Merge branch 'lmcafee/retro' into 'main'

Retro

See merge request ADLR/megatron-lm!489

.gitignore

@@ -3,3 +3,4 @@ __pycache__
 build
 .coverage_*
 *.egg-info
+*~

README.md

@@ -42,7 +42,9 @@ The following table shows both model (MFU) and hardware (HFU) FLOPs utilization
       * [Distributed Pretraining](#distributed-pretraining)
       * [Activation Checkpointing and Recomputation](#activation-checkpointing-and-recomputation)
       * [Distributed Optimizer](#distributed-optimizer)
+      * [FlashAttention](#flashattention)
       * [GPT-3 Example](#gpt-3-example)
+      * [Retro](#retro)
    * [Evaluation and Tasks](#evaluation-and-tasks)
       * [GPT Text Generation](#gpt-text-generation)
       * [GPT Evaluation](#gpt-evaluation)
@@ -323,6 +325,19 @@ In `examples/pretrain_gpt3_175B.sh` we have provided an example of how to config
 With full global batch size of 1536 on 1024 A100 GPUs, each iteration takes around 32 seconds resulting in 138 teraFLOPs per GPU which is 44% of the theoretical peak FLOPs.
 
 
+## Retro
+
+See:
+
+- `tools/retro/README.md` for an overview.
+- `tools/retro/examples/get_preprocess_cmd.sh` for an example of common preprocessing arguments.
+- `tools/retro/examples/preprocess_data.sh` for an example of how to preprocess data.
+- `tools/retro/examples/pretrain_model.sh` for an example of how to pretrain a model.
+
+Retro is a retrieval-enhanced model that is based on GPT. As described in [Improving language models by retrieving from trillions of tokens](https://arxiv.org/abs/2112.04426), Retro retrieves from a database of document chunks by performing locality search using a sample's tokens. The retrieval database can be large -- often billions or even trillions of tokens -- and provides a more efficient storage mechanism of factual knowledge, when compared to storing factual knowledge implicitly within the network's parameters.
+
+Using Retro requires two steps: 1) preprocessing the retrieval database and pretraining neighbors, and 2) pretraining a model using this data. Please see `tools/retro/README.md` for a detailed overview.
+
 <!--
 ## REALM Pipeline
 We are working on implementing the [REALM](https://arxiv.org/pdf/2002.08909.pdf) system. The following sections (will) reflect the three stages of training it. For now it's just the ICT code.

examples/detxoify_lm/finetune_gpt.py

@@ -11,10 +11,10 @@ import sys
 sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
                                              os.path.pardir, os.path.pardir)))
 from megatron import get_args
-from megatron import print_rank_0
 from megatron import get_timers
 from megatron import get_tokenizer
-from megatron import mpu
+from megatron import print_rank_0
+from megatron.core import mpu
 from megatron.data.blendable_dataset import BlendableDataset
 from megatron.data.gpt_dataset import build_train_valid_test_datasets
 from megatron.model import GPTModel, ModelType

examples/detxoify_lm/generate_samples_gpt.py

@@ -10,10 +10,10 @@ sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
                                              os.path.pardir, os.path.pardir)))
 import torch
 from megatron import get_args
-from megatron import print_rank_0
 from megatron import get_tokenizer
-from megatron import mpu
+from megatron import print_rank_0
 from megatron.checkpointing import load_checkpoint
+from megatron.core import mpu
 from megatron.initialize import initialize_megatron
 from megatron.model import GPTModel
 from megatron.training import get_model

megatron/__init__.py

 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
 import torch
 
-from .global_vars import get_args
+from .global_vars import get_args, get_retro_args
 from .global_vars import get_current_global_batch_size
 from .global_vars import get_num_microbatches
 from .global_vars import get_signal_handler

megatron/arguments.py

@@ -3,9 +3,14 @@
 """Megatron arguments."""
 
 import argparse
+import json
 import os
-
 import torch
+import types
+
+from megatron.global_vars import set_retro_args, get_retro_args
+from tools.retro.utils import get_args_path as get_retro_args_path
+
 
 def parse_args(extra_args_provider=None, ignore_unknown_args=False):
     """Parse all arguments."""
@@ -29,6 +34,7 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False):
     parser = _add_logging_args(parser)
     parser = _add_inference_args(parser)
     parser = _add_transformer_engine_args(parser)
+    parser = _add_retro_args(parser)
 
     # Custom arguments.
     if extra_args_provider is not None:
@@ -333,7 +339,6 @@ def validate_args(args, defaults={}):
     if args.sequence_parallel:
         args.async_tensor_model_parallel_allreduce = False
 
-
     if os.environ.get('CUDA_DEVICE_MAX_CONNECTIONS') != "1":
         if args.sequence_parallel:
             raise RuntimeError(
@@ -344,15 +349,31 @@ def validate_args(args, defaults={}):
                 "Using async gradient all reduce requires setting the environment "
                 "variable CUDA_DEVICE_MAX_CONNECTIONS to 1")
 
+    # Load retro args.
+    if args.retro_workdir:
+        retro_args_path = get_retro_args_path(args.retro_workdir)
+        if os.path.exists(retro_args_path):
+            with open(retro_args_path) as f:
+                retro_args = types.SimpleNamespace(**json.load(f))
+                retro_args.retro_return_doc_ids = args.retro_return_doc_ids
+                retro_args.retro_gpt_retrieved_length = \
+                    args.retro_num_retrieved_chunks * \
+                    retro_args.retro_gpt_chunk_length
+                set_retro_args(retro_args)
+
+    # Print arguments.
+    _print_args("arguments", args)
+    retro_args = get_retro_args()
+    if retro_args and args != retro_args:
+        _print_args("retro arguments", types.SimpleNamespace(**{k:v for k,v in vars(retro_args).items() if k.startswith("retro")}, rank=args.rank))
 
-    _print_args(args)
     return args
 
 
-def _print_args(args):
+def _print_args(title, args):
     """Print arguments."""
     if args.rank == 0:
-        print('------------------------ arguments ------------------------',
+        print(f'------------------------ {title} ------------------------',
               flush=True)
         str_list = []
         for arg in vars(args):
@@ -360,7 +381,7 @@ def _print_args(args):
             str_list.append('  {} {} {}'.format(arg, dots, getattr(args, arg)))
         for arg in sorted(str_list, key=lambda x: x.lower()):
             print(arg, flush=True)
-        print('-------------------- end of arguments ---------------------',
+        print(f'-------------------- end of {title} ---------------------',
               flush=True)
 
 
@@ -403,12 +424,64 @@ def _add_inference_args(parser):
                        help='During inference, if batch-size times '
                        'sequence-length is smaller than this threshold '
                        'then we will not use pipelining, otherwise we will.')
-    
     group.add_argument('--max-tokens-to-oom',
                        type=int, default=12000,
                        help='Maximum number of tokens during inference'
                        'tokens here is # in prompt + # to generate'
                        'Allows us to throw an error before OOM crashes server')
+    group.add_argument('--output-bert-embeddings', action='store_true',
+                       help='Output Bert embeddings (via mean pooling) from '
+                       'model, rather than its binary head output or entire '
+                       'hidden batch.')
+    group.add_argument('--bert-embedder-type', default="megatron",
+                       choices=["megatron", "huggingface"],
+                       help='Select either Megatron or Huggingface as the '
+                       'Bert embedder.')
+
+    return parser
+
+
+def _add_retro_args(parser):
+    group = parser.add_argument_group(title='retro')
+
+    group.add_argument('--retro-workdir', default=None,
+                       help='Retro working directory, which contains the '
+                       'preprocessed data for for pretraining. This directory '
+                       'is built during preprocessing (see '
+                       'tools/retro/README.md), and contains subdirectories '
+                       'for the chunk database and pretraining neighbors.')
+    group.add_argument('--retro-add-retriever',
+                       action='store_true', default=False,
+                       help='Add a retriever to the transformer, for use in '
+                       'pretraining a Retro model.')
+    group.add_argument('--retro-cyclic-train-iters', type=int, default=None,
+                       help='Set number of training iterations for cyclic '
+                       'Retro training.')
+    group.add_argument('--retro-encoder-layers', type=int, default=2,
+                       help='Number of layers to use for the retrieval '
+                       'encoder.')
+    group.add_argument('--retro-encoder-hidden-dropout',
+                       type=float, default=0.1, help='Hidden dropout for '
+                       'retrieval encoder.')
+    group.add_argument('--retro-encoder-attention-dropout',
+                       type=float, default=0.1, help='Attention dropout for '
+                       'retrieval encoder.')
+    group.add_argument("--retro-num-neighbors", type=int, default=2,
+                       help='Number of neighbors to retrieve during '
+                       'pretraining.')
+    group.add_argument("--retro-num-retrieved-chunks", type=int, default=2,
+                       help='Number of chunks to retrieve from the retrieval '
+                       'database.')
+    group.add_argument("--retro-return-doc-ids", action="store_true",
+                       help="Turn this on when preprocessing retro data.")
+
+    # Enforce argument naming convention.
+    for action in group._group_actions:
+        prefix = action.dest.split("_")[0]
+        assert prefix == "retro", \
+            "Retro args must be prefixed with '--retro-*', for consistent " \
+            "styling. Please fix '%s'." % ", ".join(action.option_strings)
+
     return parser
 
 
@@ -775,6 +848,10 @@ def _add_checkpointing_args(parser):
     group.add_argument('--use-checkpoint-args', action='store_true',
                        help='Override any command line arguments with arguments '
                        'from the checkpoint')
+    group.add_argument('--exit-on-missing-checkpoint', action='store_true',
+                       help="If '--load' is set, but checkpoint is not found "
+                       "(e.g., path typo), then exit instead of random "
+                       "initialization.")
 
