Spaces:
Running
Running
File size: 20,663 Bytes
9fd1204 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 |
# torchrun --nnodes=1 --nproc_per_node=1 -m pytest -s tests/trainer/test_sft_trainer.py
import json
import os
import pathlib
import tempfile
import time
import unittest
import pytest
import torch
from diffusers.utils import export_to_video
from parameterized import parameterized
from PIL import Image
from finetrainers import BaseArgs, SFTTrainer, TrainingType, get_logger
os.environ["WANDB_MODE"] = "disabled"
os.environ["FINETRAINERS_LOG_LEVEL"] = "INFO"
from ..models.cogvideox.base_specification import DummyCogVideoXModelSpecification # noqa
from ..models.cogview4.base_specification import DummyCogView4ModelSpecification # noqa
from ..models.flux.base_specification import DummyFluxModelSpecification # noqa
from ..models.hunyuan_video.base_specification import DummyHunyuanVideoModelSpecification # noqa
from ..models.ltx_video.base_specification import DummyLTXVideoModelSpecification # noqa
from ..models.wan.base_specification import DummyWanModelSpecification # noqa
logger = get_logger()
@pytest.fixture(autouse=True)
def slow_down_tests():
yield
# Sleep between each test so that process groups are cleaned and resources are released.
# Not doing so seems to randomly trigger some test failures, which wouldn't fail if run individually.
# !!!Look into this in future!!!
time.sleep(5)
class SFTTrainerFastTestsMixin:
model_specification_cls = None
num_data_files = 4
num_frames = 4
height = 64
width = 64
def setUp(self):
self.tmpdir = tempfile.TemporaryDirectory()
self.data_files = []
for i in range(self.num_data_files):
data_file = pathlib.Path(self.tmpdir.name) / f"{i}.mp4"
export_to_video(
[Image.new("RGB", (self.width, self.height))] * self.num_frames, data_file.as_posix(), fps=2
)
self.data_files.append(data_file.as_posix())
csv_filename = pathlib.Path(self.tmpdir.name) / "metadata.csv"
with open(csv_filename.as_posix(), "w") as f:
f.write("file_name,caption\n")
for i in range(self.num_data_files):
prompt = f"A cat ruling the world - {i}"
f.write(f'{i}.mp4,"{prompt}"\n')
dataset_config = {
"datasets": [
{
"data_root": self.tmpdir.name,
"dataset_type": "video",
"id_token": "TEST",
"video_resolution_buckets": [[self.num_frames, self.height, self.width]],
"reshape_mode": "bicubic",
}
]
}
self.dataset_config_filename = pathlib.Path(self.tmpdir.name) / "dataset_config.json"
with open(self.dataset_config_filename.as_posix(), "w") as f:
json.dump(dataset_config, f)
def tearDown(self):
self.tmpdir.cleanup()
# For some reason, if the process group is not destroyed, the tests that follow will fail. Just manually
# make sure to destroy it here.
if torch.distributed.is_initialized():
torch.distributed.destroy_process_group()
time.sleep(3)
def get_base_args(self) -> BaseArgs:
args = BaseArgs()
args.dataset_config = self.dataset_config_filename.as_posix()
args.train_steps = 10
args.max_data_samples = 25
args.batch_size = 1
args.gradient_checkpointing = True
args.output_dir = self.tmpdir.name
args.checkpointing_steps = 6
args.enable_precomputation = False
args.precomputation_items = self.num_data_files
args.precomputation_dir = os.path.join(self.tmpdir.name, "precomputed")
args.compile_scopes = "regional" # This will only be in effect when `compile_modules` is set
# args.attn_provider_training = ["transformer:_native_cudnn"]
# args.attn_provider_inference = ["transformer:_native_cudnn"]
return args
def get_args(self) -> BaseArgs:
raise NotImplementedError("`get_args` must be implemented in the subclass.")
