Merge staging-realm into hashing

generate_samples_gpt2.py → tools/generate_samples_gpt2.py

@@ -15,6 +15,11 @@
 
 """Sample Generate GPT2"""
 
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
+                                             os.path.pardir)))
+
 from megatron import get_args
 from megatron import get_tokenizer
 from megatron import print_rank_0

merge_mp_partitions.py → tools/merge_mp_partitions.py

+# coding=utf-8
+# Copyright (c) 2019, 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.
+
+"""Merge model parallel partitions."""
 
 import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
+                                             os.path.pardir)))
+
 import torch
 
-from arguments import get_args
 from megatron import mpu
-from megatron.utils import ensure_directory_exists
-from megatron.utils import get_checkpoint_name
-from megatron.utils import get_checkpoint_tracker_filename
-from megatron.utils import vocab_size_with_padding
+from megatron.checkpointing import ensure_directory_exists
+from megatron.checkpointing import get_checkpoint_name
+from megatron.checkpointing import get_checkpoint_tracker_filename
+from megatron.global_vars import rebuild_tokenizer
+from megatron.global_vars import _parse_args
 
 
 def split_into_partitions(tensor, num_partitions, partition_dim, stride):
@@ -84,21 +104,26 @@ def merge_partitions(merged, partitions, partition_dim, stride):
     return
 
 
-def get_model(model_type, args):
+def get_model(model_type):
 
     if model_type == 'BERT':
-        from pretrain_albert import model_provider
-        args.tokentype_size = 2
-    elif  model_type == 'GPT':
+        from pretrain_bert import model_provider
+    elif model_type == 'GPT2':
         from pretrain_gpt2 import model_provider
+    elif model_type == 'RACE':
+        from tasks.race.finetune import model_provider
+    elif model_type == ['MNLI', 'QQP']:
+        num_classes = 2
+        if model_type == 'MNLI':
+            num_classes = 3
+        from megatron.model.classification import Classification
+        def model_provider():
+            return Classification(num_classes=num_classes, num_tokentypes=2)
     else:
         raise Exception('unrecognized model type: {}'.format(model_type))
 
-    orig_vocab_size = args.vocab_size
-    args.vocab_size = vocab_size_with_padding(args.vocab_size, args)
-    model = model_provider(args)
+    model = model_provider()
     model = model.half()
-    args.vocab_size = orig_vocab_size
 
     return model
 
@@ -147,17 +172,32 @@ def test_split_merge():
     print('  > max error (should be zero): {}'.format(max_error))
 
 
-def main(model_type):
+def get_mp_merge_args(parser):
+    """Provide extra arguments required for merging."""
+    group = parser.add_argument_group(title='mp merge')
+
+    group.add_argument('--model-type', type=str, required=True,
+                       choices=['BERT', 'GPT2', 'RACE', 'MNLI', 'QQP'],
+                       help='Type of the mdoel.')
+
+    return parser
+
+
+def main():
 
     # Args
-    args = get_args()
+    args = _parse_args(extra_args_provider=get_mp_merge_args)
+    model_type = args.model_type
+    orig_model_parallel_size = args.model_parallel_size
+    args.model_parallel_size = 1
+    tokenizer = rebuild_tokenizer(args)
 
     print('\n merging model parallel partitions ...')
-    assert args.vocab_size is not None
-    print(' > number of partitions: {}'.format(args.model_parallel_size))
+    print(' > number of partitions: {}'.format(orig_model_parallel_size))
     print(' > checkpoint path: {}'.format(args.load))
     print(' > model parameters:')
-    print('    number of tokens ................ {} '.format(args.vocab_size))
+    print('    number of tokens ................ {} '.format(
+        tokenizer.vocab_size))
     print('    number of layers ................ {}'.format(args.num_layers))
     print('    hidden sise ..................... {}'.format(args.hidden_size))
     print('    number of attention heads ....... {}'.format(
@@ -169,17 +209,19 @@ def main(model_type):
     print('> building the full model ...')
     mpu.initialize.set_model_parallel_world_size(1)
     mpu.initialize.set_model_parallel_rank(0)
-    merged_model = get_model(model_type, args)
+    merged_model = get_model(model_type)
 