     return parser
 
@@ -835,6 +912,8 @@ def _add_distributed_args(parser):
     group.add_argument('--distributed-backend', default='nccl',
                        choices=['nccl', 'gloo'],
                        help='Which backend to use for distributed training.')
+    group.add_argument('--distributed-timeout-minutes', type=int, default=10,
+                       help='Timeout minutes for torch.distributed.')
     group.add_argument('--DDP-impl', default='local',
                        choices=['local', 'torch'],
                        help='which DistributedDataParallel implementation '

megatron/checkpointing.py

@@ -508,7 +508,16 @@ def load_checkpoint(model, optimizer, opt_param_scheduler, load_arg='load', stri
                               use_distributed_optimizer=args.use_distributed_optimizer,
                               rank0=False)
 
+    # Checkpoint not loaded.
     if model_state_dict is None:
+
+        # Conditionally exit at this point.
+        if args.exit_on_missing_checkpoint:
+            print_rank_0(">> '--exit-on-missing-checkpoint' set ... exiting. <<")
+            torch.distributed.barrier()
+            sys.exit()
+
+        # Iteration defaults to 0.
         return 0
 
     # set checkpoint version

megatron/data/bert_dataset.py

@@ -19,7 +19,6 @@ from megatron.data.dataset_utils import (
     create_masked_lm_predictions
 )
 
-
 class BertDataset(torch.utils.data.Dataset):
 
     def __init__(self, name, indexed_dataset, data_prefix,
@@ -156,7 +155,9 @@ def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
     # Some checks.
     num_tokens = len(tokens)
     padding_length = max_seq_length - num_tokens
-    assert padding_length >= 0
+    assert padding_length >= 0, \
+        f"num_tokens ({num_tokens}) is greater than " \
+        "max_seq_length ({max_seq_length})."
     assert len(tokentypes) == num_tokens
     assert len(masked_positions) == len(masked_labels)
 

megatron/data/blendable_dataset.py

@@ -50,4 +50,7 @@ class BlendableDataset(torch.utils.data.Dataset):
     def __getitem__(self, idx):
         dataset_idx = self.dataset_index[idx]
         sample_idx = self.dataset_sample_index[idx]
-        return self.datasets[dataset_idx][sample_idx]
+        return {
+            "dataset_idx" : dataset_idx,
+            **self.datasets[dataset_idx][sample_idx],
+        }

megatron/data/dataset_utils.py

@@ -452,7 +452,8 @@ def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
             prefixes[i], data_impl, splits_string,
             datasets_train_valid_test_num_samples[i],
             max_seq_length, masked_lm_prob, short_seq_prob,
-            seed, skip_warmup, binary_head, dataset_type=dataset_type)
+            seed, skip_warmup, binary_head, max_seq_length_dec,
+            dataset_type=dataset_type)
         if train_ds:
             train_datasets.append(train_ds)
         if valid_ds:

megatron/data/gpt_dataset.py

@@ -16,15 +16,18 @@ from megatron.data.dataset_utils import get_train_valid_test_split_
 from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
 
 
-def build_train_valid_test_datasets(data_prefix, data_impl,
-                                    splits_string, train_valid_test_num_samples,
+def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                    train_valid_test_num_samples,
                                     seq_length, seed, skip_warmup,
-                                    train_data_prefix=None, valid_data_prefix=None,
-                                    test_data_prefix=None,):
+                                    train_data_prefix=None,
+                                    valid_data_prefix=None,
+                                    test_data_prefix=None,
+                                    return_doc_ids=False):
     """Build train, valid, and test datasets."""
 
     if data_prefix:
         print_rank_0("Single data path provided for train, valid & test")
+
         # Single dataset.
         if len(data_prefix) == 1:
             return _build_train_valid_test_datasets(data_prefix[0],
@@ -46,7 +49,8 @@ def build_train_valid_test_datasets(data_prefix, data_impl,
             train_ds, valid_ds, test_ds = _build_train_valid_test_datasets(
                 prefixes[i], data_impl, splits_string,
                 datasets_train_valid_test_num_samples[i],
-                seq_length, seed, skip_warmup)
+                seq_length, seed, skip_warmup,
+                return_doc_ids)
             if train_ds:
                 train_datasets.append(train_ds)
             if valid_ds:
@@ -67,6 +71,7 @@ def build_train_valid_test_datasets(data_prefix, data_impl,
 
         return (blending_train_dataset, blending_valid_dataset,
                 blending_test_dataset)
+
     else:
         print_rank_0("Separate data paths provided for train, valid & test. Split string will be ignored.")
 
@@ -74,23 +79,69 @@ def build_train_valid_test_datasets(data_prefix, data_impl,
         # Single dataset.
         if train_data_prefix is not None:
             train_dataset = build_dataset("train", train_data_prefix, data_impl,
-                                        train_valid_test_num_samples[0], seq_length, seed,
-                                        skip_warmup)
+                                          train_valid_test_num_samples[0],
+                                          seq_length, seed, skip_warmup)
 
         if valid_data_prefix is not None:
             valid_dataset = build_dataset("valid", valid_data_prefix, data_impl,
-                                    train_valid_test_num_samples[1], seq_length, seed,
-                                    False)
+                                          train_valid_test_num_samples[1],
+                                          seq_length, seed, False)
 
         if test_data_prefix is not None:
             test_dataset = build_dataset("test", test_data_prefix, data_impl,
-                                    train_valid_test_num_samples[2], seq_length, seed,
-                                    False)
+                                         train_valid_test_num_samples[2],
+                                         seq_length, seed, False)
+
+        return (train_dataset, valid_dataset, test_dataset)
+
+
+def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                     train_valid_test_num_samples,
+                                     seq_length, seed, skip_warmup,
+                                     return_doc_ids=False):
+    """Build train, valid, and test datasets."""
+
+    # Indexed dataset.
+    indexed_dataset = get_indexed_dataset_(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+
+    total_num_of_documents = indexed_dataset.sizes.shape[0]
+    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
+
+    # Print stats about the splits.
+    print_rank_0(' > dataset split:')
+
+    def print_split_stats(name, index):
+        print_rank_0('    {}:'.format(name))
+        print_rank_0('     document indices in [{}, {}) total of {} '
+                     'documents'.format(splits[index], splits[index + 1],
+                                        splits[index + 1] - splits[index]))
+    print_split_stats('train', 0)
+    print_split_stats('validation', 1)
+    print_split_stats('test', 2)
+
+    def build_dataset(index, name):
+        dataset = None
+        if splits[index + 1] > splits[index]:
+            documents = np.arange(start=splits[index], stop=splits[index + 1],
+                                  step=1, dtype=np.int32)
+            dataset = GPTDataset(name, data_prefix,
+                                 documents, indexed_dataset,
+                                 train_valid_test_num_samples[index],
+                                 seq_length, seed,
+                                 return_doc_ids)
+        return dataset
+
+    train_dataset = build_dataset(0, 'train')
+    valid_dataset = build_dataset(1, 'valid')
+    test_dataset = build_dataset(2, 'test')
 
     return (train_dataset, valid_dataset, test_dataset)
 
 
-def build_dataset(dataset_name, data_prefix, data_impl, num_samples, seq_length, seed, skip_warmup):
+def build_dataset(dataset_name, data_prefix, data_impl, num_samples,
+                  seq_length, seed, skip_warmup):
     dataset = None
     if len(data_prefix) == 1:
         dataset = _build_dataset(dataset_name,
@@ -146,49 +197,6 @@ def _build_dataset(dataset_name, data_prefix, data_impl,
     return dataset
 
 
-def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
-                                     train_valid_test_num_samples,
-                                     seq_length, seed, skip_warmup):
-    """Build train, valid, and test datasets."""
-
-    # Indexed dataset.
-    indexed_dataset = get_indexed_dataset_(data_prefix,
-                                           data_impl,
-                                           skip_warmup)
-
-    total_num_of_documents = indexed_dataset.sizes.shape[0]
-    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
-
-    # Print stats about the splits.
-    print_rank_0(' > dataset split:')
-
-    def print_split_stats(name, index):
-        print_rank_0('    {}:'.format(name))
-        print_rank_0('     document indices in [{}, {}) total of {} '
-                     'documents'.format(splits[index], splits[index + 1],
-                                        splits[index + 1] - splits[index]))
-    print_split_stats('train', 0)
-    print_split_stats('validation', 1)
-    print_split_stats('test', 2)
-
-    def build_dataset(index, name):
-        dataset = None
-        if splits[index + 1] > splits[index]:
-            documents = np.arange(start=splits[index], stop=splits[index + 1],
-                                  step=1, dtype=np.int32)
-            dataset = GPTDataset(name, data_prefix,
-                                  documents, indexed_dataset,
-                                  train_valid_test_num_samples[index],
-                                  seq_length, seed)
-        return dataset
-
-    train_dataset = build_dataset(0, 'train')
-    valid_dataset = build_dataset(1, 'valid')
-    test_dataset = build_dataset(2, 'test')
-
-    return (train_dataset, valid_dataset, test_dataset)
-
-
 def get_indexed_dataset_(data_prefix, data_impl, skip_warmup):
     """Build indexed dataset."""
     print_rank_0(' > building dataset index ...')
@@ -208,20 +216,24 @@ def get_indexed_dataset_(data_prefix, data_impl, skip_warmup):
 class GPTDataset(torch.utils.data.Dataset):
 
     def __init__(self, name, data_prefix, documents, indexed_dataset,
-                 num_samples, seq_length, seed):
+                 num_samples, seq_length, seed,
+                 return_doc_ids=False):
 
         self.name = name
         self.indexed_dataset = indexed_dataset
+        self.return_doc_ids = return_doc_ids
 
