File size: 7,451 Bytes
12001a9 |
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 |
import os
from unittest.mock import MagicMock
import requests
from torch.utils.data import IterableDataset
def train_tokenizer(destination_path):
destination_path.mkdir(parents=True, exist_ok=True)
# download the tiny shakespeare dataset
input_file_path = destination_path / "input.txt"
if not input_file_path.exists():
data_url = "https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt"
with open(input_file_path, "w") as f:
f.write(requests.get(data_url).text)
from lit_llama import Tokenizer
Tokenizer.train(
input=input_file_path,
destination=destination_path,
vocab_size=100,
)
return destination_path / "tokenizer.model"
def test_packed_dataset(tmp_path):
tokenizer_path = train_tokenizer(tmp_path)
from lit_llama import Tokenizer
tokenizer = Tokenizer(tokenizer_path)
texts = [
"The moment of truth is upon us.",
"Time to open the fridge."
]
from lit_llama.packed_dataset import PackedDatasetBuilder, PackedDataset, HDR_SIZE
block_size = 10
n_blocks = 2
chunk_size = block_size * n_blocks
builder = PackedDatasetBuilder(
outdir=tmp_path,
prefix="packed_dataset",
chunk_size=chunk_size,
sep_token=tokenizer.bos_id,
dtype="auto",
vocab_size=100,
)
text_ids = []
for text in texts:
text_ids = tokenizer.encode(text)
assert text_ids[0] == tokenizer.bos_id
builder.add_array(text_ids)
filenames = builder.filenames
assert len(filenames) == 2
assert os.path.basename(filenames[0]) == "packed_dataset_0000000000.bin"
assert os.path.basename(filenames[1]) == "packed_dataset_0000000001.bin"
import numpy as np
ex_tokenized = [
tokenizer.encode(text).numpy().astype(builder.dtype)
for text in texts
]
ex_tokenized = np.concatenate(ex_tokenized)
ex_tokenized = ex_tokenized[:2 * chunk_size]
for filename, el in zip(filenames, np.array_split(ex_tokenized, 2)):
mmap = np.memmap(filename, mode="r", order="C", offset=HDR_SIZE)
count = len(mmap) // np.dtype(builder.dtype).itemsize
arr = np.frombuffer(
mmap, dtype=builder.dtype, count=count, offset=0
)
where_bos = np.where(arr == tokenizer.bos_id)
# we expect two BOS tokens, one per file
assert len(where_bos) == 1
assert np.array_equal(arr, el)
dataset = PackedDataset(filenames=filenames, n_chunks=2, block_size=block_size, shuffle=False)
ex_split = np.array_split(ex_tokenized, ex_tokenized.shape[0] // block_size)
for item, el in zip(dataset, ex_split):
assert np.array_equal(item, el)
dataset = PackedDataset(filenames=filenames, n_chunks=2, block_size=block_size, seed=12345)
for i, item in enumerate(dataset):
block_idxs = iter(dataset)._block_idxs
assert np.array_equal(item, ex_split[block_idxs[i]])
dataset = PackedDataset(filenames=filenames, n_chunks=2, block_size=block_size, seed=12345, wrap=True)
for i, item in enumerate(dataset):
if i > 24:
break
dataset = PackedDataset(filenames=filenames, n_chunks=1, block_size=block_size, seed=12345)
for i, item in enumerate(dataset):
block_idxs = iter(dataset)._block_idxs
chunk_idx = i // n_blocks * n_blocks
assert np.array_equal(item, ex_split[chunk_idx + block_idxs[i % n_blocks]])
block_size_ = block_size // 2
ex_split = np.array_split(ex_tokenized, ex_tokenized.shape[0] // block_size_)
dataset = PackedDataset(filenames=filenames, n_chunks=2, block_size=block_size_, seed=12345)
for i, item in enumerate(dataset):
block_idxs = iter(dataset)._block_idxs