     # Build and load partitions.
     partitions = []
     iteration = 0
+    args.model_parallel_size = orig_model_parallel_size
+    tokenizer = rebuild_tokenizer(args)
     mpu.initialize.set_model_parallel_world_size(args.model_parallel_size)
     for rank in range(args.model_parallel_size):
         mpu.initialize.set_model_parallel_rank(rank)
         checkpoint_name, iteration = get_parallel_checkpoint_name(args.load)
         print('> loading {} ...'.format(checkpoint_name))
-        model_ = get_model(model_type, args)
+        model_ = get_model(model_type)
         sd = torch.load(checkpoint_name, map_location='cpu')
         model_.load_state_dict(sd['model'])
         partitions.append(model_)
@@ -225,6 +267,7 @@ def main(model_type):
 
 
     # Save the model.
+    args.model_parallel_size = 1
     mpu.initialize.set_model_parallel_rank(0)
     sd = {}
     sd['model'] = merged_model.state_dict_for_save_checkpoint()
@@ -240,4 +283,4 @@ def main(model_type):
 
 if __name__ == '__main__':
 
-    main('BERT')
+    main()

tools/preprocess_data.py

+# coding=utf-8
+# Copyright (c) 2019, 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.
+
+"""Processing data for pretraining."""
+
+import argparse
+import json
+import multiprocessing
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
+                                             os.path.pardir)))
+import time
+
+import torch
+try:
+    import nltk
+    nltk_available = True
+except ImportError:
+    nltk_available = False
+
+from megatron.tokenizer import build_tokenizer
+from megatron.data import indexed_dataset
+
+
+# https://stackoverflow.com/questions/33139531/preserve-empty-lines-with-nltks-punkt-tokenizer
+class CustomLanguageVars(nltk.tokenize.punkt.PunktLanguageVars):
+
+    _period_context_fmt = r"""
+        \S*                          # some word material
+        %(SentEndChars)s             # a potential sentence ending
+        \s*                       #  <-- THIS is what I changed
+        (?=(?P<after_tok>
+            %(NonWord)s              # either other punctuation
+            |
+            (?P<next_tok>\S+)     #  <-- Normally you would have \s+ here
+        ))"""
+
+class IdentitySplitter(object):
+    def tokenize(self, *text):
+        return text
+
+class Encoder(object):
+    def __init__(self, args):
+        self.args = args
+
+    def initializer(self):
+        # Use Encoder class as a container for global data
+        Encoder.tokenizer = build_tokenizer(self.args)
+        if self.args.split_sentences:
+            if not nltk_available:
+                print("NLTK is not available to split sentences.")
+                exit()
+            splitter = nltk.load("tokenizers/punkt/english.pickle")
+            if self.args.keep_newlines:
+                # this prevents punkt from eating newlines after sentences
+                Encoder.splitter = nltk.tokenize.punkt.PunktSentenceTokenizer(
+                    train_text = splitter._params,
+                    lang_vars = CustomLanguageVars())
+            else:
+                Encoder.splitter = splitter
+
+        else:
+            Encoder.splitter = IdentitySplitter()
+
+    def encode(self, json_line):
+        data = json.loads(json_line)
+        ids = {}
+        for key in self.args.json_keys:
+            text = data[key]
+            doc_ids = []
+            for sentence in Encoder.splitter.tokenize(text):
+                sentence_ids = Encoder.tokenizer.tokenize(sentence)
+                if len(sentence_ids) > 0:
+                    doc_ids.append(sentence_ids)
+            if self.args.append_eod:
+                doc_ids[-1].append(Encoder.tokenizer.eod)
+            ids[key] = doc_ids
+        return ids, len(json_line)
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    group = parser.add_argument_group(title='input data')
+    group.add_argument('--input', type=str, required=True,
+                       help='Path to input JSON')
+    group.add_argument('--json-keys', nargs='+', default=['text'],
+                       help='space separate listed of keys to extract from json')
+    group.add_argument('--split-sentences', action='store_true',