         # Checks
         assert np.min(documents) >= 0
         assert np.max(documents) < indexed_dataset.sizes.shape[0]
 
         # Build index mappings.
-        self.doc_idx, self.sample_idx, self.shuffle_idx = _build_index_mappings(
-            self.name, data_prefix, documents, self.indexed_dataset.sizes,
+        self.doc_idx, self.sample_idx, self.shuffle_idx, self.index_prefix = \
+            _build_index_mappings(self.name, data_prefix,
+                                  documents, self.indexed_dataset.sizes,
                                   num_samples, seq_length, seed)
 
+
     def __len__(self):
         # -1 is due to data structure used to retieve the index:
         #    sample i --> [sample_idx[i], sample_idx[i+1])
@@ -236,23 +248,32 @@ class GPTDataset(torch.utils.data.Dataset):
         offset_f = self.sample_idx[idx][1]
         offset_l = self.sample_idx[idx + 1][1]
         # If we are within the same document, just extract the chunk.
+        doc_ids = []
         if doc_index_f == doc_index_l:
+            doc_ids.append(self.doc_idx[doc_index_f])
             sample = self.indexed_dataset.get(self.doc_idx[doc_index_f],
                                               offset=offset_f,
                                               length=offset_l - offset_f + 1)
         else:
             # Otherwise, get the rest of the initial document.
+            doc_ids.append(self.doc_idx[doc_index_f])
             sample_list = [self.indexed_dataset.get(self.doc_idx[doc_index_f],
                                                     offset=offset_f)]
             # Loop over all in between documents and add the entire document.
             for i in range(doc_index_f + 1, doc_index_l):
+                doc_ids.append(self.doc_idx[i])
                 sample_list.append(self.indexed_dataset.get(self.doc_idx[i]))
             # And finally add the relevant portion of last document.
+            doc_ids.append(self.doc_idx[doc_index_l])
             sample_list.append(self.indexed_dataset.get(
                 self.doc_idx[doc_index_l],
                 length=offset_l + 1))
             sample = np.concatenate(sample_list)
 
+        if self.return_doc_ids: # for retro preprocessing
+            return {'text': np.array(sample, dtype=np.int64),
+                    'doc_ids': np.array(doc_ids, dtype=np.int64)}
+        else:
             return {'text': np.array(sample, dtype=np.int64)}
 
 
@@ -267,15 +288,16 @@ def _build_index_mappings(name, data_prefix, documents, sizes,
     # Number of tokens in each epoch and number of required epochs.
     tokens_per_epoch = _num_tokens(documents, sizes)
     num_epochs = _num_epochs(tokens_per_epoch, seq_length, num_samples)
+
     # rng state
     np_rng = np.random.RandomState(seed=seed)
 
     # Filename of the index mappings.
-    _filename = data_prefix
-    _filename += '_{}_indexmap'.format(name)
-    _filename += '_{}ns'.format(num_samples)
-    _filename += '_{}sl'.format(seq_length)
-    _filename += '_{}s'.format(seed)
+    index_prefix = '{}_indexmap'.format(name)
+    index_prefix += '_{}ns'.format(num_samples)
+    index_prefix += '_{}sl'.format(seq_length)
+    index_prefix += '_{}s'.format(seed)
+    _filename = data_prefix + '_' + index_prefix
     doc_idx_filename = _filename + '_doc_idx.npy'
     sample_idx_filename = _filename + '_sample_idx.npy'
     shuffle_idx_filename = _filename + '_shuffle_idx.npy'
@@ -343,8 +365,6 @@ def _build_index_mappings(name, data_prefix, documents, sizes,
             assert sizes.dtype == np.int32
             sample_idx = helpers.build_sample_idx(sizes, doc_idx, seq_length,
                                                   num_epochs, tokens_per_epoch)
-            # sample_idx = _build_sample_idx(sizes, doc_idx, seq_length,
-            #                               num_epochs, tokens_per_epoch)
             np.save(sample_idx_filename, sample_idx, allow_pickle=True)
             print_rank_0(' > elasped time to build and save sample-idx mapping '
                          '(seconds): {:4f}'.format(time.time() - start_time))
@@ -389,7 +409,7 @@ def _build_index_mappings(name, data_prefix, documents, sizes,
         sample_idx.shape[0]))
     print_rank_0('    total number of epochs: {}'.format(num_epochs))
 
-    return doc_idx, sample_idx, shuffle_idx
+    return doc_idx, sample_idx, shuffle_idx, index_prefix
 
 
 def _num_tokens(documents, sizes):

megatron/data/indexed_dataset.py

@@ -460,7 +460,7 @@ class MMapIndexedDataset(torch.utils.data.Dataset):
         return self._path
 
     def __setstate__(self, state):
-        self._do_init(state)
+        self._do_init(state, skip_warmup=True)
 
     def _do_init(self, path, skip_warmup):
         self._path = path

megatron/global_vars.py

@@ -12,6 +12,7 @@ from .microbatches import build_num_microbatches_calculator
 from .timers import Timers
 
 _GLOBAL_ARGS = None
+_GLOBAL_RETRO_ARGS = None
 _GLOBAL_NUM_MICROBATCHES_CALCULATOR = None
 _GLOBAL_TOKENIZER = None
 _GLOBAL_TENSORBOARD_WRITER = None
@@ -25,6 +26,11 @@ def get_args():
     return _GLOBAL_ARGS
 
 
+def get_retro_args():
+    """Return retro arguments."""
+    return _GLOBAL_RETRO_ARGS
+
+
 def get_num_microbatches():
     return _GLOBAL_NUM_MICROBATCHES_CALCULATOR.get()
 
@@ -98,6 +104,11 @@ def set_args(args):
     _GLOBAL_ARGS = args
 
 
+def set_retro_args(retro_args):
+    global _GLOBAL_RETRO_ARGS
+    _GLOBAL_RETRO_ARGS = retro_args
+
+
 def _build_num_microbatches_calculator(args):
 
     global _GLOBAL_NUM_MICROBATCHES_CALCULATOR

megatron/initialize.py

@@ -174,7 +174,7 @@ def _initialize_distributed():
     torch.distributed.init_process_group(
         backend=args.distributed_backend,
         world_size=args.world_size, rank=args.rank,
-        timeout=timedelta(minutes=10))
+        timeout=timedelta(minutes=args.distributed_timeout_minutes))
 
     # Set the tensor model-parallel, pipeline model-parallel, and
     # data-parallel communicators.

megatron/model/bert_model.py

@@ -16,6 +16,7 @@ from megatron.model.utils import init_method_normal
 from megatron.model.utils import scaled_init_method_normal
 from .module import MegatronModule
 
+
 def bert_extended_attention_mask(attention_mask):
     # We create a 3D attention mask from a 2D tensor mask.
     # [b, 1, s]
@@ -137,6 +138,10 @@ class BertModel(MegatronModule):
         self.pre_process = pre_process
         self.post_process = post_process
 
+        self.return_embeddings = args.output_bert_embeddings
+        if self.return_embeddings:
+            assert self.post_process and self.add_binary_head
+
         init_method = init_method_normal(args.init_method_std)
         scaled_init_method = scaled_init_method_normal(args.init_method_std,
                                                        args.num_layers)
@@ -182,6 +187,24 @@ class BertModel(MegatronModule):
 
         if self.post_process and self.add_binary_head:
             lm_output, pooled_output = lm_output
+
+            # Return pooled output (e.g., when computing Bert embeddings).
+            if self.return_embeddings:
+
+                # Sum attention mask.
+                embeddings = torch.transpose(lm_output, 0, 1)
+                masks = torch.sum(attention_mask, dim=1)
+
+                # Collect masked embeddings.
+                output = torch.zeros(
+                    size=(embeddings.shape[0], embeddings.shape[2]),
+                    dtype=torch.float32,
+                    device=torch.cuda.current_device())
+                for i, (embedding, mask) in enumerate(zip(embeddings, masks)):
+                    output[i, :] = torch.mean(embedding[1: mask - 1], dim=0)
+
+                return output
+
         else:
             pooled_output = None
 

megatron/model/gpt_model.py

@@ -74,13 +74,17 @@ class GPTModel(MegatronModule):
         """See megatron.model.transformer.set_input_tensor()"""
         self.language_model.set_input_tensor(input_tensor)
 
-    def forward(self, input_ids, position_ids, attention_mask, labels=None,
-                tokentype_ids=None, inference_params=None):
+    def forward(self, input_ids, position_ids, attention_mask,
+                ret_input_ids=None, ret_position_ids=None, ret_attn_mask=None,
+                labels=None, tokentype_ids=None, inference_params=None):
 
         lm_output = self.language_model(
             input_ids,
             position_ids,
             attention_mask,
+            ret_input_ids=ret_input_ids,
+            ret_position_ids=ret_position_ids,
+            ret_attn_mask=ret_attn_mask,
             inference_params=inference_params)
 
         if self.post_process:

megatron/model/language_model.py

@@ -7,11 +7,13 @@ import torch.nn.functional as F
 
 from megatron import get_args
 from megatron.core import mpu, tensor_parallel
+
+from .enums import LayerType, AttnMaskType
 from .module import MegatronModule
-from megatron.model.enums import LayerType, AttnMaskType
-from megatron.model.transformer import ParallelTransformer
-from megatron.model.utils import get_linear_layer
-from megatron.model.utils import init_method_normal, scaled_init_method_normal
+from .retro_transformer import ParallelRetroEncoder, ParallelRetroTransformer
+from .transformer import ParallelTransformer
+from .utils import get_linear_layer
+from .utils import init_method_normal, scaled_init_method_normal
 
 
 def parallel_lm_logits(input_, word_embeddings_weight, parallel_output,
@@ -349,16 +351,38 @@ class TransformerLanguageModel(MegatronModule):
                                        self.num_tokentypes)
             self._embedding_key = 'embedding'
 
