Spaces:
Runtime error
Runtime error
File size: 15,737 Bytes
ee21b96 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import contextlib
import logging
import json
import os
import tempfile
import unittest
from io import StringIO
import torch
from fairseq import options
from fairseq_cli import train
from tests.utils import (
create_dummy_data,
generate_main,
preprocess_lm_data,
preprocess_translation_data,
train_translation_model,
)
@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU")
class TestTranslationGPU(unittest.TestCase):
def setUp(self):
logging.disable(logging.CRITICAL)
def tearDown(self):
logging.disable(logging.NOTSET)
def test_fp16_multigpu(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_fp16") as data_dir:
log = os.path.join(data_dir, "train.log")
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
train_translation_model(
data_dir,
"fconv_iwslt_de_en",
["--fp16", "--log-file", log],
world_size=min(torch.cuda.device_count(), 2),
)
generate_main(data_dir)
assert os.path.exists(log)
@staticmethod
def parse_logs(logfile):
logs = []
for ln in open(logfile, "r").readlines():
try:
logs.append(json.loads(ln))
except json.JSONDecodeError:
continue
return logs
def test_resume_training_fsdp(self):
self._test_resume_training(["--ddp-backend", "fully_sharded"])
def test_resume_training_fsdp_sharded_state(self):
self._test_resume_training(["--ddp-backend", "fully_sharded", "--use-sharded-state"])
def test_resume_training_noc10d(self):
self._test_resume_training([])
def _test_resume_training(self, extra_clargs, arch="fconv_iwslt_de_en"):
flags = [
"--fp16",
"--log-format",
"json",
"--max-update",
"10",
"--save-interval-updates",
"2",
"--log-interval",
"1",
] + extra_clargs
world_size = min(torch.cuda.device_count(), 2)
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_fp16") as data_dir:
log = os.path.join(data_dir, "train.log")
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
train_translation_model(
data_dir, arch, flags + ["--log-file", log], world_size=world_size,
)
log2 = os.path.join(data_dir, "resume.log")
restore_file = os.path.join(data_dir, "checkpoint_1_2.pt")
train_translation_model(
data_dir,
arch,
flags + ["--log-file", log2, "--restore-file", restore_file],
world_size=world_size,
)
l1 = self.parse_logs(log)
l2 = self.parse_logs(log2)
assert int(l2[0]["num_updates"]) == 3, f"{l1}\n\n {l2}"
for k in [
"train_loss",
"train_num_updates",
"train_ppl",
"train_gnorm",
]:
from_scratch, resumed = l1[-1][k], l2[-1][k]
assert (
from_scratch == resumed
), f"difference at {k} {from_scratch} != {resumed}"
def test_memory_efficient_fp16(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_memory_efficient_fp16") as data_dir:
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
train_translation_model(
data_dir, "fconv_iwslt_de_en", ["--memory-efficient-fp16"]
)
generate_main(data_dir)
def test_transformer_fp16(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_transformer") as data_dir:
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
train_translation_model(
data_dir,
"transformer_iwslt_de_en",
[
"--encoder-layers",
"2",
"--decoder-layers",
"2",
"--encoder-embed-dim",
"64",
"--decoder-embed-dim",
"64",
"--fp16",
],
run_validation=True,
)
generate_main(data_dir)
@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU")
def test_amp(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_amp") as data_dir:
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
train_translation_model(data_dir, "fconv_iwslt_de_en", ["--amp"])
generate_main(data_dir)
@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU")
def test_transformer_amp(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_transformer") as data_dir:
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
train_translation_model(
data_dir,
"transformer_iwslt_de_en",
[
"--encoder-layers",
"2",
"--decoder-layers",
"2",
"--encoder-embed-dim",
"64",
"--decoder-embed-dim",
"64",
"--amp",
],
run_validation=True,
)
generate_main(data_dir)
@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU")
def test_levenshtein_transformer(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory(
"test_levenshtein_transformer"
) as data_dir:
create_dummy_data(data_dir)
preprocess_translation_data(data_dir, ["--joined-dictionary"])
train_translation_model(
data_dir,
"levenshtein_transformer",
[
"--apply-bert-init",
"--early-exit",
"6,6,6",
"--criterion",
"nat_loss",
],
task="translation_lev",
)
gen_config = [
"--task",
"translation_lev",
"--iter-decode-max-iter",
"9",
"--iter-decode-eos-penalty",
"0",
"--print-step",
]
# non-ensemble generation
generate_main(data_dir, gen_config)
# ensemble generation
generate_main(
data_dir,
gen_config,
path=os.pathsep.join(
[