def _test_training(self, args: BaseArgs):
model_specification = self.model_specification_cls()
trainer = SFTTrainer(args, model_specification)
trainer.run()
# =============== <ACCELERATE> ===============
class SFTTrainerLoRATestsMixin___Accelerate(SFTTrainerFastTestsMixin):
def get_args(self) -> BaseArgs:
args = self.get_base_args()
args.parallel_backend = "accelerate"
args.training_type = TrainingType.LORA
args.rank = 4
args.lora_alpha = 4
args.target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
return args
@parameterized.expand([(False,), (True,)])
def test___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(True,)])
def test___layerwise_upcasting___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
args.layerwise_upcasting_modules = ["transformer"]
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
class SFTTrainerFullFinetuneTestsMixin___Accelerate(SFTTrainerFastTestsMixin):
def get_args(self) -> BaseArgs:
args = self.get_base_args()
args.parallel_backend = "accelerate"
args.training_type = TrainingType.FULL_FINETUNE
return args
@parameterized.expand([(False,), (True,)])
def test___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
class SFTTrainerCogVideoXLoRATests___Accelerate(SFTTrainerLoRATestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyCogVideoXModelSpecification
class SFTTrainerCogVideoXFullFinetuneTests___Accelerate(
SFTTrainerFullFinetuneTestsMixin___Accelerate, unittest.TestCase
):
model_specification_cls = DummyCogVideoXModelSpecification
class SFTTrainerCogView4LoRATests___Accelerate(SFTTrainerLoRATestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyCogView4ModelSpecification
class SFTTrainerCogView4FullFinetuneTests___Accelerate(
SFTTrainerFullFinetuneTestsMixin___Accelerate, unittest.TestCase
):
model_specification_cls = DummyCogView4ModelSpecification
class SFTTrainerFluxLoRATests___Accelerate(SFTTrainerLoRATestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyFluxModelSpecification
class SFTTrainerFluxFullFinetuneTests___Accelerate(SFTTrainerFullFinetuneTestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyFluxModelSpecification
class SFTTrainerHunyuanVideoLoRATests___Accelerate(SFTTrainerLoRATestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyHunyuanVideoModelSpecification
class SFTTrainerHunyuanVideoFullFinetuneTests___Accelerate(
SFTTrainerFullFinetuneTestsMixin___Accelerate, unittest.TestCase
):
model_specification_cls = DummyHunyuanVideoModelSpecification
class SFTTrainerLTXVideoLoRATests___Accelerate(SFTTrainerLoRATestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyLTXVideoModelSpecification
class SFTTrainerLTXVideoFullFinetuneTests___Accelerate(
SFTTrainerFullFinetuneTestsMixin___Accelerate, unittest.TestCase
):
model_specification_cls = DummyLTXVideoModelSpecification
class SFTTrainerWanLoRATests___Accelerate(SFTTrainerLoRATestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyWanModelSpecification
class SFTTrainerWanFullFinetuneTests___Accelerate(SFTTrainerFullFinetuneTestsMixin___Accelerate, unittest.TestCase):
model_specification_cls = DummyWanModelSpecification
# =============== </ACCELERATE> ===============
# =============== <PTD> ===============
class SFTTrainerLoRATestsMixin___PTD(SFTTrainerFastTestsMixin):
def get_args(self) -> BaseArgs:
args = self.get_base_args()
args.parallel_backend = "ptd"
args.training_type = TrainingType.LORA
args.rank = 4
args.lora_alpha = 4
args.target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
return args
@parameterized.expand([(False,), (True,)])
def test___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(True,)])
def test___layerwise_upcasting___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
args.layerwise_upcasting_modules = ["transformer"]
self._test_training(args)
@parameterized.expand([(True,)])
def test___compile___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
args.compile_modules = ["transformer"]
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_degree_1___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 2
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(True,)])
def test___layerwise_upcasting___dp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
args.layerwise_upcasting_modules = ["transformer"]
self._test_training(args)
@parameterized.expand([(True,)])
def test___compile___dp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
args.compile_modules = ["transformer"]
self._test_training(args)
@parameterized.expand([(True,)])
def test___dp_degree_2___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 2
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_shards_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_shards = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(True,)])
def test___dp_shards_2___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.dp_shards = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_degree_2___dp_shards_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.dp_shards = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___tp_degree_2___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.tp_degree = 2
args.batch_size = 2
args.enable_precomputation = enable_precomputation
self._test_training(args)
@unittest.skip(
"TODO: The model specifications for CP with cudnn/flash/efficient backend require the attention head dim to be a multiple with 8. Land math backend first for fast tests and then enable this test."
