Merge branch 'main' into main_retriver_merge_dpr

README.md

-[Megatron](https://arxiv.org/pdf/1909.08053.pdf) is a large, powerful transformer developed by the Applied Deep Learning Research team at NVIDIA. This repository is for ongoing research on training large transformer language models at scale. We developed efficient, model-parallel (tensor and pipeline), and multi-node pre-training of [GPT](https://arxiv.org/abs/2005.14165) and [BERT](https://arxiv.org/pdf/1810.04805.pdf) using mixed precision.
+Megatron ([1](https://arxiv.org/pdf/1909.08053.pdf) and [2](https://arxiv.org/pdf/2104.04473.pdf)) is a large, powerful transformer developed by the Applied Deep Learning Research team at NVIDIA. This repository is for ongoing research on training large transformer language models at scale. We developed efficient, model-parallel (tensor and pipeline), and multi-node pre-training of [GPT](https://arxiv.org/abs/2005.14165) and [BERT](https://arxiv.org/pdf/1810.04805.pdf) using mixed precision.
 
 Below are some of the projects where we have directly used Megatron:
 * [BERT and GPT Studies Using Megatron](https://arxiv.org/pdf/1909.08053.pdf)
@@ -8,6 +8,7 @@ Below are some of the projects where we have directly used Megatron:
 * [Local Knowledge Powered Conversational Agents](https://arxiv.org/abs/2010.10150)
 * [MEGATRON-CNTRL: Controllable Story Generation with External Knowledge Using Large-Scale Language Models](https://www.aclweb.org/anthology/2020.emnlp-main.226.pdf)
 * [RACE Reading Comprehension Dataset Leaderboard](http://www.qizhexie.com/data/RACE_leaderboard.html)
+* [Scaling Language Model Training to a Trillion Parameters Using Megatron](https://arxiv.org/pdf/2104.04473.pdf)
 * [Training Question Answering Models From Synthetic Data](https://www.aclweb.org/anthology/2020.emnlp-main.468.pdf)
 
 Our codebase is capable of efficiently training very large (hundreds of billions of parameters) language models with both model and data parallelism. To demonstrate how the code scales with multiple GPUs and model sizes, we consider GPT models from 1 billion all the way to 1 trillion parameters. All models use a vocabulary size of 51,200 and a sequence length of 2048. We vary hidden size, number of attention heads, and number of layers to arrive at a specifc model size. As the model size increases, we also modestly increase the batch size. We leverage [NVIDIA's Selene supercomputer](https://www.top500.org/system/179842/) to perform scaling studies and use up to 3072 [A100](https://www.nvidia.com/en-us/data-center/a100/) GPUs for the largest model. The table below shows the model configurations along with the achieved FLOPs (both per GPU and aggregate over all GPUs). Note that the FLOPs are measured for end-to-end training, i.e., includes all operations including data loading, optimization, and even logging.

megatron/arguments.py

@@ -625,6 +625,9 @@ def _add_data_args(parser):
                        help='Path to the vocab file.')
     group.add_argument('--merge-file', type=str, default=None,
                        help='Path to the BPE merge file.')
+    group.add_argument('--vocab-extra-ids', type=int, default=0,
+                       help='Number of additional vocabulary tokens. '
+                            'They are used for span masking in the T5 model')
     group.add_argument('--seq-length', type=int, default=None,
                        help='Maximum sequence length to process.')
     group.add_argument('--encoder-seq-length', type=int, default=None,

megatron/data/bert_dataset.py

@@ -15,24 +15,25 @@
 
 """BERT Style dataset."""
 
-import os
-import time
-
 import numpy as np
 import torch
-from torch.utils.data import Dataset
 
-from megatron import get_tokenizer, get_args
-from megatron import print_rank_0
-from megatron import mpu
-from megatron.data.dataset_utils import get_a_and_b_segments
-from megatron.data.dataset_utils import truncate_segments
-from megatron.data.dataset_utils import create_tokens_and_tokentypes
-from megatron.data.dataset_utils import pad_and_convert_to_numpy
-from megatron.data.dataset_utils import create_masked_lm_predictions
+from megatron import (
+    get_args,
+    get_tokenizer,
+    mpu,
+    print_rank_0
+)
+from megatron.data.dataset_utils import (
+    get_samples_mapping,
+    get_a_and_b_segments,
+    truncate_segments,
+    create_tokens_and_tokentypes,
+    create_masked_lm_predictions
+)
 
 
-class BertDataset(Dataset):
+class BertDataset(torch.utils.data.Dataset):
 
     def __init__(self, name, indexed_dataset, data_prefix,
                  num_epochs, max_num_samples, masked_lm_prob,
@@ -49,11 +50,11 @@ class BertDataset(Dataset):
         self.indexed_dataset = indexed_dataset
 
         # Build the samples mapping.
-        self.samples_mapping = get_samples_mapping_(self.indexed_dataset,
+        self.samples_mapping = get_samples_mapping(self.indexed_dataset,
                                                    data_prefix,
                                                    num_epochs,
                                                    max_num_samples,
-                                                    self.max_seq_length,
+                                                   self.max_seq_length - 3, # account for added tokens
                                                    short_seq_prob,
                                                    self.seed,
                                                    self.name,
@@ -87,91 +88,6 @@ class BertDataset(Dataset):
                                      self.binary_head)
 
 
-def get_samples_mapping_(indexed_dataset,
-                         data_prefix,
-                         num_epochs,
-                         max_num_samples,
-                         max_seq_length,
-                         short_seq_prob,
-                         seed,
-                         name,
-                         binary_head):
-    if not num_epochs:
-        if not max_num_samples:
-            raise ValueError("Need to specify either max_num_samples "
-                             "or num_epochs")
-        num_epochs = np.iinfo(np.int32).max - 1
-    if not max_num_samples:
-        max_num_samples = np.iinfo(np.int64).max - 1
-
-    # Filename of the index mapping
-    indexmap_filename = data_prefix
-    indexmap_filename += '_{}_indexmap'.format(name)
-    if num_epochs != (np.iinfo(np.int32).max - 1):
-        indexmap_filename += '_{}ep'.format(num_epochs)
-    if max_num_samples != (np.iinfo(np.int64).max - 1):
-        indexmap_filename += '_{}mns'.format(max_num_samples)
-    indexmap_filename += '_{}msl'.format(max_seq_length)
-    indexmap_filename += '_{:0.2f}ssp'.format(short_seq_prob)
-    indexmap_filename += '_{}s'.format(seed)
-    indexmap_filename += '.npy'
-
-    # Build the indexed mapping if not exist.
-    if torch.distributed.get_rank() == 0 and \
-       not os.path.isfile(indexmap_filename):
-        print(' > WARNING: could not find index map file {}, building '
-              'the indices on rank 0 ...'.format(indexmap_filename))
-
-        # Make sure the types match the helpers input types.
-        assert indexed_dataset.doc_idx.dtype == np.int64
-        assert indexed_dataset.sizes.dtype == np.int32
-
-        # Build samples mapping
-        verbose = torch.distributed.get_rank() == 0
-        start_time = time.time()
-        print_rank_0(' > building sapmles index mapping for {} ...'.format(
-            name))
-        # First compile and then import.
-        from megatron.data import helpers
-        samples_mapping = helpers.build_mapping(
-            indexed_dataset.doc_idx,
-            indexed_dataset.sizes,
-            num_epochs,
-            max_num_samples,
-            max_seq_length - 3,  # account for added tokens
-            short_seq_prob,
-            seed,
-            verbose,
-            2 if binary_head else 1)
-        print_rank_0(' > done building sapmles index maping')
-        np.save(indexmap_filename, samples_mapping, allow_pickle=True)
-        print_rank_0(' > saved the index mapping in {}'.format(
-            indexmap_filename))
-        # Make sure all the ranks have built the mapping
-        print_rank_0(' > elasped time to build and save samples mapping '
-                     '(seconds): {:4f}'.format(
-                         time.time() - start_time))
-    # This should be a barrier but nccl barrier assumes
-    # device_index=rank which is not the case for model
-    # parallel case
-    counts = torch.cuda.LongTensor([1])
-    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
-    torch.distributed.all_reduce(counts, group=mpu.get_pipeline_model_parallel_group())
-    assert counts[0].item() == (
-        torch.distributed.get_world_size() //
-        torch.distributed.get_world_size(group=mpu.get_tensor_model_parallel_group()))
-
-    # Load indexed dataset.
-    print_rank_0(' > loading indexed mapping from {}'.format(
-        indexmap_filename))
-    start_time = time.time()
-    samples_mapping = np.load(indexmap_filename, allow_pickle=True, mmap_mode='r')
-    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
-        time.time() - start_time))
-    print_rank_0('    total number of samples: {}'.format(
-        samples_mapping.shape[0]))
-
-    return samples_mapping
 
 
 def build_training_sample(sample,
@@ -225,7 +141,7 @@ def build_training_sample(sample,
 
     # Masking.
     max_predictions_per_seq = masked_lm_prob * max_num_tokens
-    (tokens, masked_positions, masked_labels, _) = create_masked_lm_predictions(
+    (tokens, masked_positions, masked_labels, _, _) = create_masked_lm_predictions(
         tokens, vocab_id_list, vocab_id_to_token_dict, masked_lm_prob,
         cls_id, sep_id, mask_id, max_predictions_per_seq, np_rng)
 
@@ -244,3 +160,35 @@ def build_training_sample(sample,
         'truncated': int(truncated)}
     return train_sample
 
+
+def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
+                             masked_labels, pad_id, max_seq_length):
+    """Pad sequences and convert them to numpy."""
+
+    # Some checks.
+    num_tokens = len(tokens)
+    padding_length = max_seq_length - num_tokens
+    assert padding_length >= 0
+    assert len(tokentypes) == num_tokens
+    assert len(masked_positions) == len(masked_labels)
+
+    # Tokens and token types.
+    filler = [pad_id] * padding_length
+    tokens_np = np.array(tokens + filler, dtype=np.int64)
+    tokentypes_np = np.array(tokentypes + filler, dtype=np.int64)
+
+    # Padding mask.
+    padding_mask_np = np.array([1] * num_tokens + [0] * padding_length,
+                               dtype=np.int64)
+
+    # Lables and loss mask.
+    labels = [-1] * max_seq_length
+    loss_mask = [0] * max_seq_length
+    for i in range(len(masked_positions)):
+        assert masked_positions[i] < num_tokens
+        labels[masked_positions[i]] = masked_labels[i]
+        loss_mask[masked_positions[i]] = 1
+    labels_np = np.array(labels, dtype=np.int64)
+    loss_mask_np = np.array(loss_mask, dtype=np.int64)
+
+    return tokens_np, tokentypes_np, labels_np, padding_mask_np, loss_mask_np

megatron/data/dataset_utils.py

@@ -19,18 +19,26 @@
 # with some modifications.
 
 import math
+import os
 import time
 import collections
 
 import numpy as np
-from megatron import get_args, print_rank_0
+import torch
+
+from megatron import (
+    get_args,
+    mpu,
+    print_rank_0
+)
 from megatron.data.blendable_dataset import BlendableDataset
 from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
 
-DSET_TYPE_STD = 'standard_bert'
+DSET_TYPE_BERT = 'standard_bert'
 DSET_TYPE_ICT = 'ict'
+DSET_TYPE_T5  = 't5'
 
-DSET_TYPES = [DSET_TYPE_ICT, DSET_TYPE_STD]
+DSET_TYPES = [DSET_TYPE_BERT, DSET_TYPE_ICT, DSET_TYPE_T5]
 
 
 def get_datasets_weights_and_num_samples(data_prefix,
@@ -179,7 +187,9 @@ def create_masked_lm_predictions(tokens,
                                  max_ngrams=3,
                                  do_whole_word_mask=True,
                                  favor_longer_ngram=False,
-                                 do_permutation=False):
+                                 do_permutation=False,
+                                 geometric_dist=False,
+                                 masking_style="bert"):
     """Creates the predictions for the masked LM objective.
     Note: Tokens here are vocab ids and not text tokens."""
 
