github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-10-29 04:40:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

[legacy] clean up legacy code (#4743)

Browse Source 
* [legacy] remove outdated codes of pipeline (#4692)

* [legacy] remove cli of benchmark and update optim (#4690)

* [legacy] remove cli of benchmark and update optim

* [doc] fix cli doc test

* [legacy] fix engine clip grad norm

* [legacy] remove outdated colo tensor (#4694)

* [legacy] remove outdated colo tensor

* [test] fix test import

* [legacy] move outdated zero to legacy (#4696)

* [legacy] clean up utils (#4700)

* [legacy] clean up utils

* [example] update examples

* [legacy] clean up amp

* [legacy] fix amp module

* [legacy] clean up gpc (#4742)

* [legacy] clean up context

* [legacy] clean core, constants and global vars

* [legacy] refactor initialize

* [example] fix examples ci

* [example] fix examples ci

* [legacy] fix tests

* [example] fix gpt example

* [example] fix examples ci

* [devops] fix ci installation

* [example] fix examples ci
This commit is contained in:
Hongxin Liu
2023-09-18 16:31:06 +08:00
committed by GitHub
parent 32e7f99416
commit b5f9e37c70
342 changed files with 2919 additions and 4182 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
examples/tutorial/sequence_parallel/data/__init__.py
View File

@@ -1,11 +1,13 @@
from colossalai.context.parallel_context import ParallelContext
from colossalai.core import global_context as gpc
from colossalai.logging import get_dist_logger
from colossalai.context import ParallelMode
from .datasets.data_samplers import build_pretraining_data_loader
from .datasets.builder import build_train_valid_test_datasets
import torch
from colossalai.legacy.context import ParallelMode
from colossalai.legacy.context.parallel_context import ParallelContext
from colossalai.legacy.core import global_context as gpc
from colossalai.logging import get_dist_logger
from .datasets.builder import build_train_valid_test_datasets
from .datasets.data_samplers import build_pretraining_data_loader
def cyclic_iter(iter):
while True:
@@ -18,8 +20,7 @@ def build_train_valid_test_data_iterators(train_iters,
eval_interval,
eval_iters,
dataloader_type='single',
**kwargs
):
**kwargs):
(train_dataloader, valid_dataloader, test_dataloader) = (None, None, None)
logger = get_dist_logger()
@@ -42,9 +43,7 @@ def build_train_valid_test_data_iterators(train_iters,
train_samples = train_iters * global_batch_size
eval_iters_ = (train_iters // eval_interval + 1) * eval_iters
test_iters = eval_iters
train_val_test_num_samples = [train_samples,
eval_iters_ * global_batch_size,
test_iters * global_batch_size]
train_val_test_num_samples = [train_samples, eval_iters_ * global_batch_size, test_iters * global_batch_size]
logger.info(' > datasets target sizes (minimum size):')
logger.info(' train: {}'.format(train_val_test_num_samples[0]), ranks=[0])
logger.info(' validation: {}'.format(train_val_test_num_samples[1]), ranks=[0])
@@ -56,19 +55,20 @@ def build_train_valid_test_data_iterators(train_iters,
# Build dataloaders.
dp_size = gpc.get_world_size(ParallelMode.DATA)
train_dataloader = build_pretraining_data_loader(
train_ds, consumed_samples=0, micro_batch_size=global_batch_size//dp_size)
valid_dataloader = build_pretraining_data_loader(
valid_ds, consumed_samples=0, micro_batch_size=global_batch_size//dp_size)
test_dataloader = build_pretraining_data_loader(test_ds, 0, micro_batch_size=global_batch_size//dp_size)
train_dataloader = build_pretraining_data_loader(train_ds,
consumed_samples=0,
micro_batch_size=global_batch_size // dp_size)
valid_dataloader = build_pretraining_data_loader(valid_ds,
consumed_samples=0,
micro_batch_size=global_batch_size // dp_size)
test_dataloader = build_pretraining_data_loader(test_ds, 0, micro_batch_size=global_batch_size // dp_size)
# Flags to know if we need to do training/validation/testing.
do_train = train_dataloader is not None and train_iters > 0
do_valid = valid_dataloader is not None and eval_iters > 0
do_test = test_dataloader is not None and eval_iters > 0
# Need to broadcast num_tokens and num_type_tokens.
flags = torch.cuda.LongTensor(
[int(do_train), int(do_valid), int(do_test)])
flags = torch.cuda.LongTensor([int(do_train), int(do_valid), int(do_test)])
else:
flags = torch.cuda.LongTensor([0, 0, 0])

