizer.param_groups[0]
+        if MMCV_FULL_AVAILABLE:
+            param_names = [
+                'param1', 'conv1.weight', 'conv2.weight', 'conv2.bias',
+                'bn.weight', 'bn.bias', 'sub.param1', 'sub.conv1.weight',
+                'sub.conv1.bias', 'sub.gn.weight', 'sub.gn.bias', 'dcn.weight',
+                'dcn.conv_offset.weight', 'dcn.conv_offset.bias'
+            ]
+        else:
+            param_names = [
+                'param1', 'conv1.weight', 'conv2.weight', 'conv2.bias',
+                'bn.weight', 'bn.bias', 'sub.param1', 'sub.conv1.weight',
+                'sub.conv1.bias', 'sub.gn.weight', 'sub.gn.bias'
+            ]
+        param_dict = dict(model.named_parameters())
+        self.assertEqual(len(param_groups['params']), len(param_names))
+        for i in range(len(param_groups['params'])):
+            assert torch.equal(param_groups['params'][i],
+                               param_dict[param_names[i]])
+
+    def test_build_zero_redundancy_optimizer(self):
+        from torch.distributed.optim import ZeroRedundancyOptimizer
+        self._init_dist_env(self.rank, self.world_size)
+        model = ExampleModel()
+        self.base_lr = 0.01
+        self.momentum = 0.0001
+        self.base_wd = 0.9
+
+        # test build function
+        optim_wrapper_cfg = dict(
+            optimizer=dict(
+                type='ZeroRedundancyOptimizer',
+                optimizer_type='SGD',
+                lr=self.base_lr,
+                weight_decay=self.base_wd,
+                momentum=self.momentum))
+        optim_wrapper = build_optim_wrapper(model, optim_wrapper_cfg)
+        self.assertIsInstance(optim_wrapper.optimizer, ZeroRedundancyOptimizer)
+        self._check_default_optimizer(optim_wrapper.optimizer, model)
+
+        # test build optimizer without ``optimizer_type``
+        with self.assertRaises(TypeError):
+            optim_wrapper_cfg = dict(
+                optimizer=dict(
+                    type='ZeroRedundancyOptimizer',
+                    lr=self.base_lr,
+                    weight_decay=self.base_wd,
+                    momentum=self.momentum))
+            optim_wrapper = build_optim_wrapper(model, optim_wrapper_cfg)
+
+    def _init_dist_env(self, rank, world_size):
+        """Initialize the distributed environment."""
+        os.environ['MASTER_ADDR'] = '127.0.0.1'
+        os.environ['MASTER_PORT'] = '29510'
+        os.environ['RANK'] = str(rank)
+        torch.distributed.init_process_group(
+            backend='gloo', rank=rank, world_size=world_size)