@@ -219,12 +229,12 @@ def create_masked_lm_predictions(tokens,
     num_to_predict = min(max_predictions_per_seq,
                          max(1, int(round(len(tokens) * masked_lm_prob))))
 
+    ngrams = np.arange(1, max_ngrams + 1, dtype=np.int64)
+    if not geometric_dist:
         # Note(mingdachen):
         # By default, we set the probilities to favor shorter ngram sequences.
-    ngrams = np.arange(1, max_ngrams + 1, dtype=np.int64)
         pvals = 1. / np.arange(1, max_ngrams + 1)
         pvals /= pvals.sum(keepdims=True)
-
         if favor_longer_ngram:
             pvals = pvals[::-1]
 
@@ -237,7 +247,7 @@ def create_masked_lm_predictions(tokens,
 
     np_rng.shuffle(ngram_indexes)
 
-    masked_lms = []
+    (masked_lms, masked_spans) = ([], [])
     covered_indexes = set()
     for cand_index_set in ngram_indexes:
         if len(masked_lms) >= num_to_predict:
@@ -251,9 +261,16 @@ def create_masked_lm_predictions(tokens,
                 if index in covered_indexes:
                     continue
 
+        if not geometric_dist:
             n = np_rng.choice(ngrams[:len(cand_index_set)],
                               p=pvals[:len(cand_index_set)] /
                               pvals[:len(cand_index_set)].sum(keepdims=True))
+        else:
+            # Sampling "n" from the geometric distribution and clipping it to
+            # the max_ngrams. Using p=0.2 default from the SpanBERT paper
+            # https://arxiv.org/pdf/1907.10529.pdf (Sec 3.1)
+            n = min(np_rng.geometric(0.2), max_ngrams)
+
         index_set = sum(cand_index_set[n - 1], [])
         n -= 1
         # Note(mingdachen):
@@ -277,8 +294,8 @@ def create_masked_lm_predictions(tokens,
             continue
         for index in index_set:
             covered_indexes.add(index)
-
             masked_token = None
+            if masking_style == "bert":
                 # 80% of the time, replace with [MASK]
                 if np_rng.random() < 0.8:
                     masked_token = mask_id
@@ -289,12 +306,19 @@ def create_masked_lm_predictions(tokens,
                     # 10% of the time, replace with random word
                     else:
                         masked_token = vocab_id_list[np_rng.randint(0, len(vocab_id_list))]
+            elif masking_style == "t5":
+                masked_token = mask_id
+            else:
+                raise ValueError("invalid value of masking style")
 
             output_tokens[index] = masked_token
-
             masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
-    assert len(masked_lms) <= num_to_predict
 
+        masked_spans.append(MaskedLmInstance(
+            index=index_set,
+            label=[tokens[index] for index in index_set]))
+
+    assert len(masked_lms) <= num_to_predict
     np_rng.shuffle(ngram_indexes)
 
     select_indexes = set()
@@ -347,12 +371,13 @@ def create_masked_lm_predictions(tokens,
             masked_lms.append(MaskedLmInstance(index=src_i, label=orig_token[src_i]))
 
     masked_lms = sorted(masked_lms, key=lambda x: x.index)
+    # Sort the spans by the index of the first span
+    masked_spans = sorted(masked_spans, key=lambda x: x.index[0])
 
     for p in masked_lms:
         masked_lm_positions.append(p.index)
         masked_lm_labels.append(p.label)
-
-    return (output_tokens, masked_lm_positions, masked_lm_labels, token_boundary)
+    return (output_tokens, masked_lm_positions, masked_lm_labels, token_boundary, masked_spans)
 
 
 def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
@@ -390,9 +415,10 @@ def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
 
 def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
                                     train_valid_test_num_samples,
-                                    max_seq_length, masked_lm_prob,
-                                    short_seq_prob, seed, skip_warmup,
-                                    binary_head,
+                                    max_seq_length,
+                                    masked_lm_prob, short_seq_prob, seed,
+                                    skip_warmup, binary_head=False,
+                                    max_seq_length_dec=None,
                                     dataset_type='standard_bert'):
 
     if len(data_prefix) == 1:
@@ -403,6 +429,7 @@ def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
                                                 short_seq_prob, seed,
                                                 skip_warmup,
                                                 binary_head,
+                                                max_seq_length_dec,
                                                 dataset_type=dataset_type)
     # Blending dataset.
     # Parse the values.
@@ -444,9 +471,10 @@ def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
 
 def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
                                      train_valid_test_num_samples,
-                                     max_seq_length, masked_lm_prob,
-                                     short_seq_prob, seed, skip_warmup,
-                                     binary_head,
+                                     max_seq_length,
+                                     masked_lm_prob, short_seq_prob, seed,
+                                     skip_warmup, binary_head,
+                                     max_seq_length_dec,
                                      dataset_type='standard_bert'):
 
     if dataset_type not in DSET_TYPES:
@@ -489,6 +517,7 @@ def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
     def build_dataset(index, name):
         from megatron.data.bert_dataset import BertDataset
         from megatron.data.ict_dataset import ICTDataset
+        from megatron.data.t5_dataset import T5Dataset
         dataset = None
         if splits[index + 1] > splits[index]:
             # Get the pointer to the original doc-idx so we can set it later.
@@ -507,7 +536,6 @@ def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
                 max_num_samples=train_valid_test_num_samples[index],
                 max_seq_length=max_seq_length,
                 seed=seed,
-                binary_head=binary_head
             )
 
             if dataset_type == DSET_TYPE_ICT:
@@ -517,15 +545,27 @@ def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
                     title_dataset=title_dataset,
                     query_in_block_prob=args.query_in_block_prob,
                     use_one_sent_docs=args.use_one_sent_docs,
+                    binary_head=binary_head,
                     **kwargs
                 )
-            else:
+            elif dataset_type == DSET_TYPE_T5:
+                dataset = T5Dataset(
+                    indexed_dataset=indexed_dataset,
+                    masked_lm_prob=masked_lm_prob,
+                    max_seq_length_dec=max_seq_length_dec,
+                    short_seq_prob=short_seq_prob,
+                    **kwargs
+                )
+            elif dataset_type == DSET_TYPE_BERT:
                 dataset = BertDataset(
                     indexed_dataset=indexed_dataset,
                     masked_lm_prob=masked_lm_prob,
                     short_seq_prob=short_seq_prob,
+                    binary_head=binary_head,
                     **kwargs
                 )
+            else:
+                raise NotImplementedError("Dataset type not fully implemented.")
 
             # Set the original pointer so dataset remains the main dataset.
             indexed_dataset.set_doc_idx(doc_idx_ptr)
@@ -590,4 +630,90 @@ def get_train_valid_test_split_(splits_string, size):
     assert splits_index[-1] == size
     return splits_index
 
+def get_samples_mapping(indexed_dataset,
+                        data_prefix,
+                        num_epochs,
+                        max_num_samples,
+                        max_seq_length,
+                        short_seq_prob,
+                        seed,
+                        name,
+                        binary_head):
+    """Get a list that maps a sample index to a starting sentence index, end sentence index, and length"""
+
+    if not num_epochs:
+        if not max_num_samples:
+            raise ValueError("Need to specify either max_num_samples "
+                             "or num_epochs")
+        num_epochs = np.iinfo(np.int32).max - 1
+    if not max_num_samples:
+        max_num_samples = np.iinfo(np.int64).max - 1
+
+    # Filename of the index mapping
+    indexmap_filename = data_prefix
+    indexmap_filename += '_{}_indexmap'.format(name)
+    if num_epochs != (np.iinfo(np.int32).max - 1):
+        indexmap_filename += '_{}ep'.format(num_epochs)
+    if max_num_samples != (np.iinfo(np.int64).max - 1):
+        indexmap_filename += '_{}mns'.format(max_num_samples)
+    indexmap_filename += '_{}msl'.format(max_seq_length)
+    indexmap_filename += '_{:0.2f}ssp'.format(short_seq_prob)
+    indexmap_filename += '_{}s'.format(seed)
+    indexmap_filename += '.npy'
+
+    # Build the indexed mapping if not exist.
+    if torch.distributed.get_rank() == 0 and \
+       not os.path.isfile(indexmap_filename):
+        print(' > WARNING: could not find index map file {}, building '
+              'the indices on rank 0 ...'.format(indexmap_filename))
+
+        # Make sure the types match the helpers input types.
+        assert indexed_dataset.doc_idx.dtype == np.int64
+        assert indexed_dataset.sizes.dtype == np.int32
+
+        # Build samples mapping
+        verbose = torch.distributed.get_rank() == 0
+        start_time = time.time()
+        print_rank_0(' > building sapmles index mapping for {} ...'.format(
+            name))
+        # First compile and then import.
+        from megatron.data import helpers
+        samples_mapping = helpers.build_mapping(
+            indexed_dataset.doc_idx,
+            indexed_dataset.sizes,
+            num_epochs,
+            max_num_samples,
+            max_seq_length,
+            short_seq_prob,
+            seed,
+            verbose,
+            2 if binary_head else 1)
+        print_rank_0(' > done building sapmles index maping')
+        np.save(indexmap_filename, samples_mapping, allow_pickle=True)
+        print_rank_0(' > saved the index mapping in {}'.format(
+            indexmap_filename))
+        # Make sure all the ranks have built the mapping
+        print_rank_0(' > elasped time to build and save samples mapping '
+                     '(seconds): {:4f}'.format(
+                         time.time() - start_time))
+    # This should be a barrier but nccl barrier assumes
+    # device_index=rank which is not the case for model
+    # parallel case
+    counts = torch.cuda.LongTensor([1])
+    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
+    torch.distributed.all_reduce(counts, group=mpu.get_pipeline_model_parallel_group())
+    assert counts[0].item() == (
+        torch.distributed.get_world_size() //
+        torch.distributed.get_world_size(group=mpu.get_tensor_model_parallel_group()))
+
+    # Load indexed dataset.
+    print_rank_0(' > loading indexed mapping from {}'.format(
+        indexmap_filename))
+    start_time = time.time()
+    samples_mapping = np.load(indexmap_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
+        time.time() - start_time))
+    print_rank_0('    total number of samples: {}'.format(
+        samples_mapping.shape[0]))
 