-        # Transformer.
+        # Retriever (bi-directional transformer with cross attention)
+        if args.retro_add_retriever:
+            self.retriever = ParallelRetroEncoder(
+                self.init_method,
+                output_layer_init_method,
+                self_attn_mask_type=AttnMaskType.padding,
+                pre_process=self.pre_process,
+                post_process=False,
+            )
+            self._retriever_key = 'retriever'
+        else:
+            self.retriever = None
+
         # Encoder (usually set to True, False if part of an encoder-decoder
         # architecture and in encoder-only stage).
         if self.add_encoder:
+            if args.retro_add_retriever:
+                self.encoder = ParallelRetroTransformer(
+                    self.init_method,
+                    output_layer_init_method,
+                    self_attn_mask_type=self.encoder_attn_mask_type,
+                    pre_process=self.pre_process,
+                    post_process=self.post_process,
+                    retriever=self.retriever,
+                )
+            else:
                 self.encoder = ParallelTransformer(
                     self.init_method,
                     output_layer_init_method,
                     self_attn_mask_type=self.encoder_attn_mask_type,
                     pre_process=self.pre_process,
-                post_process=self.post_process
+                    post_process=self.post_process,
                 )
             self._encoder_key = 'encoder'
         else:
@@ -414,11 +438,19 @@ class TransformerLanguageModel(MegatronModule):
 
     def forward(self, enc_input_ids, enc_position_ids, enc_attn_mask,
                 dec_input_ids=None, dec_position_ids=None, dec_attn_mask=None,
+                ret_input_ids=None, ret_position_ids=None, ret_attn_mask=None,
                 enc_dec_attn_mask=None, tokentype_ids=None,
                 inference_params=None,
                 pooling_sequence_index=0,
                 enc_hidden_states=None, output_enc_hidden=False):
 
+        # Retriever embedding.
+        if self.retriever and self.pre_process:
+            retriever_input = self.embedding(ret_input_ids, ret_position_ids,
+                                             tokentype_ids=tokentype_ids)
+        else:
+            retriever_input = None
+
         # Encoder embedding.
         if self.pre_process:
             encoder_input = self.embedding(enc_input_ids, enc_position_ids,
@@ -429,6 +461,14 @@ class TransformerLanguageModel(MegatronModule):
         # Run encoder.
         if enc_hidden_states is None:
             if self.encoder is not None:
+                if self.retriever:
+                    encoder_output = self.encoder(
+                        encoder_input,
+                        enc_attn_mask,
+                        retriever_output=retriever_input,
+                        retriever_attn_mask=ret_attn_mask,
+                        inference_params=inference_params)
+                else:
                     encoder_output = self.encoder(
                         encoder_input,
                         enc_attn_mask,

