Spaces:
Running on Zero
Running on Zero
File size: 10,806 Bytes
d1f1097 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 | # Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
##
# Compile megatron.core.datasets.helpers_cpp dependencies before BlendedDataset import
##
import os
import tempfile
from collections import defaultdict
from typing import Dict, Optional
import numpy
import pytest
import torch
from megatron.core.datasets.blended_megatron_dataset_builder import BlendedMegatronDatasetBuilder
from megatron.core.datasets.blended_megatron_dataset_config import BlendedMegatronDatasetConfig
from megatron.core.datasets.megatron_dataset import LowLevelDataset, MegatronDataset
from megatron.core.datasets.utils import Split, compile_helpers, get_blend_from_list
from tests.unit_tests.test_utilities import Utils
_NUM_DATASETS = 10
_SEQUENCE_LENGTH = 10
_SIZES = {}
for split in Split:
_SIZES[split] = []
for i in range(_NUM_DATASETS):
_SIZES[split].append({Split.train: 1000, Split.valid: 100, Split.test: 10}[split] * (i + 1))
_MARGIN = 0.005
def do_setup(odir):
paths = defaultdict(list)
for i in range(_NUM_DATASETS):
path_to_data = os.path.join(odir, str(i))
os.mkdir(path_to_data)
for split in _SIZES:
data = numpy.zeros((_SIZES[split][i], _SEQUENCE_LENGTH))
path = os.path.join(path_to_data, f"{split.name}.npy")
numpy.save(path, data)
paths[split].append(path)
return paths
def test_builder():
if torch.distributed.is_available():
Utils.initialize_distributed()
if torch.distributed.get_rank() == 0:
compile_helpers()
torch.distributed.barrier()
else:
compile_helpers()
# Define the class here to avoid pytest warnings
class TestDataset(MegatronDataset):
def __init__(
self,
dataset: LowLevelDataset,
dataset_path: Optional[str],
indices: numpy.ndarray,
num_samples: Optional[int],
index_split: Split,
config: BlendedMegatronDatasetConfig,
) -> None:
super().__init__(dataset, dataset_path, indices, num_samples, index_split, config)
if self.num_samples is None:
self.num_samples = len(self.indices)
self.sample_index = numpy.random.choice(self.indices, size=self.num_samples)
@staticmethod
def numel_low_level_dataset(low_level_dataset: LowLevelDataset) -> int:
return len(low_level_dataset)
@staticmethod
def build_low_level_dataset(
dataset_path: str, config: BlendedMegatronDatasetConfig
) -> LowLevelDataset:
return numpy.load(dataset_path)
def __len__(self) -> int:
return len(self.sample_index)
def __getitem__(self, idx: int) -> Dict[str, numpy.ndarray]:
return {"text": self.dataset[self.sample_index[idx]]}
with tempfile.TemporaryDirectory() as temp_dir:
paths = do_setup(temp_dir)
blends = {
split: get_blend_from_list(
[
weight_or_path
for pair in zip(list(range(1, len(paths[split]) + 1, 1)), paths[split])
for weight_or_path in pair
]
)
for split in Split
}
blends_unweighted = {split: (blends[split][0], None) for split in blends}
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend_per_split=[blends[Split.train], None, None],
mid_level_dataset_surplus=0.005,
)
try:
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [None, None, None], lambda: True, config
).build()
raise RuntimeError
except AssertionError:
pass
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend_per_split=[get_blend_from_list([paths[Split.train][0]]), None, None],
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [1000, None, None], lambda: True, config
).build()
assert len(datasets[0]) == 1000 and isinstance(datasets[0], TestDataset)
assert datasets[1] is None
assert datasets[2] is None
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend_per_split=[
blends_unweighted[Split.train],
blends_unweighted[Split.valid],
blends_unweighted[Split.test],
],
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [1000, 1000, 1000], lambda: True, config
).build()
assert len(datasets[0]) == 1000
assert len(datasets[1]) == 1000
assert len(datasets[2]) == sum(_SIZES[Split.test])
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend_per_split=[
blends_unweighted[Split.train],
blends_unweighted[Split.valid],
blends_unweighted[Split.test],
],
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [None, None, None], lambda: True, config
).build()
assert len(datasets[0]) == sum(_SIZES[Split.train])
assert numpy.all(
numpy.array(datasets[0].weights)
== numpy.unique(datasets[0].dataset_index, return_counts=True)[1]
)
assert len(datasets[1]) == sum(_SIZES[Split.valid])
assert numpy.all(
numpy.array(datasets[1].weights)
== numpy.unique(datasets[1].dataset_index, return_counts=True)[1]
)
assert len(datasets[2]) == sum(_SIZES[Split.test])
assert numpy.all(
numpy.array(datasets[2].weights)
== numpy.unique(datasets[2].dataset_index, return_counts=True)[1]
)
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend_per_split=[blends_unweighted[Split.train], None, None],
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [1000, None, None], lambda: True, config
).build()
assert len(datasets[0]) == 1000
for i in range(_NUM_DATASETS):
assert len(datasets[0].datasets[i]) == _SIZES[Split.train][i]
assert datasets[1] is None
assert datasets[2] is None
# This build used to fail when building datasets without a sample buffer
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend_per_split=[blends[Split.train], None, None],
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [1000, None, None], lambda: True, config
).build()
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend=blends_unweighted[Split.train],
split="100,0,0",
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [None, None, None], lambda: True, config
).build()
assert len(datasets[0]) == sum(_SIZES[Split.train])
assert numpy.all(
numpy.array(datasets[0].weights)
== numpy.unique(datasets[0].dataset_index, return_counts=True)[1]
)
assert datasets[1] is None
assert datasets[2] is None
if torch.distributed.is_initialized():
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend=blends_unweighted[Split.train],
split="100,0,0",
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset,
[None, None, None],
lambda: torch.distributed.get_rank() % 2 == 0,
config,
).build()
if torch.distributed.get_rank() % 2 == 0:
assert len(datasets[0]) == sum(_SIZES[Split.train])
assert numpy.all(
numpy.array(datasets[0].weights)
== numpy.unique(datasets[0].dataset_index, return_counts=True)[1]
)
else:
assert datasets[0] is None
assert datasets[1] is None
assert datasets[2] is None
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend=blends_unweighted[Split.train],
split="50,50,0",
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [1000, 0, None], lambda: True, config
).build()
assert len(datasets[0]) == 1000
assert sum(map(len, datasets[0].datasets)) == sum(_SIZES[Split.train]) / 2
assert sum(map(len, datasets[1].datasets)) == sum(_SIZES[Split.train]) / 2
assert datasets[1] is not None and len(datasets[1]) == 0
assert datasets[2] is None
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend=blends_unweighted[Split.train],
split="50,50,0",
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset,
[int(sum(_SIZES[Split.train]) / 4), int(sum(_SIZES[Split.train])), None],
lambda: True,
config,
).build()
assert len(datasets[0]) == sum(_SIZES[Split.train]) / 4
assert len(datasets[1]) == sum(_SIZES[Split.train]) / 2
assert datasets[2] is None
# This build used to fail when building datasets without a sample buffer
config = BlendedMegatronDatasetConfig(
random_seed=1234,
sequence_length=_SEQUENCE_LENGTH,
blend=blends[Split.train],
split="990,9,1",
mid_level_dataset_surplus=0.005,
)
datasets = BlendedMegatronDatasetBuilder(
TestDataset, [100000, 1000, 1], lambda: True, config
).build()
if __name__ == "__main__":
test_builder()
|