+    return samples_mapping

megatron/data/t5_dataset.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""T5 Style dataset."""
+
+import collections
+
+import numpy as np
+import torch
+
+from megatron import get_tokenizer
+from megatron.data.dataset_utils import (
+    create_masked_lm_predictions,
+    get_samples_mapping
+)
+
+class T5Dataset(torch.utils.data.Dataset):
+
+    def __init__(self, name, indexed_dataset, data_prefix,
+                 num_epochs, max_num_samples, masked_lm_prob,
+                 max_seq_length, max_seq_length_dec,
+                 short_seq_prob, seed):
+
+        # Params to store.
+        self.name = name
+        self.seed = seed
+        self.masked_lm_prob = masked_lm_prob
+        self.max_seq_length = max_seq_length
+        self.max_seq_length_dec = max_seq_length_dec
+
+        # Dataset.
+        self.indexed_dataset = indexed_dataset
+
+        # Build the samples mapping.
+        self.samples_mapping = get_samples_mapping(self.indexed_dataset,
+                                                   data_prefix,
+                                                   num_epochs,
+                                                   max_num_samples,
+                                                   self.max_seq_length - 2, # account for added tokens
+                                                   short_seq_prob,
+                                                   self.seed,
+                                                   self.name,
+                                                   False)
+
+        # Vocab stuff.
+        tokenizer = get_tokenizer()
+        self.vocab_id_list = list(tokenizer.inv_vocab.keys())
+        self.vocab_id_to_token_dict = tokenizer.inv_vocab
+        self.cls_id = tokenizer.cls
+        self.sep_id = tokenizer.sep
+        self.mask_id = tokenizer.mask
+        self.pad_id = tokenizer.pad
+        self.bos_id = tokenizer.bos_token_id
+        self.eos_id = tokenizer.eos_token_id
+        self.sentinel_tokens = tokenizer.additional_special_tokens_ids
+        assert len(self.sentinel_tokens) > 0, "Provide the argument --vocab-extra-ids 100 to the script"
+
+    def __len__(self):
+        return self.samples_mapping.shape[0]
+
+    def __getitem__(self, idx):
+
+        start_index, end_index, seq_length = self.samples_mapping[idx]
+        sample = []
+        for index in range(start_index, end_index):
+            sample.append(self.indexed_dataset[index])
+        # Note that this rng state should be numpy and not python since
+        # python randint is inclusive whereas the numpy one is exclusive.
+        np_rng = np.random.RandomState(seed=(self.seed + idx))
+        return build_training_sample(sample, seq_length,
+                                     self.max_seq_length,  # needed for padding
+                                     self.max_seq_length_dec,
+                                     self.vocab_id_list,
+                                     self.vocab_id_to_token_dict,
+                                     self.cls_id, self.sep_id,
+                                     self.mask_id, self.pad_id,
+                                     self.masked_lm_prob, np_rng,
+                                     self.bos_id, self.eos_id,
+                                     self.sentinel_tokens)
+
+
+def build_training_sample(sample, target_seq_length,
+                          max_seq_length, max_seq_length_dec,
+                          vocab_id_list, vocab_id_to_token_dict,
+                          cls_id, sep_id, mask_id, pad_id,
+                          masked_lm_prob, np_rng, bos_id=None,
+                          eos_id=None, sentinel_tokens=None):
+    """Build training sample.
+
+    Arguments:
+        sample: A list of sentences in which each sentence is a list token ids.
+        target_seq_length: Desired sequence length.
+        max_seq_length: Maximum length of the sequence. All values are padded to
+            this length.
+        vocab_id_list: List of vocabulary ids. Used to pick a random id.
+        vocab_id_to_token_dict: A dictionary from vocab ids to text tokens.
+        cls_id: Start of example id.
+        sep_id: Separator id.
+        mask_id: Mask token id.
+        pad_id: Padding token id.
+        masked_lm_prob: Probability to mask tokens.
+        np_rng: Random number genenrator. Note that this rng state should be
+              numpy and not python since python randint is inclusive for
+              the opper bound whereas the numpy one is exclusive.
+        bos_id: start of decoder example id
+        eos_id: end of generation id
+        sentinel_tokens: unique value to be substituted for every replaced span
+    """
+
+    assert target_seq_length <= max_seq_length
+
+    # flatten sentences into one list
+    tokens = [token for sentence in sample for token in sentence]
+
+    # Truncate to `target_sequence_length`.
+    max_num_tokens = target_seq_length
+    truncated = len(tokens) > max_num_tokens
+    tokens = tokens[:max_num_tokens]
+
+    # Masking.
+    max_predictions_per_seq = masked_lm_prob * max_num_tokens
+    (tokens, masked_positions, masked_labels, _, masked_spans) = create_masked_lm_predictions(
+        tokens, vocab_id_list, vocab_id_to_token_dict, masked_lm_prob,
+        cls_id, sep_id, mask_id, max_predictions_per_seq, np_rng,
+        max_ngrams=10, geometric_dist=True, masking_style="t5")
+
+    # Padding.
+    tokens_enc, tokens_dec_in, labels, enc_mask, \
+    dec_mask, enc_dec_mask, loss_mask \
+        = pad_and_convert_to_numpy(tokens, masked_positions,
+                                   masked_labels, pad_id, max_seq_length,
+                                   max_seq_length_dec, masked_spans,
+                                   bos_id, eos_id, sentinel_tokens)
+
+    train_sample = {
+        'text_enc': tokens_enc,
+        'text_dec': tokens_dec_in,
+        'labels': labels,
+        'loss_mask': loss_mask,
+        'truncated': int(truncated),
+        'enc_mask': enc_mask,
+        'dec_mask': dec_mask,
+        'enc_dec_mask': enc_dec_mask,
+    }
+    return train_sample
+
+
+def pad_and_convert_to_numpy(tokens, masked_positions,
+                             masked_labels, pad_id,
+                             max_seq_length, max_seq_length_dec,
+                             masked_spans=None, bos_id=None,
+                             eos_id=None, sentinel_tokens=None):
+    """Pad sequences and convert them to numpy."""
+
+    sentinel_tokens = collections.deque(sentinel_tokens)
+    t5_input = []
+    (t5_decoder_in, t5_decoder_out) = ([bos_id], [])
+    (start_index, end_index) = (0, None)
+    for span in masked_spans:
+        flag = sentinel_tokens.popleft()
+
+        # Append the same tokens in decoder input and output
+        t5_decoder_in.append(flag)
+        t5_decoder_in.extend(span.label)
+        t5_decoder_out.append(flag)
+        t5_decoder_out.extend(span.label)
+
+        end_index = span.index[0]
+        t5_input.extend(tokens[start_index: end_index])
+        t5_input.append(flag)
+
+        # the next start index is the token after the last span token
+        start_index = span.index[-1] + 1
+
+    # Add <eos> token to the t5_decoder_out
+    t5_decoder_out.append(eos_id)
+
+    # Add the remaining tokens to the t5 input
+    t5_input.extend(tokens[start_index:])
+
+    # assert (len(t5_input) - len(masked_spans)) + \
+    #        (len(t5_decoder_in) - (len(masked_spans) + 1)) == len(tokens)
+
+    # Some checks.
+
+    # Encoder-side padding mask.
+    num_tokens = len(t5_input)
+    padding_length = max_seq_length - num_tokens
+    assert padding_length >= 0
+    assert len(masked_positions) == len(masked_labels)
+
+    # Tokens..
+    filler = [pad_id] * padding_length
+    tokens_enc = np.array(t5_input + filler, dtype=np.int64)
+
+    # Decoder-side padding mask.
+    num_tokens_dec = len(t5_decoder_in)
+    padding_length_dec = max_seq_length_dec - num_tokens_dec
+    assert padding_length_dec >= 0
+    filler_dec = [pad_id] * padding_length_dec
+    tokens_dec_in = np.array(t5_decoder_in + filler_dec, dtype=np.int64)
+
+    # Create attention masks
+    enc_mask = make_attention_mask(tokens_enc, tokens_enc)
+    enc_dec_mask = make_attention_mask(tokens_dec_in, tokens_enc)
+    dec_mask = make_attention_mask(tokens_dec_in, tokens_dec_in)
+    dec_mask = dec_mask * make_history_mask(tokens_dec_in)
+
+    # Labels mask.
+    labels = t5_decoder_out + ([-1] * padding_length_dec)
+    labels = np.array(labels, dtype=np.int64)
+
+    # Loss mask
+    loss_mask = ([1] * num_tokens_dec) + ([0] * padding_length_dec)
+    loss_mask = np.array(loss_mask, dtype=np.int64)
+
+    return tokens_enc, tokens_dec_in, labels, enc_mask, \
+           dec_mask, enc_dec_mask, loss_mask
+
+
+def make_attention_mask(source_block, target_block):
+    """
+    Returns a 2-dimensional (2-D) attention mask
+    :param source_block: 1-D array
+    :param target_block: 1-D array
+    """
+    mask = (target_block[None, :] >= 1) * (source_block[:, None] >= 1)
+    mask = mask.astype(np.int64)
+    # (source_length, target_length)
+    return mask
+
+
+def make_attention_mask_3d(source_block, target_block):
+    """
+    Returns a 3-dimensional (3-D) attention mask
+    :param source_block: 1-D array
+    :param target_block: 1-D array
+    """
+    mask = (target_block[:, None, :] >= 1) * (source_block[:, :, None] >= 1)
+    # (batch, source_length, target_length)
+    # mask = mask.astype(np.int64)
+    return mask
+
+
+def make_history_mask(block):
+    length = block.shape[0]
+    arange = np.arange(length)
+    history_mask = (arange[None, ] <= arange[:, None])
+    history_mask = history_mask.astype(np.int64)
+    return history_mask
+
+
+def make_history_mask_3d(block):
+    batch, length = block.shape
+    arange = torch.arange(length, device=block.device)
+    history_mask = (arange[None, ] <= arange[:, None])[None, ]
+    history_mask = history_mask.expand(batch, length, length)
+    return history_mask

megatron/model/__init__.py

@@ -18,5 +18,6 @@ from .fused_layer_norm import MixedFusedLayerNorm as LayerNorm
 from .distributed import DistributedDataParallel
 from .bert_model import BertModel
 from .gpt_model import GPTModel
+from .t5_model import T5Model
 from .language_model import get_language_model
 from .module import Float16Module

