Spaces:
Running
Running
File size: 18,263 Bytes
2a0bc63 |
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 |
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from copy import deepcopy
import torch
import torch.nn.functional as F
from torch.optim import AdamW
from torch.optim.lr_scheduler import LambdaLR
from torch.utils.data import DataLoader
from accelerate.accelerator import Accelerator, GradientAccumulationPlugin
from accelerate.state import GradientState
from accelerate.test_utils import RegressionDataset, RegressionModel
from accelerate.utils import DistributedType, set_seed
def check_model_parameters(model_a, model_b, did_step, iteration, **kwargs):
for param, grad_param in zip(model_a.parameters(), model_b.parameters()):
if not param.requires_grad:
continue
if not did_step:
# Grads should not be in sync
assert (
torch.allclose(param.grad, grad_param.grad, **kwargs) is False
), f"Gradients in sync when they should not be at iteration {iteration}:\nmodel_a grad ({param.grad}) == model_b grad ({grad_param.grad})"
else:
# Grads should be in sync
assert (
torch.allclose(param.grad, grad_param.grad, **kwargs) is True
), f"Gradients not in sync when they should be at iteration {iteration}:\nmodel_a grad ({param.grad}) != model_b grad ({grad_param.grad})"
def step_model(model, input, target, accelerator, do_backward=True):
model.train()
output = model(input)
loss = F.mse_loss(output, target.to(output.device))
if not do_backward:
loss /= accelerator.gradient_accumulation_steps
loss.backward()
else:
accelerator.backward(loss)
def get_training_setup(accelerator, sched=False):
"Returns everything needed to perform basic training"
set_seed(42)
model = RegressionModel()
ddp_model = deepcopy(model)
dset = RegressionDataset(length=80)
dataloader = DataLoader(dset, batch_size=16)
model.to(accelerator.device)
if sched:
opt = AdamW(params=model.parameters(), lr=1e-3)
ddp_opt = AdamW(params=ddp_model.parameters(), lr=1e-3)
sched = LambdaLR(opt, lr_lambda=lambda epoch: epoch**0.65)
ddp_sched = LambdaLR(ddp_opt, lr_lambda=lambda epoch: epoch**0.65)
# Make a copy of `model`
if sched:
ddp_model, ddp_opt, ddp_sched, dataloader = accelerator.prepare(ddp_model, ddp_opt, ddp_sched, dataloader)
else:
ddp_model, dataloader = accelerator.prepare(ddp_model, dataloader)
if sched:
return (model, opt, sched, dataloader, ddp_model, ddp_opt, ddp_sched)
return model, ddp_model, dataloader
def test_noop_sync(accelerator):
# Test when on a single CPU or GPU that the context manager does nothing
model, ddp_model, dataloader = get_training_setup(accelerator)
# Use a single batch
ddp_input, ddp_target = next(iter(dataloader)).values()
for iteration in range(3):
# Gather the distributed inputs and targs for the base model
input, target = accelerator.gather((ddp_input, ddp_target))
input, target = input.to(accelerator.device), target.to(accelerator.device)
# Perform our initial ground truth step in non "DDP"
step_model(model, input, target, accelerator)
# Do "gradient accumulation" (noop)
if iteration % 2 == 0:
# Accumulate grads locally
with accelerator.no_sync(ddp_model):
step_model(ddp_model, ddp_input, ddp_target, accelerator)
else:
# Sync grads
step_model(ddp_model, ddp_input, ddp_target, accelerator)
# Since `no_sync` is a noop, `ddp_model` and `model` grads should always be in sync
check_model_parameters(model, ddp_model, True, iteration)
for param, ddp_param in zip(model.parameters(), ddp_model.parameters()):
if not param.requires_grad:
continue
assert torch.allclose(
param.grad, ddp_param.grad
), f"Gradients not in sync when they should be:\nModel grad ({param.grad}) != DDP grad ({ddp_param.grad})"
