Merge branch 'main' into main_dedup

megatron/indexer.py

+import sys
 import torch
 import torch.distributed as dist
 
 from megatron import get_args
 from megatron import mpu
-from megatron.checkpointing import load_ict_checkpoint
-from megatron.data.ict_dataset import get_ict_dataset
-from megatron.data.realm_dataset_utils import get_one_epoch_dataloader
-from megatron.data.realm_index import detach, BlockData
-from megatron.data.realm_dataset_utils import get_ict_batch
-from megatron.model.realm_model import general_ict_model_provider
+from megatron.checkpointing import load_biencoder_checkpoint
+from megatron.data.orqa_wiki_dataset import get_open_retrieval_wiki_dataset
+from megatron.data.orqa_wiki_dataset import get_open_retrieval_batch
+from megatron.data.biencoder_dataset_utils import get_one_epoch_dataloader
+from megatron.data.realm_index import detach, OpenRetreivalDataStore
+from megatron.model.biencoder_model import biencoder_model_provider
 from megatron.training import get_model
 
 
 class IndexBuilder(object):
-    """Object for taking one pass over a dataset and creating a BlockData of its embeddings"""
+    """
+    Object for taking one pass over a dataset and creating a BlockData of its
+    embeddings
+    """
     def __init__(self):
         args = get_args()
         self.model = None
         self.dataloader = None
-        self.block_data = None
+        self.evidence_embedder_obj = None
+        self.biencoder_shared_query_context_model = \
+            args.biencoder_shared_query_context_model
 
-        # need to know whether we're using a REALM checkpoint (args.load) or ICT checkpoint
+        # need to know whether we're using a REALM checkpoint (args.load)
+        # or ICT checkpoint
         assert not (args.load and args.ict_load)
-        self.using_realm_chkpt = args.ict_load is None
+        #self.using_realm_chkpt = args.ict_load is None
 
         self.log_interval = args.indexer_log_interval
         self.batch_size = args.indexer_batch_size
@@ -33,59 +40,88 @@ class IndexBuilder(object):
         self.iteration = self.total_processed = 0
 
     def load_attributes(self):
-        """Load the necessary attributes: model, dataloader and empty BlockData"""
-        model = get_model(lambda: general_ict_model_provider(only_block_model=True))
-        self.model = load_ict_checkpoint(model, only_block_model=True, from_realm_chkpt=self.using_realm_chkpt)
-        self.model.eval()
-        self.dataset = get_ict_dataset()
-        self.dataloader = iter(get_one_epoch_dataloader(self.dataset, self.batch_size))
-        self.block_data = BlockData(load_from_path=False)
+        """
+        Load the necessary attributes: model, dataloader and empty BlockData
+        """
+        only_context_model = True
+        if self.biencoder_shared_query_context_model:
+            only_context_model = False
+
+        model = get_model(lambda: biencoder_model_provider(only_context_model \
+            = only_context_model, biencoder_shared_query_context_model = \
+            self.biencoder_shared_query_context_model))
+
+        self.model = load_biencoder_checkpoint(model,
+                only_context_model=only_context_model)
+
+        assert len(self.model) == 1
+        self.model[0].eval()
+
+        self.dataset = get_open_retrieval_wiki_dataset()
+        self.dataloader = iter(get_one_epoch_dataloader(self.dataset, \
+            self.batch_size))
+
+        self.evidence_embedder_obj = OpenRetreivalDataStore( \
+            load_from_path=False)
 
     def track_and_report_progress(self, batch_size):
-        """Utility function for tracking progress"""
+        """
+        Utility function for tracking progress
+        """
         self.iteration += 1
         self.total_processed += batch_size * self.num_total_builders
         if self.is_main_builder and self.iteration % self.log_interval == 0:
-            print('Batch {:10d} | Total {:10d}'.format(self.iteration, self.total_processed), flush=True)
+            print('Batch {:10d} | Total {:10d}'.format(self.iteration,
+                self.total_processed), flush=True)
 
     def build_and_save_index(self):
-        """Goes through one epoch of the dataloader and adds all data to this instance's BlockData.
+        """
+        Goes through one epoch of the dataloader and adds all data to this
+        instance's BlockData.
 
-        The copy of BlockData is saved as a shard, which when run in a distributed setting will be
-        consolidated by the rank 0 process and saved as a final pickled BlockData.
+        The copy of BlockData is saved as a shard, which when run in a
+        distributed setting will be consolidated by the rank 0 process
+        and saved as a final pickled BlockData.
         """
+        assert len(self.model) == 1
+        unwrapped_model = self.model[0]
+        while not hasattr(unwrapped_model, 'embed_text'):
+            unwrapped_model = unwrapped_model.module
 
         while True:
             try:
                 # batch also has query_tokens and query_pad_data
-                _, _, block_tokens, block_pad_mask, block_sample_data = get_ict_batch(self.dataloader)
+                row_id, context_tokens, context_mask, context_types, \
+                    context_pad_mask = get_open_retrieval_batch( \
+                    self.dataloader)
             except (StopIteration, IndexError):
                 break
 
-            unwrapped_model = self.model
-            while not hasattr(unwrapped_model, 'embed_block'):
-                unwrapped_model = unwrapped_model.module
-
+            # TODO: can we add with torch.no_grad() to reduce memory usage
             # detach, separate fields and add to BlockData
-            block_logits = detach(unwrapped_model.embed_block(block_tokens, block_pad_mask))
-            detached_data = detach(block_sample_data)
-
-            # block_sample_data is a 2D array [batch x 4]
-            # with columns [start_idx, end_idx, doc_idx, block_idx] same as class BlockSampleData
-            block_indices = detached_data[:, 3]
-            block_metas = detached_data[:, :3]
-
-            self.block_data.add_block_data(block_indices, block_logits, block_metas)
-            self.track_and_report_progress(batch_size=block_tokens.shape[0])
-
-        # This process signals to finalize its shard and then synchronize with the other processes
-        self.block_data.save_shard()
+            assert context_mask.dtype == torch.bool
+            context_logits = unwrapped_model.embed_text(
+                unwrapped_model.context_model, context_tokens, context_mask,
+                context_types)
+            context_logits = detach(context_logits)
+            row_id = detach(row_id)
+
+            self.evidence_embedder_obj.add_block_data(row_id, context_logits)
+            self.track_and_report_progress(batch_size=len(row_id))
+
+        # This process signals to finalize its shard and then synchronize with
+        # the other processes
+        self.evidence_embedder_obj.save_shard()
         torch.distributed.barrier()
         del self.model
 
         # rank 0 process builds the final copy
         if self.is_main_builder:
-            self.block_data.merge_shards_and_save()
+            self.evidence_embedder_obj.merge_shards_and_save()
             # make sure that every single piece of data was embedded
-            assert len(self.block_data.embed_data) == len(self.dataset)
-        self.block_data.clear()
+            assert len(self.evidence_embedder_obj.embed_data) == \
+                len(self.dataset)
+        self.evidence_embedder_obj.clear()
+
+        # complete building the final copy
+        torch.distributed.barrier()

megatron/initialize.py

@@ -17,16 +17,20 @@
 
 import random
 import os
+import time
 
 import numpy as np
 import torch
 
+from megatron import fused_kernels
 from megatron import get_adlr_autoresume
 from megatron import get_args
 from megatron import get_tensorboard_writer
 from megatron import mpu
 from megatron.global_vars import set_global_variables
-from megatron.mpu import set_tensor_model_parallel_rank, set_tensor_model_parallel_world_size
+from megatron.mpu import (set_tensor_model_parallel_rank,
+                          set_tensor_model_parallel_world_size)
+
 
 def initialize_megatron(extra_args_provider=None, args_defaults={},
                         ignore_unknown_args=False, allow_no_cuda=False):
@@ -37,8 +41,7 @@ def initialize_megatron(extra_args_provider=None, args_defaults={},
     what you are doing.
     Returns a function to finalize distributed env initialization 
     (optionally, only when args.lazy_mpu_init == True)
-
-"""
+    """
     if not allow_no_cuda:
         # Make sure cuda is available.
         assert torch.cuda.is_available(), 'Megatron requires CUDA.'
@@ -66,7 +69,8 @@ def initialize_megatron(extra_args_provider=None, args_defaults={},
         # delayed initialization of DDP-related stuff
         # We only set basic DDP globals    
         set_tensor_model_parallel_world_size(args.tensor_model_parallel_size)
-        # and return function for external DDP manager to call when it has DDP initialized
+        # and return function for external DDP manager
+        # to call when it has DDP initialized
         set_tensor_model_parallel_rank(args.rank)    
         return finish_mpu_init
     else:
@@ -79,16 +83,71 @@ def initialize_megatron(extra_args_provider=None, args_defaults={},
         # Autoresume.
         _init_autoresume()
 