megatron/model/t5_model.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""T5 model."""
+
+import torch
+
+from megatron import (
+    get_args,
+    mpu
+)
+from megatron.model.enums import AttnMaskType
+from megatron.model.language_model import parallel_lm_logits, get_language_model
+from megatron.model.transformer import LayerNorm
+from megatron.model.utils import (
+    openai_gelu,
+    get_linear_layer,
+    init_method_normal,
+    scaled_init_method_normal
+)
+from .module import MegatronModule
+
+
+def t5_extended_attention_mask(attention_mask_list):
+
+    def attn_mask_postprocess(attn_mask):
+        # [b, 1, s, s]
+        extended_attention_mask = attn_mask.unsqueeze(1)
+        return extended_attention_mask
+
+    return [attn_mask_postprocess(attn_mask) for attn_mask in attention_mask_list]
+
+
+def t5_position_ids(token_ids):
+    # Create position ids
+    seq_length = token_ids.size(1)
+    position_ids = torch.arange(seq_length, dtype=torch.long,
+                                device=token_ids.device)
+    position_ids = position_ids.unsqueeze(0).expand_as(token_ids)
+
+    return position_ids
+
+
+class T5LMHead(MegatronModule):
+    """Masked LM head for T5
+
+    Arguments:
+        mpu_vocab_size: model parallel size of vocabulary.
+        hidden_size: hidden size
+        init_method: init method for weight initialization
+        layernorm_epsilon: tolerance for layer norm divisions
+        parallel_output: wether output logits being distributed or not.
+    """
+
+    def __init__(self, mpu_vocab_size, parallel_output):
+        super(T5LMHead, self).__init__()
+
+        args = get_args()
+
+        self.bias = torch.nn.Parameter(torch.zeros(mpu_vocab_size))
+        self.bias.model_parallel = True
+        self.bias.partition_dim = 0
+        self.bias.stride = 1
+        self.parallel_output = parallel_output
+
+    def forward(self, hidden_states, word_embeddings_weight):
+        output = parallel_lm_logits(hidden_states,
+                                    word_embeddings_weight,
+                                    self.parallel_output,
+                                    bias=self.bias)
+        return output
+
+
+class T5Model(MegatronModule):
+    """T5 Language model."""
+
+    def __init__(self, num_tokentypes=0, parallel_output=True):
+        super(T5Model, self).__init__()
+        args = get_args()
+
+        self.fp16_lm_cross_entropy = args.fp16_lm_cross_entropy
+        self.parallel_output = parallel_output
+        init_method = init_method_normal(args.init_method_std)
+        scaled_init_method = scaled_init_method_normal(args.init_method_std,
+                                                       args.num_layers)
+
+        self.language_model, self._language_model_key = get_language_model(
+            num_tokentypes=num_tokentypes,
+            add_pooler=False,
+            add_decoder=True,
+            encoder_attn_mask_type=AttnMaskType.padding,
+            init_method=init_method,
+            scaled_init_method=scaled_init_method)
+
+        self.lm_head = T5LMHead(
+            self.language_model.embedding.word_embeddings.weight.size(0),
+            parallel_output)
+        self._lm_head_key = 'lm_head'
+
+    def set_input_tensor(self, input_tensor):
+        """See megatron.model.transformer.set_input_tensor()"""
+        self.language_model.set_input_tensor(input_tensor)
+
+    def forward(self, encoder_input_ids, decoder_input_ids, encoder_attn_mask,
+                decoder_attn_mask, encoder_decoder_attn_mask,
+                tokentype_ids=None, lm_labels=None, enc_hidden_states=None):
+
+        # Converting the attention masks to proper parameter settings
+        encoder_attn_mask, decoder_attn_mask, encoder_decoder_attn_mask = t5_extended_attention_mask(
+            [encoder_attn_mask, decoder_attn_mask, encoder_decoder_attn_mask])
+
+        encoder_position_ids = t5_position_ids(encoder_input_ids)
+        decoder_position_ids = t5_position_ids(decoder_input_ids)
+
+        lm_output = self.language_model(encoder_input_ids,
+                                        encoder_position_ids,
+                                        encoder_attn_mask,
+                                        decoder_input_ids,
+                                        decoder_position_ids,
+                                        decoder_attn_mask,
+                                        encoder_decoder_attn_mask,
+                                        tokentype_ids=tokentype_ids,
+                                        enc_hidden_states=enc_hidden_states)
+
+        decoder_output, encoder_output = lm_output
+
+        # Output.
+        lm_logits = self.lm_head(decoder_output,
+                                 self.language_model.embedding.word_embeddings.weight)
+
+        if lm_labels is None:
+            return lm_logits, encoder_output
+        else:
+            if self.fp16_lm_cross_entropy:
+                assert lm_logits.dtype == torch.half
+                lm_loss = mpu.vocab_parallel_cross_entropy(lm_logits, lm_labels)
+            else:
+                lm_loss = mpu.vocab_parallel_cross_entropy(lm_logits.float(),
+                                                           lm_labels)
+            return lm_loss, encoder_output
+
+    def state_dict_for_save_checkpoint(self, destination=None, prefix='',
+                                       keep_vars=False):
+        """For easy load when model is combined with other heads,
+        add an extra key."""
+
+        state_dict_ = {}
+        state_dict_[self._language_model_key] \
+            = self.language_model.state_dict_for_save_checkpoint(
+            destination, prefix, keep_vars)
+        state_dict_[self._lm_head_key] \
+            = self.lm_head.state_dict_for_save_checkpoint(
+            destination, prefix, keep_vars)
+        return state_dict_
+
+    def load_state_dict(self, state_dict, strict=True):
+        """Customized load."""
+
+        self.language_model.load_state_dict(
+            state_dict[self._language_model_key], strict=strict)
+        self.lm_head.load_state_dict(state_dict[self._lm_head_key],
+                                     strict=strict)

megatron/text_generation_utils.py

@@ -190,6 +190,37 @@ def generate_samples_input_from_file(model):
             raw_text = None
             context_count += 1
 
+# We added this function to support the tasks evaluation such as squad
+# and drop in the https://github.com/EleutherAI/lm-evaluation-harness 
+# codebase. The lm-evaluation-harness code can now call this function
+# similar to their current generate function call used for gpt style models.
+def generate_samples_eval(model, context, max_gen_length, eos_token_id):
+    # Generate samples for lm evaluation
+    # NEED TO THINK ABOUT eos token
+
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    raw_text_len = len(context)
+    model.eval()
+
+    context_tokens = tokenizer.tokenize(context)
+    args.out_seq_length = max_gen_length + len(context_tokens)
+    args.eos_id = eos_token_id
+
+    with torch.no_grad():
+        token_stream = get_token_stream(model, [context_tokens])
+        for counter, decode_tokens in enumerate(token_stream):
+            if counter == args.out_seq_length:
+                break
+
+    decode_tokens, _ = decode_tokens
+    decode_tokens = decode_tokens[0].cpu().numpy().tolist()
+    trim_decode_tokens = tokenizer.detokenize(
+        decode_tokens)[raw_text_len:]
+ 
+    return trim_decode_tokens
+
 
 def generate_samples_interactive(model, print_frequency=24):
 
@@ -438,6 +469,12 @@ def sample_sequence_batch(model, context_tokens, context_lengths,
     model.eval()
     with torch.no_grad():
         context_length = context_lengths.min().item()
+
+        # added eos_id to support the function generate_samples_eval that passes
+        # eos_id as an argument and needs termination when that id id found.
+        if hasattr(args, 'eos_id'):
+            eos_id = args.eos_id
+        else:
             eos_id = tokenizer.eod
 
         counter = 0

megatron/tokenizer/tokenizer.py

@@ -32,10 +32,12 @@ def build_tokenizer(args):
     assert args.vocab_file is not None
     if args.tokenizer_type == 'BertWordPieceLowerCase':
         tokenizer = _BertWordPieceTokenizer(vocab_file=args.vocab_file,
-                                            lower_case=True)
+                                            lower_case=True,
+                                            vocab_extra_ids=args.vocab_extra_ids)
     elif args.tokenizer_type == 'BertWordPieceCase':
         tokenizer = _BertWordPieceTokenizer(vocab_file=args.vocab_file,
-                                            lower_case=False)
+                                            lower_case=False,
+                                            vocab_extra_ids=args.vocab_extra_ids)
     elif args.tokenizer_type == 'GPT2BPETokenizer':
         assert args.merge_file is not None
         tokenizer = _GPT2BPETokenizer(args.vocab_file, args.merge_file)
@@ -127,7 +129,7 @@ class AbstractTokenizer(ABC):
 class _BertWordPieceTokenizer(AbstractTokenizer):
     """Original BERT wordpiece tokenizer."""
 
-    def __init__(self, vocab_file, lower_case=True):
+    def __init__(self, vocab_file, lower_case=True, vocab_extra_ids=0):
         if lower_case:
             name = 'BERT Lower Case'
         else:
@@ -138,6 +140,37 @@ class _BertWordPieceTokenizer(AbstractTokenizer):
         self.sep_id = self.tokenizer.vocab['[SEP]']
         self.pad_id = self.tokenizer.vocab['[PAD]']
         self.mask_id = self.tokenizer.vocab['[MASK]']
+        self._additional_special_tokens = []
+
+        # (dsachan) Add BOS and EOS tokens
+        SPECIAL_TOKENS = {'eos_token': '[EOS]',
+                          'bos_token': '[BOS]'}
+        self._bos_token = '[BOS]'
+        self.add_token(self._bos_token)
+        self._bos_token_id = self.vocab.get(self._bos_token)
+
+        self._eos_token = '[EOS]'
+        self.add_token(self._eos_token)
+        self._eos_token_id = self.vocab.get(self._eos_token)
+
+        # (dsachan) Add additional special tokens
+        # These can be used as sentinel tokens in T5 model inputs
+        additional_special_tokens = []
+        additional_special_tokens.extend(
+            ["<extra_id_{}>".format(i) for i in range(vocab_extra_ids)])
+        self.add_additional_special_tokens(additional_special_tokens)
+
+    def add_token(self, token):
+        if token not in self.vocab:
+            self.inv_vocab[self.vocab_size] = token
+            # self.vocab_size comes from len(vocab)
+            # and it will increase as we add elements
+            self.vocab[token] = self.vocab_size
+
+    def add_additional_special_tokens(self, tokens_list):
+        setattr(self, "additional_special_tokens", tokens_list)
+        for value in tokens_list:
+            self.add_token(value)
 
     @property
     def vocab_size(self):
@@ -189,6 +222,40 @@ class _BertWordPieceTokenizer(AbstractTokenizer):
     def mask(self):
         return self.mask_id
 
+    @property
+    def bos_token(self):
+        """ Beginning of sentence token id """
+        return self._bos_token
+
+    @property
+    def eos_token(self):
+        """ End of sentence token id """
+        return self._eos_token
+
+    @property
+    def additional_special_tokens(self):
+        """ All the additional special tokens you may want to use (list of strings)."""
+        return self._additional_special_tokens
+
+    @property
+    def bos_token_id(self):
+        """ Id of the beginning of sentence token in the vocabulary."""
+        return self._bos_token_id
+
+    @property
+    def eos_token_id(self):
+        """ Id of the end of sentence token in the vocabulary."""
+        return self._eos_token_id
+
+    @property
+    def additional_special_tokens_ids(self):
+        """ Ids of all the additional special tokens in the vocabulary (list of integers)."""
+        return [self.vocab.get(token) for token in self._additional_special_tokens]
+
+    @additional_special_tokens.setter
+    def additional_special_tokens(self, value):
+        self._additional_special_tokens = value
+
 
 class _GPT2BPETokenizer(AbstractTokenizer):
     """Original GPT2 BPE tokenizer."""

pretrain_bert.py

@@ -15,9 +15,11 @@
 
 """Pretrain BERT"""
 
+from functools import partial
+
 import torch
 import torch.nn.functional as F
-from functools import partial
+
 from megatron import get_args
 from megatron import print_rank_0
 from megatron import get_timers