23
examples/tutorial/sequence_parallel/data/bert_helper.py
View File

@@ -1,7 +1,8 @@
from colossalai.core import global_context as gpc
from colossalai.context import ParallelMode
import torch
from colossalai.legacy.context import ParallelMode
from colossalai.legacy.core import global_context as gpc
_MAX_DATA_DIM = 5
@@ -22,7 +23,8 @@ def _build_key_size_numel_dictionaries(keys, data):
# Move to GPU and broadcast.
sizes_cuda = torch.cuda.LongTensor(sizes)
torch.distributed.broadcast(sizes_cuda, gpc.get_ranks_in_group(ParallelMode.TENSOR)[0],
torch.distributed.broadcast(sizes_cuda,
gpc.get_ranks_in_group(ParallelMode.TENSOR)[0],
group=gpc.get_group(ParallelMode.TENSOR))
# Move back to cpu and unpack.
@@ -60,19 +62,15 @@ def broadcast_data(keys, data, datatype):
"""
# Build (key, size) and (key, number of elements) dictionaries along
# with the total number of elements on all ranks.
key_size, key_numel, total_numel = _build_key_size_numel_dictionaries(keys,
data)
key_size, key_numel, total_numel = _build_key_size_numel_dictionaries(keys, data)
# Pack on rank zero.
if not gpc.is_initialized(ParallelMode.TENSOR) or gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# Check that all keys have the same data type.
# Flatten the data associated with the keys
flatten_data = torch.cat(
[data[key].contiguous().view(-1) for key in keys], dim=0).cuda()
flatten_data = torch.cat([data[key].contiguous().view(-1) for key in keys], dim=0).cuda()
else:
flatten_data = torch.empty(total_numel,
device=torch.cuda.current_device(),
dtype=datatype)
flatten_data = torch.empty(total_numel, device=torch.cuda.current_device(), dtype=datatype)
# Broadcast
torch.distributed.broadcast(flatten_data,
@@ -139,7 +137,7 @@ def get_batch_for_sequence_parallel(data_iterator):
seq_length = data_b['text'].size(1)
sub_seq_length = seq_length // local_world_size
sub_seq_start = local_rank * sub_seq_length
sub_seq_end = (local_rank+1) * sub_seq_length
sub_seq_end = (local_rank + 1) * sub_seq_length
#
# # Unpack.
tokens = data_b['text'][:, sub_seq_start:sub_seq_end].long()
@@ -156,10 +154,9 @@ class SequenceParallelDataIterator:
def __init__(self, data_iter):
self.data_iter = data_iter
def __iter__(self):
return self.data_iter
def __next__(self):
return get_batch_for_sequence_parallel(self.data_iter)
return get_batch_for_sequence_parallel(self.data_iter)

4
examples/tutorial/sequence_parallel/data/datasets/bert_dataset.py
View File

@@ -21,8 +21,8 @@ import numpy as np
import torch
from torch.utils.data import Dataset
from colossalai.context import ParallelMode
from colossalai.core import global_context as gpc
from colossalai.legacy.context import ParallelMode
from colossalai.legacy.core import global_context as gpc
from colossalai.logging import get_dist_logger
from ..tokenizer import get_tokenizer

8
examples/tutorial/sequence_parallel/data/datasets/data_samplers.py
View File

@@ -14,10 +14,12 @@
# limitations under the License.
"""Dataloaders."""
import torch
import random
from colossalai.core import global_context as gpc
from colossalai.context import ParallelMode
import torch
from colossalai.legacy.context import ParallelMode
from colossalai.legacy.core import global_context as gpc
def build_pretraining_data_loader(dataset, consumed_samples, micro_batch_size, dataloader_type='single', num_workers=0):

30
examples/tutorial/sequence_parallel/data/tokenizer/tokenizer.py
View File

@@ -12,13 +12,12 @@
# 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.
"""Megatron tokenizers."""
from abc import ABC
from abc import abstractmethod
from colossalai.core import global_context as gpc
from colossalai.context import ParallelMode
from abc import ABC, abstractmethod
from colossalai.legacy.context import ParallelMode
from colossalai.legacy.core import global_context as gpc
from .bert_tokenization import FullTokenizer as FullBertTokenizer
@@ -26,18 +25,13 @@ from .bert_tokenization import FullTokenizer as FullBertTokenizer
def build_tokenizer(vocab_file, tokenizer_type, vocab_extra_ids=0):
"""Initialize tokenizer."""
if not gpc.is_initialized(ParallelMode.GLOBAL) or gpc.get_global_rank() == 0:
print('> building {} tokenizer ...'.format(tokenizer_type),
flush=True)
print('> building {} tokenizer ...'.format(tokenizer_type), flush=True)
# Select and instantiate the tokenizer.
if tokenizer_type == 'BertWordPieceLowerCase':
tokenizer = _BertWordPieceTokenizer(vocab_file=vocab_file,
lower_case=True,
vocab_extra_ids=vocab_extra_ids)
tokenizer = _BertWordPieceTokenizer(vocab_file=vocab_file, lower_case=True, vocab_extra_ids=vocab_extra_ids)
elif tokenizer_type == 'BertWordPieceCase':
tokenizer = _BertWordPieceTokenizer(vocab_file=vocab_file,
lower_case=False,
vocab_extra_ids=vocab_extra_ids)
tokenizer = _BertWordPieceTokenizer(vocab_file=vocab_file, lower_case=False, vocab_extra_ids=vocab_extra_ids)
else:
raise NotImplementedError('{} tokenizer is not '
'implemented.'.format(tokenizer_type))
@@ -62,8 +56,8 @@ def _vocab_size_with_padding(orig_vocab_size, make_vocab_size_divisible_by=128):
after += 1
if not gpc.is_initialized(ParallelMode.GLOBAL) or gpc.get_global_rank() == 0:
print(' > padded vocab (size: {}) with {} dummy tokens '
'(new size: {})'.format(
orig_vocab_size, after - orig_vocab_size, after), flush=True)
'(new size: {})'.format(orig_vocab_size, after - orig_vocab_size, after),
flush=True)
return after
@@ -142,8 +136,7 @@ class _BertWordPieceTokenizer(AbstractTokenizer):
self._additional_special_tokens = []
# (dsachan) Add BOS and EOS tokens
SPECIAL_TOKENS = {'eos_token': '[EOS]',
'bos_token': '[BOS]'}
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)
@@ -155,8 +148,7 @@ class _BertWordPieceTokenizer(AbstractTokenizer):
# (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)])
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):