-        # Compile dataset C++ code.
-        if torch.distributed.get_rank() == 0:
-            from megatron.data.dataset_utils import compile_helper
-            compile_helper()
-        # Simple barrier
-        torch.distributed.barrier()
-        
+        # Compile dependencies.
+        _compile_dependencies()
+
         # No continuation function
         return None
-        
+
+
+def _compile_dependencies():
+
+    args = get_args()
+
+    # =========================
+    # Compile dataset C++ code.
+    # =========================
+    # TODO: move this to ninja
+    if torch.distributed.get_rank() == 0:
+        start_time = time.time()
+        print('> compiling dataset index builder ...')
+        from megatron.data.dataset_utils import compile_helper
+        compile_helper()
+        print('>>> done with dataset index builder. Compilation time: {:.3f} '
+              'seconds'.format(time.time() - start_time), flush=True)
+
+    # ==================
+    # Load fused kernels
+    # ==================
+
+    # Custom kernel constraints check.
+    seq_len = args.seq_length
+    attn_batch_size = \
+        (args.num_attention_heads / args.tensor_model_parallel_size) * \
+        args.micro_batch_size
+    # Constraints on sequence length and attn_batch_size to enable warp based
+    # optimization and upper triangular optimization (for causal mask)
+    custom_kernel_constraint = seq_len > 16 and seq_len <=2048 and \
+        seq_len % 4 == 0 and attn_batch_size % 4 == 0
+    # Print a warning.
+    if not ((args.fp16 or args.bf16) and
+            custom_kernel_constraint and
+            args.masked_softmax_fusion):
+        if args.rank == 0:
+            print('WARNING: constraints for invoking optimized'
+                  ' fused softmax kernel are not met. We default'
+                  ' back to unfused kernel invocations.', flush=True)
+    
+    # Always build on rank zero first.
+    if torch.distributed.get_rank() == 0:
+        start_time = time.time()
+        print('> compiling and loading fused kernels ...', flush=True)
+        fused_kernels.load(args)
+        torch.distributed.barrier()
+    else:
+        torch.distributed.barrier()
+        fused_kernels.load(args)
+    # Simple barrier to make sure all ranks have passed the
+    # compilation phase successfully before moving on to the
+    # rest of the program. We think this might ensure that
+    # the lock is released.
+    torch.distributed.barrier()
+    if torch.distributed.get_rank() == 0:
+        print('>>> done with compiling and loading fused kernels. '
+              'Compilation time: {:.3f} seconds'.format(
+                  time.time() - start_time), flush=True)
+
+
 
 def _initialize_distributed():
     """Initialize torch.distributed and mpu."""
@@ -133,7 +192,8 @@ def _initialize_distributed():
             print('model parallel is already initialized')
         else:
             mpu.initialize_model_parallel(args.tensor_model_parallel_size,
-                                          args.pipeline_model_parallel_size)
+                                          args.pipeline_model_parallel_size,
+                                          args.virtual_pipeline_model_parallel_size)
 
 
 def _init_autoresume():

megatron/model/__init__.py

@@ -13,34 +13,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-_LAYER_NORM = None
-
-
-def import_layernorm(fp32_residual_connection):
-
-    global _LAYER_NORM
-    if not _LAYER_NORM:
-        if fp32_residual_connection:
-            from .fused_layer_norm import MixedFusedLayerNorm as LayerNorm
-        else:
-            from apex.normalization.fused_layer_norm import FusedLayerNorm as LayerNorm
-        _LAYER_NORM = LayerNorm
-            
-    return _LAYER_NORM
-
+from .fused_layer_norm import MixedFusedLayerNorm as LayerNorm
 
 from .distributed import *
 from .bert_model import (BertModel,
                          BertModelFirstStage,
                          BertModelIntermediateStage,
                          BertModelLastStage)
-from .realm_model import ICTBertModel
 from .gpt_model import (GPTModel,
                         GPTModelFirstStage,
                         GPTModelIntermediateStage,
                         GPTModelLastStage)
 from .language_model import get_language_model
-from .module import FP16Module
-from .realm_model import ICTBertModel
+from .module import Float16Module
 
 

megatron/model/bert_model.py

@@ -22,7 +22,7 @@ from megatron import mpu
 from megatron.model.enums import AttnMaskType
 from megatron.model.language_model import parallel_lm_logits
 from megatron.model.language_model import get_language_model
-from megatron.model import import_layernorm
+from megatron.model import LayerNorm
 from megatron.model.utils import openai_gelu, erf_gelu
 from megatron.model.utils import get_linear_layer
 from megatron.model.utils import init_method_normal
@@ -78,7 +78,6 @@ class BertLMHead(MegatronModule):
         self.parallel_output = parallel_output
 
         self.dense = get_linear_layer(hidden_size, hidden_size, init_method)
-        LayerNorm = import_layernorm(args.fp32_residual_connection)
         self.layernorm = LayerNorm(hidden_size, eps=layernorm_epsilon)
         self.gelu = torch.nn.functional.gelu
         if args.openai_gelu:

megatron/model/biencoder_model.py

+import os
+import torch
+import sys
+
+from megatron import get_args, print_rank_0
+from megatron.checkpointing import fix_query_key_value_ordering
+from megatron.checkpointing import get_checkpoint_tracker_filename
+from megatron.checkpointing import get_checkpoint_name
+from megatron import mpu, get_tokenizer
+from megatron.model.bert_model import bert_position_ids
+from megatron.model.enums import AttnMaskType
+from megatron.model.language_model import get_language_model
+from megatron.model.utils import get_linear_layer
+from megatron.model.utils import init_method_normal
+from megatron.model.utils import scaled_init_method_normal
+from .module import MegatronModule
+
+def biencoder_model_provider(only_query_model=False,
+                             only_context_model=False,
+                             biencoder_shared_query_context_model=False):
+    """Build the model."""
+    args = get_args()
+
+    assert mpu.get_tensor_model_parallel_world_size() == 1 and \
+        mpu.get_pipeline_model_parallel_world_size() == 1, \
+        "Model parallel size > 1 not supported for ICT"
+
+    print_rank_0('building BiEncoderModel...')
+
+    # simpler to just keep using 2 tokentypes since
+    # the LM we initialize with has 2 tokentypes
+    model = BiEncoderModel(
+        num_tokentypes=2,
+        parallel_output=False,
+        only_query_model=only_query_model,
+        only_context_model=only_context_model,
+        biencoder_shared_query_context_model=\
+            biencoder_shared_query_context_model)
+
+    return model
+
+
+class BiEncoderModel(MegatronModule):
+    """Bert-based module for Biencoder model."""
+
+    def __init__(self,
+                 num_tokentypes=1,
+                 parallel_output=True,
+                 only_query_model=False,
+                 only_context_model=False,
+                 biencoder_shared_query_context_model=False):
+        super(BiEncoderModel, self).__init__()
+        args = get_args()
+
+        bert_kwargs = dict(
+            num_tokentypes=num_tokentypes,
+            parallel_output=parallel_output)
+
+        self.biencoder_shared_query_context_model = \
+            biencoder_shared_query_context_model
+        assert not (only_context_model and only_query_model)
+        self.use_context_model = not only_query_model
+        self.use_query_model = not only_context_model
+        self.biencoder_projection_dim = args.biencoder_projection_dim
+
+        if self.biencoder_shared_query_context_model:
+            self.model = PretrainedBertModel(**bert_kwargs)
+            self._model_key = 'shared_model'
+            self.query_model, self.context_model = self.model, self.model
+        else:
+            if self.use_query_model:
+                # this model embeds (pseudo-)queries - Embed_input in the paper
+                self.query_model = PretrainedBertModel(**bert_kwargs)
+                self._query_key = 'query_model'
+
+            if self.use_context_model:
+                # this model embeds evidence blocks - Embed_doc in the paper
+                self.context_model = PretrainedBertModel(**bert_kwargs)
+                self._context_key = 'context_model'
+
+    def forward(self, query_tokens, query_attention_mask, query_types,
+                context_tokens, context_attention_mask, context_types):
+        """Run a forward pass for each of the models and
+        return the respective embeddings."""
+
+        if self.use_query_model:
+            query_logits = self.embed_text(self.query_model,
+                                           query_tokens,
+                                           query_attention_mask,
+                                           query_types)
+        else:
+            raise ValueError("Cannot embed query without the query model.")
+        if self.use_context_model:
+            context_logits = self.embed_text(self.context_model,
+                                             context_tokens,
+                                             context_attention_mask,
+                                             context_types)
+        else:
+            raise ValueError("Cannot embed block without the block model.")
+        return query_logits, context_logits
+
+    @staticmethod
+    def embed_text(model, tokens, attention_mask, token_types):
+        """Embed a batch of tokens using the model"""
+        logits = model(tokens,
+                              attention_mask,
+                              token_types)
+        return logits
+
+    def state_dict_for_save_checkpoint(self, destination=None, \
+        prefix='', keep_vars=False):
+        """Save dict with state dicts of each of the models."""
+        state_dict_ = {}
+        if self.biencoder_shared_query_context_model:
+            state_dict_[self._model_key] = \
+                self.model.state_dict_for_save_checkpoint(destination,
+                                                          prefix,
+                                                          keep_vars)
+        else:
+            if self.use_query_model:
+                state_dict_[self._query_key] = \
+                    self.query_model.state_dict_for_save_checkpoint(
+                        destination, prefix, keep_vars)
+
+            if self.use_context_model:
+                state_dict_[self._context_key] = \
+                    self.context_model.state_dict_for_save_checkpoint(
+                        destination, prefix, keep_vars)
+
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Load the state dicts of each of the models"""
+        if self.biencoder_shared_query_context_model:
+            print_rank_0("Loading shared query-context model")
+            self.model.load_state_dict(state_dict[self._model_key], \
+                strict=strict)
+        else:
+            if self.use_query_model:
+                print_rank_0("Loading query model")
+                self.query_model.load_state_dict( \
+                    state_dict[self._query_key], strict=strict)
+
+            if self.use_context_model:
+                print_rank_0("Loading context model")
+                self.context_model.load_state_dict( \
+                    state_dict[self._context_key], strict=strict)
+
+    def init_state_dict_from_bert(self):
+        """Initialize the state from a pretrained BERT model
+        on iteration zero of ICT pretraining"""
+        args = get_args()
+
+        if args.bert_load is None:
+            print_rank_0("bert-load argument is None")
+            return
+
+        tracker_filename = get_checkpoint_tracker_filename(args.bert_load)
+        if not os.path.isfile(tracker_filename):
+            raise FileNotFoundError("Could not find BERT checkpoint")
+        with open(tracker_filename, 'r') as f:
+            iteration = int(f.read().strip())
+            assert iteration > 0
+
+        checkpoint_name = get_checkpoint_name(args.bert_load, iteration, False)
+        if mpu.get_data_parallel_rank() == 0:
+            print('global rank {} is loading BERT checkpoint {}'.format(
+                torch.distributed.get_rank(), checkpoint_name))
+
+        # Load the checkpoint.
+        try:
+            state_dict = torch.load(checkpoint_name, map_location='cpu')
+        except ModuleNotFoundError:
+            from megatron.fp16_deprecated import loss_scaler
+            # For backward compatibility.
+            print_rank_0(' > deserializing using the old code structure ...')
+            sys.modules['fp16.loss_scaler'] = sys.modules[
+                'megatron.fp16_deprecated.loss_scaler']
+            sys.modules['megatron.fp16.loss_scaler'] = sys.modules[
+                'megatron.fp16_deprecated.loss_scaler']
+            state_dict = torch.load(checkpoint_name, map_location='cpu')
+            sys.modules.pop('fp16.loss_scaler', None)
+            sys.modules.pop('megatron.fp16.loss_scaler', None)
+        except BaseException:
+            print_rank_0('could not load the BERT checkpoint')
+            sys.exit()
+
+        checkpoint_version = state_dict.get('checkpoint_version', 0)
+
+        # load the LM state dict into each model
+        model_dict = state_dict['model']['language_model']
+
+        if self.biencoder_shared_query_context_model:
+            self.model.language_model.load_state_dict(model_dict)
+            fix_query_key_value_ordering(self.model, checkpoint_version)
+        else:
+            if self.use_query_model:
+                self.query_model.language_model.load_state_dict(model_dict)
+                # give each model the same ict_head to begin with as well
+                if self.biencoder_projection_dim > 0:
+                    query_proj_state_dict = \
+                        self.state_dict_for_save_checkpoint()\
+                        [self._query_key]['projection_enc']
+                fix_query_key_value_ordering(self.query_model, checkpoint_version)
+
+            if self.use_context_model:
+                self.context_model.language_model.load_state_dict(model_dict)
+                if self.query_model is not None and \
+                    self.biencoder_projection_dim > 0:
+                    self.context_model.projection_enc.load_state_dict\
+                        (query_proj_state_dict)
+                fix_query_key_value_ordering(self.context_model, checkpoint_version)
+
+
+class PretrainedBertModel(MegatronModule):
+    """BERT-based encoder for queries or contexts used for
+    learned information retrieval."""
+
+    def __init__(self, num_tokentypes=2,
+            parallel_output=True):
+        super(PretrainedBertModel, self).__init__()
+
+        args = get_args()
+        tokenizer = get_tokenizer()
+        self.pad_id = tokenizer.pad
+        self.biencoder_projection_dim = args.biencoder_projection_dim
+        self.parallel_output = parallel_output
+        init_method = init_method_normal(args.init_method_std)
+        scaled_init_method = scaled_init_method_normal(
+            args.init_method_std, args.num_layers)
+
+        self.language_model, self._language_model_key = get_language_model(
+            num_tokentypes=num_tokentypes,
+            add_pooler=False,
+            encoder_attn_mask_type=AttnMaskType.padding,
+            init_method=init_method,
+            scaled_init_method=scaled_init_method)
+
+        if args.biencoder_projection_dim > 0:
+            self.projection_enc = get_linear_layer(args.hidden_size,
+                                                   args.biencoder_projection_dim,
+                                                   init_method)
+            self._projection_enc_key = 'projection_enc'
+
+    def forward(self, input_ids, attention_mask, tokentype_ids=None):
+        extended_attention_mask = attention_mask.unsqueeze(1)
+        #extended_attention_mask = bert_extended_attention_mask(attention_mask)
+        position_ids = bert_position_ids(input_ids)
+
+
+        lm_output = self.language_model(input_ids,
+                                        position_ids,
+                                        extended_attention_mask,
+                                        tokentype_ids=tokentype_ids)
+        # This mask will be used in average-pooling and max-pooling
+        pool_mask = (input_ids == self.pad_id).unsqueeze(2)
+
+        # Taking the representation of the [CLS] token of BERT
+        pooled_output = lm_output[:, 0, :]
+
+        # Converting to float16 dtype
+        pooled_output = pooled_output.to(lm_output.dtype)
+
+        # Output.
+        if self.biencoder_projection_dim:
+            pooled_output = self.projection_enc(pooled_output)
+
+        return pooled_output
+
+    def state_dict_for_save_checkpoint(self, destination=None, prefix='',
+                                       keep_vars=False):
+        """For easy load when model is combined with other heads,
+        add an extra key."""
+
+        state_dict_ = {}
+        state_dict_[self._language_model_key] \
+            = self.language_model.state_dict_for_save_checkpoint(
+            destination, prefix, keep_vars)
+
+        if self.biencoder_projection_dim > 0:
+            state_dict_[self._projection_enc_key] = \
+                self.projection_enc.state_dict(destination, prefix, keep_vars)
+
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+        print_rank_0("loading BERT weights")
+        self.language_model.load_state_dict(
+            state_dict[self._language_model_key], strict=strict)
+
+        if self.biencoder_projection_dim > 0:
+            print_rank_0("loading projection head weights")
+            self.projection_enc.load_state_dict(
+                state_dict[self._projection_enc_key], strict=strict)

megatron/model/distributed.py

@@ -13,100 +13,206 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from abc import ABC
+from abc import abstractmethod
+
 import torch
 from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
-import torch.distributed as dist
-from torch.nn.modules import Module
-from torch.autograd import Variable
 
+from megatron import get_args
 from megatron import mpu
 from .module import MegatronModule
 
 
-class DistributedDataParallel(MegatronModule):
 