pretrain_t5.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Pretrain T5"""
+
+from functools import partial
+
+import torch
+
+from megatron import (
+    get_args,
+    get_timers,
+    mpu,
+    print_rank_0
+)
+from megatron.data.dataset_utils import build_train_valid_test_datasets
+from megatron.model import T5Model
+from megatron.training import pretrain
+from megatron.utils import average_losses_across_data_parallel_group
+
+
+def model_provider(pre_process=True, post_process=True):
+    """Build the model."""
+    assert pre_process and post_process, "T5 doesn't yet support pipelining"
+
+    print_rank_0('building T5 model ...')
+    model = T5Model(num_tokentypes=0,
+                    parallel_output=True)
+    return model
+
+
+def get_batch(data_iterator):
+    """Build the batch."""
+
+    keys = ['text_enc', 'text_dec', 'labels', 'loss_mask',
+            'enc_mask', 'dec_mask', 'enc_dec_mask']
+    datatype = torch.int64
+
+    # Broadcast data.
+    if data_iterator is not None:
+        data = next(data_iterator)
+    else:
+        data = None
+    data_b = mpu.broadcast_data(keys, data, datatype)
+
+    # Unpack.
+    tokens_enc = data_b['text_enc'].long()
+    tokens_dec = data_b['text_dec'].long()
+    labels = data_b['labels'].long()
+    loss_mask = data_b['loss_mask'].float()
+
+    enc_mask = (data_b['enc_mask'] < 0.5)
+    dec_mask = (data_b['dec_mask'] < 0.5)
+    enc_dec_mask = (data_b['enc_dec_mask'] < 0.5)
+
+    return tokens_enc, tokens_dec, loss_mask, labels, \
+           enc_mask, dec_mask, enc_dec_mask
+
+
+def loss_func(loss_mask, output_tensor):
+    lm_loss_, _ = output_tensor
+
+    lm_loss_ = lm_loss_.float()
+    lm_loss = torch.sum(
+        lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
+
+    loss = lm_loss
+    averaged_losses = average_losses_across_data_parallel_group([lm_loss])
+
+    return loss, {'lm loss': averaged_losses[0]}
+
+
+def forward_step(data_iterator, model):
+    """Forward step."""
+    args = get_args()
+    timers = get_timers()
+
+    # Get the batch.
+    timers('batch generator').start()
+    tokens_enc, tokens_dec, loss_mask, lm_labels, enc_mask, dec_mask, enc_dec_mask \
+        = get_batch(data_iterator)
+    timers('batch generator').stop()
+
+    # Forward model lm_labels
+    output_tensor = model(tokens_enc,
+                          tokens_dec,
+                          enc_mask,
+                          dec_mask,
+                          enc_dec_mask,
+                          tokentype_ids=None,
+                          lm_labels=lm_labels)
+
+    return output_tensor, partial(loss_func, loss_mask)
+
+
+def train_valid_test_datasets_provider(train_val_test_num_samples):
+    """Build train, valid, and test datasets."""
+    args = get_args()
+
+    print_rank_0('> building train, validation, and test datasets '
+                 'for T5 ...')
+    train_ds, valid_ds, test_ds = build_train_valid_test_datasets(
+        data_prefix=args.data_path,
+        data_impl=args.data_impl,
+        splits_string=args.split,
+        train_valid_test_num_samples=train_val_test_num_samples,
+        max_seq_length=args.encoder_seq_length,
+        max_seq_length_dec=args.decoder_seq_length,
+        masked_lm_prob=args.mask_prob,
+        short_seq_prob=args.short_seq_prob,
+        seed=args.seed,
+        skip_warmup=(not args.mmap_warmup),
+        dataset_type='t5')
+    print_rank_0("> finished creating T5 datasets ...")
+
+    return train_ds, valid_ds, test_ds
+
+
+if __name__ == "__main__":
+
+    pretrain(train_valid_test_datasets_provider, model_provider, forward_step,
+             args_defaults={'tokenizer_type': 'BertWordPieceLowerCase'})

tools/openwebtext/README.md

@@ -26,7 +26,8 @@ python blacklist_urls.py <path to the dowloaded deduplicated URLs> <filename for
 ```
 python cleanup_dataset.py <input data file> <output cleaned data filename>
 ```
-2. Using LSH, find possible duplicates and store then in a file for later processing. This step can NOT be sharded and usually takes 12 to 24 hours for OpenWebText dataset. The code supports saving and loading fingerprints for recurrent deduplications.
+Additional cleanup (e.g. remove documents less than 512 characters or dataset specific cleaning like stories, realnews datasets) can be done using `cleanup_fix_dataset.py`. More details can be found by running `python cleanup_fix_dataset.py --help`.
+2. Using LSH, find possible duplicates and store then in a file for later processing. The code supports saving and loading fingerprints for recurrent deduplications, and is also multithreaded for faster processing. More details are can be found by `python find_duplicate.py --help`.
 ```
 python find_duplicates.py --inputs <pairlist list of input cleaned data files and keys, e.g. cc.json cc_id news.json news_id> --output <output possible duplicate urls filename>
 ```
@@ -46,10 +47,13 @@ shuf <cleaned deduped data file> -o train_data.json
 
 # Deduplicating ngrams
 
-To deduplicate the downstream tasks from the training dataset, we run the following command.
+To deduplicate the downstream tasks (e.g. lambada, squad) from the training dataset, we run the following command.
 
 ```
-python filter_ngrams.py <down stream task dataset> <training dataset to deduplicate> <output training dataset>
+python filter_ngrams.py --tasks <name of he task, e.g. lambada, squad> --dedup-dataset <training dataset to deduplicate> <json key> --output <output training dataset>
 ```
+We use 13-grams by default for the deduplication. When we find a 13-gram match in a training document, we split the document into two pieces and remove the 13-gram along with 200 characters from the both side of the 13-gram. We also remove any splitted document with less than 200 characters or if a document got splitted more than 10 times. These parameters can be changed using corresponding arguments.
 
-We use 13-grams for the deduplication. When we find a 13-gram match in a training document, we split the document into two pieces and remove the 13-gram along with 200 characters from the both side of the 13-gram. We also remove any splitted document with less than 200 characters or if a document got splitted more than 10 times.
+Only for the lambada task, we need to provide the path, `--lambada-path <path of the lambada test data>`.
+
+Several other features (e.g. save and load dictionary) have been added, look at `python filter_ngrams.py --help` for details.

tools/openwebtext/add_id.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import json
+import os
+import time
+
+"""
+This code adds id to each json object in a json file. User can add prefix
+to the ids.
+"""
+
+if __name__ == '__main__':
+
+    print('parsing the arguments ...')
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input-file', type=str, default=None, help='Input'\
+        ' json file where id needs to be added')
+    parser.add_argument('--output-file', type=str, default=None, help=\
+        'Output file name with id')
+    parser.add_argument('--id-prefix', type=str, default=None, help=\
+        'Id prefix')
+    parser.add_argument('--log-interval', type=int, default=100,
+                       help='Log interval')
+    args = parser.parse_args()
+
+    print('Adding ids to dataset ...')
+
+    f_input = open(args.input_file, 'r', encoding='utf-8')
+    f_output = open(args.output_file, 'wb')
+
+    unique_ids = 1
+    start_time = time.time()
+    for row in f_input:
+        each_row = json.loads(row)
+        adlr_id_string = args.id_prefix + '-{:010d}'.format(int(unique_ids))
+        each_row['adlr_id'] = adlr_id_string
+        myjson = json.dumps(each_row, ensure_ascii=False)
+
+        f_output.write(myjson.encode('utf-8'))
+        f_output.write('\n'.encode('utf-8'))
+
+        if unique_ids % args.log_interval == 0:
+            print('    processed {:9d} documents in {:.2f} seconds ...'.format( \
+                    unique_ids, time.time() - start_time), flush=True)
+
+        unique_ids += 1
+
+    # Close the file.
+    f_input.close()
+    f_output.close()
+    
+    print('done :-)', flush=True)

tools/openwebtext/cleanup_fix_dataset.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Filter and clean documents:
+Capable to clean docs with less than 512 characters, less than
+256 characters and contains javascript, fix text and dataset specific
+cleaning like stories and realnews datasets.
+Program arguments have the details.
+"""
+
+import argparse
+from functools import partial
+import glob
+import ftfy
+import json
+from langdetect import detect
+import multiprocessing
+import os
+from pathlib import Path
+import re
+import time
+
+def process_doc(json_line, args):
+
+    # Read the line.
+    document = json.loads(json_line)
+    text = document['text']
+
+    output = {'remove_512': False, 'remove_256_javascript': False, \
+        'remove_512_non_english': False, 'ftfy_fix_text': False, \
+        'general_cleaning': False}
+
+    try:
+        # Reomove all docs with less than 512 characters
+        if "remove_512" in args.tasks:
+            if len(text) < 512:
+                output['remove_512'] = True
+                return output, text, document, True
+
+        # Remove docs if less than 256 character length and contains Javascript
+        if "remove_256_javascript" in args.tasks:
+            if len(text) < 256 and 'javascript' in text.lower():
+                output['remove_256_javascript'] = True
+                return output, text, document, True
+
+        # Remove docs < 512 and nonenglish
+        if "remove_512_non_english" in args.tasks:
+            if len(text) < 512 and detect(text) != 'en':
+                output['remove_512_non_english'] = True
+                return output, text, document, True
+
+        # Fix the text using ftfy, don't remove the text, hence return False
+        if "ftfy_fix_text" in args.tasks:
+            fixed_text = ftfy.fix_text(text)
+            output['ftfy_fix_text'] = True
+            return output, fixed_text, document, False
+
+        # Cleaning extra spaces and newlines
+        if "general_cleaning" in args.tasks:
+            cleaned_text = re.sub(r"  +|\b\n+ |\b\n+", " ", text)
+            #cleaned_text = re.sub(r"\n\n+", "\n\n", text) # used this for Gutenberg dataset
+            #cleaned_text = re.sub(r"\n", "\n\n", text) # Used this for realnews
+
+            # stories datasets
+            #cleaned_text = re.sub(r" \'", "'", text)
+            #cleaned_text = re.sub(r" \!", "!", cleaned_text)
+            #cleaned_text = re.sub(r" \.", ".", cleaned_text)
+            #cleaned_text = re.sub(r" \?", "?", cleaned_text)
+            #cleaned_text = re.sub(r" - ", "-", cleaned_text)
+            ##cleaned_text = re.sub(r"\" ", "\"", cleaned_text)
+            #cleaned_text = re.sub(r" @ ", "@", cleaned_text)
+
+            output['general_cleaning'] = True
+            return output, cleaned_text, document, False
+
+    except Exception as e:
+        print('Error: *************************\n{}\ntext: {}'.format(e, \
+            text), flush=True)
+        return output, text, document, True
+
+    # don't remove
+    return output, text, document, False
+
+
+def process_set(args, input_file, output_f_cleaned, output_f_filtered):
+
+    print(' > working on {} ...'.format(input_file), flush=True)
+    
+    num_docs = num_remove_512 = num_remove_java = num_remove_512_non_english \
+        = num_ftfy_fix_text = num_general_cleaning = 0
+
+    # Output file and counters.
+    output_cleaned = open(output_f_cleaned, 'wb')
+    output_filtered = open(output_f_filtered, 'wb')
+
+    start_time = time.time()
+
+    # Setup multi-processing.
+    num_workers = 40
+    fin = open(input_file, 'r', encoding='utf-8')
+    pool = multiprocessing.Pool(num_workers)
+    process_doc_partial = partial(process_doc, args=args)
+    processed_docs = pool.imap(process_doc_partial, fin, 500)
+
+    # Process documents.
+    for output, text, document, to_filter in processed_docs:
+        num_docs += 1
+
+        num_remove_512 += 1 if output['remove_512'] else 0
+        num_remove_java += 1 if output['remove_256_javascript'] else 0
+        num_remove_512_non_english += 1 if output['remove_512_non_english'] \
+            else 0
+        num_ftfy_fix_text += 1 if output['ftfy_fix_text'] else 0
+        num_general_cleaning += 1 if output['general_cleaning'] else 0
+
+        document['text'] = text
+        myjson = json.dumps(document, ensure_ascii=False)
+
+        if to_filter:
+            output_filtered.write(myjson.encode('utf-8'))
+            output_filtered.write('\n'.encode('utf-8'))
+        else:
+            output_cleaned.write(myjson.encode('utf-8'))
+            output_cleaned.write('\n'.encode('utf-8'))
+
+        if num_docs % args.log_interval == 0:
+            print('    processed {:9d} documents in {:.2f} seconds ...'.format(
+                num_docs, time.time() - start_time), flush=True)
+
+    # Close the file.
+    output_cleaned.close()
+    output_filtered.close()
+    fin.close()
+
+    # Print stats.
+    print('  >> total docs: {} remove_512 {} remove_256_javascript {} '\
+        'remove_512_non_english {} ftfy_fix_text {} general_cleaning {}'.\
+        format(num_docs, num_remove_512, num_remove_java,\
+        num_remove_512_non_english, num_ftfy_fix_text, \
+        num_general_cleaning), flush=True)
+
+if __name__ == '__main__':
+
+
+    print('parsing the arguments ...')
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input-files', nargs = '*', required=True, default=\
+                        None, help = 'Input json files that needs to be'\
+                        ' cleaned')
+    parser.add_argument('--tasks', nargs = '*', required=True, default=None,\
+                        help = 'Tasks to perform on the input files, ' \
+                        'such as remove_512, remove_256_javascript, ' \
+                        'remove_512_non_english, ftfy_fix_text, and ' \
+                        'general_cleaning. 256 or 512 means the number' \
+                        ' of characters.')
+
+    parser.add_argument('--output-path', type=str, default=None,
+                       help='Directory where the output should go')
+    parser.add_argument('--log-interval', type=int, default=100,
+                       help='Log interval')
+
+    args = parser.parse_args()
+
+    print('cleanup dataset ...')
+
+    for input_file in args.input_files:
+        input_filename, input_filename_ext = os.path.splitext(Path(input_file)\
+            .name)
+
+        output_f_cleaned = os.path.join(args.output_path, input_filename + \
+            "_cleaned" + input_filename_ext)
+        output_f_filtered = os.path.join(args.output_path, input_filename + \
+            "_filtered" + input_filename_ext)
+
+        process_set(args, input_file, output_f_cleaned, output_f_filtered)
+
+    print('done :-)', flush=True)

tools/openwebtext/filter_ngrams.py

@@ -19,10 +19,12 @@ All split documents with less than 200 characters got filtered. Any document
 with more than 10 splits got filtered as well.
 """
 