# Shuffle ddp_input on each iteration
torch.manual_seed(1337 + iteration)
ddp_input = ddp_input[torch.randperm(len(ddp_input))]
def test_distributed_sync(accelerator):
# Test on distributed setup that context manager behaves properly
model, ddp_model, dataloader = get_training_setup(accelerator)
# Use a single batch
ddp_input, ddp_target = next(iter(dataloader)).values()
for iteration in range(3):
# Gather the distributed inputs and targs for the base model
input, target = accelerator.gather((ddp_input, ddp_target))
input, target = input.to(accelerator.device), target.to(accelerator.device)
# Perform our initial ground truth step in non "DDP"
step_model(model, input, target, accelerator)
# Do "gradient accumulation" (noop)
if iteration % 2 == 0:
# Accumulate grads locally
with accelerator.no_sync(ddp_model):
step_model(ddp_model, ddp_input, ddp_target, accelerator)
else:
# Sync grads
step_model(ddp_model, ddp_input, ddp_target, accelerator)
# DDP model and model should only be in sync when not (iteration % 2 == 0)
for param, ddp_param in zip(model.parameters(), ddp_model.parameters()):
if not param.requires_grad:
continue
if iteration % 2 == 0:
# Grads should not be in sync
assert (
torch.allclose(param.grad, ddp_param.grad) is False
), f"Gradients in sync when they should not be:\nModel grad ({param.grad}) == DDP grad ({ddp_param.grad})"
else:
# Grads should be in sync
assert (
torch.allclose(param.grad, ddp_param.grad) is True
), f"Gradients not in sync when they should be:\nModel grad ({param.grad}) != DDP grad ({ddp_param.grad})"
# Shuffle ddp_input on each iteration
torch.manual_seed(1337 + iteration)
ddp_input = ddp_input[torch.randperm(len(ddp_input))]
def test_distributed_sync_multiple_fwd(accelerator):
# Test on distributed setup that context manager behaves properly when used with multiple forwards followed by multiple backwards
model, ddp_model, dataloader = get_training_setup(accelerator)
# Do multiple forwards
losses = []
num_iterations = 3
for iteration in range(num_iterations):
ddp_input, ddp_target = next(iter(dataloader)).values()
# Gather the distributed inputs and targs for the base model
input, target = accelerator.gather((ddp_input, ddp_target))
input, target = input.to(accelerator.device), target.to(accelerator.device)
# Perform our initial ground truth step in non "DDP"
step_model(model, input, target, accelerator)
# Accumulate grads locally
with accelerator.no_sync(ddp_model):
ddp_output = ddp_model(ddp_input)
loss = F.mse_loss(ddp_output, ddp_target.to(ddp_output.device))
losses.append(loss)
# Do multiple backwards and sync only at the last backward
for iteration in range(num_iterations):
loss = losses[iteration]
if iteration < num_iterations - 1:
# Accumulate grads locally
accelerator.backward(loss)
# DDP model and model should only be in sync after last backward
for param, ddp_param in zip(model.parameters(), ddp_model.parameters()):
if not param.requires_grad:
continue
# Grads should not be in sync
assert (
torch.allclose(param.grad, ddp_param.grad) is False
), f"Gradients in sync when they should not be:\nModel grad ({param.grad}) == DDP grad ({ddp_param.grad})"
else:
# Sync grads if last backward
with accelerator.trigger_sync_in_backward(ddp_model):
accelerator.backward(loss)
# DDP model and model should only be in sync after last backward
for param, ddp_param in zip(model.parameters(), ddp_model.parameters()):
if not param.requires_grad:
continue
# Grads should be in sync
assert (
torch.allclose(param.grad, ddp_param.grad) is True
), f"Gradients not in sync when they should be:\nModel grad ({param.grad}) != DDP grad ({ddp_param.grad})"
