File size: 8,133 Bytes
e0fcef4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 |
"""data handling specific to pretraining"""
import functools
import logging
from collections import defaultdict
from typing import Callable, Dict, List, Optional
import torch
from datasets import Dataset
from torch.utils.data import RandomSampler
from transformers import PreTrainedTokenizerBase
from axolotl.utils.collators import PretrainingBatchSamplerDataCollatorForSeq2Seq
from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
from axolotl.utils.trainer import process_pretraining_datasets_for_packing
LOG = logging.getLogger("axolotl")
def encode_pretraining(
tokenizer: PreTrainedTokenizerBase, max_tokens: int, examples: List[str]
) -> Dict[str, List]:
res = tokenizer(
examples,
truncation=True,
max_length=max_tokens - 2,
add_special_tokens=True,
)
# Convert to PyTorch tensors
input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
new_input_ids = []
new_attention_mask = []
# Append EOS and PAD tokens to input_ids, and correct attention_mask
for i, _ in enumerate(input_ids):
input_ids[i] = torch.cat(
(
input_ids[i],
torch.tensor([tokenizer.eos_token_id, tokenizer.pad_token_id]),
),
dim=0,
)
attention_mask[i] = torch.cat((attention_mask[i], torch.tensor([1, 0])), dim=0)
# Concatenate tokens so that their lengths are less than max_tokens
buffer_input_ids = torch.tensor([], dtype=torch.long)
buffer_attention_mask = torch.tensor([], dtype=torch.long)
for ids, mask in zip(input_ids, attention_mask):
if buffer_input_ids.numel() == max_tokens:
new_input_ids.append(buffer_input_ids)
new_attention_mask.append(buffer_attention_mask)
buffer_input_ids = torch.tensor([], dtype=torch.long)
buffer_attention_mask = torch.tensor([], dtype=torch.long)
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
elif buffer_input_ids.numel() + ids.numel() <= max_tokens:
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
else:
buffer_input_ids = torch.cat(
(
buffer_input_ids,
torch.full(
(max_tokens - buffer_input_ids.numel(),),
tokenizer.pad_token_id,
dtype=torch.long,
),
),
dim=0,
)
buffer_attention_mask = torch.cat(
(
buffer_attention_mask,
torch.full(
(max_tokens - buffer_attention_mask.numel(),),
0,
dtype=torch.long,
),
),
dim=0,
)
new_input_ids.append(buffer_input_ids)
new_attention_mask.append(buffer_attention_mask)
buffer_input_ids = torch.tensor([], dtype=torch.long)
buffer_attention_mask = torch.tensor([], dtype=torch.long)
buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
if buffer_input_ids.numel() > 0: # for any leftover tokens
while buffer_input_ids.numel() < max_tokens: # make all sequences equal in size
buffer_input_ids = torch.cat(
(
buffer_input_ids,
torch.full(
(max_tokens - buffer_input_ids.numel(),),
tokenizer.pad_token_id,
dtype=torch.long,
),
),
dim=0,
)
buffer_attention_mask = torch.cat(
(
buffer_attention_mask,
torch.full(
(max_tokens - buffer_attention_mask.numel(),),
0,
dtype=torch.long,
),
),
dim=0,
)
new_input_ids.append(buffer_input_ids)
new_attention_mask.append(buffer_attention_mask)
ret = {
"input_ids": [seq.tolist() for seq in new_input_ids],
"labels": [seq.tolist() for seq in new_input_ids],
"attention_mask": [seq.tolist() for seq in new_attention_mask],
}
LOG.debug(len(ret["input_ids"]))
return ret
def wrap_pretraining_dataset(
dataset,
tokenizer,
cfg,
ds_wrapper_fn,
max_tokens=2048,
batch_size=1,
seed=42,
buffer_size=10_000,
):
if cfg.sample_packing:
collate_fn = PretrainingBatchSamplerDataCollatorForSeq2Seq(
tokenizer,
return_tensors="pt",
padding=True,
pad_to_multiple_of=max_tokens * batch_size,
multipack_attn=cfg.pretrain_multipack_attn,
)
encode = functools.partial(
encode_packed_pretraining,
collate_fn,
ds_wrapper_fn,
max_seq_length=max_tokens,
batch_size=batch_size,
multipack_attn=cfg.pretrain_multipack_attn,
)
# set this to 1 so downstream data_loader doesn't try to increase the batch again
cfg.micro_batch_size = 1
else:
encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
if cfg.shuffle_merged_datasets:
dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
else:
LOG.debug("NOT shuffling merged pretraining datasets")
# remove all the existing columns after mapping since they end up having
# a different length than the encoded/tokenized column
# this is empty during streaming/pretraining
remove_columns = []
if dataset.features is None:
for first_row in dataset:
remove_columns = first_row.keys()
break
else:
remove_columns = dataset.features.keys()
dataset = dataset.map(
encode,
batched=True,
batch_size=buffer_size,
# input_columns="text",
remove_columns=remove_columns,
)
return dataset
def encode_packed_pretraining(
collate_fn,
ds_wrapper: Callable,
examples: Dict[str, List],
max_seq_length: int = 2048,
batch_size: int = 4,
multipack_attn: Optional[bool] = False,
) -> Dict[str, List]:
# pylint: disable=duplicate-code
# tokenize all the examples
# rows get split with stride (overlap)
train_dataset = ds_wrapper(Dataset.from_dict(examples))[0]
train_dataset = process_pretraining_datasets_for_packing(
train_dataset,
max_seq_length,
skip_position_ids=not multipack_attn,
)
sampler = MultipackBatchSampler(
RandomSampler(train_dataset),
batch_size=1,
drop_last=True,
batch_max_len=batch_size * max_seq_length,
lengths=get_dataset_lengths(train_dataset),
)
chunked_data = defaultdict(list)
for batch in sampler:
for data in batch:
features = train_dataset[data]
if "num_truncated_tokens" in features:
del features["num_truncated_tokens"]
if "num_truncated_tokens" in features:
del features["num_truncated_tokens"]
if "overflow_to_sample_mapping" in features:
del features["overflow_to_sample_mapping"]
if "labels" not in features:
features["labels"] = features["input_ids"].copy()
collated_features = collate_fn(features)
for feature in features.keys():
if feature == "length":
continue
chunked_data[feature].append(collated_features[feature].squeeze(0))
return chunked_data
|