+import argparse
 from functools import partial
 import json
 import multiprocessing
 import nltk
+import pickle
 import re
 import string
 import sys
@@ -36,40 +38,20 @@ def get_words(text):
         positions.append(match.start())
     return words, positions
 
-def free_ngram(line, ngrams, ngram_size, filter_text_len, 
-    splits_count, split_window_each_size):
-    # remove all the ngrams
-
-    try:
-        myjson = json.loads(line)
-        text_buf = [myjson['text']]
-    except Exception as e:
-        print("Error: {}".format(e), flush=True)
-        text_buf = []
-
-    text_buf_ngram_free = []
-    while len(text_buf) > 0:
-
-        # get the first one from the buffer
-        text = text_buf.pop(0)
-        words, positions = get_words(text)
-        
-        not_ngram_free = True
-        punctuations = ".!?"
-        # find n-grams
-        for i in range(len(words) - ngram_size + 1):
-            seq = " ".join(words[i:i+ngram_size])
-            if seq in ngrams:
-
-                # splits the text
+# splits the text
+def split_text(text, start_position, remove_char_each_side, seq):
     # first part of the text
-                pos = positions[i] - split_window_each_size
+    punctuations = ".!?"
+    pos = start_position - remove_char_each_side
     text_first = ""
     while pos > 0 and not text[pos] in punctuations:
         pos -= 1
     if pos > 0:
         text_first = text[0:pos+1]
-                pos = positions[i] + split_window_each_size
+
+    # add length of seq and remove_char_each_side
+    pos = start_position + len(seq) + remove_char_each_side
+
     # last part of the text
     text_second = ""
     while pos < len(text) and not text[pos] in punctuations:
@@ -77,97 +59,434 @@ def free_ngram(line, ngrams, ngram_size, filter_text_len,
     if pos + 1 < len(text):
         text_second = text[pos+1:len(text)]
 
+    return text_first, text_second
+
+def check_and_clean_text(args, words, ngrams, text, start_position, \
+    text_buf_ngram_free, text_buf, local_ngram):
+
+    seq = " ".join(words)
+    if seq in ngrams:
+        print(" [matched]: {}".format(seq), flush=True)
+
+        if args.get_ngram_freq_only:
+            # increase freq of this seq and then only consider the later part
+            # of the text for further processing
+            if seq in local_ngram:
+                local_ngram[seq] += 1
+            else:
+                local_ngram[seq] = 1
+            #print(" [increased]: {} {}".format(seq, ngrams[seq]), flush=True)
+            if (start_position + len(seq) + 1) < len(text):
+                text_buf.append(text[start_position + len(seq) + 1:len(text)])
+            return False            
+
+        # split the text
+        text_first, text_second = split_text(text, start_position, \
+            args.remove_char_each_side, seq)
+
         # first part of ngrams free
-                if len(text_first) > filter_text_len:
+        if len(text_first) > args.filter_text_char_len:
             text_buf_ngram_free.append(text_first)
 
         # add second part for further processing
-                if len(text_second) > filter_text_len:
+        if len(text_second) > args.filter_text_char_len:
             text_buf.append(text_second)
-                not_ngram_free = False
+
+        return False # not ngram free
+
+    # ngram free
+    return True
+
+
+def free_ngram(line, args, key, ngrams, ngrams_freq_sorted):
+    # remove all the ngrams
+
+    try:
+        myjson = json.loads(line)
+        text_buf = [myjson[key]]
+    except Exception as e:
+        print("Error: {}".format(e), flush=True)
+        text_buf = []
+
+    text_buf_ngram_free = []
+    local_ngram = {}
+    while len(text_buf) > 0:
+
+        # get the first one from the buffer
+        text = text_buf.pop(0)
+        words, positions = get_words(text)
+        
+        ngram_free = True
+        # find each max n-grams and check dictionary
+        for i in range(len(words) - args.max_ngram_size + 1):
+            check_ngram_free = check_and_clean_text(args, words[i:\
+                i+args.max_ngram_size], ngrams, text, positions[i], \
+                text_buf_ngram_free, text_buf, local_ngram)
+
+            # the seq is ngram free? if yes, break
+            if not check_ngram_free:
+                ngram_free = False
                 break
 
-        # text are ngram free
-        if not_ngram_free:
+            # if max ngrams doesn't match, check if any other lower n-grams
+            # within max ngram macthes
+            for ngram_len, _ in ngrams_freq_sorted:
+                check_ngram_free = check_and_clean_text(args, words[i:\
+                    i+ngram_len], ngrams, text, positions[i], \
+                    text_buf_ngram_free, text_buf, local_ngram)
+
+                # same check as above
+                if not check_ngram_free:
+                    ngram_free = False
+                    break
+
+            # check break from lower than max ngram loop above
+            if not ngram_free:
+                break
+
+        # for the last max n-gram, check all the lower ngrams in it
+        if ngram_free and len(words) - args.max_ngram_size > 0:
+            # get the last words of the lax max ngram
+            last_seq_words = words[(len(words)-args.max_ngram_size):len(words)]
+            last_seq_start_position = len(words) - args.max_ngram_size
+
+            # check all n-grams lower than the max
+            for pos, (ngram_len, _) in enumerate(ngrams_freq_sorted):
+
+                # ignore the max ngram as has been considered already
+                if ngram_len == args.max_ngram_size:
+                    continue
+
+                # find each ngram of ngram_len in max n-grams and check
+                for i in range(len(last_seq_words) - ngram_len + 1):
+                    check_ngram_free = check_and_clean_text(args, \
+                        last_seq_words[i:i+ngram_len], ngrams, text,\
+                        positions[last_seq_start_position+i], \
+                        text_buf_ngram_free, text_buf, local_ngram)
+
+                    if not check_ngram_free:
+                        ngram_free = False
+                        break
+
+                if not ngram_free:
+                    break
+
+        # texts are ngram free
+        if ngram_free and not args.get_ngram_freq_only:
             text_buf_ngram_free.append(text)
 
-    return text_buf_ngram_free
+    # check if the text has only been trimmed
+    trimmed = 0
+    if not args.get_ngram_freq_only and len(text_buf_ngram_free) == 1 and \
+        len(text_buf_ngram_free[0]) < len(myjson[key]):
+        trimmed = 1
 
+    return text_buf_ngram_free, trimmed, myjson, local_ngram
 
-if __name__ == '__main__':
+# insert word sequence into dictionary
+def insert_dict(words, ngrams, pos):
+    seq = " ".join(words)
+    if seq not in ngrams:
+        ngrams[seq] = 0
+        #ngrams[seq] = pos
 
-    print('finding possible duplicate content ...')
-    main_file = sys.argv[1] # lambada file
-    dedup_file = sys.argv[2] # Book corpus
-    output_file = sys.argv[3] #Filtered book corpus
-    ngrams = {}
-    id_prefix = "lambada"
+# insert each ngram from text into the ngrams dictionary
+def compute_ngrams_insert_dict(args, text, ngrams):
+    words, positions = get_words(text)
+    if len(words) < args.min_ngram_size:
+        return
 
-    # we use 13-grams, any text less than 200 characters got removed
-    # any text splitted more than 10 got removed as well
-    ngram_size = 13
-    filter_text_len = 200
-    splits_count = 10
-    split_window_each_size = 200
+    if len(words) < args.max_ngram_size:
+        insert_dict(words, ngrams, positions[0])
+
+    for i in range(len(words) - args.max_ngram_size+1):
+        insert_dict(words[i:i+args.max_ngram_size], ngrams, positions[i])
 
-    print('Reading file {} and computing ngrams'.format(main_file))
-    with open(main_file, 'r') as f:
+
+# Build ngrams for the lambada dataset
+def process_task_lambda(args, task_file, ngrams):
+    print(' reading from {} and computing ngrams'.format(task_file))
+    with open(task_file, 'r') as f:
         for line in f:
             try:
                 myjson = json.loads(line)
-                words, positions = get_words(myjson['text'])
-                for i in range(len(words) - ngram_size+1):
-                    seq = " ".join(words[i:i+ngram_size])
-                    if seq not in ngrams:
-                        ngrams[seq] = positions[i]
+                text = myjson['text']
+                compute_ngrams_insert_dict(args, text, ngrams)
             except Exception as e:
                 print('Error:', e)
-    print("ngrams size {}".format(len(ngrams)))
+    print(" Entities in ngrams {}".format(len(ngrams)), flush=True)
+
+
+# Build ngrams for the dataset of the given task
+def process_task(args, task_name, ngrams):
+
+    print(' reading from {} and computing ngrams'.format('import datasets'))
+    print(" Current entities in ngrams {}".format(len(ngrams)), flush=True)
+    # using validation/test data from datasets
+    from datasets import load_dataset
+
+    entities_in_ngrams = len(ngrams)
+
+    # load the dataset
+    if task_name == 'squad':
+        dataset = load_dataset('squad_v2', split='validation')
+    elif task_name == 'natural_questions':
+        dataset = load_dataset('natural_questions', split='validation')
+    elif task_name == 'triviaqa':
+        dataset = load_dataset('trivia_qa', 'unfiltered', split='test')
+    elif task_name == 'webqa':
+        dataset = load_dataset('web_questions', split='test')
+    elif task_name == 'race':
+        dataset = load_dataset('race', 'all', split='test')
+    elif task_name == 'drop':
+        dataset = load_dataset('drop', split='validation')
+    elif task_name == 'coqa':
+        dataset = load_dataset('coqa', split='validation')
+    elif task_name == 'piqa':
+        dataset = load_dataset('piqa', split='test')
+    else:
+        print("Invalid task name: {}".format(task_name), flush=True)
+        return
+
+    # read the dataset and add to ngrams
+    for line in dataset:
+        try:
+            if task_name in ['squad', 'triviaqa', 'webqa', 'race', 'drop']:
+                text = line['question']
+                compute_ngrams_insert_dict(args, text, ngrams)
+            elif task_name == 'natural_questions':
+                text = line['question']['text']
+                compute_ngrams_insert_dict(args, text, ngrams)
+            elif task_name == 'coqa':
+                all_questions = line['questions']
+                for question in all_questions:
+                    compute_ngrams_insert_dict(args, question, ngrams)
+            elif task_name == 'piqa':
+                text = line['goal']
+                compute_ngrams_insert_dict(args, text, ngrams)
+        except Exception as e:
+            print('Error:', e)
+
+    print(" After task {} entities in ngrams {}, added {}".format(task_name, \
+            len(ngrams), len(ngrams) - entities_in_ngrams), flush=True)
+
+def compute_tasks_ngrams(args, ngrams):
+    start_time = time.time()
+    for _, task_name in enumerate(args.tasks):
+        print('Task: {}'.format(task_name), flush=True)
+        if task_name == 'lambada':
+            assert args.lambada_path is not None
+            process_task_lambda(args, args.lambada_path, ngrams)
+        else:
+            process_task(args, task_name, ngrams)
+    print(" Taken time to compute ngrams {:.2f}".format(time.time() - \
+        start_time), flush=True)
+
+def compute_ngram_freq_sorted(args, ngrams):
+    ngrams_freq = {}
+    for ngram_key in ngrams.keys():
+        length = len(ngram_key.split())
+        ngrams_freq[length] = ngrams_freq[length] + 1 if length in \
+            ngrams_freq else 1
+
+    ngrams_freq_sorted = sorted(ngrams_freq.items(), key=lambda item: item[0])
+    print(" Ngram frequencies: {}".format(ngrams_freq_sorted), flush=True)
+    print(" Entities in ngrams {} min_ngram_size {} max_ngram_size {}".format(\
+            len(ngrams), ngrams_freq_sorted[0][0], ngrams_freq_sorted[len(\
+            ngrams_freq_sorted) -1 ][0]), flush=True)
+    return ngrams_freq_sorted
+
+def get_ngrams_below_threshold(args, ngrams, ngrams_below_threshold, \
+    dedup_file, dedup_key, ngrams_freq_sorted):
 
