Merge branch 'torchddp' into 'master'

Use DDP command line argument instead of source flag in pretrain_bert.py.

See merge request ADLR/megatron-lm!2

pretrain_bert.py

@@ -15,9 +15,6 @@
 
 """Pretrain BERT"""
 
-# Flag to use Pytorch ddp which uses overlapping communication and computation.
-USE_TORCH_DDP = False
-
 from datetime import datetime
 import os
 import random
@@ -33,10 +30,7 @@ from learning_rates import AnnealingLR
 from model import BertModel
 from model import get_params_for_weight_decay_optimization
 from model import gpt2_get_params_for_weight_decay_optimization
-if USE_TORCH_DDP:
-    from torch.nn.parallel.distributed import DistributedDataParallel as DDP
-else:
-    from model import DistributedDataParallel as DDP
+from model import DistributedDataParallel as LocalDDP
 import mpu
 from apex.optimizers import FusedAdam as Adam
 from utils import Timers
@@ -78,12 +72,18 @@ def get_model(args):
                     _module.float()
 
     # Wrap model for distributed training.
-    if USE_TORCH_DDP:
+    if args.DDP_impl == 'torch':
         i = torch.cuda.current_device()
-        model = DDP(model, device_ids=[i], output_device=i,
-                    process_group=mpu.get_data_parallel_group())
+        args.DDP_type = torch.nn.parallel.distributed.DistributedDataParallel
+        model = args.DDP_type(model, device_ids=[i], output_device=i,
+                              process_group=mpu.get_data_parallel_group())
+    elif args.DDP_impl == 'local':
+        args.DDP_type = LocalDDP
+        model = args.DDP_type(model)
     else:
-        model = DDP(model)
+        print_rank_0('Unknown DDP implementation specified: {}. '
+                     'Exiting.'.format(args.DDP_impl))
+        exit()
 
     return model
 
@@ -92,7 +92,7 @@ def get_optimizer(model, args):
     """Set up the optimizer."""
 
     # Build parameter groups (weight decay and non-decay).
-    while isinstance(model, (DDP, FP16_Module)):
+    while isinstance(model, (args.DDP_type, FP16_Module)):
         model = model.module
     layers = model.model.bert.encoder.layer
     pooler = model.model.bert.pooler
@@ -232,7 +232,7 @@ def forward_step(data_iterator, model, args, timers):
     return lm_loss, nsp_loss
 
 
-def backward_step(optimizer, model, lm_loss, nsp_loss, args):
+def backward_step(optimizer, model, lm_loss, nsp_loss, args, timers):
     """Backward step."""
 
     # Total loss.
@@ -252,9 +252,11 @@ def backward_step(optimizer, model, lm_loss, nsp_loss, args):
     reduced_losses = torch.cat((lm_loss.view(1), nsp_loss.view(1)))
     torch.distributed.all_reduce(reduced_losses.data)
     reduced_losses.data = reduced_losses.data / args.world_size
-    if not USE_TORCH_DDP:
+    if args.DDP_impl == 'local':
+        timers('allreduce').start()
         model.allreduce_params(reduce_after=False,
                                fp32_allreduce=args.fp32_allreduce)
+        timers('allreduce').stop()
     lm_loss_reduced = reduced_losses[0]
     nsp_loss_reduced = reduced_losses[1]
 
@@ -285,7 +287,7 @@ def train_step(data_iterator, model, optimizer, lr_scheduler,
     # Calculate gradients, reduce across processes, and clip.
     timers('backward').start()
     lm_loss_reduced, nsp_loss_reduced = backward_step(optimizer, model, lm_loss,
-                                                      nsp_loss, args)
+                                                      nsp_loss, args, timers)
     timers('backward').stop()
 
     # Update parameters.
@@ -338,8 +340,12 @@ def train(model, optimizer, lr_scheduler,
 
         # Logging.
 
-        timers_to_log = ['forward', 'backward', 'optimizer',
+        if args.DDP_impl == 'torch':
+            timers_to_log = ['forward', 'backward', 'optimizer',
                             'batch generator', 'data loader']
+        else:
+            timers_to_log = ['forward', 'backward', 'allreduce', 'optimizer',
+                             'batch generator', 'data loader']
 
         learning_rate = optimizer.param_groups[0]['lr']
 
@@ -425,7 +431,7 @@ def evaluate(data_iterator, model, args, timers, verbose = False):
             lm_loss, nsp_loss = forward_step(data_iterator, model,
                                              args, timers)
             # Reduce across processes.
-            if isinstance(model, DDP):
+            if isinstance(model, args.DDP_type):
                 reduced_losses = torch.cat((lm_loss.view(1), nsp_loss.view(1)))
                 torch.distributed.all_reduce(reduced_losses.data)
                 reduced_losses.data = reduced_losses.data/args.world_size