megatron/model/retro_transformer.py

+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+
+"""Retro Transformer.
+
+** Special note about this file **
+Many classes and methods in this file directly parallel those in transformer.py
+in name and utility. However, due to 1) subtle changes in the code over time
+(i.e., transposes and contexts), and 2) other code that is soon to be merged,
+this file will *temporarily* remain as is, until a larger integration is
+complete.
+"""
+
+import math
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from megatron import get_args, get_retro_args, get_tensorboard_writer
+from megatron.core import parallel_state
+from megatron.core import tensor_parallel
+from megatron.core import utils as core_utils
+from megatron.model.enums import AttnMaskType, ModelType, LayerType, AttnType
+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, init_method_normal
+
+from .module import MegatronModule
+from .transformer import _get_num_layers, ParallelMLP, NoopTransformerLayer
+
+""" We use the following notation throughout this file:
+     h: hidden size
+     n: number of attention heads
+     p: number of model parallel partitions
+     np: n/p
+     hp: h/p
+     hn: h/n
+     b: batch size
+     s: sequence length
+     l: number of layers
+    Transformer takes input of size [s, b, h] and returns a
+    tensor of the same size. We use the following arguments:
+        hyperparameters: transformer hyperparameters
+"""
+
+
+class DropPath(MegatronModule):
+    """Drop paths (Stochastic Depth) per sample
+    (when applied in main path of residual blocks).
+
+    *Note: differs from transformer.py/DropPath in hidden_state transpose.
+    """
+
+    def __init__(self, drop_prob=0.):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, hidden_state):
+        if self.drop_prob == 0. or not self.training:
+            return hidden_state
+        keep_prob = 1 - self.drop_prob
+        # work with diff dim tensors, not just 2D ConvNets
+        shape = (hidden_state.shape[0],) + (1,) * (hidden_state.ndim - 1)
+        random_tensor = keep_prob + \
+            torch.rand(shape, dtype=hidden_state.dtype, device=hidden_state.device)
+        random_tensor.floor_() # binarize
+        output = hidden_state.div(keep_prob) * random_tensor
+        return output
+
+
+class SwitchMLP(MegatronModule):
+    """
+    Routes input to one of N MLP "experts"
+    """
+    def __init__(self, init_method, output_layer_init_method):
+        super(SwitchMLP, self).__init__()
+        args = get_args()
+        self.router = torch.nn.Linear(args.hidden_size, args.num_experts)
+        self.experts = torch.nn.ModuleList()
+        for i in range(args.num_experts):
+            self.experts.append(ParallelMLP(init_method, output_layer_init_method))
+
+    def forward(self, hidden_states):
+        # hidden_states: [b, s, h]
+        b = hidden_states.size(0)
+        s = hidden_states.size(1)
+        h = hidden_states.size(2)
+        route = self.router(hidden_states)
+        route = torch.nn.functional.softmax(route, dim=2)
+        max_prob, max_ind = torch.max(route, dim=2)
+        max_prob = torch.unsqueeze(max_prob, 2) # [b s 1]
+
+        # TODO (rprenger) TODO this could be made easier to read
+        # Converting [b, s, h] to [b*s, h].
+        # Each vector could be routed differently
+        hidden_states = hidden_states.view(-1, hidden_states.size(2)) # [b*s h]
+        max_prob = max_prob.view(-1, max_prob.size(2)) # [b*s 1]
+        max_ind = max_ind.view(-1) # [b*s]
+
+        output_total = torch.empty_like(hidden_states)
+        output_bias_total = torch.empty_like(hidden_states)
+        #TODO (rprenger) This does each expert in serial, but it could be parallelized
+
+        for expert_num, expert in enumerate(self.experts):
+            local_indices = (max_ind == expert_num).nonzero()
+            hidden = hidden_states[local_indices,:]
+            output, output_bias = expert(hidden)
+            output_bias = output_bias.expand_as(output)
+            output_total[local_indices,:] = output
+            output_bias_total[local_indices,:] = output_bias
+
+        output_total = output_total*max_prob
+        output_bias_total = output_bias_total*max_prob
+        output_total = output_total.view(b, s, h)
+        output_bias_total = output_bias_total.view(b, s, h)
+
+        return output_total, output_bias_total
+
+class ParallelAttention(MegatronModule):
+    """Parallel self-attention layer abstract class.
+
+    Self-attention layer takes input with size [b, s, h]
+    and returns output of the same size.
+    """
+
+    def __init__(self, init_method,
+                 output_layer_init_method, layer_number,
+                 attention_type=AttnType.self_attn,
+                 attn_mask_type=AttnMaskType.padding):
+        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
+        if self.apply_query_key_layer_scaling:
+            self.attention_softmax_in_fp32 = True
+        self.layer_number = max(1, layer_number)
+        self.attention_type = attention_type
+        self.attn_mask_type = attn_mask_type
+        self.params_dtype = args.params_dtype
+
+        projection_size = args.kv_channels * args.num_attention_heads
+
+        # Per attention head and per partition values.
+        world_size = parallel_state.get_tensor_model_parallel_world_size()
+        self.hidden_size_per_partition = core_utils.divide(projection_size,
+                                                           world_size)
+        self.hidden_size_per_attention_head = core_utils.divide(
+            projection_size, args.num_attention_heads)
+        self.num_attention_heads_per_partition = core_utils.divide(
+            args.num_attention_heads, world_size)
+
+        # Strided linear layer.
+        if attention_type == AttnType.self_attn:
+            self.query_key_value = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                3 * projection_size,
+                gather_output=False,
+                init_method=init_method)
+        else:
+            assert attention_type == AttnType.cross_attn
+            self.query = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                projection_size,
+                gather_output=False,
+                init_method=init_method)
+
+            self.key_value = tensor_parallel.ColumnParallelLinear(
+                args.hidden_size,
+                2 * projection_size,
+                gather_output=False,
+                init_method=init_method)
+
+        coeff = None
+        self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
+        if self.apply_query_key_layer_scaling:
+            coeff = self.layer_number
+            self.norm_factor *= coeff
+
+        self.scale_mask_softmax = FusedScaleMaskSoftmax(
+            self.fp16, self.bf16,
+            self.attn_mask_type,
+            args.masked_softmax_fusion,
+            attention_mask_func,
+            self.attention_softmax_in_fp32,
+            coeff)
+
+        # Dropout. Note that for a single iteration, this layer will generate
+        # different outputs on different number of parallel partitions but
+        # on average it should not be partition dependent.
+        self.attention_dropout = torch.nn.Dropout(args.attention_dropout)
+
+        # Output.
+        self.dense = tensor_parallel.RowParallelLinear(
+            projection_size,
+            args.hidden_size,
+            input_is_parallel=True,
+            init_method=output_layer_init_method,
+            skip_bias_add=True)
+
+    def _allocate_memory(self, inference_max_sequence_len, batch_size):
+        return torch.empty(
+            inference_max_sequence_len,
+            batch_size,
+            self.num_attention_heads_per_partition,
+            self.hidden_size_per_attention_head,
+            dtype=self.params_dtype,
+            device=torch.cuda.current_device())
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, inference_params=None):
+        # hidden_states: [sq, b, h]
+
+        # =================================================
+        # Pre-allocate memory for key-values for inference.
+        # =================================================
+        if inference_params:
+            if self.layer_number not in inference_params.key_value_memory_dict:
+                inf_max_seq_len = inference_params.max_sequence_len
+                inf_max_batch_size = inference_params.max_batch_size
+                inference_key_memory = self._allocate_memory(
+                    inf_max_seq_len, inf_max_batch_size)
+                inference_value_memory = self._allocate_memory(
+                    inf_max_seq_len, inf_max_batch_size)
+                inference_params.key_value_memory_dict[self.layer_number] = (
+                    inference_key_memory, inference_value_memory)
+            else:
+                inference_key_memory, inference_value_memory = \
+                    inference_params.key_value_memory_dict[self.layer_number]
+
+        # =====================
+        # Query, Key, and Value
+        # =====================
+
+        if self.attention_type == AttnType.self_attn:
+            # Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
+            mixed_x_layer, _ = self.query_key_value(hidden_states)
+
+            # [sq, b, (np * 3 * hn)] --> [sq, b, np, 3 * hn]
+            new_tensor_shape = mixed_x_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 3 * self.hidden_size_per_attention_head)
+            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
+
+            # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
+            (query_layer,
+             key_layer,
+             value_layer) = tensor_parallel \
+                 .split_tensor_along_last_dim(mixed_x_layer, 3)
+        else:
+            # Attention heads [sk, b, h] --> [sk, b, (np * 2 * hn)]
+            mixed_kv_layer, _ = self.key_value(encoder_output)
+
+            # [sk, b, (np * 2 * hn)] --> [sk, b, np, 2 * hn]
+            new_tensor_shape = mixed_kv_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 2 * self.hidden_size_per_attention_head)
+            mixed_kv_layer = mixed_kv_layer.view(*new_tensor_shape)
+
+            # [sk, b, np, 2 * hn] --> 2 [sk, b, np, hn]
+            (key_layer,
+             value_layer) = tensor_parallel \
+                 .split_tensor_along_last_dim(mixed_kv_layer, 2)
+
+            # Attention head [sq, b, h] --> [sq, b, hp]
+            query_layer, _ = self.query(hidden_states)
+            # [sq, b, hp] --> [sq, b, np, hn]
+            new_tensor_shape = query_layer.size()[:-1] + \
+                (self.num_attention_heads_per_partition,
+                 self.hidden_size_per_attention_head)
+            query_layer = query_layer.view(*new_tensor_shape)
+
+        # ==================================
+        # Adjust key and value for inference
+        # ==================================
+
+        if inference_params:
+            batch_start = inference_params.batch_size_offset
+            batch_end = batch_start + key_layer.size(1)
+            assert batch_end <= inference_key_memory.size(1)
+            sequence_start = inference_params.sequence_len_offset
+            sequence_end = sequence_start + key_layer.size(0)
+            assert sequence_end <= inference_key_memory.size(0)
+            # Copy key and values.
+            inference_key_memory[sequence_start:sequence_end,
+                                 batch_start:batch_end, ...] = key_layer
+            inference_value_memory[sequence_start:sequence_end,
+                                   batch_start:batch_end, ...] = value_layer
+            key_layer = inference_key_memory[
+                :sequence_end, batch_start:batch_end, ...]
+            value_layer = inference_value_memory[
+                :sequence_end, batch_start:batch_end, ...]
+
+        # ===================================
+        # Raw attention scores. [b, np, s, s]
+        # ===================================
+
+        # [b, np, sq, sk]
+        output_size = (query_layer.size(1),
+                       query_layer.size(2),
+                       query_layer.size(0),
+                       key_layer.size(0))
+
+        # [sq, b, np, hn] -> [sq, b * np, hn]
+        query_layer = query_layer.view(output_size[2],
+                                       output_size[0] * output_size[1], -1)
+        # [sk, b, np, hn] -> [sk, b * np, hn]
+        key_layer = key_layer.view(output_size[3],
+                                   output_size[0] * output_size[1], -1)
+
+        # preallocting result tensor: [b * np, sq, sk]
+        matmul_result = torch.empty(
+            output_size[0]*output_size[1],
+            output_size[2],
+            output_size[3],
+            dtype=query_layer.dtype,
+            device=torch.cuda.current_device())
+
+        # Raw attention scores. [b * np, sq, sk]
+        matmul_result = torch.baddbmm(
+            matmul_result,
+            query_layer.transpose(0, 1),   # [b * np, sq, hn]
+            key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
+            beta=0.0, alpha=(1.0/self.norm_factor))
+
+        # change view to [b, np, sq, sk]
+        attention_scores = matmul_result.view(*output_size)
+
+        # ===========================
+        # Attention probs and dropout
+        # ===========================
+
+        # attention scores and attention mask [b, np, sq, sk]
+        attention_probs = self.scale_mask_softmax(attention_scores,
+                                                  attention_mask)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        with tensor_parallel.get_cuda_rng_tracker().fork():