+                       help='Split documents into sentences.')
+    group.add_argument('--keep-newlines', action='store_true',
+                       help='Keep newlines between sentences when splitting.')
+
+    group = parser.add_argument_group(title='tokenizer')
+    group.add_argument('--tokenizer-type', type=str, required=True,
+                       choices=['BertWordPieceLowerCase',
+                                'GPT2BPETokenizer'],
+                       help='What type of tokenizer to use.')
+    group.add_argument('--vocab-file', type=str, default=None,
+                       help='Path to the vocab file')
+    group.add_argument('--merge-file', type=str, default=None,
+                       help='Path to the BPE merge file (if necessary).')
+    group.add_argument('--append-eod', action='store_true',
+                       help='Append an <eod> token to the end of a document.')
+
+
+    group = parser.add_argument_group(title='output data')
+    group.add_argument('--output-prefix', type=str, required=True,
+                       help='Path to binary output file without suffix')
+    group.add_argument('--dataset-impl', type=str, default='mmap',
+                       choices=['lazy', 'cached', 'mmap'])
+
+    group = parser.add_argument_group(title='runtime')
+    group.add_argument('--workers', type=int, default=1,
+                       help='Number of worker processes to launch')
+    group.add_argument('--log-interval', type=int, default=100,
+                       help='Interval between progress updates')
+    args = parser.parse_args()
+    args.keep_empty = False
+
+    if args.tokenizer_type.lower().startswith('bert'):
+        if not args.split_sentences:
+            print("Bert tokenizer detected, are you sure you don't want to split sentences?")
+
+    # some default/dummy values for the tokenizer
+    args.rank = 0
+    args.make_vocab_size_divisible_by = 128
+    args.model_parallel_size = 1
+
+    return args
+
+def main():
+    args = get_args()
+    startup_start = time.time()
+
+    print("Opening", args.input)
+    fin = open(args.input, 'r', encoding='utf-8')
+
+    if nltk_available and args.split_sentences:
+        nltk.download("punkt", quiet=True)
+
+    encoder = Encoder(args)
+    tokenizer = build_tokenizer(args)
+    pool = multiprocessing.Pool(args.workers, initializer=encoder.initializer)
+    encoded_docs = pool.imap(encoder.encode, fin, 25)
+    #encoded_docs = map(encoder.encode, fin)
+
+    level = "document"
+    if args.split_sentences:
+        level = "sentence"
+
+    print(f"Vocab size: {tokenizer.vocab_size}")
+    print(f"Output prefix: {args.output_prefix}")
+    output_bin_files = {}
+    output_idx_files = {}
+    builders = {}
+    for key in args.json_keys:
+        output_bin_files[key] = "{}_{}_{}.bin".format(args.output_prefix,
+                                                      key, level)
+        output_idx_files[key] = "{}_{}_{}.idx".format(args.output_prefix,
+                                                      key, level)
+        builders[key] = indexed_dataset.make_builder(output_bin_files[key],
+                                               impl=args.dataset_impl,
+                                               vocab_size=tokenizer.vocab_size)
+
+    startup_end = time.time()
+    proc_start = time.time()
+    total_bytes_processed = 0
+    print("Time to startup:", startup_end - startup_start)
+
+    for i, (doc, bytes_processed) in enumerate(encoded_docs, start=1):
+        total_bytes_processed += bytes_processed
+        for key, sentences in doc.items():
+            for sentence in sentences:
+                builders[key].add_item(torch.IntTensor(sentence))
+            builders[key].end_document()
+        if i % args.log_interval == 0:
+            current = time.time()
+            elapsed = current - proc_start
+            mbs = total_bytes_processed/elapsed/1024/1024
+            print(f"Processed {i} documents",
+                  f"({i/elapsed} docs/s, {mbs} MB/s).",
+                  file=sys.stderr)
+
+    for key in args.json_keys:
+        builders[key].finalize(output_idx_files[key])
+
+if __name__ == '__main__':
+    main()