-    print('Reading file {} and deduping n-grams'.format(dedup_file))
-    counter = 0
     start_time = time.time()
-    out_f = open(output_file, 'wb')
-    splitted, ignored, split_mt_thld = 0, 0, 0
+    # get the ngrams frequency
+    args.get_ngram_freq_only = True
  
-    # Setup multi-processing.
-    num_workers = 40
+    # Open the large file to process in parallel
+    num_workers = args.num_threads 
+    pool = multiprocessing.Pool(num_workers)
     fin = open(dedup_file, 'r', encoding='utf-8')
+    free_ngram_abt_partial=partial(free_ngram, args=args, key=dedup_key, \
+        ngrams=ngrams, ngrams_freq_sorted=ngrams_freq_sorted)
+    free_ngrams_abt = pool.imap(free_ngram_abt_partial, fin, 500)
+ 
+    counter = 0
+    for _, _, _, local_ngram in free_ngrams_abt:
+        counter += 1
+        if counter % 1000 == 0:
+            print(' [compute_stat]> processed {} documents in {:.2f} seconds ...'.
+                    format(counter, time.time() - start_time), flush=True)
+        for local_key in local_ngram:
+            if local_key in ngrams:
+                ngrams[local_key] += 1
+        local_ngram = {}
+
+    print(' Time taken to compute statistics {:.2f} seconds'.format(time.time() - \
+        start_time), flush=True)
+    pool.close()
+    pool.join()
+
+    start_time = time.time()
+    counter_threshold = 0
+    # Get ngram below theadhold
+    for local_key, local_val in ngrams.items():
+        if ngrams[local_key] < args.key_threshold:
+            print(" [threshold] {} {}".format(local_key, local_val), flush=True)
+            counter_threshold += 1
+            ngrams_below_threshold[local_key] = 1
+            
+    print(' Ngrams below threshold {}'.format(counter_threshold), flush=True)
+    fin.close()
+
+def clean_ngrams_below_threshold(args, ngrams_below_threshold, dedup_file, \
+    dedup_key):
+
+    start_time = time.time()
+    # Now actually filter the dataset
+    args.get_ngram_freq_only = False
+    #id_prefix = '-'.join(args.tasks[::2])
+    id_prefix = '-'.join(args.tasks[::1])
+
+    # get the range of the size of the ngrams
+    ngrams_freq_sorted = compute_ngram_freq_sorted(args, ngrams_below_threshold)
+
+    # Open the large file to process in parallel
+    counter = splitted = ignored = split_mt_thld = trimmed_count = 0
+    num_workers = args.num_threads
     pool = multiprocessing.Pool(num_workers)
-    free_ngram_x=partial(free_ngram, ngrams=ngrams, ngram_size=ngram_size, 
-        filter_text_len=filter_text_len, splits_count=splits_count,
-        split_window_each_size=split_window_each_size)
-    free_ngrams = pool.imap(free_ngram_x, fin, 25)
+    fin = open(dedup_file, 'r', encoding='utf-8')
+    free_ngram_clean_partial=partial(free_ngram, args=args, key=dedup_key, \
+        ngrams=ngrams_below_threshold, ngrams_freq_sorted=ngrams_freq_sorted)
+    free_ngrams_clean = pool.imap(free_ngram_clean_partial, fin, 500)
+ 
+    out_f = open(args.output, 'wb')
 
-    for text_buf_ngram_free in free_ngrams:
+    for text_buf_ngram_free, trimmed, myjson, _ in free_ngrams_clean:
         counter += 1
         try:
 
+            trimmed_count += trimmed
+
             if len(text_buf_ngram_free) > 1:
-                splitted += (len(text_buf_ngram_free) - 1)
+                splitted += 1
             if len(text_buf_ngram_free) == 0:
                 ignored += 1
             # more than 10 splits ignored
-            if len(text_buf_ngram_free) > splits_count:
+            if len(text_buf_ngram_free) > args.splits_count:
                 text_buf_ngram_free = []
                 split_mt_thld += 1
 
+            if args.output is not None:
+                if "split_id" in myjson:
+                    use_prefix = myjson["split_id"] + "-"
+                else:
+                    use_prefix = ""
+
                 for i in range(len(text_buf_ngram_free)):
-                split_id_string = id_prefix + '-{:010d}'.format(int(counter)) \
-                    + '-{:010d}'.format(int(i))
-                outjson = json.dumps({"text":text_buf_ngram_free[i], 
-                    id_prefix+"_split_id":split_id_string},
-                    ensure_ascii=False)
+                    split_id_string = id_prefix + '-{:010d}'.format(int(\
+                        counter)) + '-{:04d}'.format(int(i))
+                    myjson[dedup_key] = text_buf_ngram_free[i]
+                    myjson["split_id"] = use_prefix + split_id_string
+                    outjson = json.dumps(myjson, ensure_ascii=False)
+                    #outjson = json.dumps({"text":text_buf_ngram_free[i],
+                    #    id_prefix+"_split_id":split_id_string},
+                    #    ensure_ascii=False)
                     out_f.write(outjson.encode('utf-8'))
                     out_f.write('\n'.encode('utf-8'))
 
             if counter % 1000 == 0:
-                print(' [search]> processed {} documents in {:.2f} seconds ...'.
+                print(' [final]> processed {} documents in {:.2f} seconds ...'.
                     format(counter, time.time() - start_time), flush=True)
         except Exception as e:
             print('Error:', e)
 
-    print("Deduped file written to: {}".format(output_file), flush=True)
-    print("Total docs {} splitted {} ignored {} docs with many splits {}".\
-        format(counter, splitted, ignored, split_mt_thld), flush=True)
+    print(' [final]> processed {} documents in {:.2f} seconds ...'.
+        format(counter, time.time() - start_time), flush=True)
+    
+    print(' Total docs {} splitted {} ignored {} splits > theshold {} trimmed'\
+        ' {}'.format(counter, splitted, ignored, split_mt_thld, trimmed_count)\
+        , flush=True)
+
+    pool.close()
+    pool.join()
+
+    out_f.close()
+    fin.close()
+
+if __name__ == '__main__':
+
+    # we use 13-grams, any text less than 200 characters got removed
+    # any text splitted more than 10 got removed as well
+
+    print('parsing the arguments ...')
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--tasks', nargs = '*', required=True, default=None, \
+                        help = 'Tasks to use for deduplication: currently '
+                        ' suuport [lambada, squad, natural_questions,'
+                        ' triviaqa, webqa, race, drop, coqa, and piqa]')
+    parser.add_argument('--lambada-path', type=str, default=None,
+                       help='Only Lambada task needs the path')
+    parser.add_argument('--dedup-dataset', nargs = '*', default=None,
+                       help='Dataset to deduplicate with the key to use'
+                        ' e.g. cc.json text')
+    parser.add_argument('--output', type=str, default=None,
+                       help='Output file name to save dedup dataset')
+    parser.add_argument('--num-threads', type=int, default=40,
+                       help='Number of threads to use')
+    # Default dedup values
+    parser.add_argument('--max-ngram-size', type=int, default=13,
+                       help='Maximum size of ngram to use.')
+    parser.add_argument('--min-ngram-size', type=int, default=8,
+                       help='Minimum size of ngram to use.')
+    parser.add_argument('--filter-text-char-len', type=int, default=200,
+                       help='Remove any text below this length.')
+    parser.add_argument('--key-threshold', type=int, default=10,
+                       help='Number of keys to consider as threshold')
+    parser.add_argument('--save-dictionary', type=str, default=None,
+                       help='Save the dictionary')
+    parser.add_argument('--load-dictionary', type=str, default=None,
+                       help='Load the dictionary')
+    parser.add_argument('--splits-count', type=int, default=10,
+                       help='Remove any documents more than this many splits')
+    parser.add_argument('--remove-char-each-side', type=int, default=200,
+                       help='Maximum size of ngram to use.')
+
+    args = parser.parse_args()
+
+    assert len(args.dedup_dataset) == 2
+    dedup_file = args.dedup_dataset[0]
+    dedup_key = args.dedup_dataset[1]
+
+    # Setup multi-processing
+    num_workers = args.num_threads
+    if args.load_dictionary is None:
+
+        # Build ngrams
+        ngrams = {}
+        compute_tasks_ngrams(args, ngrams)
+
+        # get the range of the size of the ngrams
+        ngrams_freq_sorted = compute_ngram_freq_sorted(args, ngrams)
+
+        # get ngram freq from large file in parallel
+        # get ngrams below threshold
+        ngrams_below_threshold = {}
+        get_ngrams_below_threshold(args, ngrams, ngrams_below_threshold, \
+            dedup_file, dedup_key, ngrams_freq_sorted)
+
+        # save the dictionary if needed
+        if args.save_dictionary is not None:
+            with open(args.save_dictionary, 'wb') as save_dict_handle:
+                pickle.dump(ngrams_below_threshold, save_dict_handle)
+    else:
+        with open(args.load_dictionary, 'rb') as load_dict_handle:
+            ngrams_below_threshold = pickle.load(load_dict_handle)
+
+    # filter the large file
+    if args.output is not None:
+        clean_ngrams_below_threshold(args, ngrams_below_threshold, \
+            dedup_file, dedup_key)
+
     print('done :-)')

tools/openwebtext/find_duplicates.py

@@ -14,13 +14,16 @@
 # limitations under the License.
 
 import argparse
+from functools import partial
 import itertools
 import json
 from lsh import cache, minhash
+import multiprocessing
 import numpy as np
 import time
 import pickle
 import sys
+import os
 
 # This function is adapted from:
 #   https://github.com/mattilyra/LSH/blob/master/examples/Introduction.ipynb
@@ -31,11 +34,158 @@ def shingles(text, char_ngram=5):
 
 # This function is adapted from:
 #  https://github.com/mattilyra/LSH/blob/master/examples/Introduction.ipynb
-def jaccard(set_a, set_b):
+def jaccard(set_a, set_b, args):
+    if len(set_a) < 1 or len(set_b) < 1:
+        return 0.0
+
     intersection = set_a & set_b
     union = set_a | set_b
+
+    if args.jaccard == 'min':
+        return len(intersection) / min(len(set_a), len(set_b))
+    elif args.jaccard == 'max':
+        return len(intersection) / max(len(set_a), len(set_b))
+    else:
         return len(intersection) / len(union)
 