-    def __init__(self, module):
-        super(DistributedDataParallel, self).__init__()
-        self.warn_on_half = True if dist._backend == dist.dist_backend.GLOO else False
+class MemoryBuffer:
+
+    def __init__(self, numel, dtype):
+        self.numel = numel
+        self.dtype = dtype
+        self.data = torch.zeros(self.numel,
+                                dtype=self.dtype,
+                                device=torch.cuda.current_device(),
+                                requires_grad=False)
+
+
+    def zero(self):
+        """Reset the buffer to zero."""
+        self.data.zero_()
+
+
+    def get(self, shape, start_index):
+        """Return a tensor with the input `shape` as a view into the
+        1-D data starting at `start_index`."""
+        end_index = start_index + shape.numel()
+        assert end_index <= self.numel, \
+            'requested tensor is out of the buffer range.'
+        buffer_tensor = self.data[start_index:end_index]
+        buffer_tensor = buffer_tensor.view(shape)
+        return buffer_tensor
+
+
 
+class DistributedDataParallelBase(MegatronModule, ABC):
+    """Abstract class for DDP."""
+
+    def __init__(self, module):
+        super(DistributedDataParallelBase, self).__init__()
+        # Keep a pointer to the model.
         self.module = module
-        self.data_parallel_group = mpu.get_data_parallel_group()
-
-        def allreduce_params(reduce_after=True, no_scale=False, fp32_allreduce=False):
-            if(self.needs_reduction):
-                self.needs_reduction = False
-                buckets = {}
-                for name, param in self.module.named_parameters():
-                    if param.requires_grad and param.grad is not None:
-                        tp = (param.data.type())
-                        if tp not in buckets:
-                            buckets[tp] = []
-                        buckets[tp].append(param)
-                if self.warn_on_half:
-                    if torch.cuda.HalfTensor in buckets:
-                        print("WARNING: gloo dist backend for half parameters may be extremely slow." +
-                              " It is recommended to use the NCCL backend in this case.")
-                        self.warn_on_half = False
-                for tp in buckets:
-                    bucket = buckets[tp]
-                    grads = [param.grad.data for param in bucket]
-                    coalesced = _flatten_dense_tensors(grads)
-                    if fp32_allreduce:
-                        coalesced = coalesced.float()
-                    if not no_scale and not reduce_after:
-                        coalesced /= dist.get_world_size(group=self.data_parallel_group)
-                    dist.all_reduce(coalesced, group=self.data_parallel_group)
-                    torch.cuda.synchronize()
-                    if not no_scale and reduce_after:
-                        coalesced /= dist.get_world_size(group=self.data_parallel_group)
-                    for buf, synced in zip(grads, _unflatten_dense_tensors(coalesced, grads)):
-                        buf.copy_(synced)
-        self.hook_handles = []
-        self.hooks = []
-        for param in list(self.module.parameters()):
-            def allreduce_hook(*unused):
-                Variable._execution_engine.queue_callback(allreduce_params)
-        #    handle = param.register_hook(allreduce_hook)
-            # self.hooks.append(allreduce_hook)
-            # self.hook_handles.append(handle)
-        self.allreduce_params = allreduce_params
+
+
+    @abstractmethod
+    def allreduce_gradients(self):
+        pass
+
 
     def forward(self, *inputs, **kwargs):
-        self.needs_reduction = True
         return self.module(*inputs, **kwargs)
 
+
     def state_dict(self, destination=None, prefix='', keep_vars=False):
-        #[h.remove() for h in self.hook_handles]
-        sd = self.module.state_dict(destination, prefix, keep_vars)
-       # for handle, hook in zip(self.hook_handles, self.hooks):
-       #     d = handle.hooks_dict_ref()
-       #     d[handle.id] = hook
+        return self.module.state_dict(destination, prefix, keep_vars)
 
-        return sd
 
     def state_dict_for_save_checkpoint(self, destination=None, prefix='',
                                        keep_vars=False):
         return self.module.state_dict_for_save_checkpoint(destination, prefix,
                                                           keep_vars)
 
+
     def load_state_dict(self, state_dict, strict=True):
         self.module.load_state_dict(state_dict, strict=strict)
 
-    '''
-    def _sync_buffers(self):
-        buffers = list(self.module._all_buffers())
-        if len(buffers) > 0:
-            # cross-node buffer sync
-            flat_buffers = _flatten_dense_tensors(buffers)
-            dist.broadcast(flat_buffers, 0)
-            for buf, synced in zip(buffers, _unflatten_dense_tensors(flat_buffers, buffers)):
-                buf.copy_(synced)
-    def train(self, mode=True):
-        # Clear NCCL communicator and CUDA event cache of the default group ID,
-        # These cache will be recreated at the later call. This is currently a
-        # work-around for a potential NCCL deadlock.
-        if dist._backend == dist.dist_backend.NCCL:
-            dist._clear_group_cache()
-        super(DistributedDataParallel, self).train(mode)
-        self.module.train(mode)
-    '''
+
+
+class DistributedDataParallel(DistributedDataParallelBase):
+    """DDP with contiguous buffers options to storre and accumulate gradients.
+    This class:
+        - has the potential to reduce memory fragmentation.
+        - provides the option to do the gradient accumulation
+          in a type other than the params type (for example fp32)
+
+    Arguments:
+        module: input model.
+        accumulate_allreduce_grads_in_fp32: if true do the gradient accumulation
+            and the gradient all-reduce all in in float32. If this option is
+            true, we require `use_contiguous_buffers` to be true too.
+        use_contiguous_buffers: if true, use a contiguous buffer to store the
+            gradients.
+    """
+
+    def __init__(self, module,
+                 accumulate_allreduce_grads_in_fp32,
+                 use_contiguous_buffers):
+
+        super(DistributedDataParallel, self).__init__(module)
+
+        self.accumulate_allreduce_grads_in_fp32 \
+            = accumulate_allreduce_grads_in_fp32
+        self.use_contiguous_buffers = use_contiguous_buffers
+        # If we are using fp32-accumulate-allreduce explicitly
+        # this means we need main grads in a continous buffer.
+        if self.accumulate_allreduce_grads_in_fp32:
+            assert self.use_contiguous_buffers
+
+        # ===================================
+        # Rest of this part applies only to
+        # the case we use continuous buffers.
+        # ===================================
+        self._grad_buffers = None
+        if self.use_contiguous_buffers:
+            self._grad_buffers = {}
+
+            # Simple function to define buffer type.
+            def _get_buffer_type(param):
+                return torch.float if \
+                    self.accumulate_allreduce_grads_in_fp32 else param.dtype
+
+            # First calculate total number of elements per type.
+            type_num_elements = {}
+            for param in self.module.parameters():
+                if param.requires_grad:
+                    dtype = _get_buffer_type(param)
+                    type_num_elements[dtype] = type_num_elements.get(dtype, 0) \
+                                               + param.data.nelement()
+
+            # Allocate the buffer.
+            for dtype, num_elements in type_num_elements.items():
+                self._grad_buffers[dtype] = MemoryBuffer(num_elements, dtype)
+
+            # Assume the back prop order is reverse the params order,
+            # store the start index for the gradients.
+            for param in self.module.parameters():
+                if param.requires_grad:
+                    dtype = _get_buffer_type(param)
+                    type_num_elements[dtype] -= param.data.nelement()
+                    param.main_grad = self._grad_buffers[dtype].get(
+                        param.data.shape, type_num_elements[dtype])
+
+            # Backward hook.
+            # Accumalation function for the gradients. We need
+            # to store them so they don't go out of scope.
+            self.grad_accs = []
+            # Loop over all the parameters in the model.
+            for param in self.module.parameters():
+                if param.requires_grad:
+                    # Expand so we get access to grad_fn.
+                    param_tmp = param.expand_as(param)
+                    # Get the gradient accumulator functtion.
+                    grad_acc = param_tmp.grad_fn.next_functions[0][0]
+                    grad_acc.register_hook(self._make_param_hook(param))
+                    self.grad_accs.append(grad_acc)
+
+
+    def _make_param_hook(self, param):
+        """Create the all-reduce hook for backprop."""
+        # Hook used for back-prop.
+        def param_hook(*unused):
+            # Add the gradient to the buffer.
+            if param.grad.data is not None:
+                param.main_grad.add_(param.grad.data)
+                # Now we can deallocate grad memory.
+                param.grad = None
+        return param_hook
+
+
+    def zero_grad_buffer(self):
+        """Set the grad buffer data to zero. Needs to be called at the
+        begining of each iteration."""
+        assert self._grad_buffers is not None, 'buffers are not initialized.'
+        for _, buffer_ in self._grad_buffers.items():
+            buffer_.zero()
+
+
+    def allreduce_gradients(self):
+        """Reduce gradients across data parallel ranks."""
+        # If we have buffers, simply reduce the data in the buffer.
+        if self._grad_buffers is not None:
+            for _, buffer_ in self._grad_buffers.items():
+                buffer_.data /= mpu.get_data_parallel_world_size()
+                torch.distributed.all_reduce(
+                    buffer_.data, group=mpu.get_data_parallel_group())
+        else:
+            # Otherwise, bucketize and all-reduce
+            buckets = {}
+            # Pack the buckets.
+            for param in self.module.parameters():
+                if param.requires_grad and param.grad is not None:
+                    tp = param.data.type()
+                    if tp not in buckets:
+                        buckets[tp] = []
+                    buckets[tp].append(param)
+                    param.main_grad = param.grad
+
+            # For each bucket, all-reduce and copy all-reduced grads.
+            for tp in buckets:
+                bucket = buckets[tp]
+                grads = [param.grad.data for param in bucket]
+                coalesced = _flatten_dense_tensors(grads)
+                coalesced /= mpu.get_data_parallel_world_size()
+                torch.distributed.all_reduce(
+                    coalesced, group=mpu.get_data_parallel_group())
+                for buf, synced in zip(grads, _unflatten_dense_tensors(
+                        coalesced, grads)):
+                    buf.copy_(synced)

megatron/model/fused_layer_norm.py

@@ -15,29 +15,23 @@
 
 """This code is copied fron NVIDIA apex:
       https://github.com/NVIDIA/apex
-   with minor changes. """
+   with some changes. """
 