+            attention_probs = self.attention_dropout(attention_probs)
+
+        # =========================
+        # Context layer. [sq, b, hp]
+        # =========================
+
+        # value_layer -> context layer.
+        # [sk, b, np, hn] --> [b, np, sq, hn]
+
+        # context layer shape: [b, np, sq, hn]
+        output_size = (value_layer.size(1),
+                       value_layer.size(2),
+                       query_layer.size(0),
+                       value_layer.size(3))
+
+        # change view [sk, b * np, hn]
+        value_layer = value_layer.view(value_layer.size(0),
+                                       output_size[0] * output_size[1], -1)
+
+        # change view [b * np, sq, sk]
+        attention_probs = attention_probs.view(output_size[0] * output_size[1],
+                                               output_size[2], -1)
+
+        # matmul: [b * np, sq, hn]
+        context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
+
+        # change view [b, np, sq, hn]
+        context_layer = context_layer.view(*output_size)
+
+        # [b, np, sq, hn] --> [sq, b, np, hn]
+        context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
+
+        # [sq, b, np, hn] --> [sq, b, hp]
+        new_context_layer_shape = context_layer.size()[:-2] + \
+            (self.hidden_size_per_partition,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        # =================
+        # Output. [sq, b, h]
+        # =================
+
+        output, bias = self.dense(context_layer)
+
+        return output, bias
+
+
+def bias_dropout_add(x, bias, residual, prob, training):
+    # type: (Tensor, Tensor, Tensor, float, bool) -> Tensor
+    out = torch.nn.functional.dropout(x + bias, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+def get_bias_dropout_add(training):
+    def _bias_dropout_add(x, bias, residual, prob):
+        return bias_dropout_add(x, bias, residual, prob, training)
+    return _bias_dropout_add
+
+
+@torch.jit.script
+def bias_dropout_add_fused_train(x: torch.Tensor,
+                                 bias: torch.Tensor,
+                                 residual: torch.Tensor,
+                                 prob: float) -> torch.Tensor:
+    return bias_dropout_add(x, bias, residual, prob, True)
+
+
+@torch.jit.script
+def bias_dropout_add_fused_inference(x: torch.Tensor,
+                                     bias: torch.Tensor,
+                                     residual: torch.Tensor,
+                                     prob: float) -> torch.Tensor:
+    return bias_dropout_add(x, bias, residual, prob, False)
+
+
+class ParallelRetroTransformerEncoderLayer(MegatronModule):
+    """A single transformer layer for Retro Decoder with an retriever encoder inside and cross attention.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0., retriever=None):
+        args = get_args()
+
+        super(ParallelRetroTransformerEncoderLayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        self.apply_residual_connection_post_layernorm \
+            = args.apply_residual_connection_post_layernorm
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+
+        # Retro Encoder
+        self.retriever = retriever
+
+        # Layernorm on the input data.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.hidden_dropout = args.hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        self.inter_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.cross_attn)
+
+        # Layernorm on the attention output.
+        self.post_inter_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [s, b, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.  # [ns, bs, d]
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        """
+        notations:
+            l: number of chunks
+            m: number of token per chunk
+            bs: batch size
+            d: hidden size
+            k: number of neighbors
+            r: number of tokens per neighbors (neighbors + continuation)
+        """
+
+        args = get_args()
+        retro_args = get_retro_args()
+
+        chunk_length = retro_args.retro_gpt_chunk_length
+        retrieved_length = retro_args.retro_gpt_retrieved_length
+        num_neighbors = args.retro_num_neighbors
+
+        ns, bs, d = layernorm_output.shape
+        l = int(np.ceil(ns / chunk_length))
+        first_ns = ns % chunk_length
+        if first_ns > 0:
+            first_chunk, rest_chunk = \
+                layernorm_output[:first_ns], layernorm_output[first_ns:]
+            first_chunk = torch.nn.functional.pad(
+                first_chunk,
+                (0, 0, 0, 0, 0, chunk_length - first_ns),
+                'constant',
+                0)
+            chunked_output = \
+                torch.cat((first_chunk, rest_chunk), dim=0) # [l * m, bs, d]
+        else:
+            chunked_output = layernorm_output # [l * m, bs, d]
+        chunked_output = chunked_output \
+            .reshape(l, chunk_length, bs, d) \
+            .permute(1, 2, 0, 3) \
+            .reshape(chunk_length, bs * l, d) \
+            .contiguous()
+
+        # Get Encoder Output
+        retriever_output = self.retriever(
+            retriever_output,
+            retriever_attn_mask,
+            retriever_output=chunked_output,
+            retriever_attn_mask=retriever_attn_mask,
+            inference_params=inference_params) # [r, k * bs * l , d]
+        retriever_output = retriever_output.reshape(
+            retrieved_length * num_neighbors, bs * l, d) # [r * k, bs * l, d]
+
+        # Chunked Cross attention with Retriever Encoder
+        pad = (ns - 1) % chunk_length
+        attending_chunks = layernorm_output[pad:] # [ns - m + 1, bs, d]
+        padded_chunks = torch.nn.functional.pad(
+            attending_chunks,
+            (0, 0, 0, 0, 0, chunk_length-1),
+            'constant', 0) # [ns, bs, d]
+        padded_chunked_output = padded_chunks \
+            .reshape(l, chunk_length, bs, d) \
+            .permute(1, 2, 0, 3)
+        padded_chunked_output = padded_chunked_output.reshape(
+            chunk_length, bs * l, d).contiguous() # [m, bs * l, d]
+
+        # attention_output: [m, bs * l, d]
+        # attention_bias: [d]
+        attention_output, attention_bias = \
+            self.inter_attention(
+                padded_chunked_output, # Q: main model embedding
+                None,
+                encoder_output=retriever_output) # KV: retriever output embedding
+
+        # Residual connection
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        # Re-enable torch grad to enable fused optimization.
+        with torch.enable_grad():
+            layernorm_input = bias_dropout_add_func(
+                attention_output,
+                attention_bias.expand_as(attention_output),
+                torch.zeros_like(attention_output),
+                self.hidden_dropout)
+            layernorm_input = layernorm_input \
+                .reshape(chunk_length, bs, l, d) \
+                .permute(2, 0, 1, 3)  # [l, m, bs, d]
+            layernorm_input = layernorm_input.reshape(chunk_length * l, bs, d)
+            layernorm_input = torch.nn.functional.pad(
+                layernorm_input,
+                (0, 0, 0, 0, pad, 0),
+                'constant', 0)[:ns] # [ns, b, d]
+            layernorm_input = layernorm_input + residual
+
+        # Layer norm post the decoder attention
+        layernorm_output = self.post_inter_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output, retriever_output
+
+
+class ParallelRetroTransformerLayer(MegatronModule):
+    """A single transformer layer for Retro Decoder with cross attention.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0.):
+        args = get_args()
+
+        super(ParallelRetroTransformerLayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        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.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.hidden_dropout = args.hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        self.inter_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.cross_attn)
+        # Layernorm on the attention output.
+        self.post_inter_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [b, s, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        args = get_args()
+        retro_args = get_retro_args()
+        chunk_length = retro_args.retro_gpt_chunk_length
+
+        ns, bs, d = layernorm_output.shape
+        l = int(np.ceil(ns / chunk_length))
+        pad = (ns - 1) % chunk_length
+        attending_chunks = layernorm_output[pad:]
+        padded_chunks = torch.nn.functional.pad(
+            attending_chunks,
+            (0, 0, 0, 0, 0, chunk_length - 1),
+            'constant', 0)
+        padded_chunked_output = padded_chunks \
+            .reshape(l, chunk_length, bs, d) \
+            .permute(1, 2, 0, 3)
+        padded_chunked_output = padded_chunked_output.reshape(
+            chunk_length, bs * l, d).contiguous()
+
+        # Encoder output.
+        attention_output, attention_bias = \
+            self.inter_attention(padded_chunked_output,
+                                 None,
+                                 encoder_output=retriever_output)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        # Re-enable torch grad to enable fused optimization.
+        with torch.enable_grad():
+            layernorm_input = bias_dropout_add_func(
+                attention_output,
+                attention_bias.expand_as(attention_output),
+                torch.zeros_like(attention_output),
+                self.hidden_dropout)
+            layernorm_input = layernorm_input \
+                .reshape(chunk_length, bs, l, d) \
+                .permute(2, 0, 1, 3) # [l, m, bs, d]
+            layernorm_input = layernorm_input.reshape(chunk_length * l, bs, d)
+            layernorm_input = torch.nn.functional.pad(
+                layernorm_input,
+                (0, 0, 0, 0, pad, 0),
+                'constant', 0)[:ns]
+            layernorm_input = layernorm_input + residual
+
+        # Layer norm post the decoder attention
+        layernorm_output = self.post_inter_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output
+
+
+class ParallelRetroEncoderTransformerCALayer(MegatronModule):
+    """A single transformer layer for Retro Encoder with cross attention.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0.):
+        args = get_args()
+
+        super(ParallelRetroEncoderTransformerCALayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        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.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.self_attention.attention_dropout = \
+            torch.nn.Dropout(args.retro_encoder_attention_dropout)
+        self.hidden_dropout = args.retro_encoder_hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        self.inter_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.cross_attn)
+        # Layernorm on the attention output.
+        self.post_inter_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [s, b, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        # Neighbors.
+        args = get_args()
+        retro_args = get_retro_args()
+
+        retrieved_length = retro_args.retro_gpt_retrieved_length
+        num_neighbors = args.retro_num_neighbors
+
+        ns, bs, d = layernorm_output.shape # [r, bs * l * k, d]
+        chunked_outputs = layernorm_output.reshape(retrieved_length, -1,
+                                                   num_neighbors, d)
+        chunked_outputs_before_layer_norm = \
+            layernorm_input.reshape(retrieved_length, -1,
+                                    num_neighbors, d) # [r, bs * l, k, d]
+
+        layernorm_inputs = []
+        layernorm_outputs = []
+        for k in range(num_neighbors):
+            chunked_output = chunked_outputs[:,:,k].contiguous()
+            attention_output, attention_bias = \
+                self.inter_attention(
+                    chunked_output, # Q (neighbor embedding)
+                    None,
+                    encoder_output=retriever_output) # K, V (hidden act)
+
+            # Residual connection.
+            if self.apply_residual_connection_post_layernorm:
+                residual = chunked_output
+            else:
+                residual = chunked_outputs_before_layer_norm[:,:,k]
+
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+
+                layernorm_inputs.append(layernorm_input)
+
+            # Layer norm post the decoder attention
+            layernorm_output = \
+                self.post_inter_attention_layernorm(layernorm_input)
+            layernorm_outputs.append(layernorm_output)
+
+        # layernorm_input : [r, k * bs * l, d]
+        # layernorm_output : [r, k * bs * l, d]
+        layernorm_input = \
+            torch.stack(layernorm_inputs, dim=1).reshape(ns, bs, d)
+        layernorm_output = \
+            torch.stack(layernorm_outputs, dim=1).reshape(ns, bs, d)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # Re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output
+
+
+class ParallelTransformerLayer(MegatronModule):
+    """A single transformer layer.
+
+    Transformer layer takes input with size [b, s, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_number, layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 drop_path_rate=0.):
+        args = get_args()
+
+        super(ParallelTransformerLayer, self).__init__()
+        self.layer_number = layer_number
+        self.layer_type = layer_type
+
+        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.