+def compute_fingerprint(line, key):
+    try:
+        myjson = json.loads(line)
+        url = myjson[key]
+        text = myjson['text']
+        fingerprint = hasher.fingerprint(text)
+    except Exception as e:
+        print('Error:', e)
+        return None, None, None, False
+
+    return url, text, fingerprint, True
+
+def url_pairs_to_remove(args, bucket_urls, url_doc):
+    remove_urls_list = []
+    deduped_local, counter_local = 0, 0
+    iteration = 0
+    while len(bucket_urls) > 1:
+        if args.heuristic_iter != -1 and \
+            iteration == args.heuristic_iter:
+            break
+
+        items = list(bucket_urls)
+        remove_urls = []
+        main_url = items[np.random.randint(0, len(items))]
+        main_dhingles = shingles(url_doc[main_url])
+
+        for i in range(0, len(items)):
+            counter_local += 1
+            other_url = items[i]
+            if other_url == main_url:
+                continue
+            other_shingles = shingles(url_doc[other_url])
+            try:
+                jaccard_sim = jaccard(main_dhingles, other_shingles, args)
+            except Exception as e:
+                print('Error:', e)
+                jaccard_sim = 0.0
+            if jaccard_sim > 0.5:
+                remove_urls.append({other_url: jaccard_sim})
+                deduped_local += 1
+                bucket_urls.remove(other_url)
+
+        bucket_urls.remove(main_url)
+        if len(remove_urls) > 0:
+            remove_urls_list.append({main_url: remove_urls})
+        iteration += 1
+    return remove_urls_list, deduped_local, counter_local
+
+def write_remove_urls_list(remove_urls_list, f_out):
+    if len(remove_urls_list) > 0:
+        for each_url_remove in remove_urls_list:
+            myjson = json.dumps(each_url_remove, ensure_ascii=False)
+            f_out.write(myjson.encode('utf-8'))
+            f_out.write('\n'.encode('utf-8'))
+
+def compute_jaccard(each_bin, num_bins, start_time_local):
+
+    remove_urls_list = []
+    deduped_local, counter_local, bucket_local = 0, 0, 0
+
+    for bucket_id in each_bin:
+        bucket_local += 1
+        if os.getpid() % num_bins == 0 and bucket_local % 100000 == 0:
+            print("Counter {}, progress {:.2f} time {:.2f}".\
+                format(bucket_local, float(bucket_local)/float(len(each_bin)),\
+                time.time() - start_time_local), flush=True)
+
+        if len(each_bin[bucket_id]) <= 1:
+            continue
+
+        bucket_urls = each_bin[bucket_id].copy()
+        remove_urls_list_sub, deduped_local_sub, counter_local_sub = \
+            url_pairs_to_remove(args, bucket_urls, url_doc)
+
+        deduped_local += deduped_local_sub
+        counter_local += counter_local_sub
+        if len(remove_urls_list_sub) > 0:
+            remove_urls_list.extend(remove_urls_list_sub)
+
+    return remove_urls_list, deduped_local, counter_local
+
+def find_pair_urls_parallel(args, lshcache, url_doc):
+    start_time = time.time()
+    f_out = open(args.output, 'wb')
+    deduped, counter = 0, 0
+
+    # compute jaccards of buckets in bin in parallel (parallelism
+    # limited to # of bins)
+    num_bins = len(lshcache.bins)
+    pool = multiprocessing.Pool(num_bins)
+    compute_jaccard_partial = partial(compute_jaccard, num_bins=num_bins, \
+        start_time_local=start_time)
+    # don't need to pass args and url_doc as they are already shared
+    compute_jaccard_iter = pool.imap(compute_jaccard_partial, lshcache.bins)
+
+    print("multiprocessing init took {:.2f}".format(time.time() - start_time),\
+        flush=True)
+    for remove_urls_list, deduped_local, counter_local in compute_jaccard_iter:
+        deduped += deduped_local
+        counter += counter_local
+        write_remove_urls_list(remove_urls_list, f_out)
+        print(' [write]> processed {} documents in {:.2f} '
+            'seoncds and deduped {} documents ...'.format(counter, time.time()\
+            - start_time, deduped), flush=True)
+
+    pool.close()
+    pool.join()
+    f_out.close()
+
+    print(' Taken time for jaccard similariries {:.2f} seconds'.format(\
+        time.time() - start_time), flush=True)
+
+def find_pair_urls_sequential(args, lshcache, url_doc):
+    start_time = time.time()
+    f_out = open(args.output, 'wb')
+    deduped, counter = 0, 0
+    for b in lshcache.bins:
+        for bucket_id in b:
+            if len(b[bucket_id]) <= 1:
+                continue
+
+            bucket_urls = b[bucket_id].copy()
+            remove_urls_list_sub, deduped_local_sub, counter_local_sub = \
+                url_pairs_to_remove(args, bucket_urls, url_doc)
+
+            deduped += deduped_local_sub
+            counter += counter_local_sub
+            write_remove_urls_list(remove_urls_list_sub, f_out)
+            if counter % 10000 == 0:
+                print(' [write]> processed {} documents in {:.2f} '
+                    'seoncds and deduped {} documents ...'.
+                    format(counter, time.time() - start_time,
+                    deduped), flush=True)
+    f_out.close()
+    print(' [write]> processed {} documents in {:.2f} '
+        'seoncds and deduped {} documents ...'.
+        format(counter, time.time() - start_time,
+        deduped), flush=True)
 
 if __name__ == '__main__':
 
@@ -55,17 +205,30 @@ if __name__ == '__main__':
     parser.add_argument('--output', type=str, default=None,
                        help='Output file name that consists of all ids'
                         ' with matching similarities')
+    parser.add_argument('--jaccard', type=str, default='union',
+                        choices=['union', 'min', 'max'], help='Jaccard'\
+                        ' similarity computation')
+    parser.add_argument('--heuristic-iter', type=int, default=1,
+                       help='Number of iterations to run the heuristics'
+                        ': use -1 for exact')
+    parser.add_argument('--num-bands', type=int, default=10,
+                       help='Number of bands to use in cache')
+    parser.add_argument('--num-seeds', type=int, default=100,
+                       help='Number of seeds to use for minhash. Note that'
+                        ' this value should be divisible by num-bands')
+    parser.add_argument('--jaccard-parallel', action='store_true',
+                       help='Use this to process large number of documents.')
     args = parser.parse_args()
 
     print('finding possible duplicate content ...')
 
     # set seed and get an array of seeds of 100 integers
     np.random.seed(args.seed)
-    seeds = np.random.randint(0, 1e6, size=100)
+    seeds = np.random.randint(0, 1e6, size=args.num_seeds)
 
     # initialize minhash and lsh cache
     hasher = minhash.MinHasher(seeds=seeds, char_ngram=5, hashbytes=4)
-    lshcache = cache.Cache(bands=10, hasher=hasher)
+    lshcache = cache.Cache(num_bands=args.num_bands, hasher=hasher)
 
     url_doc = {}
 
@@ -91,32 +254,37 @@ if __name__ == '__main__':
     counter = 0
     start_time = time.time()
 
-    print("Computing fingerprints", flush=True)
-
     # compute finger prints of the inputs if any
     # input file and the key to use as id
     if args.inputs is not None:
+        print("Computing fingerprints", flush=True)
         assert len(args.inputs) % 2 == 0
         for input_file, key in zip(args.inputs[::2], args.inputs[1::2]):
             print(' document processing {} with key {}'.format(input_file, key),
                 flush=True)
+
+            # compute fingerprints in parallel
+            num_workers = 40
+            pool = multiprocessing.Pool(num_workers)
+            fin = open(input_file, 'r', encoding='utf-8')
+            compute_fingerprint_partial = partial(compute_fingerprint, key=key)
+            compute_fingerprint_iter = pool.imap(compute_fingerprint_partial,
+                                                    fin, 512)
             # traverse all the texts and add fingerprints
-            with open(input_file, 'r') as f_input:
-                for line in f_input:
-                    try:
-                        myjson = json.loads(line)
-                        url = myjson[key]
-                        text = myjson['text']
+            for url, text, fingerprint, flag in compute_fingerprint_iter:
                 counter += 1
+                if flag:
                     url_doc[url] = text
-                        lshcache.add_fingerprint(hasher.fingerprint(text), url)
-                    except Exception as e:
-                        print('Error:', e)
+                    lshcache.add_fingerprint(fingerprint, url)
                 if counter % 10000 == 0:
                     print(' [read]> processed {} documents in {:.2f} '
                         'seconds ...'.format(counter, time.time() - \
                         start_time), flush=True)
 
+            fin.close()
+            pool.close()
+            pool.join()
+
     # Save the fingerprints if needed
     if args.save_fingerprints is not None:
         print("Saving fingerprints to pickle file {}".format(
@@ -125,40 +293,13 @@ if __name__ == '__main__':
             pickle.dump(lshcache, f_save)
             pickle.dump(url_doc, f_save)
 
-    counter = 0
-    start_time = time.time()
-    deduped = 0
     # compute jaccard index of the input texts and write to file if needed
     if args.output is not None:
-        f_out = open(args.output, 'wb')
-        for b in lshcache.bins:
-            for bucket_id in b:
-                if len(b[bucket_id]) > 1:
-                    items = list(b[bucket_id])
-                    main_url = items[0]
-                    main_dhingles = shingles(url_doc[main_url])
-                    remove_urls = []
-                    for i in range(1, len(items)):
-                        counter += 1
-                        other_url= items[i]
-                        other_shingles = shingles(url_doc[other_url])
-                        try:
-                            jaccard_sim = jaccard(main_dhingles, other_shingles)
-                        except Exception as e:
-                            print('Error:', e)
-                        if jaccard_sim > 0.5:
-                            remove_urls.append({other_url: jaccard_sim})
-                            deduped += 1
-                        if counter % 10000 == 0:
-                            print(' [write]> processed {} documents in {:.2f} '
-                                  'seoncds and deduped {} documents ...'.
-                                  format(counter, time.time() - start_time,
-                                         deduped), flush=True)
-                    if len(remove_urls) > 0:
-                        myjson = json.dumps({main_url: remove_urls},
-                                            ensure_ascii=False)
-                        f_out.write(myjson.encode('utf-8'))
-                        f_out.write('\n'.encode('utf-8'))
-        f_out.close()
+        print("Compute jaccard similarity", flush=True)
+        if args.jaccard_parallel:
+            find_pair_urls_parallel(args, lshcache, url_doc)
+        else:
+            find_pair_urls_sequential(args, lshcache, url_doc)
 
     print('done :-)')
+ 

tools/openwebtext/group_duplicate_url.py

@@ -18,10 +18,6 @@ import time
 import sys
 
 
-def is_similar(jaccard_similarity):
-    return (jaccard_similarity >= 0.7)
-
-
 if __name__ == '__main__':
 
 
@@ -29,6 +25,10 @@ if __name__ == '__main__':
 
     input = sys.argv[1]
     output = sys.argv[2]
+    if len(sys.argv) > 3:
+        jaccard_similarity_threshold = float(sys.argv[3])
+    else:
+        jaccard_similarity_threshold = 0.7
 
     url_to_index = {}
     index_to_urls = []
@@ -43,7 +43,7 @@ if __name__ == '__main__':
                 urls.append(main_url)
                 for value in myjson[main_url]:
                     for other_url, js in value.items():
-                        if is_similar(js):
+                        if js >= jaccard_similarity_threshold:
                             urls.append(other_url)
             current_index = -1
             other_indices = set()