-
-import math
-import torch
 import numbers
+import torch
 from torch.nn.parameter import Parameter
 from torch.nn import init
-from torch.nn import functional as F
 import importlib
 
-global fused_layer_norm_cuda
-fused_layer_norm_cuda = None
 global fused_mix_prec_layer_norm_cuda
 fused_mix_prec_layer_norm_cuda = None
 
+
 class FusedLayerNormAffineFunction(torch.autograd.Function):
 
   @staticmethod
   def forward(ctx, input, weight, bias, normalized_shape, eps):
-    global fused_mix_prec_layer_norm_cuda
-    if fused_mix_prec_layer_norm_cuda is None:
-        fused_mix_prec_layer_norm_cuda = importlib.import_module("fused_mix_prec_layer_norm_cuda")
+
     ctx.normalized_shape = normalized_shape
     ctx.eps = eps
     input_ = input.contiguous()
@@ -46,134 +40,51 @@ class FusedLayerNormAffineFunction(torch.autograd.Function):
     output, mean, invvar = fused_mix_prec_layer_norm_cuda.forward_affine(
         input_, ctx.normalized_shape, weight_, bias_, ctx.eps)
     ctx.save_for_backward(input_, weight_, bias_, mean, invvar)
+
     return output
 
+
   @staticmethod
   def backward(ctx, grad_output):
+
     input_, weight_, bias_, mean, invvar = ctx.saved_tensors
     grad_input = grad_weight = grad_bias = None
-    grad_input, grad_weight, grad_bias = fused_mix_prec_layer_norm_cuda.backward_affine(
+    grad_input, grad_weight, grad_bias \
+      = fused_mix_prec_layer_norm_cuda.backward_affine(
         grad_output.contiguous(), mean, invvar,
         input_, ctx.normalized_shape,
         weight_, bias_, ctx.eps)
-    return grad_input, grad_weight, grad_bias, None, None
-
-class FusedLayerNormFunction(torch.autograd.Function):
 
-  @staticmethod
-  def forward(ctx, input, normalized_shape, eps):
-    global fused_layer_norm_cuda
-    if fused_layer_norm_cuda is None:
-        fused_layer_norm_cuda = importlib.import_module("fused_layer_norm_cuda")
-    ctx.normalized_shape = normalized_shape
-    ctx.eps = eps
-    input_ = input.contiguous()
-    output, mean, invvar = fused_layer_norm_cuda.forward(
-        input_, ctx.normalized_shape, ctx.eps)
-    ctx.save_for_backward(input_, mean, invvar)
-    return output
-
-  @staticmethod
-  def backward(ctx, grad_output):
-    input_, mean, invvar = ctx.saved_tensors
-    grad_input = None
-    grad_input = fused_layer_norm_cuda.backward(
-        grad_output.contiguous(), mean, invvar,
-        input_, ctx.normalized_shape,
-        ctx.eps)
-    return grad_input, None, None
+    return grad_input, grad_weight, grad_bias, None, None
 
-def fused_layer_norm_affine(input, normalized_shape, weight, bias, eps=1e-6):
-    return FusedLayerNormAffineFunction.apply(input, weight, bias, normalized_shape, eps)
 
-def fused_layer_norm(input, normalized_shape, eps=1e-6):
-    return FusedLayerNormFunction.apply(input, normalized_shape, eps)
 
 class MixedFusedLayerNorm(torch.nn.Module):
-    r"""Applies Layer Normalization over a mini-batch of inputs as described in
-    the paper `Layer Normalization`_ .
-    Currently only runs on cuda() tensors.
-    .. math::
-        y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta
-    The mean and standard-deviation are calculated separately over the last
-    certain number dimensions which have to be of the shape specified by
-    :attr:`normalized_shape`.
-    :math:`\gamma` and :math:`\beta` are learnable affine transform parameters of
-    :attr:`normalized_shape` if :attr:`elementwise_affine` is ``True``.
-    .. note::
-        Unlike Batch Normalization and Instance Normalization, which applies
-        scalar scale and bias for each entire channel/plane with the
-        :attr:`affine` option, Layer Normalization applies per-element scale and
-        bias with :attr:`elementwise_affine`.
-    This layer uses statistics computed from input data in both training and
-    evaluation modes.
-    Args:
-        normalized_shape (int or list or torch.Size): input shape from an expected input
-            of size
-            .. math::
-                [* \times \text{normalized}\_\text{shape}[0] \times \text{normalized}\_\text{shape}[1]
-                    \times \ldots \times \text{normalized}\_\text{shape}[-1]]
-            If a single integer is used, it is treated as a singleton list, and this module will
-            normalize over the last dimension which is expected to be of that specific size.
-        eps: a value added to the denominator for numerical stability. Default: 1e-5
-        elementwise_affine: a boolean value that when set to ``True``, this module
-            has learnable per-element affine parameters initialized to ones (for weights)
-            and zeros (for biases). Default: ``True``.
-    Shape:
-        - Input: :math:`(N, *)`
-        - Output: :math:`(N, *)` (same shape as input)
-    Examples::
-        >>> input = torch.randn(20, 5, 10, 10)
-        >>> # With Learnable Parameters
-        >>> m = apex.normalization.FusedLayerNorm(input.size()[1:])
-        >>> # Without Learnable Parameters
-        >>> m = apex.normalization.FusedLayerNorm(input.size()[1:], elementwise_affine=False)
-        >>> # Normalize over last two dimensions
-        >>> m = apex.normalization.FusedLayerNorm([10, 10])
-        >>> # Normalize over last dimension of size 10
-        >>> m = apex.normalization.FusedLayerNorm(10)
-        >>> # Activating the module
-        >>> output = m(input)
-    .. _`Layer Normalization`: https://arxiv.org/abs/1607.06450
-    """
-    def __init__(self, normalized_shape, eps=1e-5, elementwise_affine=True):
+
+  def __init__(self, normalized_shape, eps=1e-5):
         super(MixedFusedLayerNorm, self).__init__()
 
-        global fused_layer_norm_cuda
-        fused_layer_norm_cuda = importlib.import_module("fused_layer_norm_cuda")
         global fused_mix_prec_layer_norm_cuda
-        fused_mix_prec_layer_norm_cuda = importlib.import_module("fused_mix_prec_layer_norm_cuda")
-
+        fused_mix_prec_layer_norm_cuda = importlib.import_module(
+          "fused_mix_prec_layer_norm_cuda")
 
         if isinstance(normalized_shape, numbers.Integral):
             normalized_shape = (normalized_shape,)
         self.normalized_shape = torch.Size(normalized_shape)
         self.eps = eps
-        self.elementwise_affine = elementwise_affine
-        if self.elementwise_affine:
-            self.weight = Parameter(torch.Tensor(*normalized_shape))
-            self.bias = Parameter(torch.Tensor(*normalized_shape))
-        else:
-            self.register_parameter('weight', None)
-            self.register_parameter('bias', None)
+        self.weight = Parameter(torch.Tensor(*normalized_shape))
+        self.bias = Parameter(torch.Tensor(*normalized_shape))
         self.reset_parameters()
 
