added buffer sync to naive amp model wrapper (#291)

colossalai/amp/naive_amp/naive_amp.py

@@ -3,12 +3,15 @@
 
 import torch
 import torch.nn as nn
+import torch.distributed as dist
 from torch import Tensor
-from typing import Union, List, Any, Dict
+from typing import Any
 from torch.optim import Optimizer
-import torch.cuda.amp as torch_amp
-
+from torch.distributed import ReduceOp
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
 from colossalai.nn.optimizer import ColossalaiOptimizer
+from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
 from ._fp16_optimizer import FP16Optimizer
 
 
@@ -49,10 +52,30 @@ class NaiveAMPModel(nn.Module):
 
     def __init__(self,
                  model: nn.Module,
-                 output_to_fp32: bool = True):
+                 output_to_fp32: bool = True,
+                 parallel_mode: ParallelMode = ParallelMode.DATA,
+                 sync_buffer: bool = True):
         super().__init__()
         self.model = model.half()
         self._output_to_fp32 = output_to_fp32
+        self._sync_buf = sync_buffer
+
+        if gpc.is_initialized(parallel_mode) and gpc.get_world_size(parallel_mode) > 1:
+            self._process_group = gpc.get_group(parallel_mode)
+            self._world_size = gpc.get_world_size(parallel_mode)
+        else:
+            self._process_group = None
+            self._world_size = 1
+            self._sync_buf = False
+        self._first_eval_run = False
+
+    @property
+    def sync_buffer(self):
+        return self._sync_buf
+
+    @sync_buffer.setter
+    def sync_buffer(self, state: bool):
+        self._sync_buf = state
 
     def _convert_to_fp16(self, input_: Any):
         if isinstance(input_, Tensor) and input_.dtype == torch.float32:
@@ -64,7 +87,46 @@ class NaiveAMPModel(nn.Module):
             input_ = input_.float()
         return input_
 
+    def _reduce_module_buffer(self):
+        """
+        All-reduce the buffers (e.g. running stats of batch normalization) across
+        data parallel ranks so that all the ranks will produce consistent results
+        when given the same input
+        """
+        buf_list = []
+
+        # find valid buffers
+        for buf in self.model.buffers():
+            if buf is not None:
+                buf_list.append(buf)
+
+        # reduce buffers across data parallel ranks
+        if buf_list:
+            coalesced_buf = _flatten_dense_tensors(buf_list)
+            coalesced_buf.div_(self._world_size)
+            dist.all_reduce(coalesced_buf, op=ReduceOp.SUM, group=self._process_group)
+            unflattened_buf_list = _unflatten_dense_tensors(coalesced_buf, buf_list)
+            for old, new in zip(buf_list, unflattened_buf_list):
+                old.copy_(new)
+
+    def eval(self):
+        self.model.eval()
+
+        # we only sync buffer in the first eval iteration
+        # so that future eval iterations can be done without communication
+        self._first_eval_run = True
+
     def forward(self, *args, **kwargs):
+        # reduce buffers after forward will lead to error
+        # as we cannot change the variables needed for gradient computation after forward
+        # so we sync buffer before forward
+        if (self.training or self._first_eval_run) and self._sync_buf:
+            with torch.no_grad():
+                self._reduce_module_buffer()
+
+            if self._first_eval_run:
+                self._first_eval_run = False
+
         if args:
             args = [self._convert_to_fp16(arg) for arg in args]
         if kwargs:

colossalai/initialize.py

@@ -16,6 +16,7 @@ from torch.optim.optimizer import Optimizer
 from torch.utils.data import DataLoader
 
 from colossalai.amp import AMP_TYPE, convert_to_amp
+from colossalai.amp.naive_amp import NaiveAMPModel
 from colossalai.builder.builder import build_gradient_handler
 from colossalai.context import Config, ConfigException, ParallelMode
 from colossalai.core import global_context as gpc
@@ -23,8 +24,7 @@ from colossalai.engine import Engine
 from colossalai.global_variables import moe_env
 from colossalai.logging import get_dist_logger
 from colossalai.nn.optimizer.colossalai_optimizer import ColossalaiOptimizer
-from colossalai.utils import (accumulate_gradient, get_current_device,
-                              is_using_ddp, is_using_pp, is_using_sequence,
+from colossalai.utils import (accumulate_gradient, get_current_device, is_using_ddp, is_using_pp, is_using_sequence,
                               sync_model_param)
 from colossalai.zero import convert_to_zero, ShardedOptimizer
 from colossalai.engine.ophooks import register_ophooks_recursively, BaseOpHook
@@ -39,21 +39,12 @@ def get_default_parser():
     """
     parser = argparse.ArgumentParser()
     parser.add_argument('--config', type=str, help='path to the config file')
-    parser.add_argument('--host',
-                        type=str,
-                        help='the master address for distributed training')
-    parser.add_argument('--port',
-                        type=int,
-                        help='the master port for distributed training')
+    parser.add_argument('--host', type=str, help='the master address for distributed training')
+    parser.add_argument('--port', type=int, help='the master port for distributed training')
     parser.add_argument('--world_size', type=int, help='world size for distributed training')
     parser.add_argument('--rank', type=int, help='rank for the default process group')
-    parser.add_argument('--local_rank',
-                        type=int,
-                        help='local rank on the node')
-    parser.add_argument('--backend',
-                        type=str,
-                        default='nccl',
-                        help='backend for distributed communication')
+    parser.add_argument('--local_rank', type=int, help='local rank on the node')
+    parser.add_argument('--backend', type=str, default='nccl', help='backend for distributed communication')
     return parser
 