os.path.join(data_dir, "checkpoint_last.pt"),
os.path.join(data_dir, "checkpoint_last.pt"),
]
),
)
def test_fsdp_checkpoint_generate(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_fsdp_sharded") as data_dir:
log = os.path.join(data_dir, "train.log")
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
world_size = min(torch.cuda.device_count(), 2)
train_translation_model(
data_dir,
"fconv_iwslt_de_en",
["--log-file", log, "--ddp-backend", "fully_sharded"],
world_size=world_size,
)
generate_main(data_dir)
assert os.path.exists(log)
def test_fsdp_sharded_checkpoint_generate(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_fsdp_sharded") as data_dir:
log = os.path.join(data_dir, "train.log")
create_dummy_data(data_dir)
preprocess_translation_data(data_dir)
world_size = min(torch.cuda.device_count(), 2)
train_translation_model(
data_dir,
"fconv_iwslt_de_en",
["--log-file", log, "--ddp-backend", "fully_sharded", "--use-sharded-state"],
world_size=world_size,
)
generate_main(data_dir, ["--checkpoint-shard-count", str(world_size)])
assert os.path.exists(log)
def _quantize_language_model(data_dir, arch, extra_flags=None, run_validation=False):
train_parser = options.get_training_parser()
train_args = options.parse_args_and_arch(
train_parser,
[
"--task",
"language_modeling",
data_dir,
"--arch",
arch,
"--optimizer",
"adam",
"--lr",
"0.0001",
"--criterion",
"adaptive_loss",
"--adaptive-softmax-cutoff",
"5,10,15",
"--max-tokens",
"500",
"--tokens-per-sample",
"500",
"--save-dir",
data_dir,
"--max-epoch",
"1",
"--no-progress-bar",
"--distributed-world-size",
"1",
"--ddp-backend",
"no_c10d",
"--num-workers",
"0",
]
+ (extra_flags or []),
)
train.main(train_args)
# try scalar quantization
scalar_quant_train_parser = options.get_training_parser()
scalar_quant_train_args = options.parse_args_and_arch(
scalar_quant_train_parser,
[
"--task",
"language_modeling",
data_dir,
"--arch",
arch,
"--optimizer",
"adam",
"--lr",
"0.0001",
"--criterion",
"adaptive_loss",
"--adaptive-softmax-cutoff",
"5,10,15",
"--max-tokens",
"500",
"--tokens-per-sample",
"500",
"--save-dir",
data_dir,
"--max-update",
"3",
"--no-progress-bar",
"--distributed-world-size",
"1",
"--ddp-backend",
"no_c10d",
"--num-workers",
"0",
"--quant-noise-scalar",
"0.5",
]
+ (extra_flags or []),
)
train.main(scalar_quant_train_args)
# try iterative PQ quantization
quantize_parser = options.get_training_parser()
quantize_args = options.parse_args_and_arch(
quantize_parser,
[
"--task",
"language_modeling",
data_dir,
"--arch",
arch,
"--optimizer",
"adam",
"--lr",
"0.0001",
"--criterion",
"adaptive_loss",
"--adaptive-softmax-cutoff",
"5,10,15",
"--max-tokens",
"50",
"--tokens-per-sample",
"50",
"--max-update",
"6",
"--no-progress-bar",
"--distributed-world-size",
"1",
"--ddp-backend",
"no_c10d",
"--num-workers",
"0",
"--restore-file",
os.path.join(data_dir, "checkpoint_last.pt"),
"--reset-optimizer",
"--quantization-config-path",
os.path.join(
os.path.dirname(__file__), "transformer_quantization_config.yaml"
),
]
+ (extra_flags or []),
)
train.main(quantize_args)
@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU")
class TestQuantization(unittest.TestCase):
def setUp(self):
logging.disable(logging.CRITICAL)
def tearDown(self):
logging.disable(logging.NOTSET)
def test_quantization(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_quantization") as data_dir:
create_dummy_data(data_dir)
preprocess_lm_data(data_dir)
# tests both scalar and iterative PQ quantization
_quantize_language_model(data_dir, "transformer_lm")
@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU")
class TestOptimizersGPU(unittest.TestCase):
def setUp(self):
logging.disable(logging.CRITICAL)
def tearDown(self):
logging.disable(logging.NOTSET)
def test_flat_grads(self):
with contextlib.redirect_stdout(StringIO()):
with tempfile.TemporaryDirectory("test_flat_grads") as data_dir:
# Use just a bit of data and tiny model to keep this test runtime reasonable
create_dummy_data(data_dir, num_examples=10, maxlen=5)
preprocess_translation_data(data_dir)
with self.assertRaises(RuntimeError):
# adafactor isn't compatible with flat grads, which
# are used by default with --fp16
train_translation_model(
data_dir,
"lstm",
[
"--required-batch-size-multiple",
"1",
"--encoder-layers",
"1",
"--encoder-hidden-size",
"32",
"--decoder-layers",
"1",
"--optimizer",
"adafactor",
"--fp16",
],
)
# but it should pass once we set --fp16-no-flatten-grads
train_translation_model(
data_dir,
"lstm",
[
"--required-batch-size-multiple",
"1",
"--encoder-layers",
"1",
"--encoder-hidden-size",
"32",
"--decoder-layers",
"1",
"--optimizer",
"adafactor",
"--fp16",
"--fp16-no-flatten-grads",
],
)
if __name__ == "__main__":
unittest.main()
|