+        self.input_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # Self attention.
+        self.self_attention = ParallelAttention(
+            init_method,
+            output_layer_init_method,
+            layer_number,
+            attention_type=AttnType.self_attn,
+            attn_mask_type=self_attn_mask_type)
+        self.hidden_dropout = args.hidden_dropout
+        self.bias_dropout_fusion = args.bias_dropout_fusion
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 \
+            else None
+
+        # Layernorm on the attention output
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size,
+            eps=args.layernorm_epsilon,
+            no_persist_layer_norm=args.no_persist_layer_norm)
+
+        if self.layer_type == LayerType.decoder:
+            self.inter_attention = ParallelAttention(
+                init_method,
+                output_layer_init_method,
+                layer_number,
+                attention_type=AttnType.cross_attn)
+            # Layernorm on the attention output.
+            self.post_inter_attention_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm)
+
+        # MLP
+        if args.num_experts is not None:
+            self.mlp = SwitchMLP(init_method, output_layer_init_method)
+        else:
+            self.mlp = ParallelMLP(init_method, output_layer_init_method)
+
+    def forward(self, hidden_states, attention_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+        # hidden_states: [b, s, h]
+
+        # Layer norm at the beginning of the transformer layer.
+        layernorm_output = self.input_layernorm(hidden_states)
+        # Self attention.
+        attention_output, attention_bias = \
+            self.self_attention(
+                layernorm_output,
+                attention_mask,
+                inference_params=inference_params)
+
+        # Residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        if self.drop_path is None:
+            # jit scripting for a nn.module (with dropout) is not
+            # trigerring the fusion kernel. For now, we use two
+            # different nn.functional routines to account for varying
+            # dropout semantics during training and inference phases.
+            if self.bias_dropout_fusion:
+                if self.training:
+                    bias_dropout_add_func = bias_dropout_add_fused_train
+                else:
+                    bias_dropout_add_func = bias_dropout_add_fused_inference
+            else:
+                bias_dropout_add_func = get_bias_dropout_add(self.training)
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(attention_output + attention_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            layernorm_input = residual + self.drop_path(out)
+
+        # Layer norm post the self attention.
+        layernorm_output = self.post_attention_layernorm(layernorm_input)
+
+        if self.layer_type == LayerType.decoder:
+            attention_output, attention_bias = \
+                self.inter_attention(layernorm_output,
+                                     enc_dec_attn_mask,
+                                     encoder_output=encoder_output)
+            # residual connection
+            if self.apply_residual_connection_post_layernorm:
+                residual = layernorm_output
+            else:
+                residual = layernorm_input
+
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                layernorm_input = bias_dropout_add_func(
+                    attention_output,
+                    attention_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+
+            # Layer norm post the decoder attention
+            layernorm_output = self.post_inter_attention_layernorm(layernorm_input)
+
+        # MLP.
+        mlp_output, mlp_bias = self.mlp(layernorm_output)
+
+        # Second residual connection.
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
+        if self.drop_path is None:
+            # re-enable torch grad to enable fused optimization.
+            with torch.enable_grad():
+                output = bias_dropout_add_func(
+                    mlp_output,
+                    mlp_bias.expand_as(residual),
+                    residual,
+                    self.hidden_dropout)
+        else:
+            out = torch.nn.functional.dropout(mlp_output + mlp_bias,
+                                              p=self.hidden_dropout,
+                                              training=self.training)
+            output = residual + self.drop_path(out)
+
+        return output
+
+
+class ParallelRetroEncoder(MegatronModule):
+    """ Retro Transformer class for encoder ."""
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 pre_process=True, post_process=True,
+                 drop_path_rate=0.0):
+        super(ParallelRetroEncoder, self).__init__()
+        args = get_args()
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.input_tensor = None
+        self.drop_path_rate = drop_path_rate
+
+        # Store activation checkpoiting flag.
+        self.recompute_granularity = args.recompute_granularity
+        self.recompute_method = args.recompute_method
+        self.recompute_num_layers = args.recompute_num_layers
+        self.distribute_saved_activations = \
+            args.distribute_saved_activations and not args.sequence_parallel
+
+        self.sequence_parallel = args.sequence_parallel
+
+        # Number of layers.
+        self.num_layers = args.retro_encoder_layers
+
+        self.drop_path_rates = [rate.item() for rate in torch.linspace(0, self.drop_path_rate, args.num_layers)]
+
+        if args.retro_add_retriever:
+            self.P = [1]
+
+        # Transformer layers.
+        assert args.retro_add_retriever
+        def build_layer(layer_number):
+            if layer_number in self.P:
+                return ParallelRetroEncoderTransformerCALayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+            else:
+                layer = ParallelTransformerLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+                layer.self_attention.attention_dropout = \
+                    torch.nn.Dropout(args.retro_encoder_attention_dropout)
+                layer.hidden_dropout = args.retro_encoder_hidden_dropout
+                return layer
+
+        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'
+            assert args.model_type != ModelType.encoder_and_decoder
+
+            # 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 = parallel_state.get_virtual_pipeline_model_parallel_rank() * (
+                args.num_layers // args.virtual_pipeline_model_parallel_size) + \
+                (parallel_state.get_pipeline_model_parallel_rank() * self.num_layers)
+        else:
+            # Each stage gets a contiguous set of layers.
+            if args.model_type == ModelType.encoder_and_decoder and \
+                    parallel_state.get_pipeline_model_parallel_world_size() > 1:
+                pipeline_rank = parallel_state.get_pipeline_model_parallel_rank()
+                if layer_type == LayerType.encoder:
+                    offset = pipeline_rank * self.num_layers
+                else:
+                    num_ranks_in_enc = args.pipeline_model_parallel_split_rank
+                    offset = (pipeline_rank - num_ranks_in_enc) * self.num_layers
+            else:
+                offset = parallel_state.get_pipeline_model_parallel_rank() * self.num_layers
+
+        if self.num_layers == 0:
+            # When a standalone embedding stage is used (e.g.,
+            # args.standalone_embedding_stage == True), virtual pipeline ranks
+            # on pipeline rank 0 will have zero transformer layers assigned to
+            # them. This results in the model's input and output tensors to be
+            # the same, which will cause failure for certain output tensor
+            # optimizations (e.g., pipeline output deallocation). To remedy
+            # this, we assign a 'no-op' layer on these ranks, which will
+            # disconnect the input tensor from the output tensor.
+            self.num_layers = 1
+            self.layers = torch.nn.ModuleList([ NoopTransformerLayer(1) ])
+        else:
+            self.layers = torch.nn.ModuleList(
+                [build_layer(i + 1 + offset) for i in range(self.num_layers)])
+
+        if self.post_process:
+            # Final layer norm before output.
+            self.final_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm)
+
+    def _get_layer(self, layer_number):
+        return self.layers[layer_number]
+
+    def _checkpointed_forward(self, hidden_states, attention_mask,
+                              encoder_output, enc_dec_attn_mask):
+        """Forward method with activation checkpointing."""
+        def custom(start, end):
+            def custom_forward(*inputs):
+                x_ = inputs[0]
+                attention_mask = inputs[1]
+                encoder_output = inputs[2]
+                enc_dec_attn_mask = inputs[3]
+                for index in range(start, end):
+                    layer = self._get_layer(index)
+                    x_ = layer(x_, attention_mask, encoder_output, enc_dec_attn_mask)
+                return x_
+            return custom_forward
+
+        if self.activations_checkpoint_method == 'uniform':
+            # Uniformly divide the total number of Transformer layers and
+            # checkpoint the input activation of each divided chunk.
+            # A method to further reduce memory usage reducing checkpoints.
+            l = 0
+            while l < self.num_layers:
+                hidden_states = parallel_state.checkpoint(
+                    custom(l, l + self.activations_checkpoint_num_layers),
+                    self.distribute_checkpointed_activations,
+                    hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                l += self.activations_checkpoint_num_layers
+        elif self.activations_checkpoint_method == 'block':
+            # Checkpoint the input activation of only a set number of individual
+            # Transformer layers and skip the rest.
+            # A method fully use the device memory removing redundant re-computation.
+            for l in range(self.num_layers):
+                if l < self.activations_checkpoint_num_layers:
+                    hidden_states = parallel_state.checkpoint(
+                        custom(l, l + 1),
+                        self.distribute_checkpointed_activations,
+                        hidden_states, attention_mask,
+                        encoder_output, enc_dec_attn_mask)
+                else:
+                    hidden_states = custom(l, l + 1)(
+                        hidden_states, attention_mask,
+                        encoder_output, enc_dec_attn_mask)
+        else:
+            raise ValueError("Invalid activation checkpoint method.")
+
+        return hidden_states
+
+    def set_input_tensor(self, input_tensor):
+        """Set input tensor to be used instead of forward()'s input.
+
+        When doing pipeline parallelism the input from the previous
+        stage comes from communication, not from the input, so the
+        model's forward_step_func won't have it. This function is thus
+        used by internal code to bypass the input provided by the
+        forward_step_func"""
+        self.input_tensor = input_tensor
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output, retriever_attn_mask,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+
+        # Checks.
+        if inference_params:
+            assert self.activations_checkpoint_method is None, \
+                'inference does not work with activation checkpointing'
+
+        if self.pre_process:
+            # Data format change to avoid explicit tranposes : [b s h] --> [s b h].
+            # If the input flag for fp32 residual connection is set, convert for float.
+            if self.fp32_residual_connection:
+                hidden_states = hidden_states.transpose(0, 1).contiguous().float()
+            # Otherwise, leave it as is.
+            else:
+                hidden_states = hidden_states.transpose(0, 1).contiguous()
+        else:
+            # See set_input_tensor()
+            hidden_states = self.input_tensor
+
+        # Viewless tensor.
+        # - We only need to create a viewless tensor in the case of micro batch
+        #   size (mbs) == 1, since in this case, 'hidden_states.transpose()'
+        #   above creates a view tensor, and '.contiguous()' is a pass-through.
+        #   For mbs >= 2, '.contiguous()' creates a new tensor, eliminating
+        #   the need to make it viewless.
+        #
+        #   However, we don't explicitly check mbs == 1 here because
+        #   make_viewless_tensor() has negligible overhead when its input
+        #   is already viewless.
+        #
+        # - For the 'else' case above, calling make_viewless_tensor() here is
+        #   likely redundant, since p2p_communication.py (likely originator)
+        #   already creates viewless tensors. That said, make_viewless_tensor()
+        #   is called here to be future-proof and corner-case-proof.
+        hidden_states = core_utils.make_viewless_tensor(
+            hidden_states,
+            requires_grad = True,
+            keep_graph = True,
+        )
+
+        # Transpose encoder output.
+        if encoder_output is not None:
+            encoder_output = encoder_output.transpose(0, 1).contiguous()
+
+        args = get_args()
+        assert not args.sequence_parallel, "if SP, need rng context."
+
+        # Forward pass.
+        if self.recompute_granularity == 'full':