-    def reset_parameters(self):
-        if self.elementwise_affine:
-            init.ones_(self.weight)
-            init.zeros_(self.bias)
-
-    def forward(self, input):
-        if not input.is_cuda:
-            return  F.layer_norm(
-                input, self.normalized_shape, self.weight, self.bias, self.eps)
-        
-        if self.elementwise_affine:
-          return FusedLayerNormAffineFunction.apply(
-              input, self.weight, self.bias, self.normalized_shape,self.eps)
-        else:
-          return FusedLayerNormFunction.apply(input, self.normalized_shape, self.eps) 
-
-    def extra_repr(self):
-        return '{normalized_shape}, eps={eps}, ' \
-            'elementwise_affine={elementwise_affine}'.format(**self.__dict__)
+
+  def reset_parameters(self):
+
+    init.ones_(self.weight)
+    init.zeros_(self.bias)
+
+
+  def forward(self, input):
+
+    return FusedLayerNormAffineFunction.apply(
+      input, self.weight, self.bias, self.normalized_shape,self.eps)
+

megatron/model/fused_softmax.py

@@ -96,6 +96,7 @@ class FusedScaleMaskSoftmax(torch.nn.Module):
     def __init__(
         self,
         input_in_fp16,
+        input_in_bf16,
         attn_mask_type,
         scaled_masked_softmax_fusion,
         mask_func,
@@ -104,6 +105,10 @@ class FusedScaleMaskSoftmax(torch.nn.Module):
     ):
         super(FusedScaleMaskSoftmax, self).__init__()
         self.input_in_fp16 = input_in_fp16
+        self.input_in_bf16 = input_in_bf16
+        assert not (self.input_in_fp16 and self.input_in_bf16),\
+            'both fp16 and bf16 flags cannot be active at the same time.'
+        self.input_in_float16 = self.input_in_fp16 or self.input_in_bf16
         self.attn_mask_type = attn_mask_type
         self.scaled_masked_softmax_fusion = scaled_masked_softmax_fusion
         self.mask_func = mask_func
@@ -128,8 +133,8 @@ class FusedScaleMaskSoftmax(torch.nn.Module):
             query_seq_len % 4 == 0 and attn_batch_size % 4 == 0
 
         # invoke custom kernel
-        if self.input_in_fp16 and mask is not None and \
-           custom_kernel_constraint and self.scaled_masked_softmax_fusion:
+        if self.input_in_float16 and mask is not None and \
+            custom_kernel_constraint and self.scaled_masked_softmax_fusion:
             scale = self.scale if self.scale is not None else 1.0
 
             if self.attn_mask_type == AttnMaskType.causal:
@@ -142,7 +147,7 @@ class FusedScaleMaskSoftmax(torch.nn.Module):
                 assert self.attn_mask_type == AttnMaskType.padding
                 probs = ScaledMaskedSoftmax.apply(input, mask, scale)
         else:
-            if self.input_in_fp16 and self.softmax_in_fp32:
+            if self.input_in_float16 and self.softmax_in_fp32:
                 input = input.float()
 
             if self.scale is not None:
@@ -150,7 +155,10 @@ class FusedScaleMaskSoftmax(torch.nn.Module):
             mask_output = self.mask_func(input, mask) if mask is not None else input
             probs = torch.nn.Softmax(dim=-1)(mask_output)
 
-            if self.input_in_fp16 and self.softmax_in_fp32:
-                probs = probs.half()
+            if self.input_in_float16 and self.softmax_in_fp32:
+                if self.input_in_fp16:
+                    probs = probs.half()
+                else:
+                    probs = probs.bfloat16()
 
         return probs

megatron/model/module.py

@@ -25,6 +25,7 @@ from megatron import mpu
 
 _FLOAT_TYPES = (torch.FloatTensor, torch.cuda.FloatTensor)
 _HALF_TYPES = (torch.HalfTensor, torch.cuda.HalfTensor)
+_BF16_TYPES = (torch.BFloat16Tensor, torch.cuda.BFloat16Tensor)
 
 
 
@@ -50,9 +51,9 @@ class MegatronModule(torch.nn.Module):
 
 
     def word_embeddings_weight(self):
-        if mpu.is_pipeline_first_stage():
+        if mpu.is_pipeline_first_stage(ignore_virtual=True):
             return self.language_model.embedding.word_embeddings.weight
-        if mpu.is_pipeline_last_stage():
+        if mpu.is_pipeline_last_stage(ignore_virtual=True):
             if not self.share_word_embeddings:
                 raise Exception('word_embeddings_weight() called for last '
                                 'stage, but share_word_embeddings is false')
@@ -109,6 +110,7 @@ class MegatronModule(torch.nn.Module):
                   "this needs to be handled manually. If you are training "
                   "something is definitely wrong.")
 
+
 def conversion_helper(val, conversion):
     """Apply conversion to val. Recursively apply conversion if `val`
     #is a nested tuple/list structure."""
@@ -120,44 +122,56 @@ def conversion_helper(val, conversion):
     return rtn
 
 
-def fp32_to_fp16(val):
-    """Convert fp32 `val` to fp16"""
+def fp32_to_float16(val, float16_convertor):
+    """Convert fp32 `val` to fp16/bf16"""
     def half_conversion(val):
         val_typecheck = val
         if isinstance(val_typecheck, (Parameter, Variable)):
             val_typecheck = val.data
         if isinstance(val_typecheck, _FLOAT_TYPES):
-            val = val.half()
+            val = float16_convertor(val)
         return val
     return conversion_helper(val, half_conversion)
 
 
-def fp16_to_fp32(val):
-    """Convert fp16 `val` to fp32"""
+def float16_to_fp32(val):
+    """Convert fp16/bf16 `val` to fp32"""
     def float_conversion(val):
         val_typecheck = val
         if isinstance(val_typecheck, (Parameter, Variable)):
             val_typecheck = val.data
-        if isinstance(val_typecheck, _HALF_TYPES):
+        if isinstance(val_typecheck, (_BF16_TYPES, _HALF_TYPES)):
             val = val.float()
         return val
     return conversion_helper(val, float_conversion)
 
 
 
-class FP16Module(MegatronModule):
+class Float16Module(MegatronModule):
+
+    def __init__(self, module, args):
+        super(Float16Module, self).__init__()
+
+        if args.fp16:
+            self.add_module('module', module.half())
+            def float16_convertor(val):
+                return val.half()
+        elif args.bf16:
+            self.add_module('module', module.bfloat16())
+            def float16_convertor(val):
+                return val.bfloat16()
+        else:
+            raise Exception('should not be here')
 
-    def __init__(self, module):
-        super(FP16Module, self).__init__()
-        self.add_module('module', module.half())
+        self.float16_convertor = float16_convertor
 
 
     def forward(self, *inputs, **kwargs):
         if mpu.is_pipeline_first_stage():
-            inputs = fp32_to_fp16(inputs)
+            inputs = fp32_to_float16(inputs, self.float16_convertor)
         outputs = self.module(*inputs, **kwargs)
         if mpu.is_pipeline_last_stage():
-            outputs = fp16_to_fp32(outputs)
+            outputs = float16_to_fp32(outputs)
         return outputs
 
 

megatron/model/transformer.py

@@ -22,7 +22,7 @@ from megatron import get_args
 from megatron import mpu
 from .module import MegatronModule
 from megatron.model.enums import AttnMaskType, LayerType, AttnType
-from megatron.model import import_layernorm
+from megatron.model import LayerNorm
 from megatron.model.fused_softmax import FusedScaleMaskSoftmax
 from megatron.model.fused_bias_gelu import bias_gelu_impl
 from megatron.model.utils import attention_mask_func, openai_gelu, erf_gelu
@@ -116,6 +116,7 @@ class ParallelAttention(MegatronModule):
         super(ParallelAttention, self).__init__()
         args = get_args()
         self.fp16 = args.fp16
+        self.bf16 = args.bf16
 
         self.apply_query_key_layer_scaling = args.apply_query_key_layer_scaling
         self.attention_softmax_in_fp32 = args.attention_softmax_in_fp32
@@ -164,7 +165,7 @@ class ParallelAttention(MegatronModule):
             self.norm_factor *= coeff
 
         self.scale_mask_softmax = FusedScaleMaskSoftmax(
-            self.fp16,
+            self.fp16, self.bf16,
             self.attn_mask_type,
             args.masked_softmax_fusion,
             attention_mask_func,
@@ -397,8 +398,10 @@ class ParallelTransformerLayer(MegatronModule):
         self.apply_residual_connection_post_layernorm \
             = args.apply_residual_connection_post_layernorm
 
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+
         # Layernorm on the input data.
-        LayerNorm = import_layernorm(args.fp32_residual_connection)
         self.input_layernorm = LayerNorm(
             args.hidden_size,
             eps=args.layernorm_epsilon)
@@ -533,6 +536,7 @@ class ParallelTransformer(MegatronModule):
         super(ParallelTransformer, self).__init__()
         args = get_args()
 