 
@@ -116,9 +107,11 @@ def launch(config: Union[str, Path, Config, Dict],
 
     if verbose:
         logger = get_dist_logger()
-        logger.info(f'Distributed environment is initialized, '
+        logger.info(
+            f'Distributed environment is initialized, '
             f'data parallel size: {gpc.data_parallel_size}, pipeline parallel size: {gpc.pipeline_parallel_size}, '
-                    f'tensor parallel size: {gpc.tensor_parallel_size}', ranks=[0])
+            f'tensor parallel size: {gpc.tensor_parallel_size}',
+            ranks=[0])
 
 
 def launch_from_slurm(config: Union[str, Path, Config, Dict],
@@ -261,9 +254,11 @@ def initialize(model: Union[nn.Module, List[nn.Module]],
 
     # print config
     if verbose:
-        logger.info(f"\n========== Your Config ========\n"
+        logger.info(
+            f"\n========== Your Config ========\n"
             f"{pprint.pformat(gpc.config)}\n"
-                    f"================================\n", ranks=[0])
+            f"================================\n",
+            ranks=[0])
 
     # cudnn
     cudnn_benchmark = config.get('cudnn_benchmark', True)
@@ -271,8 +266,7 @@ def initialize(model: Union[nn.Module, List[nn.Module]],
     torch.backends.cudnn.benchmark = cudnn_benchmark
     torch.backends.cudnn.deterministic = cudnn_deterministic
     if verbose:
-        logger.info(
-            f"cuDNN benchmark = {cudnn_benchmark}, deterministic = {cudnn_deterministic}", ranks=[0])
+        logger.info(f"cuDNN benchmark = {cudnn_benchmark}, deterministic = {cudnn_deterministic}", ranks=[0])
 
     # first sync model across dp ranks
     model.to(get_current_device())
@@ -321,11 +315,7 @@ def initialize(model: Union[nn.Module, List[nn.Module]],
     if zero_cfg is not None:
         cfg_ = zero_cfg.copy()
         level = cfg_.pop('level')
-        model, optimizer = convert_to_zero(model=model,
-                                           optimizer=optimizer,
-                                           level=level,
-                                           zero_config=cfg_
-                                           )
+        model, optimizer = convert_to_zero(model=model, optimizer=optimizer, level=level, zero_config=cfg_)
 
     # gradient handler
     gradient_handler_cfg = gpc.config.get('gradient_handler', None)
@@ -350,21 +340,22 @@ def initialize(model: Union[nn.Module, List[nn.Module]],
                     "added even though not specified in the configuration",
                     ranks=[0])
         elif is_using_sequence():
-            model = DDP(model, process_group=gpc.get_group(ParallelMode.SEQUENCE_DP),
+            model = DDP(model,
+                        process_group=gpc.get_group(ParallelMode.SEQUENCE_DP),
                         device_ids=[torch.cuda.current_device()])
             if verbose:
-                logger.info(
-                    'Model is using torch.nn.parallel.DistributedDataParallel for Sequence Parallelism', ranks=[0])
+                logger.info('Model is using torch.nn.parallel.DistributedDataParallel for Sequence Parallelism',
+                            ranks=[0])
         elif is_using_ddp() and not is_using_pp() and amp_mode != AMP_TYPE.NAIVE:
             model = DDP(model, process_group=gpc.get_group(ParallelMode.DATA), device_ids=[torch.cuda.current_device()])
             if verbose:
-                logger.info(
-                    'Model is using torch.nn.parallel.DistributedDataParallel for Data Parallelism', ranks=[0])
+                logger.info('Model is using torch.nn.parallel.DistributedDataParallel for Data Parallelism', ranks=[0])
         elif is_using_ddp():
             gradient_handler_cfg = [dict(type='DataParallelGradientHandler')]
             if verbose:
                 logger.info(
-                    "Data parallel training is detected when using pipeline parallel, DataParallelGradientHandler is automatically "
+                    "Data parallel training is detected when using pipeline parallel, "
+                    "DataParallelGradientHandler is automatically "
                     "added even though not specified in the configuration",
                     ranks=[0])
         # add pipeline parallel gradient handler, if pipeline shared module is detected
@@ -383,7 +374,13 @@ def initialize(model: Union[nn.Module, List[nn.Module]],
     else:
         if not isinstance(gradient_handler_cfg, list):
             raise ConfigException(
-                f"expected gradient_handler in the configuration file to be a list but got {type(gradient_handler_cfg)}")
+                f"expected gradient_handler in the configuration file to be a list but got {type(gradient_handler_cfg)}"
+            )
+
+    # turn off sync buffer for NaiveAMPModel if using torch DDP and NaiveAMPModel at the same time
+    # to avoid duplicated buffer synchronization
+    if isinstance(model, DDP) and isinstance(model.module, NaiveAMPModel):
+        model.module.sync_buffer = False
 