)
@parameterized.expand([(True,)])
def test___cp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.cp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@unittest.skip(
"TODO: The model specifications for CP with cudnn/flash/efficient backend require the attention head dim to be a multiple with 8. Land math backend first for fast tests and then enable this test."
)
@parameterized.expand([(True,)])
def test___dp_degree_2___cp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.cp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
class SFTTrainerFullFinetuneTestsMixin___PTD(SFTTrainerFastTestsMixin):
def get_args(self) -> BaseArgs:
args = self.get_base_args()
args.parallel_backend = "ptd"
args.training_type = TrainingType.FULL_FINETUNE
return args
@parameterized.expand([(False,), (True,)])
def test___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(True,)])
def test___compile___dp_degree_1___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 1
args.enable_precomputation = enable_precomputation
args.compile_modules = ["transformer"]
self._test_training(args)
@parameterized.expand([(True,)])
def test___dp_degree_1___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 1
args.batch_size = 2
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(True,)])
def test___compile___dp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
args.compile_modules = ["transformer"]
self._test_training(args)
@parameterized.expand([(True,)])
def test___dp_degree_2___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.batch_size = 2
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_shards_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_shards = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(True,)])
def test___dp_shards_2___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.dp_shards = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___dp_degree_2___dp_shards_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.dp_shards = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@parameterized.expand([(False,), (True,)])
def test___tp_degree_2___batch_size_2(self, enable_precomputation: bool):
args = self.get_args()
args.tp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@unittest.skip(
"TODO: The model specifications for CP with cudnn/flash/efficient backend require the attention head dim to be a multiple with 8. Land math backend first for fast tests and then enable this test."
)
@parameterized.expand([(True,)])
def test___cp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.cp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
@unittest.skip(
"TODO: The model specifications for CP with cudnn/flash/efficient backend require the attention head dim to be a multiple with 8. Land math backend first for fast tests and then enable this test."
)
@parameterized.expand([(True,)])
def test___dp_degree_2___cp_degree_2___batch_size_1(self, enable_precomputation: bool):
args = self.get_args()
args.dp_degree = 2
args.cp_degree = 2
args.batch_size = 1
args.enable_precomputation = enable_precomputation
self._test_training(args)
class SFTTrainerCogVideoXLoRATests___PTD(SFTTrainerLoRATestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyCogVideoXModelSpecification
class SFTTrainerCogVideoXFullFinetuneTests___PTD(SFTTrainerFullFinetuneTestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyCogVideoXModelSpecification
class SFTTrainerCogView4LoRATests___PTD(SFTTrainerLoRATestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyCogView4ModelSpecification
class SFTTrainerCogView4FullFinetuneTests___PTD(SFTTrainerFullFinetuneTestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyCogView4ModelSpecification
class SFTTrainerFluxLoRATests___PTD(SFTTrainerLoRATestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyFluxModelSpecification
class SFTTrainerFluxFullFinetuneTests___PTD(SFTTrainerFullFinetuneTestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyFluxModelSpecification
class SFTTrainerHunyuanVideoLoRATests___PTD(SFTTrainerLoRATestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyHunyuanVideoModelSpecification
class SFTTrainerHunyuanVideoFullFinetuneTests___PTD(SFTTrainerFullFinetuneTestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyHunyuanVideoModelSpecification
class SFTTrainerLTXVideoLoRATests___PTD(SFTTrainerLoRATestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyLTXVideoModelSpecification
class SFTTrainerLTXVideoFullFinetuneTests___PTD(SFTTrainerFullFinetuneTestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyLTXVideoModelSpecification
class SFTTrainerWanLoRATests___PTD(SFTTrainerLoRATestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyWanModelSpecification
class SFTTrainerWanFullFinetuneTests___PTD(SFTTrainerFullFinetuneTestsMixin___PTD, unittest.TestCase):
model_specification_cls = DummyWanModelSpecification
# =============== </PTD> ===============
|