+            hidden_states = self._checkpointed_forward(hidden_states,
+                                                       attention_mask,
+                                                       encoder_output,
+                                                       enc_dec_attn_mask)
+        else:
+            for index in range(self.num_layers):
+                layer = self._get_layer(index)
+                if index + 1 in self.P:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask,
+                        retriever_output=retriever_output,
+                        retriever_attn_mask=retriever_attn_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+                else:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+
+        # Final layer norm.
+        if self.post_process:
+            # Reverting data format change [s b h] --> [b s h].
+            hidden_states = hidden_states.transpose(0, 1).contiguous()
+            output = self.final_layernorm(hidden_states)
+        else:
+            output = hidden_states
+
+        return output
+
+
+class ParallelRetroTransformer(MegatronModule):
+    """Standard GPT Transformer class."""
+
+    def __init__(self, init_method, output_layer_init_method,
+                 layer_type=LayerType.encoder,
+                 self_attn_mask_type=AttnMaskType.padding,
+                 pre_process=True, post_process=True,
+                 drop_path_rate=0.0, retriever=None):
+        super(ParallelRetroTransformer, self).__init__()
+        args = get_args()
+
+        assert pre_process and post_process, \
+            "pipeline parallelism un-supported."
+
+        self.bf16 = args.bf16
+        self.fp32_residual_connection = args.fp32_residual_connection
+        self.pre_process = pre_process
+        self.post_process = post_process
+        self.input_tensor = None
+        self.drop_path_rate = drop_path_rate
+
+        # Store activation checkpoiting flag.
+        self.recompute_granularity = args.recompute_granularity
+        self.recompute_method = args.recompute_method
+        self.recompute_num_layers = args.recompute_num_layers
+        self.distribute_saved_activations = \
+            args.distribute_saved_activations and not args.sequence_parallel
+
+        self.sequence_parallel = args.sequence_parallel
+
+        # Number of layers.
+        self.num_layers = _get_num_layers(
+            args, args.model_type == ModelType.encoder_and_decoder)
+
+        self.drop_path_rates = [rate.item() for rate in torch.linspace(0, self.drop_path_rate, args.num_layers)]
+
+        if args.retro_add_retriever:
+            if args.num_layers == 12:
+                self.P = [6, 9, 12]
+            elif args.num_layers == 24:
+                self.P = np.arange(9, 25, 3).tolist()
+            elif args.num_layers == 40:
+                self.P = np.arange(9, 41, 3).tolist()
+                self.P.append(40)
+            self.retriever = retriever
+
+        # Transformer layers.
+        assert args.retro_add_retriever
+        def build_layer(layer_number):
+            if layer_number == min(self.P):
+                return ParallelRetroTransformerEncoderLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1],
+                    retriever=retriever
+                )
+            elif layer_number in self.P:
+                return ParallelRetroTransformerLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+            else:
+                return ParallelTransformerLayer(
+                    init_method,
+                    output_layer_init_method,
+                    layer_number,
+                    layer_type=layer_type,
+                    self_attn_mask_type=self_attn_mask_type,
+                    drop_path_rate=self.drop_path_rates[layer_number - 1])
+
+        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'
+            assert args.model_type != ModelType.encoder_and_decoder
+            # 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 = parallel_state.get_virtual_pipeline_model_parallel_rank() * (
+                args.num_layers // args.virtual_pipeline_model_parallel_size) + \
+                (parallel_state.get_pipeline_model_parallel_rank() * self.num_layers)
+        else:
+            # Each stage gets a contiguous set of layers.
+            if args.model_type == ModelType.encoder_and_decoder and \
+                    parallel_state.get_pipeline_model_parallel_world_size() > 1:
+                pipeline_rank = parallel_state.get_pipeline_model_parallel_rank()
+                if layer_type == LayerType.encoder:
+                    offset = pipeline_rank * self.num_layers
+                else:
+                    num_ranks_in_enc = args.pipeline_model_parallel_split_rank
+                    offset = (pipeline_rank - num_ranks_in_enc) * self.num_layers
+            else:
+                offset = parallel_state.get_pipeline_model_parallel_rank() * self.num_layers
+
+        if self.num_layers == 0:
+            # When a standalone embedding stage is used (e.g.,
+            # args.standalone_embedding_stage == True), virtual pipeline ranks
+            # on pipeline rank 0 will have zero transformer layers assigned to
+            # them. This results in the model's input and output tensors to be
+            # the same, which will cause failure for certain output tensor
+            # optimizations (e.g., pipeline output deallocation). To remedy
+            # this, we assign a 'no-op' layer on these ranks, which will
+            # disconnect the input tensor from the output tensor.
+            self.num_layers = 1
+            self.layers = torch.nn.ModuleList([ NoopTransformerLayer(1) ])
+        else:
+            self.layers = torch.nn.ModuleList(
+                [build_layer(i + 1 + offset) for i in range(self.num_layers)])
+
+        if self.post_process:
+            # Final layer norm before output.
+            self.final_layernorm = LayerNorm(
+                args.hidden_size,
+                eps=args.layernorm_epsilon,
+                no_persist_layer_norm=args.no_persist_layer_norm)
+
+    def _get_layer(self, layer_number):
+        return self.layers[layer_number]
+
+    def _checkpointed_forward(self, hidden_states, attention_mask,
+                              encoder_output, enc_dec_attn_mask):
+        """Forward method with activation checkpointing."""
+        def custom(start, end):
+            def custom_forward(*inputs):
+                x_ = inputs[0]
+                attention_mask = inputs[1]
+                encoder_output = inputs[2]
+                enc_dec_attn_mask = inputs[3]
+                for index in range(start, end):
+                    layer = self._get_layer(index)
+                    x_ = layer(x_, attention_mask, encoder_output, enc_dec_attn_mask)
+                return x_
+            return custom_forward
+
+        if self.activations_checkpoint_method == 'uniform':
+            # Uniformly divide the total number of Transformer layers and checkpoint
+            # the input activation of each divided chunk.
+            # A method to further reduce memory usage reducing checkpoints.
+            l = 0
+            while l < self.num_layers:
+                hidden_states = parallel_state.checkpoint(
+                    custom(l, l + self.activations_checkpoint_num_layers),
+                    self.distribute_checkpointed_activations,
+                    hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                l += self.activations_checkpoint_num_layers
+        elif self.activations_checkpoint_method == 'block':
+            # Checkpoint the input activation of only a set number of individual
+            # Transformer layers and skip the rest.
+            # A method fully use the device memory removing redundant re-computation.
+            for l in range(self.num_layers):
+                if l < self.activations_checkpoint_num_layers:
+                    hidden_states = parallel_state.checkpoint(
+                        custom(l, l + 1),
+                        self.distribute_checkpointed_activations,
+                        hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                else:
+                    hidden_states = custom(l, l + 1)(
+                        hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+        else:
+            raise ValueError("Invalid activation checkpoint method.")
+
+        return hidden_states
+
+    def set_input_tensor(self, input_tensor):
+        """Set input tensor to be used instead of forward()'s input.
+
+        When doing pipeline parallelism the input from the previous
+        stage comes from communication, not from the input, so the
+        model's forward_step_func won't have it. This function is thus
+        used by internal code to bypass the input provided by the
+        forward_step_func"""
+        self.input_tensor = input_tensor
+
+    def forward(self, hidden_states, attention_mask,
+                retriever_output=None, retriever_attn_mask=None,
+                encoder_output=None, enc_dec_attn_mask=None,
+                inference_params=None):
+
+        # Checks.
+        if inference_params:
+            assert self.recompute_granularity is None, \
+                'inference does not work with activation checkpointing'
+
+        args = get_args()
+
+        # Transpose retriever output, to match hidden_states shape.
+        retriever_output = retriever_output.transpose(0, 1).contiguous()
+
+        # Viewless tensor.
+        # - We only need to create a viewless tensor in the case of micro batch
+        #   size (mbs) == 1, since in this case, 'hidden_states.transpose()'
+        #   above creates a view tensor, and '.contiguous()' is a pass-through.
+        #   For mbs >= 2, '.contiguous()' creates a new tensor, eliminating
+        #   the need to make it viewless.
+        #
+        #   However, we don't explicitly check mbs == 1 here because
+        #   make_viewless_tensor() has negligible overhead when its input
+        #   is already viewless.
+        #
+        # - For the 'else' case above, calling make_viewless_tensor() here is
+        #   likely redundant, since p2p_communication.py (likely originator)
+        #   already creates viewless tensors. That said, make_viewless_tensor()
+        #   is called here to be future-proof and corner-case-proof.
+        hidden_states = core_utils.make_viewless_tensor(
+            hidden_states,
+            requires_grad=True,
+            keep_graph=True,
+        )
+
+        # Transpose encoder output.
+        if encoder_output is not None:
+            encoder_output = encoder_output.transpose(0, 1).contiguous()
+
+        # Forward pass.
+        assert not args.sequence_parallel, "if SP, need rng context."
+        if self.recompute_granularity == 'full':
+            hidden_states = self._checkpointed_forward(hidden_states,
+                                                       attention_mask,
+                                                       encoder_output,
+                                                       enc_dec_attn_mask)
+        else:
+            for index in range(self.num_layers):
+                layer = self._get_layer(index)
+                if args.retro_add_retriever and index + 1 == min(self.P):
+                    hidden_states, E = layer(
+                        hidden_states,
+                        attention_mask,
+                        retriever_output=retriever_output,
+                        retriever_attn_mask=retriever_attn_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+                elif args.retro_add_retriever and index + 1 in self.P:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask,
+                        retriever_output=E,
+                        retriever_attn_mask=retriever_attn_mask,
+                        encoder_output=encoder_output,
+                        enc_dec_attn_mask=enc_dec_attn_mask,
+                        inference_params=inference_params)
+                else:
+                    hidden_states = layer(
+                    hidden_states,
+                    attention_mask,
+                    encoder_output=encoder_output,
+                    enc_dec_attn_mask=enc_dec_attn_mask,
+                    inference_params=inference_params)
+
+        # Final layer norm.
+        output = self.final_layernorm(hidden_states)
+
+        return output

megatron/model/transformer.py

@@ -270,7 +270,6 @@ class CoreAttention(MegatronModule):
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.
-
         if not self.sequence_parallel:
             with tensor_parallel.get_cuda_rng_tracker().fork():
                 attention_probs = self.attention_dropout(attention_probs)
@@ -477,6 +476,7 @@ class ParallelAttention(MegatronModule):
         # =================================================
         # Pre-allocate memory for key-values for inference.
         # =================================================
+
         if inference_params:
             if self.layer_number not in inference_params.key_value_memory_dict:
                 inf_max_seq_len = inference_params.max_sequence_len

megatron/schedules.py

@@ -244,6 +244,7 @@ def forward_backward_no_pipelining(forward_step_func,
     output_tensor = forward_step(forward_step_func, data_iterator,
                                  model, input_tensor, forward_data_store,
                                  timers, collect_non_loss_data)
+
     if not forward_only:
         backward_step(optimizer, input_tensor, output_tensor,
                       output_tensor_grad, timers)