     if gradient_handler_cfg is None:
         gradient_handlers = None

colossalai/zero/__init__.py

@@ -9,10 +9,7 @@ from .sharded_model import ShardedModel
 from .sharded_optim import ShardedOptimizer
 
 
-def convert_to_zero(model: nn.Module,
-                    optimizer: Optimizer,
-                    level: int,
-                    zero_config: dict):
+def convert_to_zero(model: nn.Module, optimizer: Optimizer, level: int, zero_config: dict):
     """
     A helper function to integrate the model and optimizer with ZeRO optimizer and off-loading
 
@@ -31,11 +28,16 @@ def convert_to_zero(model: nn.Module,
     assert 1 <= level <= 3, 'Only ZERO Optimizer Level 1-3 are provided'
     if level in [1, 2]:
         if level == 2:
-            assert config['partition_grad'], 'ZeRO Optimizer requires partition_grad to be True'
+            if 'partition_grad' in zero_config:
+                assert zero_config['partition_grad'], \
+                    'Sharded Optimizer requires partition_grad to be True'
+            else:
+                zero_config['partiton_grad'] = True
         model = NaiveAMPModel(model, output_to_fp32=True)
-        optimizer = ShardedOptimizer(model.parameters(), *zero_config)
+        optimizer = ShardedOptimizer(optimizer, **zero_config)
     else:
         model = ShardedModel(module=model, **zero_config)
     return model, optimizer
 
+
 __all__ = ['convert_to_zero', 'ShardedModel', 'ShardedOptimizer']

tests/test_zero_data_parallel/test_sharded_optim_with_sync_bn.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+from functools import partial
+
+import colossalai
+import pytest
+import torch
+import torch.multiprocessing as mp
+from colossalai.utils import free_port
+from colossalai.core import global_context as gpc
+from colossalai.context.parallel_mode import ParallelMode
+from torchvision.models import resnet50
+import torch.distributed as dist
+
+
+def run_dist(rank, world_size, port):
+    # need to configure cudnn deterministic so that
+    # randomness of convolution layers will be disabled
+    colossalai.launch(config=dict(zero=dict(level=2, partition_grad=True),
+                                  cudnn_determinstic=True,
+                                  cudnn_benchmark=False),
+                      rank=rank,
+                      world_size=world_size,
+                      host='localhost',
+                      port=port,
+                      backend='nccl')
+
+    model = resnet50()
+    optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
+    criterion = torch.nn.CrossEntropyLoss()
+
+    engine, *args = colossalai.initialize(model, optimizer, criterion)
+
+    # train for dummy iterations
+    engine.train()
+    for _ in range(2):
+        data = torch.rand(4, 3, 128, 128).cuda().half()
+        label = torch.randint(0, 10, size=(4,)).cuda()
+        engine.zero_grad()
+        out = engine(data)
+        loss = engine.criterion(out, label)
+        engine.backward(loss)
+        engine.step()
+
+    # test
+    # need to make sure the batch norm stats are synchronized
+    # so that given the same input, the model will produce the same
+    # output on different ranks
+    engine.eval()
+    data = torch.rand(4, 3, 128, 128).cuda().half()
+    dist.broadcast(data, src=0, group=gpc.get_group(ParallelMode.DATA))
+
+    # predict
+    out = engine(data)
+
+    # test if results are equal
+    tensor_list = [torch.empty_like(out) for _ in range(world_size - 1)]
+    tensor_list.insert(rank, out)
+    dist.all_gather(tensor_list=tensor_list, tensor=out, group=gpc.get_group(ParallelMode.DATA))
+
+    assert torch.all(tensor_list[0] == tensor_list[1]), \
+        'expected the output from different ranks to be the same, but got different values'
+
+
+@pytest.mark.dist
+def test_sharded_optim_with_sync_bn():
+    """
+    This test is to make sure that buffers are synchronized between ranks
+    when using ZeRO. An example of module buffer is the running stats of
+    BatchNormalization layer, i.e. mean and var.
+
+    If the buffers are not synchronized, the model will produce different
+    output even though the input and parameters are the same. This is not
+    wanted if we are doing predictions.
+
+    """
+    world_size = 2
+    run_func = partial(run_dist, world_size=world_size, port=free_port())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_sharded_optim_with_sync_bn()