def test_gradient_accumulation(split_batches=False, dispatch_batches=False, sync_each_batch=False):
gradient_accumulation_plugin = GradientAccumulationPlugin(num_steps=2, sync_each_batch=sync_each_batch)
accelerator = Accelerator(
split_batches=split_batches,
dispatch_batches=dispatch_batches,
gradient_accumulation_plugin=gradient_accumulation_plugin,
)
# Test that context manager behaves properly
model, ddp_model, dataloader = get_training_setup(accelerator)
for iteration, batch in enumerate(dataloader):
ddp_input, ddp_target = batch.values()
# Gather the distributed inputs and targs for the base model
input, target = accelerator.gather((ddp_input, ddp_target))
input, target = input.to(accelerator.device), target.to(accelerator.device)
# Perform our initial ground truth step in non "DDP"
step_model(model, input, target, accelerator, False)
# Do "gradient accumulation" (noop)
with accelerator.accumulate(ddp_model):
step_model(ddp_model, ddp_input, ddp_target, accelerator)
# DDP model and model should only be in sync when not (iteration % 2 == 0)
for param, ddp_param in zip(model.parameters(), ddp_model.parameters()):
if not param.requires_grad:
continue
if ((iteration + 1) % 2 == 0) or (iteration == len(dataloader) - 1) or sync_each_batch:
# Grads should be in sync
assert (
torch.allclose(param.grad, ddp_param.grad) is True
), f"Gradients not in sync when they should be at iteration {iteration}:\nModel grad ({param.grad}) != DDP grad ({ddp_param.grad})"
else:
# Grads should not be in sync
assert (
torch.allclose(param.grad, ddp_param.grad) is False
), f"Gradients in sync when they should not be at iteration {iteration}:\nModel grad ({param.grad}) == DDP grad ({ddp_param.grad})"
# Shuffle ddp_input on each iteration
torch.manual_seed(1337 + iteration)
ddp_input = ddp_input[torch.randperm(len(ddp_input))]
GradientState._reset_state()
def test_gradient_accumulation_with_opt_and_scheduler(
split_batches=False, dispatch_batches=False, sync_each_batch=False
):
gradient_accumulation_plugin = GradientAccumulationPlugin(num_steps=2, sync_each_batch=sync_each_batch)
accelerator = Accelerator(
split_batches=split_batches,
dispatch_batches=dispatch_batches,
gradient_accumulation_plugin=gradient_accumulation_plugin,
)
# Test that context manager behaves properly
model, opt, sched, dataloader, ddp_model, ddp_opt, ddp_sched = get_training_setup(accelerator, True)
for iteration, batch in enumerate(dataloader):
ddp_input, ddp_target = batch.values()
# Gather the distributed inputs and targs for the base model
input, target = accelerator.gather((ddp_input, ddp_target))
input, target = input.to(accelerator.device), target.to(accelerator.device)
# Perform our initial ground truth step in non "DDP"
model.train()
ddp_model.train()
step_model(model, input, target, accelerator, False)
opt.step()
if ((iteration + 1) % 2 == 0) or ((iteration + 1) == len(dataloader)) or sync_each_batch:
if split_batches:
sched.step()
else:
for _ in range(accelerator.num_processes):
sched.step()
# Perform gradient accumulation under wrapper
with accelerator.accumulate(ddp_model):
step_model(ddp_model, ddp_input, ddp_target, accelerator)
ddp_opt.step()
ddp_sched.step()
# Learning rates should be the same
assert (
opt.param_groups[0]["lr"] == ddp_opt.param_groups[0]["lr"]
), f'Learning rates found in each optimizer did not align\nopt: {opt.param_groups[0]["lr"]}\nDDP opt: {ddp_opt.param_groups[0]["lr"]}\n'
did_step = (((iteration + 1) % 2) == 0) or ((iteration + 1) == len(dataloader)) or sync_each_batch
if accelerator.num_processes > 1:
check_model_parameters(
model,
ddp_model,
did_step,
iteration,
rtol=1e-3, # somehow needs a relative tolerance
)