+        self.bf16 = args.bf16
         self.fp32_residual_connection = args.fp32_residual_connection
 
         # Store activation checkpoiting flag.
@@ -552,13 +556,32 @@ class ParallelTransformer(MegatronModule):
                 layer_number,
                 layer_type=layer_type,
                 self_attn_mask_type=self_attn_mask_type)
-        offset = mpu.get_pipeline_model_parallel_rank() * self.num_layers
+        if args.virtual_pipeline_model_parallel_size is not None:
+            assert args.num_layers % args.virtual_pipeline_model_parallel_size == 0, \
+                'num_layers_per_stage must be divisible by ' \
+                'virtual_pipeline_model_parallel_size'
+            # Number of layers in each model chunk is the number of layers in the stage,
+            # divided by the number of model chunks in a stage.
+            self.num_layers = self.num_layers // args.virtual_pipeline_model_parallel_size
+            # With 8 layers, 2 stages, and 4 model chunks, we want an assignment of
+            # layers to stages like (each list is a model chunk):
+            # Stage 0: [0]  [2]  [4]  [6]
+            # Stage 1: [1]  [3]  [5]  [7]
+            # With 8 layers, 2 stages, and 2 virtual stages, we want an assignment of
+            # layers to stages like (each list is a model chunk):
+            # Stage 0: [0, 1]  [4, 5]
+            # Stage 1: [2, 3]  [6, 7]
+            offset = mpu.get_virtual_pipeline_model_parallel_rank() * (
+                    args.num_layers // args.virtual_pipeline_model_parallel_size) + \
+                (mpu.get_pipeline_model_parallel_rank() * self.num_layers)
+        else:
+            # Each stage gets a contiguous set of layers.
+            offset = mpu.get_pipeline_model_parallel_rank() * self.num_layers
         self.layers = torch.nn.ModuleList(
             [build_layer(i + 1 + offset) for i in range(self.num_layers)])
 
         if mpu.is_pipeline_last_stage():
             # Final layer norm before output.
-            LayerNorm = import_layernorm(args.fp32_residual_connection)
             self.final_layernorm = LayerNorm(
                 args.hidden_size,
                 eps=args.layernorm_epsilon)

megatron/mpu/__init__.py

@@ -38,6 +38,7 @@ from .initialize import get_pipeline_model_parallel_next_rank
 from .initialize import get_pipeline_model_parallel_prev_rank
 from .initialize import get_tensor_model_parallel_world_size, set_tensor_model_parallel_world_size
 from .initialize import get_pipeline_model_parallel_world_size, set_pipeline_model_parallel_world_size
+from .initialize import get_virtual_pipeline_model_parallel_rank, set_virtual_pipeline_model_parallel_rank
 from .initialize import initialize_model_parallel
 from .initialize import model_parallel_is_initialized
 
@@ -58,6 +59,8 @@ from .random import get_cuda_rng_tracker
 from .random import init_checkpointed_activations_memory_buffer
 from .random import model_parallel_cuda_manual_seed
 from .random import reset_checkpointed_activations_memory_buffer
+from .random import gather_split_1d_tensor
+from .random import split_tensor_into_1d_equal_chunks
 
 from .utils import divide
 from .utils import split_tensor_along_last_dim

megatron/mpu/initialize.py

@@ -32,6 +32,9 @@ _EMBEDDING_GROUP = None
 # Data parallel group that the current rank belongs to.
 _DATA_PARALLEL_GROUP = None
 
+_VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = None
+_VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
+
 # These values enable us to change the mpu sizes on the fly.
 _MPU_TENSOR_MODEL_PARALLEL_WORLD_SIZE = None
 _MPU_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = None
@@ -48,7 +51,8 @@ def is_unitialized():
 
 
 def initialize_model_parallel(tensor_model_parallel_size_=1,
-                              pipeline_model_parallel_size_=1):
+                              pipeline_model_parallel_size_=1,
+                              virtual_pipeline_model_parallel_size_=None):
     """
     Initialize model data parallel groups.
 
@@ -91,6 +95,12 @@ def initialize_model_parallel(tensor_model_parallel_size_=1,
     num_pipeline_model_parallel_groups = world_size // pipeline_model_parallel_size
     num_data_parallel_groups = world_size // data_parallel_size
 
+    if virtual_pipeline_model_parallel_size_ is not None:
+        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+        global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+        _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = 0
+        _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE = virtual_pipeline_model_parallel_size_
+
     rank = torch.distributed.get_rank()
 
     # Build the data-parallel groups.
@@ -258,17 +268,46 @@ def get_pipeline_model_parallel_rank():
     return torch.distributed.get_rank(group=get_pipeline_model_parallel_group())
 
 
-def is_pipeline_first_stage():
+def is_pipeline_first_stage(ignore_virtual=False):
     """Return True if in the first pipeline model-parallel stage, False otherwise."""
+    if not ignore_virtual:
+        if get_virtual_pipeline_model_parallel_world_size() is not None and \
+            get_virtual_pipeline_model_parallel_rank() != 0:
+            return False
     return get_pipeline_model_parallel_rank() == 0
 
 
-def is_pipeline_last_stage():
+def is_pipeline_last_stage(ignore_virtual=False):
     """Return True if in the last pipeline model-parallel stage, False otherwise."""
+    if not ignore_virtual:
+        virtual_pipeline_model_parallel_world_size = \
+            get_virtual_pipeline_model_parallel_world_size()
+        if virtual_pipeline_model_parallel_world_size is not None and \
+            get_virtual_pipeline_model_parallel_rank() != (
+                virtual_pipeline_model_parallel_world_size - 1):
+            return False
     return get_pipeline_model_parallel_rank() == (
         get_pipeline_model_parallel_world_size() - 1)
 
 
+def get_virtual_pipeline_model_parallel_rank():
+    """Return the virtual pipeline-parallel rank."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+
+
+def set_virtual_pipeline_model_parallel_rank(rank):
+    """Set the virtual pipeline-parallel rank."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK
+    _VIRTUAL_PIPELINE_MODEL_PARALLEL_RANK = rank
+
+
+def get_virtual_pipeline_model_parallel_world_size():
+    """Return the virtual pipeline-parallel world size."""
+    global _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+    return _VIRTUAL_PIPELINE_MODEL_PARALLEL_WORLD_SIZE
+
+
 def get_tensor_model_parallel_src_rank():
     """Calculate the global rank corresponding to the first local rank
     in the tensor model parallel group."""
@@ -276,11 +315,13 @@ def get_tensor_model_parallel_src_rank():
     local_world_size = get_tensor_model_parallel_world_size()
     return (global_rank // local_world_size) * local_world_size
 
+
 def get_pipeline_model_parallel_first_rank():
     assert _PIPELINE_GLOBAL_RANKS is not None, \
         "Pipeline parallel group is not initialized"
     return _PIPELINE_GLOBAL_RANKS[0]
 
+
 def get_pipeline_model_parallel_last_rank():
     assert _PIPELINE_GLOBAL_RANKS is not None, \
         "Pipeline parallel group is not initialized"
@@ -294,6 +335,7 @@ def get_pipeline_model_parallel_next_rank():
     world_size = get_pipeline_model_parallel_world_size()
     return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline + 1) % world_size]
 
+
 def get_pipeline_model_parallel_prev_rank():
     assert _PIPELINE_GLOBAL_RANKS is not None, \
         "Pipeline parallel group is not initialized"
@@ -301,6 +343,7 @@ def get_pipeline_model_parallel_prev_rank():
     world_size = get_pipeline_model_parallel_world_size()
     return _PIPELINE_GLOBAL_RANKS[(rank_in_pipeline - 1) % world_size]
 
+
 def get_data_parallel_world_size():
     """Return world size for the data parallel group."""
     return torch.distributed.get_world_size(group=get_data_parallel_group())

megatron/optimizer/__init__.py

@@ -17,33 +17,32 @@ from apex.optimizers import FusedAdam as Adam
 from apex.optimizers import FusedSGD as SGD
 
 from megatron import get_args
-from megatron.model import import_layernorm
+from megatron.model import LayerNorm
 
 from .grad_scaler import ConstantGradScaler, DynamicGradScaler
-from .optimizer import FP16OptimizerWithFP16Params, FP32Optimizer
+from .optimizer import Float16OptimizerWithFloat16Params, FP32Optimizer
 