assert np.array_equal(item, ex_split[block_idxs[i]])
block_size_ = block_size // 3
n_chunks = 2
ex_chunks = np.split(ex_tokenized, n_chunks)
n_splits = ex_tokenized.shape[0] // n_chunks // block_size_
ex_splits = [np.split(el[:n_splits * block_size_], n_splits) for el in ex_chunks]
ex_split = sum(ex_splits, [])
dataset = PackedDataset(filenames=filenames, n_chunks=n_chunks, block_size=block_size_, seed=12345)
for i, item in enumerate(dataset):
block_idxs = iter(dataset)._block_idxs
assert np.array_equal(item, ex_split[block_idxs[i]])
class SimpleDataset(IterableDataset):
def __init__(self, start, end):
super().__init__()
self._start = start
self._end = end
def __iter__(self):
return iter(range(self._start, self._end))
def test_combined_dataset(tmp_path):
from lit_llama.packed_dataset import CombinedDataset
dataset1 = SimpleDataset(0, 10)
dataset2 = SimpleDataset(10, 20)
dataset = CombinedDataset(datasets=[dataset1, dataset2], weights=[1.0, 0.0], seed=12345)
res = [el for el in dataset]
assert res == list(range(0, 10))
dataset1 = SimpleDataset(0, 10)
dataset2 = SimpleDataset(10, 20)
dataset = CombinedDataset(datasets=[dataset1, dataset2], weights=[0.0, 1.0], seed=12345)
res = [el for el in dataset]
assert res == list(range(10, 20))
dataset1 = SimpleDataset(0, 10)
dataset2 = SimpleDataset(10, 20)
dataset = CombinedDataset(datasets=[dataset1, dataset2], weights=[0.5, 0.5], seed=12345)
res = [el for el in dataset]
assert 9 in res or 19 in res
if len(res) > 10:
assert 0 in res and 10 in res
def test_sharded_packed_dataset(monkeypatch):
import lit_llama.packed_dataset
from lit_llama.packed_dataset import PackedDataset
dataset_iterator_mock = MagicMock()
monkeypatch.setattr(lit_llama.packed_dataset, "PackedDatasetIterator", dataset_iterator_mock)
filenames = [str(i) for i in range(10)]
# world_size = 1, rank = 0
iter(PackedDataset(filenames=filenames, n_chunks=2, block_size=2))
assert dataset_iterator_mock.call_args[1]["filenames"] == filenames
dataset_iterator_mock.reset_mock()
# world_size = 2, rank = 0
iter(PackedDataset(filenames=filenames, n_chunks=2, block_size=2, num_processes=2, process_rank=0))
assert dataset_iterator_mock.call_args[1]["filenames"] == ["0", "2", "4", "6", "8"]
dataset_iterator_mock.reset_mock()
# world_size = 2, rank = 1
iter(PackedDataset(filenames=filenames, n_chunks=2, block_size=2, num_processes=2, process_rank=1))
assert dataset_iterator_mock.call_args[1]["filenames"] == ["1", "3", "5", "7", "9"]
dataset_iterator_mock.reset_mock()
# world_size = 3, rank = 0 (dataset size not cleanly divisible by world size)
iter(PackedDataset(filenames=filenames, n_chunks=2, block_size=2, num_processes=3, process_rank=0))
assert dataset_iterator_mock.call_args[1]["filenames"] == ["0", "3", "6"]
dataset_iterator_mock.reset_mock()
# world_size = 3, rank = 1 (dataset size not cleanly divisible by world size)
iter(PackedDataset(filenames=filenames, n_chunks=2, block_size=2, num_processes=3, process_rank=1))
assert dataset_iterator_mock.call_args[1]["filenames"] == ["1", "4", "7"]
dataset_iterator_mock.reset_mock()
# world_size = 3, rank = 2 (dataset size not cleanly divisible by world size)
iter(PackedDataset(filenames=filenames, n_chunks=2, block_size=2, num_processes=3, process_rank=2))
assert dataset_iterator_mock.call_args[1]["filenames"] == ["2", "5", "8"]
|