if ((iteration + 1) % 2 == 0) or ((iteration + 1) == len(dataloader)) or sync_each_batch:
opt.zero_grad() # needs to be guarded by logic as to when we should zero grads
ddp_opt.zero_grad()
# Shuffle ddp_input on each iteration
torch.manual_seed(1337 + iteration)
GradientState._reset_state()
def test_dataloader_break():
accelerator = Accelerator()
first_dset = RegressionDataset(length=80)
first_dataloader = DataLoader(first_dset, batch_size=16)
second_dset = RegressionDataset(length=96)
second_dataloader = DataLoader(second_dset, batch_size=16)
first_dataloader, second_dataloader = accelerator.prepare(first_dataloader, second_dataloader)
assert accelerator.gradient_state.active_dataloader is None
for iteration, _ in enumerate(first_dataloader):
assert id(accelerator.gradient_state.active_dataloader) == id(first_dataloader)
if iteration < len(first_dataloader) - 1:
assert not accelerator.gradient_state.end_of_dataloader
if iteration == 1:
for batch_num, _ in enumerate(second_dataloader):
assert id(accelerator.gradient_state.active_dataloader) == id(second_dataloader)
if batch_num < len(second_dataloader) - 1:
assert not accelerator.gradient_state.end_of_dataloader
else:
assert accelerator.gradient_state.end_of_dataloader
else:
assert accelerator.gradient_state.end_of_dataloader
assert accelerator.gradient_state.active_dataloader is None
def main():
accelerator = Accelerator()
state = accelerator.state
if state.local_process_index == 0:
print("**Test `accumulate` gradient accumulation with dataloader break**")
if state.distributed_type != DistributedType.XLA:
test_dataloader_break()
if state.distributed_type == DistributedType.NO:
if state.local_process_index == 0:
print("**Test NOOP `no_sync` context manager**")
test_noop_sync(accelerator)
if state.distributed_type in (
DistributedType.MULTI_GPU,
DistributedType.MULTI_NPU,
DistributedType.MULTI_MLU,
DistributedType.MULTI_CPU,
):
if state.local_process_index == 0:
print("**Test Distributed `no_sync` context manager**")
test_distributed_sync(accelerator)
if state.local_process_index == 0:
print("**Test Distributed `no_sync` context manager with multiple forwards**")
test_distributed_sync_multiple_fwd(accelerator)
if state.distributed_type in (DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.MULTI_MLU):
for split_batch in [True, False]:
for dispatch_batches in [True, False]:
for sync_each_batch in [True, False]:
if state.local_process_index == 0:
print(
"**Test `accumulate` gradient accumulation, ",
f"`split_batches={split_batch}` and `dispatch_batches={dispatch_batches}` and `sync_each_batch={sync_each_batch}`**",
)
test_gradient_accumulation(split_batch, dispatch_batches, sync_each_batch)
# Currently will break on torch 2.0 +, need to investigate why
if state.local_process_index == 0:
print(
"**Test `accumulate` gradient accumulation with optimizer and scheduler, ",
"`split_batches=False`, `dispatch_batches=False`, `sync_each_batch=False`**",
)
test_gradient_accumulation_with_opt_and_scheduler()
if state.distributed_type in (DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.MULTI_MLU):
for split_batch in [True, False]:
for dispatch_batches in [True, False]:
for sync_each_batch in [True, False]:
if not split_batch and not dispatch_batches and not sync_each_batch:
continue
if state.local_process_index == 0:
print(
"**Test `accumulate` gradient accumulation with optimizer and scheduler, ",
f"`split_batches={split_batch}` and `dispatch_batches={dispatch_batches}` and `sync_each_batch={sync_each_batch}`**",
)
test_gradient_accumulation_with_opt_and_scheduler(split_batch, dispatch_batches, sync_each_batch)
def _mp_fn(index):
# For xla_spawn (TPUs)
main()
if __name__ == "__main__":
main()
|