 
-def _get_params_for_weight_decay_optimization(module):
+def _get_params_for_weight_decay_optimization(modules):
     """Divide params into with-weight-decay and without-weight-decay groups.
     Layernorms and baises will have no weight decay but the rest will.
     """
-    args = get_args()
-    LayerNorm = import_layernorm(args.fp32_residual_connection)
 
     weight_decay_params = {'params': []}
     no_weight_decay_params = {'params': [], 'weight_decay': 0.0}
-    for module_ in module.modules():
-        if isinstance(module_, LayerNorm):
-            no_weight_decay_params['params'].extend(
-                [p for p in list(module_._parameters.values())
-                 if p is not None])
-        else:
-            weight_decay_params['params'].extend(
-                [p for n, p in list(module_._parameters.items())
-                 if p is not None and n != 'bias'])
-            no_weight_decay_params['params'].extend(
-                [p for n, p in list(module_._parameters.items())
-                 if p is not None and n == 'bias'])
+    for module in modules:
+        for module_ in module.modules():
+            if isinstance(module_, LayerNorm):
+                no_weight_decay_params['params'].extend(
+                    [p for p in list(module_._parameters.values())
+                     if p is not None])
+            else:
+                weight_decay_params['params'].extend(
+                    [p for n, p in list(module_._parameters.items())
+                     if p is not None and n != 'bias'])
+                no_weight_decay_params['params'].extend(
+                    [p for n, p in list(module_._parameters.items())
+                     if p is not None and n == 'bias'])
 
     return weight_decay_params, no_weight_decay_params
 
@@ -66,24 +65,45 @@ def get_megatron_optimizer(model):
                         momentum=args.sgd_momentum)
     else:
         raise Exception('{} optimizer is not supported.'.format(
-                args.optimizer))
+            args.optimizer))
+
+    # Determine whether the params have main-grad field.
+    params_have_main_grad = False
+    if args.DDP_impl == 'local':
+        params_have_main_grad = True
 
-    if args.fp16:
+    if args.fp16 or args.bf16:
+
+        # Grad scaler:
+        #    if loss-scale is provided, instantiate the constant scaler.
+        #    if we are using fp16 and loss-scale is not present, use a
+        #       dynamic scaler.
+        #    otherwise we are running in bf16 with no loss-scale so
+        #       leave it as None.
+        grad_scaler = None
         # Constant loss scale.
         if args.loss_scale:
             grad_scaler = ConstantGradScaler(args.loss_scale)
         # Dynamic loss scale.
         else:
-            grad_scaler = DynamicGradScaler(
-                initial_scale=args.initial_loss_scale,
-                min_scale=args.min_loss_scale,
-                growth_factor=2.0,
-                backoff_factor=0.5,
-                growth_interval=args.loss_scale_window,
-                hysteresis=args.hysteresis)
+            if args.fp16:
+                grad_scaler = DynamicGradScaler(
+                    initial_scale=args.initial_loss_scale,
+                    min_scale=args.min_loss_scale,
+                    growth_factor=2.0,
+                    backoff_factor=0.5,
+                    growth_interval=args.loss_scale_window,
+                    hysteresis=args.hysteresis)
+
         # Megatron optimizer.
-        return FP16OptimizerWithFP16Params(optimizer, grad_scaler,
-                                           args.clip_grad)
+        return Float16OptimizerWithFloat16Params(optimizer,
+                                                 args.clip_grad,
+                                                 args.log_num_zeros_in_grad,
+                                                 params_have_main_grad,
+                                                 args.bf16,
+                                                 grad_scaler)
 
     # FP32.
-    return FP32Optimizer(optimizer, args.clip_grad)
+    return FP32Optimizer(optimizer, args.clip_grad,
+                         args.log_num_zeros_in_grad,
+                         params_have_main_grad)

megatron/optimizer/clip_grads.py

@@ -118,3 +118,31 @@ def clip_grad_norm_fp32(parameters, max_norm, norm_type=2):
                              clip_coeff)
 
     return total_norm
+
+
+def count_zeros_fp32(parameters):
+
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+
+    # Filter parameters based on:
+    #   - grad should not be none
+    #   - parameter should not be shared
+    #   - should not be a replica due to tensor model parallelism
+    total_num_zeros = 0.0
+    for param in parameters:
+        grad_not_none = param.grad is not None
+        is_not_shared = param_is_not_shared(param)
+        is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
+        if grad_not_none and is_not_shared and is_not_tp_duplicate:
+            grad = param.grad.detach()
+            num_zeros = grad.numel() - torch.count_nonzero(grad)
+            total_num_zeros = num_zeros + total_num_zeros
+
+    # Sum across all model-parallel GPUs.
+    torch.distributed.all_reduce(total_num_zeros,
+                                 op=torch.distributed.ReduceOp.SUM,
+                                 group=mpu.get_model_parallel_group())
+    total_num_zeros = total_num_zeros.item()
+
+    return total_num_zeros

megatron/utils.py

@@ -18,6 +18,7 @@
 import sys
 
 import torch
+from torch.nn.parallel import DistributedDataParallel as torchDDP
 
 from apex.multi_tensor_apply import multi_tensor_applier
 import amp_C
@@ -26,20 +27,41 @@ from megatron import get_args
 from megatron import print_rank_0
 from megatron import get_adlr_autoresume
 from megatron import mpu
-from megatron.checkpointing import save_checkpoint
 from megatron.model.module import param_is_not_shared
 from megatron.mpu.layers import param_is_not_tensor_parallel_duplicate
 
 
+def unwrap_model(model, module_instances=(torchDDP)):
+    return_list = True
+    if not isinstance(model, list):
+        model = [model]
+        return_list = False
+    unwrapped_model = []
+    for model_module in model:
+        while isinstance(model_module, module_instances):
+            model_module = model_module.module
+        unwrapped_model.append(model_module)
+    if not return_list:
+        return unwrapped_model[0]
+    return unwrapped_model
+
+
 def calc_params_l2_norm(model):
     """Calculate l2 norm of parameters """
+    args = get_args()
+    if not isinstance(model, list):
+        model = [model]
     # Remove duplicate params.
     params_data = []
-    for param in model.parameters():
-        is_not_shared = param_is_not_shared(param)
-        is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
-        if is_not_shared and is_not_tp_duplicate:
-            params_data.append(param.data)
+    for model_ in model:
+        for param in model_.parameters():
+            is_not_shared = param_is_not_shared(param)
+            is_not_tp_duplicate = param_is_not_tensor_parallel_duplicate(param)
+            if is_not_shared and is_not_tp_duplicate:
+                if args.bf16:
+                    params_data.append(param.data.float())
+                else:
+                    params_data.append(param.data)
     # Calculate norm
     dummy_overflow_buf = torch.cuda.IntTensor([0])
     norm, _ = multi_tensor_applier(
@@ -106,6 +128,8 @@ def print_params_min_max_norm(optimizer, iteration):
 def check_adlr_autoresume_termination(iteration, model,
                                       optimizer, lr_scheduler):
     """Check for autoresume signal and exit if it is received."""
+    from megatron.checkpointing import save_checkpoint
+
     args = get_args()
     autoresume = get_adlr_autoresume()
     # Add barrier to ensure consistnecy.

pretrain_bert.py

@@ -38,7 +38,7 @@ def model_provider():
 
     args = get_args()
     num_tokentypes = 2 if args.bert_binary_head else 0
-    if mpu.get_pipeline_model_parallel_world_size() > 1:
+    def model_provider_pipelined():
         # Determine model based on position of stage in pipeline.
         if mpu.is_pipeline_first_stage():
             model = BertModelFirstStage(
@@ -51,6 +51,17 @@ def model_provider():
         else:
             model = BertModelIntermediateStage(
                 num_tokentypes=num_tokentypes)
+        return model
+
+    args = get_args()
+    if mpu.get_pipeline_model_parallel_world_size() > 1:
+        if args.virtual_pipeline_model_parallel_size is not None:
+            model = []
+            for i in range(args.virtual_pipeline_model_parallel_size):
+                mpu.set_virtual_pipeline_model_parallel_rank(i)
+                model.append(model_provider_pipelined())
+        else:
+            model = model_provider_pipelined()
     else:
         model = BertModel(
             num_tokentypes=num_tokentypes,
@@ -92,8 +103,8 @@ def forward_step(data_iterator, model, input_tensor):
 
     # Get the batch.
     timers('batch-generator').start()
-    tokens, types, sentence_order, loss_mask, lm_labels, padding_mask \
-        = get_batch(data_iterator)
+    tokens, types, sentence_order, loss_mask, lm_labels, padding_mask = get_batch(
+        data_iterator)
     timers('batch-generator').stop()
 
     if not args.bert_binary_head: