muellerzr HF staff commited on
Commit
7a76c8f
1 Parent(s): ecb9f74

Try again?

Browse files
src/accelerator.py CHANGED
@@ -87,6 +87,7 @@ class Accelerator:
87
  - **sync_gradients** (`bool`) -- Whether the gradients are currently being synced across all processes.
88
  - **use_distributed** (`bool`) -- Whether the current configuration is for distributed training.
89
  """
 
90
  def __init__(
91
  self,
92
  device_placement: bool = True,
@@ -296,44 +297,56 @@ class Accelerator:
296
  self.flag_tensor = None
297
  check_os_kernel()
298
  @property
 
299
  def use_distributed(self):
300
  """
301
  Whether the Accelerator is configured for distributed training
302
  """
303
  return self.state.use_distributed
304
  @property
 
305
  def distributed_type(self):
306
  return self.state.distributed_type
307
  @property
 
308
  def num_processes(self):
309
  return self.state.num_processes
310
  @property
 
311
  def process_index(self):
312
  return self.state.process_index
313
  @property
 
314
  def local_process_index(self):
315
  return self.state.local_process_index
316
  @property
 
317
  def device(self):
318
  return self.state.device
319
  @property
 
320
  def project_dir(self):
321
  return self.project_configuration.project_dir
322
  @property
 
323
  def logging_dir(self):
324
  return self.project_configuration.logging_dir
325
  @property
 
326
  def save_iteration(self):
327
  return self.project_configuration.iteration
328
  @property
 
329
  def is_main_process(self):
330
  """True for one process only."""
331
  return self.state.is_main_process
332
  @property
 
333
  def is_local_main_process(self):
334
  """True for one process per server."""
335
  return self.state.is_local_main_process
336
  @property
 
337
  def use_fp16(self):
338
  warnings.warn(
339
  "The `use_fp16` property is deprecated and will be removed in version 1.0 of Accelerate use "
@@ -342,12 +355,15 @@ class Accelerator:
342
  )
343
  return self.mixed_precision != "no"
344
  @property
 
345
  def is_last_process(self):
346
  return self.process_index == self.num_processes - 1
347
  @property
 
348
  def mixed_precision(self):
349
  return self.state.mixed_precision
350
  @contextmanager
 
351
  def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
352
  """
353
  Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
@@ -382,6 +398,7 @@ class Accelerator:
382
  """
383
  with PartialState().split_between_processes(inputs, apply_padding=apply_padding) as inputs:
384
  yield inputs
 
385
  def on_main_process(self, function: Callable[..., Any] = None):
386
  """
387
  A decorator that will run the decorated function on the main process only. Can also be called using the
@@ -407,9 +424,11 @@ class Accelerator:
407
  raise ValueError(
408
  "The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
409
  )
410
- def _inner(*args, **kwargs):
 
411
  return PartialState().on_main_process(function)(*args, **kwargs)
412
  return _inner
 
413
  def on_local_main_process(self, function: Callable[..., Any] = None):
414
  """
415
  A decorator that will run the decorated function on the local main process only. Can also be called using the
@@ -422,7 +441,8 @@ class Accelerator:
422
  from accelerate import Accelerator
423
  accelerator = Accelerator()
424
  @accelerator.on_local_main_process
425
- def print_something():
 
426
  print("This will be printed by process 0 only on each server.")
427
  print_something()
428
  # On server 1:
@@ -439,9 +459,11 @@ class Accelerator:
439
  raise ValueError(
440
  "The `on_local_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
441
  )
442
- def _inner(*args, **kwargs):
 
443
  return PartialState().on_local_main_process(function)(*args, **kwargs)
444
  return _inner
 
445
  def on_last_process(self, function: Callable[..., Any]):
446
  """
447
  A decorator that will run the decorated function on the last process only. Can also be called using the
@@ -454,7 +476,8 @@ class Accelerator:
454
  from accelerate import Accelerator
455
  accelerator = Accelerator()
456
  @accelerator.on_last_process
457
- def print_something():
 
458
  print(f"Printed on process {accelerator.process_index}")
459
  print_something()
460
  "Printed on process 3"
@@ -468,9 +491,11 @@ class Accelerator:
468
  raise ValueError(
469
  "The `on_last_process` decorator must be called with a function on an instantiated `Accelerator` object."
470
  )
471
- def _inner(*args, **kwargs):
 
472
  return PartialState().on_last_process(function)(*args, **kwargs)
473
  return _inner
 
474
  def on_process(self, function: Callable[..., Any] = None, process_index: int = None):
475
  """
476
  A decorator that will run the decorated function on a given process index only. Can also be called using the
@@ -486,7 +511,8 @@ class Accelerator:
486
  from accelerate import Accelerator
487
  accelerator = Accelerator()
488
  @accelerator.on_process(process_index=2)
489
- def print_something():
 
490
  print(f"Printed on process {accelerator.process_index}")
491
  print_something()
492
  "Printed on process 2"
@@ -503,9 +529,11 @@ class Accelerator:
503
  raise ValueError(
504
  "The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
505
  )
506
- def _inner(*args, **kwargs):
 
507
  return PartialState().on_process(function, process_index)(*args, **kwargs)
508
  return _inner
 
509
  def on_local_process(self, function: Callable[..., Any] = None, local_process_index: int = None):
510
  """
511
  A decorator that will run the decorated function on a given local process index only. Can also be called using
@@ -521,7 +549,8 @@ class Accelerator:
521
  from accelerate import Accelerator
522
  accelerator = Accelerator()
523
  @accelerator.on_local_process(local_process_index=2)
524
- def print_something():
 
525
  print(f"Printed on process {accelerator.local_process_index}")
526
  print_something()
527
  # On server 1:
@@ -541,10 +570,12 @@ class Accelerator:
541
  raise ValueError(
542
  "The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
543
  )
544
- def _inner(*args, **kwargs):
 
545
  return PartialState().on_local_process(function, local_process_index)(*args, **kwargs)
546
  return _inner
547
  @contextmanager
 
548
  def main_process_first(self):
549
  """
550
  Lets the main process go first inside a with block.
@@ -562,6 +593,7 @@ class Accelerator:
562
  with self.state.main_process_first():
563
  yield
564
  @contextmanager
 
565
  def local_main_process_first(self):
566
  """
567
  Lets the local main process go inside a with block.
@@ -579,6 +611,7 @@ class Accelerator:
579
  with self.state.local_main_process_first():
580
  yield
581
  @contextmanager
 
582
  def no_sync(self, model):
583
  """
584
  A context manager to disable gradient synchronizations across DDP processes by calling
@@ -613,6 +646,7 @@ class Accelerator:
613
  yield
614
  @staticmethod
615
  @contextmanager
 
616
  def trigger_sync_in_backward(model):
617
  """Trigger the sync of the gradients in the next backward pass of the model after multiple forward passes under
618
  `Accelerator.no_sync` (only applicable in multi-GPU scenarios).
@@ -651,6 +685,7 @@ class Accelerator:
651
  finally:
652
  model.require_backward_grad_sync = old_require_backward_grad_sync
653
  model.require_forward_param_sync = old_require_forward_param_sync
 
654
  def _do_sync(self):
655
  "Sets the right `sync_gradients` context and either resets or increases `self.step`"
656
  if self.gradient_state.sync_with_dataloader and self.gradient_state.end_of_dataloader:
@@ -660,18 +695,23 @@ class Accelerator:
660
  self.step += 1
661
  self.gradient_state._set_sync_gradients((self.step % self.gradient_state.num_steps) == 0)
662
  @property
 
663
  def sync_gradients(self):
664
  return self.gradient_state.sync_gradients
665
  @sync_gradients.setter
 
666
  def sync_gradients(self, sync_gradients):
667
  self.gradient_state.sync_gradients = sync_gradients
668
  @property
 
669
  def gradient_accumulation_steps(self):
670
  return self.gradient_state.num_steps
671
  @gradient_accumulation_steps.setter
 
672
  def gradient_accumulation_steps(self, gradient_accumulation_steps):
673
  self.gradient_state.plugin_kwargs.update({"num_steps": gradient_accumulation_steps})
674
  @contextmanager
 
675
  def accumulate(self, *models):
676
  """
677
  A context manager that will lightly wrap around and perform gradient accumulation automatically
@@ -700,6 +740,7 @@ class Accelerator:
700
  cm_stack.enter_context(contextlib.nullcontext() if self.sync_gradients else self.no_sync(m))
701
  yield
702
  @contextmanager
 
703
  def join_uneven_inputs(self, joinables, even_batches=None):
704
  """
705
  A context manager that facilitates distributed training or evaluation on uneven inputs, which acts as a wrapper
@@ -766,6 +807,7 @@ class Accelerator:
766
  )
767
  with contextlib.nullcontext(joinables):
768
  yield
 
769
  def print(self, *args, **kwargs):
770
  """
771
  Drop in replacement of `print()` to only print once per server.
@@ -777,6 +819,7 @@ class Accelerator:
777
  ```
778
  """
779
  self.state.print(*args, **kwargs)
 
780
  def _prepare_one(self, obj, first_pass=False, device_placement=None):
781
  # First pass of preparation: DataLoader, model, optimizer
782
  if first_pass:
@@ -793,6 +836,7 @@ class Accelerator:
793
  return scheduler
794
  # Return the unprocessed object if previous criteria was not met
795
  return obj
 
796
  def prepare(self, *args, device_placement=None):
797
  """
798
  Prepare all objects passed in `args` for distributed training and mixed precision, then return them in the same
@@ -909,6 +953,7 @@ class Accelerator:
909
  ):
910
  setattr(item, "_is_accelerate_prepared", True)
911
  return result if len(result) > 1 else result[0]
 
912
  def prepare_model(self, model: torch.nn.Module, device_placement: bool = None, evaluation_mode: bool = False):
913
  """
914
  Prepares a PyTorch model for training in any distributed setup. It is recommended to use
@@ -1061,6 +1106,7 @@ class Accelerator:
1061
  raise ValueError("Using `torch.compile` requires PyTorch 2.0 or higher.")
1062
  model = torch.compile(model, **self.state.dynamo_plugin.to_kwargs())
1063
  return model
 
1064
  def _prepare_deepspeed(self, *args):
1065
  import deepspeed
1066
  deepspeed_plugin = self.state.deepspeed_plugin
@@ -1250,6 +1296,7 @@ class Accelerator:
1250
  "You can't use same `Accelerator()` instance with multiple models when using DeepSpeed"
1251
  )
1252
  return tuple(result)
 
1253
  def _prepare_megatron_lm(self, *args):
1254
  megatron_lm_plugin = self.state.megatron_lm_plugin
1255
  if not megatron_lm_plugin.megatron_dataset_flag:
@@ -1342,6 +1389,7 @@ class Accelerator:
1342
  "You can't use same `Accelerator()` instance with multiple models when using Megatron-LM"
1343
  )
1344
  return tuple(result)
 
1345
  def _prepare_ipex(self, *args):
1346
  if not is_ipex_available():
1347
  raise ImportError(
@@ -1373,6 +1421,7 @@ class Accelerator:
1373
  elif isinstance(result[i], (torch.optim.Optimizer)):
1374
  result[i] = optimizer
1375
  return tuple(result)
 
1376
  def _prepare_msamp(self, *args):
1377
  if not is_msamp_available():
1378
  raise ImportError(
@@ -1407,6 +1456,7 @@ class Accelerator:
1407
  elif isinstance(result[i], (torch.optim.Optimizer)):
1408
  result[i] = optimizer
1409
  return tuple(result)
 
1410
  def prepare_data_loader(
1411
  self, data_loader: torch.utils.data.DataLoader, device_placement=None, slice_fn_for_dispatch=None
1412
  ):
@@ -1453,6 +1503,7 @@ class Accelerator:
1453
  )
1454
  self._dataloaders.append(prepared_data_loader)
1455
  return prepared_data_loader
 
1456
  def prepare_optimizer(self, optimizer: torch.optim.Optimizer, device_placement=None):
1457
  """
1458
  Prepares a PyTorch Optimizer for training in any distributed setup. It is recommended to use
@@ -1481,6 +1532,7 @@ class Accelerator:
1481
  optimizer = AcceleratedOptimizer(optimizer, device_placement=device_placement, scaler=self.scaler)
1482
  self._optimizers.append(optimizer)
1483
  return optimizer
 
1484
  def prepare_scheduler(self, scheduler: LRScheduler):
1485
  """
1486
  Prepares a PyTorch Scheduler for training in any distributed setup. It is recommended to use
@@ -1517,6 +1569,7 @@ class Accelerator:
1517
  )
1518
  self._schedulers.append(scheduler)
1519
  return scheduler
 
1520
  def backward(self, loss, **kwargs):
1521
  """
1522
  Scales the gradients in accordance to the `GradientAccumulationPlugin` and calls the correct `backward()` based
@@ -1542,6 +1595,7 @@ class Accelerator:
1542
  self.scaler.scale(loss).backward(**kwargs)
1543
  else:
1544
  loss.backward(**kwargs)
 
1545
  def set_trigger(self):
1546
  """
1547
  Sets the internal trigger tensor to 1 on the current process. A latter check should follow using this which
@@ -1563,6 +1617,7 @@ class Accelerator:
1563
  ```
1564
  """
1565
  self.flag_tensor = torch.tensor(1, device=self.device)
 
1566
  def check_trigger(self):
1567
  """
1568
  Checks if the internal trigger tensor has been set to 1 in any of the processes. If so, will return `True` and
@@ -1591,6 +1646,7 @@ class Accelerator:
1591
  self.flag_tensor = torch.tensor(0, device=self.device)
1592
  return True
1593
  return False
 
1594
  def unscale_gradients(self, optimizer=None):
1595
  """
1596
  Unscale the gradients in mixed precision training with AMP. This is a noop in all other settings.
@@ -1624,6 +1680,7 @@ class Accelerator:
1624
  gradients = xm._fetch_gradients(opt)
1625
  self.reduce(gradients, scale=1.0 / self.num_processes)
1626
  self.scaler.unscale_(opt)
 
1627
  def clip_grad_norm_(self, parameters, max_norm, norm_type=2):
1628
  """
1629
  Should be used in place of `torch.nn.utils.clip_grad_norm_`.
@@ -1656,6 +1713,7 @@ class Accelerator:
1656
  return None
1657
  self.unscale_gradients()
1658
  return torch.nn.utils.clip_grad_norm_(parameters, max_norm, norm_type=norm_type)
 
1659
  def clip_grad_value_(self, parameters, clip_value):
1660
  """
1661
  Should be used in place of `torch.nn.utils.clip_grad_value_`.
@@ -1678,6 +1736,7 @@ class Accelerator:
1678
  raise Exception("DeepSpeed and FSDP do not support `clip_grad_value_`. Use `clip_grad_norm_` instead.")
1679
  self.unscale_gradients()
1680
  torch.nn.utils.clip_grad_value_(parameters, clip_value)
 
1681
  def gather(self, tensor):
1682
  """
1683
  Gather the values in *tensor* across all processes and concatenate them on the first dimension. Useful to
@@ -1703,6 +1762,7 @@ class Accelerator:
1703
  ```
1704
  """
1705
  return gather(tensor)
 
1706
  def gather_for_metrics(self, input_data):
1707
  """
1708
  Gathers `input_data` and potentially drops duplicates in the last batch if on a distributed system. Should be
@@ -1744,7 +1804,8 @@ class Accelerator:
1744
  return data
1745
  elif self.gradient_state.remainder > 0:
1746
  # Last batch needs to be truncated on distributed systems as it contains additional samples
1747
- def _adjust_samples(tensor):
 
1748
  return tensor[: self.gradient_state.remainder]
1749
  return recursively_apply(_adjust_samples, data)
1750
  else: # remainder is 0
@@ -1756,6 +1817,7 @@ class Accelerator:
1756
  except Exception:
1757
  # Dataset had no length or raised an error
1758
  return data
 
1759
  def reduce(self, tensor, reduction="sum", scale=1.0):
1760
  """
1761
  Reduce the values in *tensor* across all processes based on *reduction*.
@@ -1785,6 +1847,7 @@ class Accelerator:
1785
  ```
1786
  """
1787
  return reduce(tensor, reduction, scale)
 
1788
  def pad_across_processes(self, tensor, dim=0, pad_index=0, pad_first=False):
1789
  """
1790
  Recursively pad the tensors in a nested list/tuple/dictionary of tensors from all devices to the same size so
@@ -1814,6 +1877,7 @@ class Accelerator:
1814
  ```
1815
  """
1816
  return pad_across_processes(tensor, dim=dim, pad_index=pad_index, pad_first=pad_first)
 
1817
  def unwrap_model(self, model, keep_fp32_wrapper: bool = True):
1818
  """
1819
  Unwraps the `model` from the additional layer possible added by [`~Accelerator.prepare`]. Useful before saving
@@ -1840,6 +1904,7 @@ class Accelerator:
1840
  ```
1841
  """
1842
  return extract_model_from_parallel(model, keep_fp32_wrapper)
 
1843
  def wait_for_everyone(self):
1844
  """
1845
  Will stop the execution of the current process until every other process has reached that point (so this does
@@ -1861,6 +1926,7 @@ class Accelerator:
1861
  """
1862
  wait_for_everyone()
1863
  @on_main_process
 
1864
  def init_trackers(self, project_name: str, config: dict | None = None, init_kwargs: dict | None = {}):
1865
  """
1866
  Initializes a run for all trackers stored in `self.log_with`, potentially with starting configurations
@@ -1902,6 +1968,7 @@ class Accelerator:
1902
  if config is not None:
1903
  for tracker in self.trackers:
1904
  tracker.store_init_configuration(config)
 
1905
  def get_tracker(self, name: str, unwrap: bool = False):
1906
  """
1907
  Returns a `tracker` from `self.trackers` based on `name` on the main process only.
@@ -1929,6 +1996,7 @@ class Accelerator:
1929
  # Handle tracker only made on main process
1930
  return GeneralTracker(_blank=True)
1931
  @on_main_process
 
1932
  def log(self, values: dict, step: int | None = None, log_kwargs: dict | None = {}):
1933
  """
1934
  Logs `values` to all stored trackers in `self.trackers` on the main process only.
@@ -1954,6 +2022,7 @@ class Accelerator:
1954
  for tracker in self.trackers:
1955
  tracker.log(values, step=step, **log_kwargs.get(tracker.name, {}))
1956
  @on_main_process
 
1957
  def end_training(self):
1958
  """
1959
  Runs any special end training behaviors, such as stopping trackers on the main process only. Should always be
@@ -1969,6 +2038,7 @@ class Accelerator:
1969
  """
1970
  for tracker in self.trackers:
1971
  tracker.finish()
 
1972
  def save(self, obj, f, safe_serialization=False):
1973
  """
1974
  Save the object passed to disk once per machine. Use in place of `torch.save`.
@@ -1993,6 +2063,7 @@ class Accelerator:
1993
  save_on_each_node=self.project_configuration.save_on_each_node,
1994
  safe_serialization=safe_serialization,
1995
  )
 
1996
  def save_model(
1997
  self,
1998
  model: torch.nn.Module,
@@ -2081,6 +2152,7 @@ class Accelerator:
2081
  f"split in {len(shards)} checkpoint shards. You can find where each parameters has been saved in the "
2082
  f"index located at {save_index_file}."
2083
  )
 
2084
  def register_save_state_pre_hook(self, hook: Callable[..., None]) -> hooks.RemovableHandle:
2085
  """
2086
  Registers a pre hook to be run before `save_checkpoint` is called in [`Accelerator.save_state`].
@@ -2104,6 +2176,7 @@ class Accelerator:
2104
  handle = hooks.RemovableHandle(self._save_model_state_pre_hook)
2105
  self._save_model_state_pre_hook[handle.id] = hook
2106
  return handle
 
2107
  def save_state(self, output_dir: str = None, safe_serialization: bool = True, **save_model_func_kwargs):
2108
  """
2109
  Saves the current states of the model, optimizer, scaler, RNG generators, and registered objects to a folder.
@@ -2143,7 +2216,8 @@ class Accelerator:
2143
  and (len(folders) + 1 > self.project_configuration.total_limit)
2144
  and self.is_main_process
2145
  ):
2146
- def _inner(folder):
 
2147
  return list(map(int, re.findall(r"[\/]?([0-9]+)(?=[^\/]*$)", folder)))[0]
2148
  folders.sort(key=_inner)
2149
  logger.warning(
@@ -2219,6 +2293,7 @@ class Accelerator:
2219
  save_custom_state(obj, output_dir, i, save_on_each_node=self.project_configuration.save_on_each_node)
2220
  self.project_configuration.iteration += 1
2221
  return save_location
 
2222
  def register_load_state_pre_hook(self, hook: Callable[..., None]) -> hooks.RemovableHandle:
2223
  """
2224
  Registers a pre hook to be run before [`load_checkpoint`] is called in [`Accelerator.load_state`].
@@ -2241,6 +2316,7 @@ class Accelerator:
2241
  handle = hooks.RemovableHandle(self._load_model_state_pre_hook)
2242
  self._load_model_state_pre_hook[handle.id] = hook
2243
  return handle
 
2244
  def load_state(self, input_dir: str = None, **load_model_func_kwargs):
2245
  """
2246
  Loads the current states of the model, optimizer, scaler, RNG generators, and registered objects.
@@ -2274,7 +2350,8 @@ class Accelerator:
2274
  # Pick up from automatic checkpoint naming
2275
  input_dir = os.path.join(self.project_dir, "checkpoints")
2276
  folders = [os.path.join(input_dir, folder) for folder in os.listdir(input_dir)]
2277
- def _inner(folder):
 
2278
  return list(map(int, re.findall(r"[\/]?([0-9]+)(?=[^\/]*$)", folder)))[0]
2279
  folders.sort(key=_inner)
2280
  input_dir = folders[-1]
@@ -2356,6 +2433,7 @@ class Accelerator:
2356
  logger.info(f"Loading in {len(custom_checkpoints)} custom states")
2357
  for index, obj in enumerate(self._custom_objects):
2358
  load_custom_state(obj, input_dir, index)
 
2359
  def free_memory(self):
2360
  """
2361
  Will release all references to the internal objects stored and call the garbage collector. You should call this
@@ -2377,6 +2455,7 @@ class Accelerator:
2377
  self.deepspeed_engine_wrapped = None
2378
  self.step = 0
2379
  release_memory()
 
2380
  def clear(self):
2381
  """
2382
  Alias for [`Accelerate.free_memory`], releases all references to the internal objects stored and call the
@@ -2392,6 +2471,7 @@ class Accelerator:
2392
  ```
2393
  """
2394
  self.free_memory()
 
2395
  def _get_named_parameters(self, *args):
2396
  named_parameters = {}
2397
  for obj in args:
@@ -2399,6 +2479,7 @@ class Accelerator:
2399
  obj = extract_model_from_parallel(obj)
2400
  named_parameters.update({n: p for n, p in obj.named_parameters()})
2401
  return named_parameters
 
2402
  def _get_devices(self, *args):
2403
  model_device = None
2404
  optimizer_device = None
@@ -2415,6 +2496,7 @@ class Accelerator:
2415
  optimizer_device = param_group["params"][0].device
2416
  break
2417
  return (model_device, optimizer_device)
 
2418
  def get_state_dict(self, model, unwrap=True):
2419
  """
2420
  Returns the state dictionary of a model sent through [`Accelerator.prepare`] potentially without full
@@ -2461,6 +2543,7 @@ class Accelerator:
2461
  model = self.unwrap_model(model)
2462
  state_dict = model.state_dict()
2463
  return state_dict
 
2464
  def register_for_checkpointing(self, *objects):
2465
  """
2466
  Makes note of `objects` and will save or load them in during `save_state` or `load_state`.
@@ -2490,6 +2573,7 @@ class Accelerator:
2490
  raise ValueError(err)
2491
  self._custom_objects.extend(objects)
2492
  @contextmanager
 
2493
  def autocast(self, cache_enabled: bool = False, autocast_handler: AutocastKwargs = None):
2494
  """
2495
  Will apply automatic mixed-precision inside the block inside this context manager, if it is enabled. Nothing
@@ -2521,6 +2605,7 @@ class Accelerator:
2521
  yield
2522
  autocast_context.__exit__(*sys.exc_info())
2523
  @property
 
2524
  def optimizer_step_was_skipped(self):
2525
  """
2526
  Whether or not the optimizer update was skipped (because of gradient overflow in mixed precision), in which
@@ -2530,6 +2615,7 @@ class Accelerator:
2530
  if optimizer.step_was_skipped:
2531
  return True
2532
  return False
 
2533
  def skip_first_batches(self, dataloader, num_batches: int = 0):
2534
  """
2535
  Creates a new `torch.utils.data.DataLoader` that will efficiently skip the first `num_batches`.
@@ -2556,9 +2642,11 @@ class Accelerator:
2556
  ```
2557
  """
2558
  return skip_first_batches(dataloader, num_batches=num_batches)
 
2559
  def __deepcopy__(self, memo):
2560
  logger.info("Deep copying the `Accelerator` object, note that this will point to the same original object.")
2561
  return self
 
2562
  def verify_device_map(self, model: torch.nn.Module) -> bool:
2563
  """
2564
  Verifies that `model` has not been prepared with big model inference with a device-map resembling `auto`.
 
87
  - **sync_gradients** (`bool`) -- Whether the gradients are currently being synced across all processes.
88
  - **use_distributed** (`bool`) -- Whether the current configuration is for distributed training.
89
  """
90
+
91
  def __init__(
92
  self,
93
  device_placement: bool = True,
 
297
  self.flag_tensor = None
298
  check_os_kernel()
299
  @property
300
+
301
  def use_distributed(self):
302
  """
303
  Whether the Accelerator is configured for distributed training
304
  """
305
  return self.state.use_distributed
306
  @property
307
+
308
  def distributed_type(self):
309
  return self.state.distributed_type
310
  @property
311
+
312
  def num_processes(self):
313
  return self.state.num_processes
314
  @property
315
+
316
  def process_index(self):
317
  return self.state.process_index
318
  @property
319
+
320
  def local_process_index(self):
321
  return self.state.local_process_index
322
  @property
323
+
324
  def device(self):
325
  return self.state.device
326
  @property
327
+
328
  def project_dir(self):
329
  return self.project_configuration.project_dir
330
  @property
331
+
332
  def logging_dir(self):
333
  return self.project_configuration.logging_dir
334
  @property
335
+
336
  def save_iteration(self):
337
  return self.project_configuration.iteration
338
  @property
339
+
340
  def is_main_process(self):
341
  """True for one process only."""
342
  return self.state.is_main_process
343
  @property
344
+
345
  def is_local_main_process(self):
346
  """True for one process per server."""
347
  return self.state.is_local_main_process
348
  @property
349
+
350
  def use_fp16(self):
351
  warnings.warn(
352
  "The `use_fp16` property is deprecated and will be removed in version 1.0 of Accelerate use "
 
355
  )
356
  return self.mixed_precision != "no"
357
  @property
358
+
359
  def is_last_process(self):
360
  return self.process_index == self.num_processes - 1
361
  @property
362
+
363
  def mixed_precision(self):
364
  return self.state.mixed_precision
365
  @contextmanager
366
+
367
  def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
368
  """
369
  Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
 
398
  """
399
  with PartialState().split_between_processes(inputs, apply_padding=apply_padding) as inputs:
400
  yield inputs
401
+
402
  def on_main_process(self, function: Callable[..., Any] = None):
403
  """
404
  A decorator that will run the decorated function on the main process only. Can also be called using the
 
424
  raise ValueError(
425
  "The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
426
  )
427
+
428
+ def _inner(*args, **kwargs):
429
  return PartialState().on_main_process(function)(*args, **kwargs)
430
  return _inner
431
+
432
  def on_local_main_process(self, function: Callable[..., Any] = None):
433
  """
434
  A decorator that will run the decorated function on the local main process only. Can also be called using the
 
441
  from accelerate import Accelerator
442
  accelerator = Accelerator()
443
  @accelerator.on_local_main_process
444
+
445
+ def print_something():
446
  print("This will be printed by process 0 only on each server.")
447
  print_something()
448
  # On server 1:
 
459
  raise ValueError(
460
  "The `on_local_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
461
  )
462
+
463
+ def _inner(*args, **kwargs):
464
  return PartialState().on_local_main_process(function)(*args, **kwargs)
465
  return _inner
466
+
467
  def on_last_process(self, function: Callable[..., Any]):
468
  """
469
  A decorator that will run the decorated function on the last process only. Can also be called using the
 
476
  from accelerate import Accelerator
477
  accelerator = Accelerator()
478
  @accelerator.on_last_process
479
+
480
+ def print_something():
481
  print(f"Printed on process {accelerator.process_index}")
482
  print_something()
483
  "Printed on process 3"
 
491
  raise ValueError(
492
  "The `on_last_process` decorator must be called with a function on an instantiated `Accelerator` object."
493
  )
494
+
495
+ def _inner(*args, **kwargs):
496
  return PartialState().on_last_process(function)(*args, **kwargs)
497
  return _inner
498
+
499
  def on_process(self, function: Callable[..., Any] = None, process_index: int = None):
500
  """
501
  A decorator that will run the decorated function on a given process index only. Can also be called using the
 
511
  from accelerate import Accelerator
512
  accelerator = Accelerator()
513
  @accelerator.on_process(process_index=2)
514
+
515
+ def print_something():
516
  print(f"Printed on process {accelerator.process_index}")
517
  print_something()
518
  "Printed on process 2"
 
529
  raise ValueError(
530
  "The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
531
  )
532
+
533
+ def _inner(*args, **kwargs):
534
  return PartialState().on_process(function, process_index)(*args, **kwargs)
535
  return _inner
536
+
537
  def on_local_process(self, function: Callable[..., Any] = None, local_process_index: int = None):
538
  """
539
  A decorator that will run the decorated function on a given local process index only. Can also be called using
 
549
  from accelerate import Accelerator
550
  accelerator = Accelerator()
551
  @accelerator.on_local_process(local_process_index=2)
552
+
553
+ def print_something():
554
  print(f"Printed on process {accelerator.local_process_index}")
555
  print_something()
556
  # On server 1:
 
570
  raise ValueError(
571
  "The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
572
  )
573
+
574
+ def _inner(*args, **kwargs):
575
  return PartialState().on_local_process(function, local_process_index)(*args, **kwargs)
576
  return _inner
577
  @contextmanager
578
+
579
  def main_process_first(self):
580
  """
581
  Lets the main process go first inside a with block.
 
593
  with self.state.main_process_first():
594
  yield
595
  @contextmanager
596
+
597
  def local_main_process_first(self):
598
  """
599
  Lets the local main process go inside a with block.
 
611
  with self.state.local_main_process_first():
612
  yield
613
  @contextmanager
614
+
615
  def no_sync(self, model):
616
  """
617
  A context manager to disable gradient synchronizations across DDP processes by calling
 
646
  yield
647
  @staticmethod
648
  @contextmanager
649
+
650
  def trigger_sync_in_backward(model):
651
  """Trigger the sync of the gradients in the next backward pass of the model after multiple forward passes under
652
  `Accelerator.no_sync` (only applicable in multi-GPU scenarios).
 
685
  finally:
686
  model.require_backward_grad_sync = old_require_backward_grad_sync
687
  model.require_forward_param_sync = old_require_forward_param_sync
688
+
689
  def _do_sync(self):
690
  "Sets the right `sync_gradients` context and either resets or increases `self.step`"
691
  if self.gradient_state.sync_with_dataloader and self.gradient_state.end_of_dataloader:
 
695
  self.step += 1
696
  self.gradient_state._set_sync_gradients((self.step % self.gradient_state.num_steps) == 0)
697
  @property
698
+
699
  def sync_gradients(self):
700
  return self.gradient_state.sync_gradients
701
  @sync_gradients.setter
702
+
703
  def sync_gradients(self, sync_gradients):
704
  self.gradient_state.sync_gradients = sync_gradients
705
  @property
706
+
707
  def gradient_accumulation_steps(self):
708
  return self.gradient_state.num_steps
709
  @gradient_accumulation_steps.setter
710
+
711
  def gradient_accumulation_steps(self, gradient_accumulation_steps):
712
  self.gradient_state.plugin_kwargs.update({"num_steps": gradient_accumulation_steps})
713
  @contextmanager
714
+
715
  def accumulate(self, *models):
716
  """
717
  A context manager that will lightly wrap around and perform gradient accumulation automatically
 
740
  cm_stack.enter_context(contextlib.nullcontext() if self.sync_gradients else self.no_sync(m))
741
  yield
742
  @contextmanager
743
+
744
  def join_uneven_inputs(self, joinables, even_batches=None):
745
  """
746
  A context manager that facilitates distributed training or evaluation on uneven inputs, which acts as a wrapper
 
807
  )
808
  with contextlib.nullcontext(joinables):
809
  yield
810
+
811
  def print(self, *args, **kwargs):
812
  """
813
  Drop in replacement of `print()` to only print once per server.
 
819
  ```
820
  """
821
  self.state.print(*args, **kwargs)
822
+
823
  def _prepare_one(self, obj, first_pass=False, device_placement=None):
824
  # First pass of preparation: DataLoader, model, optimizer
825
  if first_pass:
 
836
  return scheduler
837
  # Return the unprocessed object if previous criteria was not met
838
  return obj
839
+
840
  def prepare(self, *args, device_placement=None):
841
  """
842
  Prepare all objects passed in `args` for distributed training and mixed precision, then return them in the same
 
953
  ):
954
  setattr(item, "_is_accelerate_prepared", True)
955
  return result if len(result) > 1 else result[0]
956
+
957
  def prepare_model(self, model: torch.nn.Module, device_placement: bool = None, evaluation_mode: bool = False):
958
  """
959
  Prepares a PyTorch model for training in any distributed setup. It is recommended to use
 
1106
  raise ValueError("Using `torch.compile` requires PyTorch 2.0 or higher.")
1107
  model = torch.compile(model, **self.state.dynamo_plugin.to_kwargs())
1108
  return model
1109
+
1110
  def _prepare_deepspeed(self, *args):
1111
  import deepspeed
1112
  deepspeed_plugin = self.state.deepspeed_plugin
 
1296
  "You can't use same `Accelerator()` instance with multiple models when using DeepSpeed"
1297
  )
1298
  return tuple(result)
1299
+
1300
  def _prepare_megatron_lm(self, *args):
1301
  megatron_lm_plugin = self.state.megatron_lm_plugin
1302
  if not megatron_lm_plugin.megatron_dataset_flag:
 
1389
  "You can't use same `Accelerator()` instance with multiple models when using Megatron-LM"
1390
  )
1391
  return tuple(result)
1392
+
1393
  def _prepare_ipex(self, *args):
1394
  if not is_ipex_available():
1395
  raise ImportError(
 
1421
  elif isinstance(result[i], (torch.optim.Optimizer)):
1422
  result[i] = optimizer
1423
  return tuple(result)
1424
+
1425
  def _prepare_msamp(self, *args):
1426
  if not is_msamp_available():
1427
  raise ImportError(
 
1456
  elif isinstance(result[i], (torch.optim.Optimizer)):
1457
  result[i] = optimizer
1458
  return tuple(result)
1459
+
1460
  def prepare_data_loader(
1461
  self, data_loader: torch.utils.data.DataLoader, device_placement=None, slice_fn_for_dispatch=None
1462
  ):
 
1503
  )
1504
  self._dataloaders.append(prepared_data_loader)
1505
  return prepared_data_loader
1506
+
1507
  def prepare_optimizer(self, optimizer: torch.optim.Optimizer, device_placement=None):
1508
  """
1509
  Prepares a PyTorch Optimizer for training in any distributed setup. It is recommended to use
 
1532
  optimizer = AcceleratedOptimizer(optimizer, device_placement=device_placement, scaler=self.scaler)
1533
  self._optimizers.append(optimizer)
1534
  return optimizer
1535
+
1536
  def prepare_scheduler(self, scheduler: LRScheduler):
1537
  """
1538
  Prepares a PyTorch Scheduler for training in any distributed setup. It is recommended to use
 
1569
  )
1570
  self._schedulers.append(scheduler)
1571
  return scheduler
1572
+
1573
  def backward(self, loss, **kwargs):
1574
  """
1575
  Scales the gradients in accordance to the `GradientAccumulationPlugin` and calls the correct `backward()` based
 
1595
  self.scaler.scale(loss).backward(**kwargs)
1596
  else:
1597
  loss.backward(**kwargs)
1598
+
1599
  def set_trigger(self):
1600
  """
1601
  Sets the internal trigger tensor to 1 on the current process. A latter check should follow using this which
 
1617
  ```
1618
  """
1619
  self.flag_tensor = torch.tensor(1, device=self.device)
1620
+
1621
  def check_trigger(self):
1622
  """
1623
  Checks if the internal trigger tensor has been set to 1 in any of the processes. If so, will return `True` and
 
1646
  self.flag_tensor = torch.tensor(0, device=self.device)
1647
  return True
1648
  return False
1649
+
1650
  def unscale_gradients(self, optimizer=None):
1651
  """
1652
  Unscale the gradients in mixed precision training with AMP. This is a noop in all other settings.
 
1680
  gradients = xm._fetch_gradients(opt)
1681
  self.reduce(gradients, scale=1.0 / self.num_processes)
1682
  self.scaler.unscale_(opt)
1683
+
1684
  def clip_grad_norm_(self, parameters, max_norm, norm_type=2):
1685
  """
1686
  Should be used in place of `torch.nn.utils.clip_grad_norm_`.
 
1713
  return None
1714
  self.unscale_gradients()
1715
  return torch.nn.utils.clip_grad_norm_(parameters, max_norm, norm_type=norm_type)
1716
+
1717
  def clip_grad_value_(self, parameters, clip_value):
1718
  """
1719
  Should be used in place of `torch.nn.utils.clip_grad_value_`.
 
1736
  raise Exception("DeepSpeed and FSDP do not support `clip_grad_value_`. Use `clip_grad_norm_` instead.")
1737
  self.unscale_gradients()
1738
  torch.nn.utils.clip_grad_value_(parameters, clip_value)
1739
+
1740
  def gather(self, tensor):
1741
  """
1742
  Gather the values in *tensor* across all processes and concatenate them on the first dimension. Useful to
 
1762
  ```
1763
  """
1764
  return gather(tensor)
1765
+
1766
  def gather_for_metrics(self, input_data):
1767
  """
1768
  Gathers `input_data` and potentially drops duplicates in the last batch if on a distributed system. Should be
 
1804
  return data
1805
  elif self.gradient_state.remainder > 0:
1806
  # Last batch needs to be truncated on distributed systems as it contains additional samples
1807
+
1808
+ def _adjust_samples(tensor):
1809
  return tensor[: self.gradient_state.remainder]
1810
  return recursively_apply(_adjust_samples, data)
1811
  else: # remainder is 0
 
1817
  except Exception:
1818
  # Dataset had no length or raised an error
1819
  return data
1820
+
1821
  def reduce(self, tensor, reduction="sum", scale=1.0):
1822
  """
1823
  Reduce the values in *tensor* across all processes based on *reduction*.
 
1847
  ```
1848
  """
1849
  return reduce(tensor, reduction, scale)
1850
+
1851
  def pad_across_processes(self, tensor, dim=0, pad_index=0, pad_first=False):
1852
  """
1853
  Recursively pad the tensors in a nested list/tuple/dictionary of tensors from all devices to the same size so
 
1877
  ```
1878
  """
1879
  return pad_across_processes(tensor, dim=dim, pad_index=pad_index, pad_first=pad_first)
1880
+
1881
  def unwrap_model(self, model, keep_fp32_wrapper: bool = True):
1882
  """
1883
  Unwraps the `model` from the additional layer possible added by [`~Accelerator.prepare`]. Useful before saving
 
1904
  ```
1905
  """
1906
  return extract_model_from_parallel(model, keep_fp32_wrapper)
1907
+
1908
  def wait_for_everyone(self):
1909
  """
1910
  Will stop the execution of the current process until every other process has reached that point (so this does
 
1926
  """
1927
  wait_for_everyone()
1928
  @on_main_process
1929
+
1930
  def init_trackers(self, project_name: str, config: dict | None = None, init_kwargs: dict | None = {}):
1931
  """
1932
  Initializes a run for all trackers stored in `self.log_with`, potentially with starting configurations
 
1968
  if config is not None:
1969
  for tracker in self.trackers:
1970
  tracker.store_init_configuration(config)
1971
+
1972
  def get_tracker(self, name: str, unwrap: bool = False):
1973
  """
1974
  Returns a `tracker` from `self.trackers` based on `name` on the main process only.
 
1996
  # Handle tracker only made on main process
1997
  return GeneralTracker(_blank=True)
1998
  @on_main_process
1999
+
2000
  def log(self, values: dict, step: int | None = None, log_kwargs: dict | None = {}):
2001
  """
2002
  Logs `values` to all stored trackers in `self.trackers` on the main process only.
 
2022
  for tracker in self.trackers:
2023
  tracker.log(values, step=step, **log_kwargs.get(tracker.name, {}))
2024
  @on_main_process
2025
+
2026
  def end_training(self):
2027
  """
2028
  Runs any special end training behaviors, such as stopping trackers on the main process only. Should always be
 
2038
  """
2039
  for tracker in self.trackers:
2040
  tracker.finish()
2041
+
2042
  def save(self, obj, f, safe_serialization=False):
2043
  """
2044
  Save the object passed to disk once per machine. Use in place of `torch.save`.
 
2063
  save_on_each_node=self.project_configuration.save_on_each_node,
2064
  safe_serialization=safe_serialization,
2065
  )
2066
+
2067
  def save_model(
2068
  self,
2069
  model: torch.nn.Module,
 
2152
  f"split in {len(shards)} checkpoint shards. You can find where each parameters has been saved in the "
2153
  f"index located at {save_index_file}."
2154
  )
2155
+
2156
  def register_save_state_pre_hook(self, hook: Callable[..., None]) -> hooks.RemovableHandle:
2157
  """
2158
  Registers a pre hook to be run before `save_checkpoint` is called in [`Accelerator.save_state`].
 
2176
  handle = hooks.RemovableHandle(self._save_model_state_pre_hook)
2177
  self._save_model_state_pre_hook[handle.id] = hook
2178
  return handle
2179
+
2180
  def save_state(self, output_dir: str = None, safe_serialization: bool = True, **save_model_func_kwargs):
2181
  """
2182
  Saves the current states of the model, optimizer, scaler, RNG generators, and registered objects to a folder.
 
2216
  and (len(folders) + 1 > self.project_configuration.total_limit)
2217
  and self.is_main_process
2218
  ):
2219
+
2220
+ def _inner(folder):
2221
  return list(map(int, re.findall(r"[\/]?([0-9]+)(?=[^\/]*$)", folder)))[0]
2222
  folders.sort(key=_inner)
2223
  logger.warning(
 
2293
  save_custom_state(obj, output_dir, i, save_on_each_node=self.project_configuration.save_on_each_node)
2294
  self.project_configuration.iteration += 1
2295
  return save_location
2296
+
2297
  def register_load_state_pre_hook(self, hook: Callable[..., None]) -> hooks.RemovableHandle:
2298
  """
2299
  Registers a pre hook to be run before [`load_checkpoint`] is called in [`Accelerator.load_state`].
 
2316
  handle = hooks.RemovableHandle(self._load_model_state_pre_hook)
2317
  self._load_model_state_pre_hook[handle.id] = hook
2318
  return handle
2319
+
2320
  def load_state(self, input_dir: str = None, **load_model_func_kwargs):
2321
  """
2322
  Loads the current states of the model, optimizer, scaler, RNG generators, and registered objects.
 
2350
  # Pick up from automatic checkpoint naming
2351
  input_dir = os.path.join(self.project_dir, "checkpoints")
2352
  folders = [os.path.join(input_dir, folder) for folder in os.listdir(input_dir)]
2353
+
2354
+ def _inner(folder):
2355
  return list(map(int, re.findall(r"[\/]?([0-9]+)(?=[^\/]*$)", folder)))[0]
2356
  folders.sort(key=_inner)
2357
  input_dir = folders[-1]
 
2433
  logger.info(f"Loading in {len(custom_checkpoints)} custom states")
2434
  for index, obj in enumerate(self._custom_objects):
2435
  load_custom_state(obj, input_dir, index)
2436
+
2437
  def free_memory(self):
2438
  """
2439
  Will release all references to the internal objects stored and call the garbage collector. You should call this
 
2455
  self.deepspeed_engine_wrapped = None
2456
  self.step = 0
2457
  release_memory()
2458
+
2459
  def clear(self):
2460
  """
2461
  Alias for [`Accelerate.free_memory`], releases all references to the internal objects stored and call the
 
2471
  ```
2472
  """
2473
  self.free_memory()
2474
+
2475
  def _get_named_parameters(self, *args):
2476
  named_parameters = {}
2477
  for obj in args:
 
2479
  obj = extract_model_from_parallel(obj)
2480
  named_parameters.update({n: p for n, p in obj.named_parameters()})
2481
  return named_parameters
2482
+
2483
  def _get_devices(self, *args):
2484
  model_device = None
2485
  optimizer_device = None
 
2496
  optimizer_device = param_group["params"][0].device
2497
  break
2498
  return (model_device, optimizer_device)
2499
+
2500
  def get_state_dict(self, model, unwrap=True):
2501
  """
2502
  Returns the state dictionary of a model sent through [`Accelerator.prepare`] potentially without full
 
2543
  model = self.unwrap_model(model)
2544
  state_dict = model.state_dict()
2545
  return state_dict
2546
+
2547
  def register_for_checkpointing(self, *objects):
2548
  """
2549
  Makes note of `objects` and will save or load them in during `save_state` or `load_state`.
 
2573
  raise ValueError(err)
2574
  self._custom_objects.extend(objects)
2575
  @contextmanager
2576
+
2577
  def autocast(self, cache_enabled: bool = False, autocast_handler: AutocastKwargs = None):
2578
  """
2579
  Will apply automatic mixed-precision inside the block inside this context manager, if it is enabled. Nothing
 
2605
  yield
2606
  autocast_context.__exit__(*sys.exc_info())
2607
  @property
2608
+
2609
  def optimizer_step_was_skipped(self):
2610
  """
2611
  Whether or not the optimizer update was skipped (because of gradient overflow in mixed precision), in which
 
2615
  if optimizer.step_was_skipped:
2616
  return True
2617
  return False
2618
+
2619
  def skip_first_batches(self, dataloader, num_batches: int = 0):
2620
  """
2621
  Creates a new `torch.utils.data.DataLoader` that will efficiently skip the first `num_batches`.
 
2642
  ```
2643
  """
2644
  return skip_first_batches(dataloader, num_batches=num_batches)
2645
+
2646
  def __deepcopy__(self, memo):
2647
  logger.info("Deep copying the `Accelerator` object, note that this will point to the same original object.")
2648
  return self
2649
+
2650
  def verify_device_map(self, model: torch.nn.Module) -> bool:
2651
  """
2652
  Verifies that `model` has not been prepared with big model inference with a device-map resembling `auto`.
src/big_modeling.py CHANGED
@@ -51,12 +51,14 @@ def init_on_device(device: torch.device, include_buffers: bool = None):
51
  old_register_parameter = nn.Module.register_parameter
52
  if include_buffers:
53
  old_register_buffer = nn.Module.register_buffer
 
54
  def register_empty_parameter(module, name, param):
55
  old_register_parameter(module, name, param)
56
  if param is not None:
57
  param_cls = type(module._parameters[name])
58
  kwargs = module._parameters[name].__dict__
59
  module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
 
60
  def register_empty_buffer(module, name, buffer, persistent=True):
61
  old_register_buffer(module, name, buffer, persistent=persistent)
62
  if buffer is not None:
@@ -69,8 +71,10 @@ def init_on_device(device: torch.device, include_buffers: bool = None):
69
  }
70
  else:
71
  tensor_constructors_to_patch = {}
 
72
  def patch_tensor_constructor(fn):
73
- def wrapper(*args, **kwargs):
 
74
  kwargs["device"] = device
75
  return fn(*args, **kwargs)
76
  return wrapper
@@ -318,9 +322,11 @@ def dispatch_model(
318
  # Attaching the hook may break tied weights, so we retie them
319
  retie_parameters(model, tied_params)
320
  # add warning to cuda and to method
321
- def add_warning(fn, model):
 
322
  @wraps(fn)
323
- def wrapper(*args, **kwargs):
 
324
  logger.warning("You shouldn't move a model when it is dispatched on multiple devices.")
325
  for param in model.parameters():
326
  if param.device == torch.device("meta"):
 
51
  old_register_parameter = nn.Module.register_parameter
52
  if include_buffers:
53
  old_register_buffer = nn.Module.register_buffer
54
+
55
  def register_empty_parameter(module, name, param):
56
  old_register_parameter(module, name, param)
57
  if param is not None:
58
  param_cls = type(module._parameters[name])
59
  kwargs = module._parameters[name].__dict__
60
  module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
61
+
62
  def register_empty_buffer(module, name, buffer, persistent=True):
63
  old_register_buffer(module, name, buffer, persistent=persistent)
64
  if buffer is not None:
 
71
  }
72
  else:
73
  tensor_constructors_to_patch = {}
74
+
75
  def patch_tensor_constructor(fn):
76
+
77
+ def wrapper(*args, **kwargs):
78
  kwargs["device"] = device
79
  return fn(*args, **kwargs)
80
  return wrapper
 
322
  # Attaching the hook may break tied weights, so we retie them
323
  retie_parameters(model, tied_params)
324
  # add warning to cuda and to method
325
+
326
+ def add_warning(fn, model):
327
  @wraps(fn)
328
+
329
+ def wrapper(*args, **kwargs):
330
  logger.warning("You shouldn't move a model when it is dispatched on multiple devices.")
331
  for param in model.parameters():
332
  if param.device == torch.device("meta"):
src/commands/config/config_args.py CHANGED
@@ -47,6 +47,7 @@ class BaseConfig:
47
  mixed_precision: str
48
  use_cpu: bool
49
  debug: bool
 
50
  def to_dict(self):
51
  result = self.__dict__
52
  # For serialization, it's best to convert Enums to strings (or their underlying value type).
@@ -58,6 +59,7 @@ class BaseConfig:
58
  result = {k: v for k, v in result.items() if v is not None}
59
  return result
60
  @classmethod
 
61
  def from_json_file(cls, json_file=None):
62
  json_file = default_json_config_file if json_file is None else json_file
63
  with open(json_file, "r", encoding="utf-8") as f:
@@ -82,11 +84,13 @@ class BaseConfig:
82
  " version or fix (and potentially remove) these keys from your config file."
83
  )
84
  return cls(**config_dict)
 
85
  def to_json_file(self, json_file):
86
  with open(json_file, "w", encoding="utf-8") as f:
87
  content = json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
88
  f.write(content)
89
  @classmethod
 
90
  def from_yaml_file(cls, yaml_file=None):
91
  yaml_file = default_yaml_config_file if yaml_file is None else yaml_file
92
  with open(yaml_file, "r", encoding="utf-8") as f:
@@ -113,9 +117,11 @@ class BaseConfig:
113
  " version or fix (and potentially remove) these keys from your config file."
114
  )
115
  return cls(**config_dict)
 
116
  def to_yaml_file(self, yaml_file):
117
  with open(yaml_file, "w", encoding="utf-8") as f:
118
  yaml.safe_dump(self.to_dict(), f)
 
119
  def __post_init__(self):
120
  if isinstance(self.compute_environment, str):
121
  self.compute_environment = ComputeEnvironment(self.compute_environment)
@@ -158,6 +164,7 @@ class ClusterConfig(BaseConfig):
158
  tpu_env: List[str] = None
159
  # args for dynamo
160
  dynamo_config: dict = None
 
161
  def __post_init__(self):
162
  if self.deepspeed_config is None:
163
  self.deepspeed_config = {}
 
47
  mixed_precision: str
48
  use_cpu: bool
49
  debug: bool
50
+
51
  def to_dict(self):
52
  result = self.__dict__
53
  # For serialization, it's best to convert Enums to strings (or their underlying value type).
 
59
  result = {k: v for k, v in result.items() if v is not None}
60
  return result
61
  @classmethod
62
+
63
  def from_json_file(cls, json_file=None):
64
  json_file = default_json_config_file if json_file is None else json_file
65
  with open(json_file, "r", encoding="utf-8") as f:
 
84
  " version or fix (and potentially remove) these keys from your config file."
85
  )
86
  return cls(**config_dict)
87
+
88
  def to_json_file(self, json_file):
89
  with open(json_file, "w", encoding="utf-8") as f:
90
  content = json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
91
  f.write(content)
92
  @classmethod
93
+
94
  def from_yaml_file(cls, yaml_file=None):
95
  yaml_file = default_yaml_config_file if yaml_file is None else yaml_file
96
  with open(yaml_file, "r", encoding="utf-8") as f:
 
117
  " version or fix (and potentially remove) these keys from your config file."
118
  )
119
  return cls(**config_dict)
120
+
121
  def to_yaml_file(self, yaml_file):
122
  with open(yaml_file, "w", encoding="utf-8") as f:
123
  yaml.safe_dump(self.to_dict(), f)
124
+
125
  def __post_init__(self):
126
  if isinstance(self.compute_environment, str):
127
  self.compute_environment = ComputeEnvironment(self.compute_environment)
 
164
  tpu_env: List[str] = None
165
  # args for dynamo
166
  dynamo_config: dict = None
167
+
168
  def __post_init__(self):
169
  if self.deepspeed_config is None:
170
  self.deepspeed_config = {}
src/commands/config/config_utils.py CHANGED
@@ -49,6 +49,7 @@ class SubcommandHelpFormatter(argparse.RawDescriptionHelpFormatter):
49
  """
50
  A custom formatter that will remove the usage line from the help message for subcommands.
51
  """
 
52
  def _format_usage(self, usage, actions, groups, prefix):
53
  usage = super()._format_usage(usage, actions, groups, prefix)
54
  usage = usage.replace("<command> [<args>] ", "")
 
49
  """
50
  A custom formatter that will remove the usage line from the help message for subcommands.
51
  """
52
+
53
  def _format_usage(self, usage, actions, groups, prefix):
54
  usage = super()._format_usage(usage, actions, groups, prefix)
55
  usage = usage.replace("<command> [<args>] ", "")
src/commands/estimate.py CHANGED
@@ -100,6 +100,7 @@ def create_ascii_table(headers: list, rows: list, title: str):
100
  formats = [f"%{column_widths[i]}s" for i in range(len(rows[0]))]
101
  pattern = f"{sep_char}{sep_char.join(formats)}{sep_char}"
102
  diff = 0
 
103
  def make_row(left_char, middle_char, right_char):
104
  return f"{left_char}{middle_char.join([in_between * n for n in column_widths])}{in_between * diff}{right_char}"
105
  separator = make_row("├", "┼", "┤")
 
100
  formats = [f"%{column_widths[i]}s" for i in range(len(rows[0]))]
101
  pattern = f"{sep_char}{sep_char.join(formats)}{sep_char}"
102
  diff = 0
103
+
104
  def make_row(left_char, middle_char, right_char):
105
  return f"{left_char}{middle_char.join([in_between * n for n in column_widths])}{in_between * diff}{right_char}"
106
  separator = make_row("├", "┼", "┤")
src/commands/launch.py CHANGED
@@ -17,6 +17,7 @@ class _CustomHelpAction(argparse._HelpAction):
17
  called. This is useful for the case where the user is using a specific platform and only wants to see the arguments
18
  for that platform.
19
  """
 
20
  def __call__(self, parser, namespace, values, option_string=None):
21
  if "accelerate" in sys.argv[0] and "launch" in sys.argv[1:]:
22
  args = sys.argv[2:]
 
17
  called. This is useful for the case where the user is using a specific platform and only wants to see the arguments
18
  for that platform.
19
  """
20
+
21
  def __call__(self, parser, namespace, values, option_string=None):
22
  if "accelerate" in sys.argv[0] and "launch" in sys.argv[1:]:
23
  args = sys.argv[2:]
src/data_loader.py CHANGED
@@ -29,10 +29,12 @@ class SeedableRandomSampler(RandomSampler):
29
  If a custom `generator` is passed, it will rely on its initial seed as well as the current iteration it is on
30
  (stored in `self.epoch`).
31
  """
 
32
  def __init__(self, *args, **kwargs):
33
  super().__init__(*args, **kwargs)
34
  self.epoch = 0
35
  self.seed = torch.random.initial_seed()
 
36
  def __iter__(self):
37
  if self.generator is None:
38
  self.generator = torch.Generator()
@@ -43,6 +45,7 @@ class SeedableRandomSampler(RandomSampler):
43
  self.generator.manual_seed(seed)
44
  yield from super().__iter__()
45
  self.set_epoch(self.epoch + 1)
 
46
  def set_epoch(self, epoch: int):
47
  "Sets the current iteration of the sampler."
48
  self.epoch = epoch
@@ -74,6 +77,7 @@ class BatchSamplerShard(BatchSampler):
74
  `BatchSampler`s with varying batch sizes are not enabled by default. To enable this behaviour, set `even_batches`
75
  equal to `False`
76
  </Tip>"""
 
77
  def __init__(
78
  self,
79
  batch_sampler: BatchSampler,
@@ -100,8 +104,10 @@ class BatchSamplerShard(BatchSampler):
100
  "are not calling this method directly, set `accelerator.even_batches=False` instead."
101
  )
102
  @property
 
103
  def total_length(self):
104
  return len(self.batch_sampler)
 
105
  def __len__(self):
106
  if self.split_batches:
107
  # Split batches does not change the length of the batch sampler
@@ -119,8 +125,10 @@ class BatchSamplerShard(BatchSampler):
119
  else:
120
  # Otherwise it depends on the process index.
121
  return length + 1 if self.process_index < len(self.batch_sampler) % self.num_processes else length
 
122
  def __iter__(self):
123
  return self._iter_with_split() if self.split_batches else self._iter_with_no_split()
 
124
  def _iter_with_split(self):
125
  initial_data = []
126
  batch_length = self.batch_sampler.batch_size // self.num_processes
@@ -141,6 +149,7 @@ class BatchSamplerShard(BatchSampler):
141
  initial_data += initial_data
142
  batch = batch + initial_data
143
  yield batch[batch_length * self.process_index : batch_length * (self.process_index + 1)]
 
144
  def _iter_with_no_split(self):
145
  initial_data = []
146
  batch_to_yield = []
@@ -212,6 +221,7 @@ class IterableDatasetShard(IterableDataset):
212
  - the shard on process 0 to yield `[0, 1, 4, 5]` and the sampler on process 1 to yield `[2, 3, 6, 7]` if
213
  this argument is set to `True`.
214
  """
 
215
  def __init__(
216
  self,
217
  dataset: IterableDataset,
@@ -232,16 +242,19 @@ class IterableDatasetShard(IterableDataset):
232
  self.num_processes = num_processes
233
  self.process_index = process_index
234
  self.split_batches = split_batches
 
235
  def set_epoch(self, epoch):
236
  self.epoch = epoch
237
  if hasattr(self.dataset, "set_epoch"):
238
  self.dataset.set_epoch(epoch)
 
239
  def __len__(self):
240
  # We will just raise the downstream error if the underlying dataset is not sized
241
  if self.drop_last:
242
  return (len(self.dataset) // (self.batch_size * self.num_processes)) * self.batch_size
243
  else:
244
  return math.ceil(len(self.dataset) / (self.batch_size * self.num_processes)) * self.batch_size
 
245
  def __iter__(self):
246
  if (
247
  not hasattr(self.dataset, "set_epoch")
@@ -281,12 +294,15 @@ class DataLoaderStateMixin:
281
  - **remainder** (`int`) -- The number of items that are remaining in the last batch, relative to the total
282
  batch size
283
  """
 
284
  def __init_subclass__(cls, **kwargs):
285
  cls.end_of_dataloader = False
286
  cls.remainder = -1
 
287
  def reset(self):
288
  self.end_of_dataloader = False
289
  self.remainder = -1
 
290
  def begin(self):
291
  "Prepares the gradient state for the current dataloader"
292
  self.reset()
@@ -295,6 +311,7 @@ class DataLoaderStateMixin:
295
  length = getattr(self.dataset, "total_dataset_length", len(self.dataset))
296
  self.remainder = length % self.total_batch_size
297
  self.gradient_state._add_dataloader(self)
 
298
  def end(self):
299
  "Cleans up the gradient state after exiting the dataloader"
300
  self.gradient_state._remove_dataloader(self)
@@ -325,6 +342,7 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
325
  number of processes
326
  - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
327
  """
 
328
  def __init__(
329
  self,
330
  dataset,
@@ -343,6 +361,7 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
343
  self.gradient_state = GradientState()
344
  self._drop_last = _drop_last
345
  self.iteration = 0
 
346
  def __iter__(self):
347
  if self.rng_types is not None:
348
  synchronize_rng_states(self.rng_types, self.synchronized_generator)
@@ -372,6 +391,7 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
372
  break
373
  self.iteration += 1
374
  self.end()
 
375
  def set_epoch(self, epoch: int):
376
  # In case it is manually passed in, the user can set it to what they like
377
  if self.iteration != epoch:
@@ -383,6 +403,7 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
383
  elif hasattr(self.dataset, "set_epoch"):
384
  self.dataset.set_epoch(epoch)
385
  @property
 
386
  def total_batch_size(self):
387
  batch_sampler = self.sampler if isinstance(self.sampler, BatchSampler) else self.batch_sampler
388
  return (
@@ -391,6 +412,7 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
391
  else (batch_sampler.batch_size * getattr(batch_sampler, "num_processes", 1))
392
  )
393
  @property
 
394
  def total_dataset_length(self):
395
  if hasattr(self.dataset, "total_length"):
396
  return self.dataset.total_length
@@ -410,22 +432,27 @@ if is_tpu_available(check_device=False):
410
  number of processes
411
  - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
412
  """
413
- def __init__(self, dataloader: DataLoaderShard, device: torch.device):
 
414
  super().__init__(dataloader, device)
415
  self._rng_types = self._loader.rng_types
416
  self._loader.rng_types = None
417
- def __iter__(self):
 
418
  if self._rng_types is not None:
419
  synchronize_rng_states(self._rng_types, self._loader.synchronized_generator)
420
  return super().__iter__()
421
  @property
422
- def total_batch_size(self):
 
423
  return self._loader.total_batch_size
424
  @property
425
- def total_dataset_length(self):
 
426
  return self._loader.total_dataset_length
427
  @property
428
- def batch_sampler(self):
 
429
  return self._loader.batch_sampler
430
  class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
431
  """
@@ -447,6 +474,7 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
447
  number of processes
448
  - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
449
  """
 
450
  def __init__(
451
  self, dataset, split_batches: bool = False, skip_batches=0, _drop_last: bool = False, slice_fn=None, **kwargs
452
  ):
@@ -466,6 +494,7 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
466
  self.skip_batches = skip_batches
467
  self.slice_fn = slice_tensors if slice_fn is None else slice_fn
468
  self.iteration = 0
 
469
  def _fetch_batches(self, iterator):
470
  batches, batch = None, None
471
  # On process 0, we gather the batch to dispatch.
@@ -502,6 +531,7 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
502
  batch_info = [None, True]
503
  broadcast_object_list(batch_info)
504
  return batch, batch_info
 
505
  def __iter__(self):
506
  self.begin()
507
  self.set_epoch(self.iteration)
@@ -568,6 +598,7 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
568
  batch_index += 1
569
  self.iteration += 1
570
  self.end()
 
571
  def set_epoch(self, epoch: int):
572
  # In case it is manually passed in, the user can set it to what they like
573
  if self.iteration != epoch:
@@ -576,6 +607,7 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
576
  self.batch_sampler.sampler.set_epoch(epoch)
577
  elif hasattr(self.dataset, "set_epoch"):
578
  self.dataset.set_epoch(epoch)
 
579
  def __len__(self):
580
  whole_length = super().__len__()
581
  if self.split_batches:
@@ -585,11 +617,13 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
585
  else:
586
  return math.ceil(whole_length / self.state.num_processes)
587
  @property
 
588
  def total_batch_size(self):
589
  return (
590
  self.dataset.batch_size if self.split_batches else (self.dataset.batch_size * self.dataset.num_processes)
591
  )
592
  @property
 
593
  def total_dataset_length(self):
594
  return len(self.dataset)
595
  def prepare_data_loader(
@@ -783,16 +817,20 @@ class SkipBatchSampler(BatchSampler):
783
  """
784
  A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
785
  """
 
786
  def __init__(self, batch_sampler, skip_batches=0):
787
  self.batch_sampler = batch_sampler
788
  self.skip_batches = skip_batches
 
789
  def __iter__(self):
790
  for index, samples in enumerate(self.batch_sampler):
791
  if index >= self.skip_batches:
792
  yield samples
793
  @property
 
794
  def total_length(self):
795
  return len(self.batch_sampler)
 
796
  def __len__(self):
797
  return len(self.batch_sampler) - self.skip_batches
798
  class SkipDataLoader(DataLoader):
@@ -806,9 +844,11 @@ class SkipDataLoader(DataLoader):
806
  kwargs:
807
  All other keyword arguments to pass to the regular `DataLoader` initialization.
808
  """
 
809
  def __init__(self, dataset, skip_batches=0, **kwargs):
810
  super().__init__(dataset, **kwargs)
811
  self.skip_batches = skip_batches
 
812
  def __iter__(self):
813
  for index, batch in enumerate(super().__iter__()):
814
  if index >= self.skip_batches:
 
29
  If a custom `generator` is passed, it will rely on its initial seed as well as the current iteration it is on
30
  (stored in `self.epoch`).
31
  """
32
+
33
  def __init__(self, *args, **kwargs):
34
  super().__init__(*args, **kwargs)
35
  self.epoch = 0
36
  self.seed = torch.random.initial_seed()
37
+
38
  def __iter__(self):
39
  if self.generator is None:
40
  self.generator = torch.Generator()
 
45
  self.generator.manual_seed(seed)
46
  yield from super().__iter__()
47
  self.set_epoch(self.epoch + 1)
48
+
49
  def set_epoch(self, epoch: int):
50
  "Sets the current iteration of the sampler."
51
  self.epoch = epoch
 
77
  `BatchSampler`s with varying batch sizes are not enabled by default. To enable this behaviour, set `even_batches`
78
  equal to `False`
79
  </Tip>"""
80
+
81
  def __init__(
82
  self,
83
  batch_sampler: BatchSampler,
 
104
  "are not calling this method directly, set `accelerator.even_batches=False` instead."
105
  )
106
  @property
107
+
108
  def total_length(self):
109
  return len(self.batch_sampler)
110
+
111
  def __len__(self):
112
  if self.split_batches:
113
  # Split batches does not change the length of the batch sampler
 
125
  else:
126
  # Otherwise it depends on the process index.
127
  return length + 1 if self.process_index < len(self.batch_sampler) % self.num_processes else length
128
+
129
  def __iter__(self):
130
  return self._iter_with_split() if self.split_batches else self._iter_with_no_split()
131
+
132
  def _iter_with_split(self):
133
  initial_data = []
134
  batch_length = self.batch_sampler.batch_size // self.num_processes
 
149
  initial_data += initial_data
150
  batch = batch + initial_data
151
  yield batch[batch_length * self.process_index : batch_length * (self.process_index + 1)]
152
+
153
  def _iter_with_no_split(self):
154
  initial_data = []
155
  batch_to_yield = []
 
221
  - the shard on process 0 to yield `[0, 1, 4, 5]` and the sampler on process 1 to yield `[2, 3, 6, 7]` if
222
  this argument is set to `True`.
223
  """
224
+
225
  def __init__(
226
  self,
227
  dataset: IterableDataset,
 
242
  self.num_processes = num_processes
243
  self.process_index = process_index
244
  self.split_batches = split_batches
245
+
246
  def set_epoch(self, epoch):
247
  self.epoch = epoch
248
  if hasattr(self.dataset, "set_epoch"):
249
  self.dataset.set_epoch(epoch)
250
+
251
  def __len__(self):
252
  # We will just raise the downstream error if the underlying dataset is not sized
253
  if self.drop_last:
254
  return (len(self.dataset) // (self.batch_size * self.num_processes)) * self.batch_size
255
  else:
256
  return math.ceil(len(self.dataset) / (self.batch_size * self.num_processes)) * self.batch_size
257
+
258
  def __iter__(self):
259
  if (
260
  not hasattr(self.dataset, "set_epoch")
 
294
  - **remainder** (`int`) -- The number of items that are remaining in the last batch, relative to the total
295
  batch size
296
  """
297
+
298
  def __init_subclass__(cls, **kwargs):
299
  cls.end_of_dataloader = False
300
  cls.remainder = -1
301
+
302
  def reset(self):
303
  self.end_of_dataloader = False
304
  self.remainder = -1
305
+
306
  def begin(self):
307
  "Prepares the gradient state for the current dataloader"
308
  self.reset()
 
311
  length = getattr(self.dataset, "total_dataset_length", len(self.dataset))
312
  self.remainder = length % self.total_batch_size
313
  self.gradient_state._add_dataloader(self)
314
+
315
  def end(self):
316
  "Cleans up the gradient state after exiting the dataloader"
317
  self.gradient_state._remove_dataloader(self)
 
342
  number of processes
343
  - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
344
  """
345
+
346
  def __init__(
347
  self,
348
  dataset,
 
361
  self.gradient_state = GradientState()
362
  self._drop_last = _drop_last
363
  self.iteration = 0
364
+
365
  def __iter__(self):
366
  if self.rng_types is not None:
367
  synchronize_rng_states(self.rng_types, self.synchronized_generator)
 
391
  break
392
  self.iteration += 1
393
  self.end()
394
+
395
  def set_epoch(self, epoch: int):
396
  # In case it is manually passed in, the user can set it to what they like
397
  if self.iteration != epoch:
 
403
  elif hasattr(self.dataset, "set_epoch"):
404
  self.dataset.set_epoch(epoch)
405
  @property
406
+
407
  def total_batch_size(self):
408
  batch_sampler = self.sampler if isinstance(self.sampler, BatchSampler) else self.batch_sampler
409
  return (
 
412
  else (batch_sampler.batch_size * getattr(batch_sampler, "num_processes", 1))
413
  )
414
  @property
415
+
416
  def total_dataset_length(self):
417
  if hasattr(self.dataset, "total_length"):
418
  return self.dataset.total_length
 
432
  number of processes
433
  - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
434
  """
435
+
436
+ def __init__(self, dataloader: DataLoaderShard, device: torch.device):
437
  super().__init__(dataloader, device)
438
  self._rng_types = self._loader.rng_types
439
  self._loader.rng_types = None
440
+
441
+ def __iter__(self):
442
  if self._rng_types is not None:
443
  synchronize_rng_states(self._rng_types, self._loader.synchronized_generator)
444
  return super().__iter__()
445
  @property
446
+
447
+ def total_batch_size(self):
448
  return self._loader.total_batch_size
449
  @property
450
+
451
+ def total_dataset_length(self):
452
  return self._loader.total_dataset_length
453
  @property
454
+
455
+ def batch_sampler(self):
456
  return self._loader.batch_sampler
457
  class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
458
  """
 
474
  number of processes
475
  - **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
476
  """
477
+
478
  def __init__(
479
  self, dataset, split_batches: bool = False, skip_batches=0, _drop_last: bool = False, slice_fn=None, **kwargs
480
  ):
 
494
  self.skip_batches = skip_batches
495
  self.slice_fn = slice_tensors if slice_fn is None else slice_fn
496
  self.iteration = 0
497
+
498
  def _fetch_batches(self, iterator):
499
  batches, batch = None, None
500
  # On process 0, we gather the batch to dispatch.
 
531
  batch_info = [None, True]
532
  broadcast_object_list(batch_info)
533
  return batch, batch_info
534
+
535
  def __iter__(self):
536
  self.begin()
537
  self.set_epoch(self.iteration)
 
598
  batch_index += 1
599
  self.iteration += 1
600
  self.end()
601
+
602
  def set_epoch(self, epoch: int):
603
  # In case it is manually passed in, the user can set it to what they like
604
  if self.iteration != epoch:
 
607
  self.batch_sampler.sampler.set_epoch(epoch)
608
  elif hasattr(self.dataset, "set_epoch"):
609
  self.dataset.set_epoch(epoch)
610
+
611
  def __len__(self):
612
  whole_length = super().__len__()
613
  if self.split_batches:
 
617
  else:
618
  return math.ceil(whole_length / self.state.num_processes)
619
  @property
620
+
621
  def total_batch_size(self):
622
  return (
623
  self.dataset.batch_size if self.split_batches else (self.dataset.batch_size * self.dataset.num_processes)
624
  )
625
  @property
626
+
627
  def total_dataset_length(self):
628
  return len(self.dataset)
629
  def prepare_data_loader(
 
817
  """
818
  A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
819
  """
820
+
821
  def __init__(self, batch_sampler, skip_batches=0):
822
  self.batch_sampler = batch_sampler
823
  self.skip_batches = skip_batches
824
+
825
  def __iter__(self):
826
  for index, samples in enumerate(self.batch_sampler):
827
  if index >= self.skip_batches:
828
  yield samples
829
  @property
830
+
831
  def total_length(self):
832
  return len(self.batch_sampler)
833
+
834
  def __len__(self):
835
  return len(self.batch_sampler) - self.skip_batches
836
  class SkipDataLoader(DataLoader):
 
844
  kwargs:
845
  All other keyword arguments to pass to the regular `DataLoader` initialization.
846
  """
847
+
848
  def __init__(self, dataset, skip_batches=0, **kwargs):
849
  super().__init__(dataset, **kwargs)
850
  self.skip_batches = skip_batches
851
+
852
  def __iter__(self):
853
  for index, batch in enumerate(super().__iter__()):
854
  if index >= self.skip_batches:
src/hooks.py CHANGED
@@ -7,6 +7,7 @@ class ModelHook:
7
  the `torch.no_grad()` context manager.
8
  """
9
  no_grad = False
 
10
  def init_hook(self, module):
11
  """
12
  To be executed when the hook is attached to the module.
@@ -14,6 +15,7 @@ class ModelHook:
14
  module (`torch.nn.Module`): The module attached to this hook.
15
  """
16
  return module
 
17
  def pre_forward(self, module, *args, **kwargs):
18
  """
19
  To be executed just before the forward method of the model.
@@ -25,6 +27,7 @@ class ModelHook:
25
  `Tuple[Tuple[Any], Dict[Str, Any]]`: A tuple with the treated `args` and `kwargs`.
26
  """
27
  return args, kwargs
 
28
  def post_forward(self, module, output):
29
  """
30
  To be executed just after the forward method of the model.
@@ -35,6 +38,7 @@ class ModelHook:
35
  `Any`: The processed `output`.
36
  """
37
  return output
 
38
  def detach_hook(self, module):
39
  """
40
  To be executed when the hook is detached from a module.
@@ -46,20 +50,25 @@ class SequentialHook(ModelHook):
46
  """
47
  A hook that can contain several hooks and iterates through them at each event.
48
  """
 
49
  def __init__(self, *hooks):
50
  self.hooks = hooks
 
51
  def init_hook(self, module):
52
  for hook in self.hooks:
53
  module = hook.init_hook(module)
54
  return module
 
55
  def pre_forward(self, module, *args, **kwargs):
56
  for hook in self.hooks:
57
  args, kwargs = hook.pre_forward(module, *args, **kwargs)
58
  return args, kwargs
 
59
  def post_forward(self, module, output):
60
  for hook in self.hooks:
61
  output = hook.post_forward(module, output)
62
  return output
 
63
  def detach_hook(self, module):
64
  for hook in self.hooks:
65
  module = hook.detach_hook(module)
@@ -95,6 +104,7 @@ def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False)
95
  module._old_forward = old_forward
96
  module = hook.init_hook(module)
97
  module._hf_hook = hook
 
98
  def new_forward(module, *args, **kwargs):
99
  args, kwargs = module._hf_hook.pre_forward(module, *args, **kwargs)
100
  if module._hf_hook.no_grad:
@@ -143,6 +153,7 @@ class AlignDevicesHook(ModelHook):
143
  place_submodules (`bool`, *optional*, defaults to `False`):
144
  Whether to place the submodules on `execution_device` during the `init_hook` event.
145
  """
 
146
  def __init__(
147
  self,
148
  execution_device: Optional[Union[int, str, torch.device]] = None,
@@ -164,12 +175,14 @@ class AlignDevicesHook(ModelHook):
164
  self.input_device = None
165
  self.param_original_devices = {}
166
  self.buffer_original_devices = {}
 
167
  def __repr__(self):
168
  return (
169
  f"AlignDevicesHook(execution_device={self.execution_device}, offload={self.offload}, "
170
  f"io_same_device={self.io_same_device}, offload_buffers={self.offload_buffers}, "
171
  f"place_submodules={self.place_submodules}, skip_keys={repr(self.skip_keys)})"
172
  )
 
173
  def init_hook(self, module):
174
  if not self.offload and self.execution_device is not None:
175
  for name, _ in named_module_tensors(module, recurse=self.place_submodules):
@@ -196,6 +209,7 @@ class AlignDevicesHook(ModelHook):
196
  for name in get_non_persistent_buffers(module, recurse=self.place_submodules):
197
  set_module_tensor_to_device(module, name, self.execution_device)
198
  return module
 
199
  def pre_forward(self, module, *args, **kwargs):
200
  if self.io_same_device:
201
  self.input_device = find_device([args, kwargs])
@@ -216,6 +230,7 @@ class AlignDevicesHook(ModelHook):
216
  return send_to_device(args, self.execution_device), send_to_device(
217
  kwargs, self.execution_device, skip_keys=self.skip_keys
218
  )
 
219
  def post_forward(self, module, output):
220
  if self.offload:
221
  for name, _ in named_module_tensors(
@@ -231,6 +246,7 @@ class AlignDevicesHook(ModelHook):
231
  if self.io_same_device and self.input_device is not None:
232
  output = send_to_device(output, self.input_device, skip_keys=self.skip_keys)
233
  return output
 
234
  def detach_hook(self, module):
235
  if self.offload:
236
  for name, device in self.original_devices.items():
@@ -462,6 +478,7 @@ class CpuOffload(ModelHook):
462
  passed, its offload method will be called just before the forward of the model to which this hook is
463
  attached.
464
  """
 
465
  def __init__(
466
  self,
467
  execution_device: Optional[Union[str, int, torch.device]] = None,
@@ -469,8 +486,10 @@ class CpuOffload(ModelHook):
469
  ):
470
  self.prev_module_hook = prev_module_hook
471
  self.execution_device = execution_device if execution_device is not None else PartialState().default_device
 
472
  def init_hook(self, module):
473
  return module.to("cpu")
 
474
  def pre_forward(self, module, *args, **kwargs):
475
  if self.prev_module_hook is not None:
476
  self.prev_module_hook.offload()
@@ -481,10 +500,13 @@ class UserCpuOffloadHook:
481
  A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
482
  or remove it entirely.
483
  """
 
484
  def __init__(self, model, hook):
485
  self.model = model
486
  self.hook = hook
 
487
  def offload(self):
488
  self.hook.init_hook(self.model)
 
489
  def remove(self):
490
  remove_hook_from_module(self.model)
 
7
  the `torch.no_grad()` context manager.
8
  """
9
  no_grad = False
10
+
11
  def init_hook(self, module):
12
  """
13
  To be executed when the hook is attached to the module.
 
15
  module (`torch.nn.Module`): The module attached to this hook.
16
  """
17
  return module
18
+
19
  def pre_forward(self, module, *args, **kwargs):
20
  """
21
  To be executed just before the forward method of the model.
 
27
  `Tuple[Tuple[Any], Dict[Str, Any]]`: A tuple with the treated `args` and `kwargs`.
28
  """
29
  return args, kwargs
30
+
31
  def post_forward(self, module, output):
32
  """
33
  To be executed just after the forward method of the model.
 
38
  `Any`: The processed `output`.
39
  """
40
  return output
41
+
42
  def detach_hook(self, module):
43
  """
44
  To be executed when the hook is detached from a module.
 
50
  """
51
  A hook that can contain several hooks and iterates through them at each event.
52
  """
53
+
54
  def __init__(self, *hooks):
55
  self.hooks = hooks
56
+
57
  def init_hook(self, module):
58
  for hook in self.hooks:
59
  module = hook.init_hook(module)
60
  return module
61
+
62
  def pre_forward(self, module, *args, **kwargs):
63
  for hook in self.hooks:
64
  args, kwargs = hook.pre_forward(module, *args, **kwargs)
65
  return args, kwargs
66
+
67
  def post_forward(self, module, output):
68
  for hook in self.hooks:
69
  output = hook.post_forward(module, output)
70
  return output
71
+
72
  def detach_hook(self, module):
73
  for hook in self.hooks:
74
  module = hook.detach_hook(module)
 
104
  module._old_forward = old_forward
105
  module = hook.init_hook(module)
106
  module._hf_hook = hook
107
+
108
  def new_forward(module, *args, **kwargs):
109
  args, kwargs = module._hf_hook.pre_forward(module, *args, **kwargs)
110
  if module._hf_hook.no_grad:
 
153
  place_submodules (`bool`, *optional*, defaults to `False`):
154
  Whether to place the submodules on `execution_device` during the `init_hook` event.
155
  """
156
+
157
  def __init__(
158
  self,
159
  execution_device: Optional[Union[int, str, torch.device]] = None,
 
175
  self.input_device = None
176
  self.param_original_devices = {}
177
  self.buffer_original_devices = {}
178
+
179
  def __repr__(self):
180
  return (
181
  f"AlignDevicesHook(execution_device={self.execution_device}, offload={self.offload}, "
182
  f"io_same_device={self.io_same_device}, offload_buffers={self.offload_buffers}, "
183
  f"place_submodules={self.place_submodules}, skip_keys={repr(self.skip_keys)})"
184
  )
185
+
186
  def init_hook(self, module):
187
  if not self.offload and self.execution_device is not None:
188
  for name, _ in named_module_tensors(module, recurse=self.place_submodules):
 
209
  for name in get_non_persistent_buffers(module, recurse=self.place_submodules):
210
  set_module_tensor_to_device(module, name, self.execution_device)
211
  return module
212
+
213
  def pre_forward(self, module, *args, **kwargs):
214
  if self.io_same_device:
215
  self.input_device = find_device([args, kwargs])
 
230
  return send_to_device(args, self.execution_device), send_to_device(
231
  kwargs, self.execution_device, skip_keys=self.skip_keys
232
  )
233
+
234
  def post_forward(self, module, output):
235
  if self.offload:
236
  for name, _ in named_module_tensors(
 
246
  if self.io_same_device and self.input_device is not None:
247
  output = send_to_device(output, self.input_device, skip_keys=self.skip_keys)
248
  return output
249
+
250
  def detach_hook(self, module):
251
  if self.offload:
252
  for name, device in self.original_devices.items():
 
478
  passed, its offload method will be called just before the forward of the model to which this hook is
479
  attached.
480
  """
481
+
482
  def __init__(
483
  self,
484
  execution_device: Optional[Union[str, int, torch.device]] = None,
 
486
  ):
487
  self.prev_module_hook = prev_module_hook
488
  self.execution_device = execution_device if execution_device is not None else PartialState().default_device
489
+
490
  def init_hook(self, module):
491
  return module.to("cpu")
492
+
493
  def pre_forward(self, module, *args, **kwargs):
494
  if self.prev_module_hook is not None:
495
  self.prev_module_hook.offload()
 
500
  A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
501
  or remove it entirely.
502
  """
503
+
504
  def __init__(self, model, hook):
505
  self.model = model
506
  self.hook = hook
507
+
508
  def offload(self):
509
  self.hook.init_hook(self.model)
510
+
511
  def remove(self):
512
  remove_hook_from_module(self.model)
src/launchers.py CHANGED
@@ -43,6 +43,7 @@ def notebook_launcher(
43
  ```python
44
  # Assume this is defined in a Jupyter Notebook on an instance with two GPUs
45
  from accelerate import notebook_launcher
 
46
  def train(*args):
47
  # Your training function here
48
  ...
 
43
  ```python
44
  # Assume this is defined in a Jupyter Notebook on an instance with two GPUs
45
  from accelerate import notebook_launcher
46
+
47
  def train(*args):
48
  # Your training function here
49
  ...
src/local_sgd.py CHANGED
@@ -12,16 +12,19 @@ class LocalSGD:
12
  Stich, Sebastian Urban. ["Local SGD Converges Fast and Communicates Little." ICLR 2019-International Conference on
13
  Learning Representations. No. CONF. 2019.](https://arxiv.org/abs/1805.09767)
14
  """
 
15
  def __enter__(self):
16
  if self.enabled:
17
  self.model_sync_obj = self.model.no_sync()
18
  self.model_sync_obj.__enter__()
19
  return self
 
20
  def __exit__(self, type, value, tb):
21
  if self.enabled:
22
  # Average all models on exit
23
  self._sync_and_avg_model_params()
24
  self.model_sync_obj.__exit__(type, value, tb)
 
25
  def __init__(self, accelerator: Accelerator, model: torch.nn.Module, local_sgd_steps: int, enabled: bool = True):
26
  """
27
  Constructor.
@@ -47,6 +50,7 @@ class LocalSGD:
47
  self.accelerator = accelerator
48
  self.model = model
49
  self.local_sgd_steps = local_sgd_steps
 
50
  def step(self):
51
  """
52
  This function makes a "step" and synchronizes model parameters if necessary.
@@ -56,6 +60,7 @@ class LocalSGD:
56
  return
57
  if self.num_steps % self.local_sgd_steps == 0:
58
  self._sync_and_avg_model_params()
 
59
  def _sync_and_avg_model_params(self):
60
  """
61
  Synchronize + Average model parameters across all GPUs
 
12
  Stich, Sebastian Urban. ["Local SGD Converges Fast and Communicates Little." ICLR 2019-International Conference on
13
  Learning Representations. No. CONF. 2019.](https://arxiv.org/abs/1805.09767)
14
  """
15
+
16
  def __enter__(self):
17
  if self.enabled:
18
  self.model_sync_obj = self.model.no_sync()
19
  self.model_sync_obj.__enter__()
20
  return self
21
+
22
  def __exit__(self, type, value, tb):
23
  if self.enabled:
24
  # Average all models on exit
25
  self._sync_and_avg_model_params()
26
  self.model_sync_obj.__exit__(type, value, tb)
27
+
28
  def __init__(self, accelerator: Accelerator, model: torch.nn.Module, local_sgd_steps: int, enabled: bool = True):
29
  """
30
  Constructor.
 
50
  self.accelerator = accelerator
51
  self.model = model
52
  self.local_sgd_steps = local_sgd_steps
53
+
54
  def step(self):
55
  """
56
  This function makes a "step" and synchronizes model parameters if necessary.
 
60
  return
61
  if self.num_steps % self.local_sgd_steps == 0:
62
  self._sync_and_avg_model_params()
63
+
64
  def _sync_and_avg_model_params(self):
65
  """
66
  Synchronize + Average model parameters across all GPUs
src/logging.py CHANGED
@@ -6,10 +6,12 @@ class MultiProcessAdapter(logging.LoggerAdapter):
6
  Does not require an `Accelerator` object to be created first.
7
  """
8
  @staticmethod
 
9
  def _should_log(main_process_only):
10
  "Check if log should be performed"
11
  state = PartialState()
12
  return not main_process_only or (main_process_only and state.is_main_process)
 
13
  def log(self, level, msg, *args, **kwargs):
14
  """
15
  Delegates logger call after checking if we should log.
@@ -38,6 +40,7 @@ class MultiProcessAdapter(logging.LoggerAdapter):
38
  self.logger.log(level, msg, *args, **kwargs)
39
  state.wait_for_everyone()
40
  @functools.lru_cache(None)
 
41
  def warning_once(self, *args, **kwargs):
42
  """
43
  This method is identical to `logger.warning()`, but will emit the warning with the same message only once
 
6
  Does not require an `Accelerator` object to be created first.
7
  """
8
  @staticmethod
9
+
10
  def _should_log(main_process_only):
11
  "Check if log should be performed"
12
  state = PartialState()
13
  return not main_process_only or (main_process_only and state.is_main_process)
14
+
15
  def log(self, level, msg, *args, **kwargs):
16
  """
17
  Delegates logger call after checking if we should log.
 
40
  self.logger.log(level, msg, *args, **kwargs)
41
  state.wait_for_everyone()
42
  @functools.lru_cache(None)
43
+
44
  def warning_once(self, *args, **kwargs):
45
  """
46
  This method is identical to `logger.warning()`, but will emit the warning with the same message only once
src/optimizer.py CHANGED
@@ -20,6 +20,7 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
20
  scaler (`torch.cuda.amp.grad_scaler.GradScaler`, *optional*):
21
  The scaler to use in the step function if training with mixed precision.
22
  """
 
23
  def __init__(self, optimizer, device_placement=True, scaler=None):
24
  self.optimizer = optimizer
25
  self.scaler = scaler
@@ -40,31 +41,41 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
40
  state_dict = move_to_device(state_dict, self.accelerator_state.device)
41
  self.optimizer.load_state_dict(state_dict)
42
  @property
 
43
  def state(self):
44
  return self.optimizer.state
45
  @state.setter
 
46
  def state(self, state):
47
  self.optimizer.state = state
48
  @property
 
49
  def param_groups(self):
50
  return self.optimizer.param_groups
51
  @param_groups.setter
 
52
  def param_groups(self, param_groups):
53
  self.optimizer.param_groups = param_groups
54
  @property
 
55
  def defaults(self):
56
  return self.optimizer.defaults
57
  @defaults.setter
 
58
  def defaults(self, defaults):
59
  self.optimizer.defaults = defaults
 
60
  def add_param_group(self, param_group):
61
  self.optimizer.add_param_group(param_group)
 
62
  def load_state_dict(self, state_dict):
63
  if self.accelerator_state.distributed_type == DistributedType.TPU and self.device_placement:
64
  xm.send_cpu_data_to_device(state_dict, self.accelerator_state.device)
65
  self.optimizer.load_state_dict(state_dict)
 
66
  def state_dict(self):
67
  return self.optimizer.state_dict()
 
68
  def zero_grad(self, set_to_none=None):
69
  if self.gradient_state.sync_gradients:
70
  accept_arg = "set_to_none" in inspect.signature(self.optimizer.zero_grad).parameters
@@ -76,6 +87,7 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
76
  if set_to_none is not None:
77
  raise ValueError("`set_to_none` for Optimizer.zero_grad` is not supported by this optimizer.")
78
  self.optimizer.zero_grad()
 
79
  def step(self, closure=None):
80
  if self.gradient_state.sync_gradients:
81
  if self.accelerator_state.distributed_type == DistributedType.TPU:
@@ -96,10 +108,12 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
96
  self._accelerate_step_called = False
97
  else:
98
  self.optimizer.step(closure)
 
99
  def _switch_parameters(self, parameters_map):
100
  for param_group in self.optimizer.param_groups:
101
  param_group["params"] = [parameters_map.get(p, p) for p in param_group["params"]]
102
  @property
 
103
  def is_overflow(self):
104
  """Whether or not the optimizer step was done, or skipped because of gradient overflow."""
105
  warnings.warn(
@@ -109,9 +123,11 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
109
  )
110
  return self._is_overflow
111
  @property
 
112
  def step_was_skipped(self):
113
  """Whether or not the optimizer step was skipped."""
114
  return self._is_overflow
 
115
  def __getstate__(self):
116
  _ignored_keys = [
117
  "_accelerate_step_called",
@@ -119,6 +135,7 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
119
  "_optimizer_patched_step_method",
120
  ]
121
  return {k: v for k, v in self.__dict__.items() if k not in _ignored_keys}
 
122
  def __setstate__(self, state):
123
  self.__dict__.update(state)
124
  if self.scaler is not None:
@@ -126,6 +143,7 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
126
  self._optimizer_original_step_method = self.optimizer.step
127
  self._optimizer_patched_step_method = patch_optimizer_step(self, self.optimizer.step)
128
  def patch_optimizer_step(accelerated_optimizer: AcceleratedOptimizer, method):
 
129
  def patched_step(*args, **kwargs):
130
  accelerated_optimizer._accelerate_step_called = True
131
  return method(*args, **kwargs)
 
20
  scaler (`torch.cuda.amp.grad_scaler.GradScaler`, *optional*):
21
  The scaler to use in the step function if training with mixed precision.
22
  """
23
+
24
  def __init__(self, optimizer, device_placement=True, scaler=None):
25
  self.optimizer = optimizer
26
  self.scaler = scaler
 
41
  state_dict = move_to_device(state_dict, self.accelerator_state.device)
42
  self.optimizer.load_state_dict(state_dict)
43
  @property
44
+
45
  def state(self):
46
  return self.optimizer.state
47
  @state.setter
48
+
49
  def state(self, state):
50
  self.optimizer.state = state
51
  @property
52
+
53
  def param_groups(self):
54
  return self.optimizer.param_groups
55
  @param_groups.setter
56
+
57
  def param_groups(self, param_groups):
58
  self.optimizer.param_groups = param_groups
59
  @property
60
+
61
  def defaults(self):
62
  return self.optimizer.defaults
63
  @defaults.setter
64
+
65
  def defaults(self, defaults):
66
  self.optimizer.defaults = defaults
67
+
68
  def add_param_group(self, param_group):
69
  self.optimizer.add_param_group(param_group)
70
+
71
  def load_state_dict(self, state_dict):
72
  if self.accelerator_state.distributed_type == DistributedType.TPU and self.device_placement:
73
  xm.send_cpu_data_to_device(state_dict, self.accelerator_state.device)
74
  self.optimizer.load_state_dict(state_dict)
75
+
76
  def state_dict(self):
77
  return self.optimizer.state_dict()
78
+
79
  def zero_grad(self, set_to_none=None):
80
  if self.gradient_state.sync_gradients:
81
  accept_arg = "set_to_none" in inspect.signature(self.optimizer.zero_grad).parameters
 
87
  if set_to_none is not None:
88
  raise ValueError("`set_to_none` for Optimizer.zero_grad` is not supported by this optimizer.")
89
  self.optimizer.zero_grad()
90
+
91
  def step(self, closure=None):
92
  if self.gradient_state.sync_gradients:
93
  if self.accelerator_state.distributed_type == DistributedType.TPU:
 
108
  self._accelerate_step_called = False
109
  else:
110
  self.optimizer.step(closure)
111
+
112
  def _switch_parameters(self, parameters_map):
113
  for param_group in self.optimizer.param_groups:
114
  param_group["params"] = [parameters_map.get(p, p) for p in param_group["params"]]
115
  @property
116
+
117
  def is_overflow(self):
118
  """Whether or not the optimizer step was done, or skipped because of gradient overflow."""
119
  warnings.warn(
 
123
  )
124
  return self._is_overflow
125
  @property
126
+
127
  def step_was_skipped(self):
128
  """Whether or not the optimizer step was skipped."""
129
  return self._is_overflow
130
+
131
  def __getstate__(self):
132
  _ignored_keys = [
133
  "_accelerate_step_called",
 
135
  "_optimizer_patched_step_method",
136
  ]
137
  return {k: v for k, v in self.__dict__.items() if k not in _ignored_keys}
138
+
139
  def __setstate__(self, state):
140
  self.__dict__.update(state)
141
  if self.scaler is not None:
 
143
  self._optimizer_original_step_method = self.optimizer.step
144
  self._optimizer_patched_step_method = patch_optimizer_step(self, self.optimizer.step)
145
  def patch_optimizer_step(accelerated_optimizer: AcceleratedOptimizer, method):
146
+
147
  def patched_step(*args, **kwargs):
148
  accelerated_optimizer._accelerate_step_called = True
149
  return method(*args, **kwargs)
src/scheduler.py CHANGED
@@ -18,12 +18,14 @@ class AcceleratedScheduler:
18
  regardless of the number of processes) or create batches on each process (so batch size is the original
19
  batch size multiplied by the number of processes).
20
  """
 
21
  def __init__(self, scheduler, optimizers, step_with_optimizer: bool = True, split_batches: bool = False):
22
  self.scheduler = scheduler
23
  self.optimizers = optimizers if isinstance(optimizers, (list, tuple)) else [optimizers]
24
  self.split_batches = split_batches
25
  self.step_with_optimizer = step_with_optimizer
26
  self.gradient_state = GradientState()
 
27
  def step(self, *args, **kwargs):
28
  if not self.step_with_optimizer:
29
  # No link between scheduler and optimizer -> just step
@@ -52,13 +54,18 @@ class AcceleratedScheduler:
52
  else:
53
  self.scheduler.step(*args, **kwargs)
54
  # Passthroughs
 
55
  def get_last_lr(self):
56
  return self.scheduler.get_last_lr()
 
57
  def state_dict(self):
58
  return self.scheduler.state_dict()
 
59
  def load_state_dict(self, state_dict):
60
  self.scheduler.load_state_dict(state_dict)
 
61
  def get_lr(self):
62
  return self.scheduler.get_lr()
 
63
  def print_lr(self, *args, **kwargs):
64
  return self.scheduler.print_lr(*args, **kwargs)
 
18
  regardless of the number of processes) or create batches on each process (so batch size is the original
19
  batch size multiplied by the number of processes).
20
  """
21
+
22
  def __init__(self, scheduler, optimizers, step_with_optimizer: bool = True, split_batches: bool = False):
23
  self.scheduler = scheduler
24
  self.optimizers = optimizers if isinstance(optimizers, (list, tuple)) else [optimizers]
25
  self.split_batches = split_batches
26
  self.step_with_optimizer = step_with_optimizer
27
  self.gradient_state = GradientState()
28
+
29
  def step(self, *args, **kwargs):
30
  if not self.step_with_optimizer:
31
  # No link between scheduler and optimizer -> just step
 
54
  else:
55
  self.scheduler.step(*args, **kwargs)
56
  # Passthroughs
57
+
58
  def get_last_lr(self):
59
  return self.scheduler.get_last_lr()
60
+
61
  def state_dict(self):
62
  return self.scheduler.state_dict()
63
+
64
  def load_state_dict(self, state_dict):
65
  self.scheduler.load_state_dict(state_dict)
66
+
67
  def get_lr(self):
68
  return self.scheduler.get_lr()
69
+
70
  def print_lr(self, *args, **kwargs):
71
  return self.scheduler.print_lr(*args, **kwargs)
src/state.py CHANGED
@@ -20,10 +20,13 @@ class ThreadLocalSharedDict(threading.local):
20
  This is required for using PyTorch/XLA with PJRT in multithreaded mode (required for TPU v2 and v3).
21
  See https://github.com/pytorch/xla/blob/r2.0/docs/pjrt.md#multithreading-on-tpu-v2v3
22
  """
 
23
  def __init__(self, thread_local: bool = False):
24
  self._storage = {}
 
25
  def __get__(self, obj, objtype=None):
26
  return self._storage
 
27
  def __set__(self, obj, value):
28
  self._storage = value
29
  # Prefer global shared dictionary, except when using TPU.
@@ -49,6 +52,7 @@ class PartialState:
49
  - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
50
  """
51
  _shared_state = SharedDict()
 
52
  def __init__(self, cpu: bool = False, **kwargs):
53
  self.__dict__ = self._shared_state
54
  if not self.initialized:
@@ -240,6 +244,7 @@ class PartialState:
240
  if self.device is None:
241
  self.device = torch.device("cpu") if cpu else self.default_device
242
  self.fork_launched = parse_flag_from_env("FORK_LAUNCHED", 0)
 
243
  def __repr__(self) -> str:
244
  return (
245
  f"Distributed environment: {self.distributed_type}{(' Backend: ' + self.backend) if self.backend else ''}\n"
@@ -249,30 +254,36 @@ class PartialState:
249
  f"Device: {self.device}\n"
250
  )
251
  @staticmethod
 
252
  def _reset_state():
253
  "Resets `_shared_state`, is used internally and should not be called"
254
  PartialState._shared_state.clear()
255
  @property
 
256
  def initialized(self) -> bool:
257
  "Returns whether the `PartialState` has been initialized"
258
  return self._shared_state != {}
259
  @property
 
260
  def use_distributed(self):
261
  """
262
  Whether the Accelerator is configured for distributed training
263
  """
264
  return self.distributed_type != DistributedType.NO and self.num_processes > 1
265
  @property
 
266
  def is_last_process(self) -> bool:
267
  "Returns whether the current process is the last one"
268
  return self.process_index == self.num_processes - 1
269
  @property
 
270
  def is_main_process(self) -> bool:
271
  "Returns whether the current process is the main process"
272
  return (
273
  self.process_index == 0 if self.distributed_type != DistributedType.MEGATRON_LM else self.is_last_process
274
  )
275
  @property
 
276
  def is_local_main_process(self) -> bool:
277
  "Returns whether the current process is the main process on the local node"
278
  return (
@@ -280,6 +291,7 @@ class PartialState:
280
  if self.distributed_type != DistributedType.MEGATRON_LM
281
  else self.is_last_process
282
  )
 
283
  def wait_for_everyone(self):
284
  """
285
  Will stop the execution of the current process until every other process has reached that point (so this does
@@ -310,6 +322,7 @@ class PartialState:
310
  torch.distributed.barrier()
311
  elif self.distributed_type == DistributedType.TPU:
312
  xm.rendezvous("accelerate.utils.wait_for_everyone")
 
313
  def _goes_first(self, is_main: bool):
314
  if not is_main:
315
  self.wait_for_everyone()
@@ -317,6 +330,7 @@ class PartialState:
317
  if is_main:
318
  self.wait_for_everyone()
319
  @contextmanager
 
320
  def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
321
  """
322
  Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
@@ -362,7 +376,8 @@ class PartialState:
362
  end_index = start_index + num_samples_per_process
363
  if (len(inputs) % self.num_processes != 0) and (self.process_index == self.num_processes - 1):
364
  end_index = length
365
- def _split_values(inputs, start_index, end_index):
 
366
  if isinstance(inputs, (list, tuple, torch.Tensor)):
367
  if start_index >= len(inputs):
368
  result = inputs[-1:]
@@ -385,6 +400,7 @@ class PartialState:
385
  return inputs
386
  yield _split_values(inputs, start_index, end_index)
387
  @contextmanager
 
388
  def main_process_first(self):
389
  """
390
  Lets the main process go first inside a with block.
@@ -401,6 +417,7 @@ class PartialState:
401
  """
402
  yield from self._goes_first(self.is_main_process)
403
  @contextmanager
 
404
  def local_main_process_first(self):
405
  """
406
  Lets the local main process go inside a with block.
@@ -416,6 +433,7 @@ class PartialState:
416
  ```
417
  """
418
  yield from self._goes_first(self.is_local_main_process)
 
419
  def on_main_process(self, function: Callable[..., Any] = None):
420
  """
421
  Decorator that only runs the decorated function on the main process.
@@ -437,6 +455,7 @@ class PartialState:
437
  if self.is_main_process or not self.use_distributed:
438
  return function
439
  return do_nothing
 
440
  def on_local_main_process(self, function: Callable[..., Any] = None):
441
  """
442
  Decorator that only runs the decorated function on the local main process.
@@ -448,7 +467,8 @@ class PartialState:
448
  from accelerate.state import PartialState
449
  state = PartialState()
450
  @state.on_local_main_process
451
- def print_something():
 
452
  print("This will be printed by process 0 only on each server.")
453
  print_something()
454
  # On server 1:
@@ -460,6 +480,7 @@ class PartialState:
460
  if self.is_local_main_process or not self.use_distributed:
461
  return function
462
  return do_nothing
 
463
  def on_last_process(self, function: Callable[..., Any]):
464
  """
465
  Decorator that only runs the decorated function on the last process.
@@ -471,7 +492,8 @@ class PartialState:
471
  from accelerate.state import PartialState
472
  state = PartialState()
473
  @state.on_last_process
474
- def print_something():
 
475
  print(f"Printed on process {state.process_index}")
476
  print_something()
477
  "Printed on process 3"
@@ -480,6 +502,7 @@ class PartialState:
480
  if self.is_last_process or not self.use_distributed:
481
  return function
482
  return do_nothing
 
483
  def on_process(self, function: Callable[..., Any] = None, process_index: int = None):
484
  """
485
  Decorator that only runs the decorated function on the process with the given index.
@@ -494,7 +517,8 @@ class PartialState:
494
  from accelerate.state import PartialState
495
  state = PartialState()
496
  @state.on_process(process_index=2)
497
- def print_something():
 
498
  print(f"Printed on process {state.process_index}")
499
  print_something()
500
  "Printed on process 2"
@@ -505,6 +529,7 @@ class PartialState:
505
  if (self.process_index == process_index) or (not self.use_distributed):
506
  return function
507
  return do_nothing
 
508
  def on_local_process(self, function: Callable[..., Any] = None, local_process_index: int = None):
509
  """
510
  Decorator that only runs the decorated function on the process with the given index on the current node.
@@ -519,7 +544,8 @@ class PartialState:
519
  from accelerate import Accelerator
520
  accelerator = Accelerator()
521
  @accelerator.on_local_process(local_process_index=2)
522
- def print_something():
 
523
  print(f"Printed on process {accelerator.local_process_index}")
524
  print_something()
525
  # On server 1:
@@ -533,10 +559,12 @@ class PartialState:
533
  if (self.local_process_index == local_process_index) or (not self.use_distributed):
534
  return function
535
  return do_nothing
 
536
  def print(self, *args, **kwargs):
537
  if self.is_local_main_process:
538
  print(*args, **kwargs)
539
  @property
 
540
  def default_device(self) -> torch.device:
541
  """
542
  Returns the default device which is:
@@ -575,6 +603,7 @@ class AcceleratorState:
575
  - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
576
  """
577
  _shared_state = SharedDict()
 
578
  def __init__(
579
  self,
580
  mixed_precision: str = None,
@@ -669,13 +698,16 @@ class AcceleratorState:
669
  torch.backends.cuda.matmul.allow_tf32 = True
670
  PartialState._shared_state["distributed_type"] = self.distributed_type
671
  @property
 
672
  def initialized(self) -> bool:
673
  return self._shared_state != PartialState._shared_state
 
674
  def __repr__(self):
675
  repr = PartialState().__repr__() + f"\nMixed precision type: {self.mixed_precision}\n"
676
  if self.distributed_type == DistributedType.DEEPSPEED:
677
  repr += f"ds_config: {self.deepspeed_plugin.deepspeed_config}\n"
678
  return repr
 
679
  def _check_initialized(self, mixed_precision=None, cpu=None):
680
  "Checks if a modification is trying to be made and the `AcceleratorState` has already been initialized"
681
  if self.initialized:
@@ -690,6 +722,7 @@ class AcceleratorState:
690
  raise ValueError(err.format(flag=f"mixed_precision='{mixed_precision}'"))
691
  # For backward compatibility
692
  @property
 
693
  def use_fp16(self):
694
  warnings.warn(
695
  "The `use_fp16` property is deprecated and will be removed in version 1.0 of Accelerate use "
@@ -698,6 +731,7 @@ class AcceleratorState:
698
  )
699
  return self._mixed_precision != "no"
700
  @property
 
701
  def mixed_precision(self):
702
  if self.distributed_type == DistributedType.DEEPSPEED:
703
  config = self.deepspeed_plugin.deepspeed_config
@@ -711,32 +745,39 @@ class AcceleratorState:
711
  mixed_precision = self._mixed_precision
712
  return mixed_precision
713
  @staticmethod
 
714
  def _reset_state(reset_partial_state: bool = False):
715
  "Resets `_shared_state`, is used internally and should not be called"
716
  AcceleratorState._shared_state.clear()
717
  if reset_partial_state:
718
  PartialState._reset_state()
719
  @property
 
720
  def use_distributed(self):
721
  """
722
  Whether the Accelerator is configured for distributed training
723
  """
724
  return PartialState().use_distributed
725
  @property
 
726
  def is_last_process(self) -> bool:
727
  "Returns whether the current process is the last one"
728
  return PartialState().is_last_process
729
  @property
 
730
  def is_main_process(self) -> bool:
731
  "Returns whether the current process is the main process"
732
  return PartialState().is_main_process
733
  @property
 
734
  def is_local_main_process(self) -> bool:
735
  "Returns whether the current process is the main process on the local node"
736
  return PartialState().is_local_main_process
 
737
  def wait_for_everyone(self):
738
  PartialState().wait_for_everyone()
739
  @contextmanager
 
740
  def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
741
  """
742
  Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
@@ -771,6 +812,7 @@ class AcceleratorState:
771
  with PartialState().split_between_processes(inputs, apply_padding=apply_padding) as inputs:
772
  yield inputs
773
  @contextmanager
 
774
  def main_process_first(self):
775
  """
776
  Lets the main process go first inside a with block.
@@ -779,6 +821,7 @@ class AcceleratorState:
779
  with PartialState().main_process_first():
780
  yield
781
  @contextmanager
 
782
  def local_main_process_first(self):
783
  """
784
  Lets the local main process go inside a with block.
@@ -786,6 +829,7 @@ class AcceleratorState:
786
  """
787
  with PartialState().local_main_process_first():
788
  yield
 
789
  def print(self, *args, **kwargs):
790
  PartialState().print(*args, **kwargs)
791
  class GradientState:
@@ -805,6 +849,7 @@ class GradientState:
805
  iteration and the number of total steps reset
806
  """
807
  _shared_state = SharedDict()
 
808
  def __init__(self, gradient_accumulation_plugin: Optional[GradientAccumulationPlugin] = None):
809
  self.__dict__ = self._shared_state
810
  if not self.initialized:
@@ -818,33 +863,40 @@ class GradientState:
818
  if gradient_accumulation_plugin is not None and self.plugin_kwargs != gradient_accumulation_plugin.to_kwargs():
819
  self.plugin_kwargs = gradient_accumulation_plugin.to_kwargs()
820
  @property
 
821
  def num_steps(self) -> int:
822
  "Returns the number of steps to accumulate over"
823
  return self.plugin_kwargs.get("num_steps", 1)
824
  @property
 
825
  def adjust_scheduler(self) -> bool:
826
  "Returns whether the scheduler should be adjusted"
827
  return self.plugin_kwargs.get("adjust_scheduler", False)
828
  @property
 
829
  def sync_with_dataloader(self) -> bool:
830
  "Returns whether the gradients should be synced at the end of the dataloader iteration and the number of total steps reset"
831
  return self.plugin_kwargs.get("sync_with_dataloader", True)
832
  @property
 
833
  def initialized(self) -> bool:
834
  "Returns whether the `GradientState` has been initialized"
835
  return GradientState._shared_state != {}
836
  @property
 
837
  def end_of_dataloader(self) -> bool:
838
  "Returns whether we have reached the end of the current dataloader"
839
  if not self.in_dataloader:
840
  return False
841
  return self.active_dataloader.end_of_dataloader
842
  @property
 
843
  def remainder(self) -> int:
844
  "Returns the number of extra samples that were added from padding the dataloader"
845
  if not self.in_dataloader:
846
  return -1
847
  return self.active_dataloader.remainder
 
848
  def __repr__(self):
849
  return (
850
  f"Sync Gradients: {self.sync_gradients}\n"
@@ -852,22 +904,27 @@ class GradientState:
852
  f"Extra samples added: {self.remainder}\n"
853
  f"Gradient accumulation plugin: {self.plugin_kwargs}\n"
854
  )
 
855
  def _set_sync_gradients(self, sync_gradients):
856
  "Private function that sets whether gradients should be synchronized. Users should not have to call this."
857
  self.sync_gradients = sync_gradients
 
858
  def _add_dataloader(self, dataloader):
859
  "Private function that adds a dataloader to `self.dataloader_references` and sets `in_dataloader` to `True`. Users should not have to call this."
860
  self.active_dataloader = dataloader
861
  self.dataloader_references.append(self.active_dataloader)
 
862
  def _remove_dataloader(self, dataloader):
863
  "Private function that removes a dataloader from `self.dataloader_references` and sets `in_dataloader` to `False` if there are no more dataloaders. Users should not have to call this."
864
  self.dataloader_references.remove(dataloader)
865
  self.active_dataloader = self.dataloader_references[-1]
866
  @property
 
867
  def in_dataloader(self) -> bool:
868
  "Returns whether the current process is in a dataloader"
869
  return self.active_dataloader is not None
870
  @staticmethod
 
871
  def _reset_state():
872
  "Resets `_shared_state`, is used internally and should not be called"
873
  GradientState._shared_state.clear()
 
20
  This is required for using PyTorch/XLA with PJRT in multithreaded mode (required for TPU v2 and v3).
21
  See https://github.com/pytorch/xla/blob/r2.0/docs/pjrt.md#multithreading-on-tpu-v2v3
22
  """
23
+
24
  def __init__(self, thread_local: bool = False):
25
  self._storage = {}
26
+
27
  def __get__(self, obj, objtype=None):
28
  return self._storage
29
+
30
  def __set__(self, obj, value):
31
  self._storage = value
32
  # Prefer global shared dictionary, except when using TPU.
 
52
  - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
53
  """
54
  _shared_state = SharedDict()
55
+
56
  def __init__(self, cpu: bool = False, **kwargs):
57
  self.__dict__ = self._shared_state
58
  if not self.initialized:
 
244
  if self.device is None:
245
  self.device = torch.device("cpu") if cpu else self.default_device
246
  self.fork_launched = parse_flag_from_env("FORK_LAUNCHED", 0)
247
+
248
  def __repr__(self) -> str:
249
  return (
250
  f"Distributed environment: {self.distributed_type}{(' Backend: ' + self.backend) if self.backend else ''}\n"
 
254
  f"Device: {self.device}\n"
255
  )
256
  @staticmethod
257
+
258
  def _reset_state():
259
  "Resets `_shared_state`, is used internally and should not be called"
260
  PartialState._shared_state.clear()
261
  @property
262
+
263
  def initialized(self) -> bool:
264
  "Returns whether the `PartialState` has been initialized"
265
  return self._shared_state != {}
266
  @property
267
+
268
  def use_distributed(self):
269
  """
270
  Whether the Accelerator is configured for distributed training
271
  """
272
  return self.distributed_type != DistributedType.NO and self.num_processes > 1
273
  @property
274
+
275
  def is_last_process(self) -> bool:
276
  "Returns whether the current process is the last one"
277
  return self.process_index == self.num_processes - 1
278
  @property
279
+
280
  def is_main_process(self) -> bool:
281
  "Returns whether the current process is the main process"
282
  return (
283
  self.process_index == 0 if self.distributed_type != DistributedType.MEGATRON_LM else self.is_last_process
284
  )
285
  @property
286
+
287
  def is_local_main_process(self) -> bool:
288
  "Returns whether the current process is the main process on the local node"
289
  return (
 
291
  if self.distributed_type != DistributedType.MEGATRON_LM
292
  else self.is_last_process
293
  )
294
+
295
  def wait_for_everyone(self):
296
  """
297
  Will stop the execution of the current process until every other process has reached that point (so this does
 
322
  torch.distributed.barrier()
323
  elif self.distributed_type == DistributedType.TPU:
324
  xm.rendezvous("accelerate.utils.wait_for_everyone")
325
+
326
  def _goes_first(self, is_main: bool):
327
  if not is_main:
328
  self.wait_for_everyone()
 
330
  if is_main:
331
  self.wait_for_everyone()
332
  @contextmanager
333
+
334
  def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
335
  """
336
  Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
 
376
  end_index = start_index + num_samples_per_process
377
  if (len(inputs) % self.num_processes != 0) and (self.process_index == self.num_processes - 1):
378
  end_index = length
379
+
380
+ def _split_values(inputs, start_index, end_index):
381
  if isinstance(inputs, (list, tuple, torch.Tensor)):
382
  if start_index >= len(inputs):
383
  result = inputs[-1:]
 
400
  return inputs
401
  yield _split_values(inputs, start_index, end_index)
402
  @contextmanager
403
+
404
  def main_process_first(self):
405
  """
406
  Lets the main process go first inside a with block.
 
417
  """
418
  yield from self._goes_first(self.is_main_process)
419
  @contextmanager
420
+
421
  def local_main_process_first(self):
422
  """
423
  Lets the local main process go inside a with block.
 
433
  ```
434
  """
435
  yield from self._goes_first(self.is_local_main_process)
436
+
437
  def on_main_process(self, function: Callable[..., Any] = None):
438
  """
439
  Decorator that only runs the decorated function on the main process.
 
455
  if self.is_main_process or not self.use_distributed:
456
  return function
457
  return do_nothing
458
+
459
  def on_local_main_process(self, function: Callable[..., Any] = None):
460
  """
461
  Decorator that only runs the decorated function on the local main process.
 
467
  from accelerate.state import PartialState
468
  state = PartialState()
469
  @state.on_local_main_process
470
+
471
+ def print_something():
472
  print("This will be printed by process 0 only on each server.")
473
  print_something()
474
  # On server 1:
 
480
  if self.is_local_main_process or not self.use_distributed:
481
  return function
482
  return do_nothing
483
+
484
  def on_last_process(self, function: Callable[..., Any]):
485
  """
486
  Decorator that only runs the decorated function on the last process.
 
492
  from accelerate.state import PartialState
493
  state = PartialState()
494
  @state.on_last_process
495
+
496
+ def print_something():
497
  print(f"Printed on process {state.process_index}")
498
  print_something()
499
  "Printed on process 3"
 
502
  if self.is_last_process or not self.use_distributed:
503
  return function
504
  return do_nothing
505
+
506
  def on_process(self, function: Callable[..., Any] = None, process_index: int = None):
507
  """
508
  Decorator that only runs the decorated function on the process with the given index.
 
517
  from accelerate.state import PartialState
518
  state = PartialState()
519
  @state.on_process(process_index=2)
520
+
521
+ def print_something():
522
  print(f"Printed on process {state.process_index}")
523
  print_something()
524
  "Printed on process 2"
 
529
  if (self.process_index == process_index) or (not self.use_distributed):
530
  return function
531
  return do_nothing
532
+
533
  def on_local_process(self, function: Callable[..., Any] = None, local_process_index: int = None):
534
  """
535
  Decorator that only runs the decorated function on the process with the given index on the current node.
 
544
  from accelerate import Accelerator
545
  accelerator = Accelerator()
546
  @accelerator.on_local_process(local_process_index=2)
547
+
548
+ def print_something():
549
  print(f"Printed on process {accelerator.local_process_index}")
550
  print_something()
551
  # On server 1:
 
559
  if (self.local_process_index == local_process_index) or (not self.use_distributed):
560
  return function
561
  return do_nothing
562
+
563
  def print(self, *args, **kwargs):
564
  if self.is_local_main_process:
565
  print(*args, **kwargs)
566
  @property
567
+
568
  def default_device(self) -> torch.device:
569
  """
570
  Returns the default device which is:
 
603
  - **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
604
  """
605
  _shared_state = SharedDict()
606
+
607
  def __init__(
608
  self,
609
  mixed_precision: str = None,
 
698
  torch.backends.cuda.matmul.allow_tf32 = True
699
  PartialState._shared_state["distributed_type"] = self.distributed_type
700
  @property
701
+
702
  def initialized(self) -> bool:
703
  return self._shared_state != PartialState._shared_state
704
+
705
  def __repr__(self):
706
  repr = PartialState().__repr__() + f"\nMixed precision type: {self.mixed_precision}\n"
707
  if self.distributed_type == DistributedType.DEEPSPEED:
708
  repr += f"ds_config: {self.deepspeed_plugin.deepspeed_config}\n"
709
  return repr
710
+
711
  def _check_initialized(self, mixed_precision=None, cpu=None):
712
  "Checks if a modification is trying to be made and the `AcceleratorState` has already been initialized"
713
  if self.initialized:
 
722
  raise ValueError(err.format(flag=f"mixed_precision='{mixed_precision}'"))
723
  # For backward compatibility
724
  @property
725
+
726
  def use_fp16(self):
727
  warnings.warn(
728
  "The `use_fp16` property is deprecated and will be removed in version 1.0 of Accelerate use "
 
731
  )
732
  return self._mixed_precision != "no"
733
  @property
734
+
735
  def mixed_precision(self):
736
  if self.distributed_type == DistributedType.DEEPSPEED:
737
  config = self.deepspeed_plugin.deepspeed_config
 
745
  mixed_precision = self._mixed_precision
746
  return mixed_precision
747
  @staticmethod
748
+
749
  def _reset_state(reset_partial_state: bool = False):
750
  "Resets `_shared_state`, is used internally and should not be called"
751
  AcceleratorState._shared_state.clear()
752
  if reset_partial_state:
753
  PartialState._reset_state()
754
  @property
755
+
756
  def use_distributed(self):
757
  """
758
  Whether the Accelerator is configured for distributed training
759
  """
760
  return PartialState().use_distributed
761
  @property
762
+
763
  def is_last_process(self) -> bool:
764
  "Returns whether the current process is the last one"
765
  return PartialState().is_last_process
766
  @property
767
+
768
  def is_main_process(self) -> bool:
769
  "Returns whether the current process is the main process"
770
  return PartialState().is_main_process
771
  @property
772
+
773
  def is_local_main_process(self) -> bool:
774
  "Returns whether the current process is the main process on the local node"
775
  return PartialState().is_local_main_process
776
+
777
  def wait_for_everyone(self):
778
  PartialState().wait_for_everyone()
779
  @contextmanager
780
+
781
  def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
782
  """
783
  Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
 
812
  with PartialState().split_between_processes(inputs, apply_padding=apply_padding) as inputs:
813
  yield inputs
814
  @contextmanager
815
+
816
  def main_process_first(self):
817
  """
818
  Lets the main process go first inside a with block.
 
821
  with PartialState().main_process_first():
822
  yield
823
  @contextmanager
824
+
825
  def local_main_process_first(self):
826
  """
827
  Lets the local main process go inside a with block.
 
829
  """
830
  with PartialState().local_main_process_first():
831
  yield
832
+
833
  def print(self, *args, **kwargs):
834
  PartialState().print(*args, **kwargs)
835
  class GradientState:
 
849
  iteration and the number of total steps reset
850
  """
851
  _shared_state = SharedDict()
852
+
853
  def __init__(self, gradient_accumulation_plugin: Optional[GradientAccumulationPlugin] = None):
854
  self.__dict__ = self._shared_state
855
  if not self.initialized:
 
863
  if gradient_accumulation_plugin is not None and self.plugin_kwargs != gradient_accumulation_plugin.to_kwargs():
864
  self.plugin_kwargs = gradient_accumulation_plugin.to_kwargs()
865
  @property
866
+
867
  def num_steps(self) -> int:
868
  "Returns the number of steps to accumulate over"
869
  return self.plugin_kwargs.get("num_steps", 1)
870
  @property
871
+
872
  def adjust_scheduler(self) -> bool:
873
  "Returns whether the scheduler should be adjusted"
874
  return self.plugin_kwargs.get("adjust_scheduler", False)
875
  @property
876
+
877
  def sync_with_dataloader(self) -> bool:
878
  "Returns whether the gradients should be synced at the end of the dataloader iteration and the number of total steps reset"
879
  return self.plugin_kwargs.get("sync_with_dataloader", True)
880
  @property
881
+
882
  def initialized(self) -> bool:
883
  "Returns whether the `GradientState` has been initialized"
884
  return GradientState._shared_state != {}
885
  @property
886
+
887
  def end_of_dataloader(self) -> bool:
888
  "Returns whether we have reached the end of the current dataloader"
889
  if not self.in_dataloader:
890
  return False
891
  return self.active_dataloader.end_of_dataloader
892
  @property
893
+
894
  def remainder(self) -> int:
895
  "Returns the number of extra samples that were added from padding the dataloader"
896
  if not self.in_dataloader:
897
  return -1
898
  return self.active_dataloader.remainder
899
+
900
  def __repr__(self):
901
  return (
902
  f"Sync Gradients: {self.sync_gradients}\n"
 
904
  f"Extra samples added: {self.remainder}\n"
905
  f"Gradient accumulation plugin: {self.plugin_kwargs}\n"
906
  )
907
+
908
  def _set_sync_gradients(self, sync_gradients):
909
  "Private function that sets whether gradients should be synchronized. Users should not have to call this."
910
  self.sync_gradients = sync_gradients
911
+
912
  def _add_dataloader(self, dataloader):
913
  "Private function that adds a dataloader to `self.dataloader_references` and sets `in_dataloader` to `True`. Users should not have to call this."
914
  self.active_dataloader = dataloader
915
  self.dataloader_references.append(self.active_dataloader)
916
+
917
  def _remove_dataloader(self, dataloader):
918
  "Private function that removes a dataloader from `self.dataloader_references` and sets `in_dataloader` to `False` if there are no more dataloaders. Users should not have to call this."
919
  self.dataloader_references.remove(dataloader)
920
  self.active_dataloader = self.dataloader_references[-1]
921
  @property
922
+
923
  def in_dataloader(self) -> bool:
924
  "Returns whether the current process is in a dataloader"
925
  return self.active_dataloader is not None
926
  @staticmethod
927
+
928
  def _reset_state():
929
  "Resets `_shared_state`, is used internally and should not be called"
930
  GradientState._shared_state.clear()
src/tracking.py CHANGED
@@ -24,6 +24,7 @@ def on_main_process(function):
24
  `PartialState`.
25
  """
26
  @wraps(function)
 
27
  def execute_on_main_process(self, *args, **kwargs):
28
  if getattr(self, "main_process_only", False):
29
  return PartialState().on_main_process(function)(self, *args, **kwargs)
@@ -46,6 +47,7 @@ class GeneralTracker:
46
  other functions should occur on the main process or across all processes (by default will use `True`)
47
  """
48
  main_process_only = True
 
49
  def __init__(self, _blank=False):
50
  if not _blank:
51
  err = ""
@@ -66,6 +68,7 @@ class GeneralTracker:
66
  f"required attributes. Please define them in the class definition: "
67
  f"{err}"
68
  )
 
69
  def store_init_configuration(self, values: dict):
70
  """
71
  Logs `values` as hyperparameters for the run. Implementations should use the experiment configuration
@@ -76,6 +79,7 @@ class GeneralTracker:
76
  `str`, `float`, `int`, or `None`.
77
  """
78
  pass
 
79
  def log(self, values: dict, step: Optional[int], **kwargs):
80
  """
81
  Logs `values` to the current run. Base `log` implementations of a tracking API should go in here, along with
@@ -87,6 +91,7 @@ class GeneralTracker:
87
  The run step. If included, the log will be affiliated with this step.
88
  """
89
  pass
 
90
  def finish(self):
91
  """
92
  Should run any finalizing functions within the tracking API. If the API should not have one, just don't
@@ -107,6 +112,7 @@ class TensorBoardTracker(GeneralTracker):
107
  name = "tensorboard"
108
  requires_logging_directory = True
109
  @on_main_process
 
110
  def __init__(self, run_name: str, logging_dir: Union[str, os.PathLike], **kwargs):
111
  try:
112
  from torch.utils import tensorboard
@@ -121,9 +127,11 @@ class TensorBoardTracker(GeneralTracker):
121
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
122
  )
123
  @property
 
124
  def tracker(self):
125
  return self.writer
126
  @on_main_process
 
127
  def store_init_configuration(self, values: dict):
128
  """
129
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment. Stores the
@@ -146,6 +154,7 @@ class TensorBoardTracker(GeneralTracker):
146
  raise
147
  logger.debug("Stored initial configuration hyperparameters to TensorBoard and hparams yaml file")
148
  @on_main_process
 
149
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
150
  """
151
  Logs `values` to the current run.
@@ -170,6 +179,7 @@ class TensorBoardTracker(GeneralTracker):
170
  self.writer.flush()
171
  logger.debug("Successfully logged to TensorBoard")
172
  @on_main_process
 
173
  def log_images(self, values: dict, step: Optional[int], **kwargs):
174
  """
175
  Logs `images` to the current run.
@@ -185,6 +195,7 @@ class TensorBoardTracker(GeneralTracker):
185
  self.writer.add_images(k, v, global_step=step, **kwargs)
186
  logger.debug("Successfully logged images to TensorBoard")
187
  @on_main_process
 
188
  def finish(self):
189
  """
190
  Closes `TensorBoard` writer
@@ -204,6 +215,7 @@ class WandBTracker(GeneralTracker):
204
  requires_logging_directory = False
205
  main_process_only = False
206
  @on_main_process
 
207
  def __init__(self, run_name: str, **kwargs):
208
  super().__init__()
209
  self.run_name = run_name
@@ -214,9 +226,11 @@ class WandBTracker(GeneralTracker):
214
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
215
  )
216
  @property
 
217
  def tracker(self):
218
  return self.run
219
  @on_main_process
 
220
  def store_init_configuration(self, values: dict):
221
  """
222
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
@@ -229,6 +243,7 @@ class WandBTracker(GeneralTracker):
229
  wandb.config.update(values, allow_val_change=True)
230
  logger.debug("Stored initial configuration hyperparameters to WandB")
231
  @on_main_process
 
232
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
233
  """
234
  Logs `values` to the current run.
@@ -244,6 +259,7 @@ class WandBTracker(GeneralTracker):
244
  self.run.log(values, step=step, **kwargs)
245
  logger.debug("Successfully logged to WandB")
246
  @on_main_process
 
247
  def log_images(self, values: dict, step: Optional[int] = None, **kwargs):
248
  """
249
  Logs `images` to the current run.
@@ -260,6 +276,7 @@ class WandBTracker(GeneralTracker):
260
  self.log({k: [wandb.Image(image) for image in v]}, step=step, **kwargs)
261
  logger.debug("Successfully logged images to WandB")
262
  @on_main_process
 
263
  def log_table(
264
  self,
265
  table_name: str,
@@ -288,6 +305,7 @@ class WandBTracker(GeneralTracker):
288
  values = {table_name: wandb.Table(columns=columns, data=data, dataframe=dataframe)}
289
  self.log(values, step=step, **kwargs)
290
  @on_main_process
 
291
  def finish(self):
292
  """
293
  Closes `wandb` writer
@@ -307,6 +325,7 @@ class CometMLTracker(GeneralTracker):
307
  name = "comet_ml"
308
  requires_logging_directory = False
309
  @on_main_process
 
310
  def __init__(self, run_name: str, **kwargs):
311
  super().__init__()
312
  self.run_name = run_name
@@ -317,9 +336,11 @@ class CometMLTracker(GeneralTracker):
317
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
318
  )
319
  @property
 
320
  def tracker(self):
321
  return self.writer
322
  @on_main_process
 
323
  def store_init_configuration(self, values: dict):
324
  """
325
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
@@ -331,6 +352,7 @@ class CometMLTracker(GeneralTracker):
331
  self.writer.log_parameters(values)
332
  logger.debug("Stored initial configuration hyperparameters to CometML")
333
  @on_main_process
 
334
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
335
  """
336
  Logs `values` to the current run.
@@ -355,6 +377,7 @@ class CometMLTracker(GeneralTracker):
355
  self.writer.log_metrics(v, step=step, **kwargs)
356
  logger.debug("Successfully logged to CometML")
357
  @on_main_process
 
358
  def finish(self):
359
  """
360
  Closes `comet-ml` writer
@@ -373,6 +396,7 @@ class AimTracker(GeneralTracker):
373
  name = "aim"
374
  requires_logging_directory = True
375
  @on_main_process
 
376
  def __init__(self, run_name: str, logging_dir: Optional[Union[str, os.PathLike]] = ".", **kwargs):
377
  self.run_name = run_name
378
  from aim import Run
@@ -383,9 +407,11 @@ class AimTracker(GeneralTracker):
383
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
384
  )
385
  @property
 
386
  def tracker(self):
387
  return self.writer
388
  @on_main_process
 
389
  def store_init_configuration(self, values: dict):
390
  """
391
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
@@ -395,6 +421,7 @@ class AimTracker(GeneralTracker):
395
  """
396
  self.writer["hparams"] = values
397
  @on_main_process
 
398
  def log(self, values: dict, step: Optional[int], **kwargs):
399
  """
400
  Logs `values` to the current run.
@@ -410,6 +437,7 @@ class AimTracker(GeneralTracker):
410
  for key, value in values.items():
411
  self.writer.track(value, name=key, step=step, **kwargs)
412
  @on_main_process
 
413
  def log_images(self, values: dict, step: Optional[int] = None, kwargs: Optional[Dict[str, dict]] = None):
414
  """
415
  Logs `images` to the current run.
@@ -437,6 +465,7 @@ class AimTracker(GeneralTracker):
437
  aim_image = aim.Image(img, caption=caption, **aim_image_kw)
438
  self.writer.track(aim_image, name=key, step=step, **track_kw)
439
  @on_main_process
 
440
  def finish(self):
441
  """
442
  Closes `aim` writer
@@ -471,6 +500,7 @@ class MLflowTracker(GeneralTracker):
471
  name = "mlflow"
472
  requires_logging_directory = False
473
  @on_main_process
 
474
  def __init__(
475
  self,
476
  experiment_name: str = None,
@@ -512,9 +542,11 @@ class MLflowTracker(GeneralTracker):
512
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
513
  )
514
  @property
 
515
  def tracker(self):
516
  return self.active_run
517
  @on_main_process
 
518
  def store_init_configuration(self, values: dict):
519
  """
520
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
@@ -537,6 +569,7 @@ class MLflowTracker(GeneralTracker):
537
  mlflow.log_params(dict(values_list[i : i + mlflow.utils.validation.MAX_PARAMS_TAGS_PER_BATCH]))
538
  logger.debug("Stored initial configuration hyperparameters to MLflow")
539
  @on_main_process
 
540
  def log(self, values: dict, step: Optional[int]):
541
  """
542
  Logs `values` to the current run.
@@ -559,6 +592,7 @@ class MLflowTracker(GeneralTracker):
559
  mlflow.log_metrics(metrics, step=step)
560
  logger.debug("Successfully logged to mlflow")
561
  @on_main_process
 
562
  def finish(self):
563
  """
564
  End the active MLflow run.
@@ -578,6 +612,7 @@ class ClearMLTracker(GeneralTracker):
578
  name = "clearml"
579
  requires_logging_directory = False
580
  @on_main_process
 
581
  def __init__(self, run_name: str = None, **kwargs):
582
  from clearml import Task
583
  current_task = Task.current_task()
@@ -590,9 +625,11 @@ class ClearMLTracker(GeneralTracker):
590
  kwargs.setdefault("task_name", os.environ.get("CLEARML_TASK", run_name))
591
  self.task = Task.init(**kwargs)
592
  @property
 
593
  def tracker(self):
594
  return self.task
595
  @on_main_process
 
596
  def store_init_configuration(self, values: dict):
597
  """
598
  Connect configuration dictionary to the Task object. Should be run at the beginning of your experiment.
@@ -602,6 +639,7 @@ class ClearMLTracker(GeneralTracker):
602
  """
603
  return self.task.connect_configuration(values)
604
  @on_main_process
 
605
  def log(self, values: Dict[str, Union[int, float]], step: Optional[int] = None, **kwargs):
606
  """
607
  Logs `values` dictionary to the current run. The dictionary keys must be strings. The dictionary values must be
@@ -634,6 +672,7 @@ class ClearMLTracker(GeneralTracker):
634
  title, series = ClearMLTracker._get_title_series(k)
635
  clearml_logger.report_scalar(title=title, series=series, value=v, iteration=step, **kwargs)
636
  @on_main_process
 
637
  def log_images(self, values: dict, step: Optional[int] = None, **kwargs):
638
  """
639
  Logs `images` to the current run.
@@ -650,6 +689,7 @@ class ClearMLTracker(GeneralTracker):
650
  title, series = ClearMLTracker._get_title_series(k)
651
  clearml_logger.report_image(title=title, series=series, iteration=step, image=v, **kwargs)
652
  @on_main_process
 
653
  def log_table(
654
  self,
655
  table_name: str,
@@ -686,6 +726,7 @@ class ClearMLTracker(GeneralTracker):
686
  title, series = ClearMLTracker._get_title_series(table_name)
687
  self.task.get_logger().report_table(title=title, series=series, table_plot=to_report, iteration=step, **kwargs)
688
  @on_main_process
 
689
  def finish(self):
690
  """
691
  Close the ClearML task. If the task was initialized externally (e.g. by manually calling `Task.init`), this
@@ -694,6 +735,7 @@ class ClearMLTracker(GeneralTracker):
694
  if self.task and not self._initialized_externally:
695
  self.task.close()
696
  @staticmethod
 
697
  def _get_title_series(name):
698
  for prefix in ["eval", "test", "train"]:
699
  if name.startswith(prefix + "_"):
@@ -717,14 +759,17 @@ class DVCLiveTracker(GeneralTracker):
717
  name = "dvclive"
718
  requires_logging_directory = False
719
  @on_main_process
 
720
  def __init__(self, run_name: Optional[str] = None, live: Optional[Any] = None, **kwargs):
721
  from dvclive import Live
722
  super().__init__()
723
  self.live = live if live is not None else Live(**kwargs)
724
  @property
 
725
  def tracker(self):
726
  return self.live
727
  @on_main_process
 
728
  def store_init_configuration(self, values: dict):
729
  """
730
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment. Stores the
@@ -736,6 +781,7 @@ class DVCLiveTracker(GeneralTracker):
736
  """
737
  self.live.log_params(values)
738
  @on_main_process
 
739
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
740
  """
741
  Logs `values` to the current run.
@@ -761,6 +807,7 @@ class DVCLiveTracker(GeneralTracker):
761
  "is incorrect so we dropped this attribute."
762
  )
763
  @on_main_process
 
764
  def finish(self):
765
  """
766
  Closes `dvclive.Live()`.
 
24
  `PartialState`.
25
  """
26
  @wraps(function)
27
+
28
  def execute_on_main_process(self, *args, **kwargs):
29
  if getattr(self, "main_process_only", False):
30
  return PartialState().on_main_process(function)(self, *args, **kwargs)
 
47
  other functions should occur on the main process or across all processes (by default will use `True`)
48
  """
49
  main_process_only = True
50
+
51
  def __init__(self, _blank=False):
52
  if not _blank:
53
  err = ""
 
68
  f"required attributes. Please define them in the class definition: "
69
  f"{err}"
70
  )
71
+
72
  def store_init_configuration(self, values: dict):
73
  """
74
  Logs `values` as hyperparameters for the run. Implementations should use the experiment configuration
 
79
  `str`, `float`, `int`, or `None`.
80
  """
81
  pass
82
+
83
  def log(self, values: dict, step: Optional[int], **kwargs):
84
  """
85
  Logs `values` to the current run. Base `log` implementations of a tracking API should go in here, along with
 
91
  The run step. If included, the log will be affiliated with this step.
92
  """
93
  pass
94
+
95
  def finish(self):
96
  """
97
  Should run any finalizing functions within the tracking API. If the API should not have one, just don't
 
112
  name = "tensorboard"
113
  requires_logging_directory = True
114
  @on_main_process
115
+
116
  def __init__(self, run_name: str, logging_dir: Union[str, os.PathLike], **kwargs):
117
  try:
118
  from torch.utils import tensorboard
 
127
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
128
  )
129
  @property
130
+
131
  def tracker(self):
132
  return self.writer
133
  @on_main_process
134
+
135
  def store_init_configuration(self, values: dict):
136
  """
137
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment. Stores the
 
154
  raise
155
  logger.debug("Stored initial configuration hyperparameters to TensorBoard and hparams yaml file")
156
  @on_main_process
157
+
158
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
159
  """
160
  Logs `values` to the current run.
 
179
  self.writer.flush()
180
  logger.debug("Successfully logged to TensorBoard")
181
  @on_main_process
182
+
183
  def log_images(self, values: dict, step: Optional[int], **kwargs):
184
  """
185
  Logs `images` to the current run.
 
195
  self.writer.add_images(k, v, global_step=step, **kwargs)
196
  logger.debug("Successfully logged images to TensorBoard")
197
  @on_main_process
198
+
199
  def finish(self):
200
  """
201
  Closes `TensorBoard` writer
 
215
  requires_logging_directory = False
216
  main_process_only = False
217
  @on_main_process
218
+
219
  def __init__(self, run_name: str, **kwargs):
220
  super().__init__()
221
  self.run_name = run_name
 
226
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
227
  )
228
  @property
229
+
230
  def tracker(self):
231
  return self.run
232
  @on_main_process
233
+
234
  def store_init_configuration(self, values: dict):
235
  """
236
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
 
243
  wandb.config.update(values, allow_val_change=True)
244
  logger.debug("Stored initial configuration hyperparameters to WandB")
245
  @on_main_process
246
+
247
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
248
  """
249
  Logs `values` to the current run.
 
259
  self.run.log(values, step=step, **kwargs)
260
  logger.debug("Successfully logged to WandB")
261
  @on_main_process
262
+
263
  def log_images(self, values: dict, step: Optional[int] = None, **kwargs):
264
  """
265
  Logs `images` to the current run.
 
276
  self.log({k: [wandb.Image(image) for image in v]}, step=step, **kwargs)
277
  logger.debug("Successfully logged images to WandB")
278
  @on_main_process
279
+
280
  def log_table(
281
  self,
282
  table_name: str,
 
305
  values = {table_name: wandb.Table(columns=columns, data=data, dataframe=dataframe)}
306
  self.log(values, step=step, **kwargs)
307
  @on_main_process
308
+
309
  def finish(self):
310
  """
311
  Closes `wandb` writer
 
325
  name = "comet_ml"
326
  requires_logging_directory = False
327
  @on_main_process
328
+
329
  def __init__(self, run_name: str, **kwargs):
330
  super().__init__()
331
  self.run_name = run_name
 
336
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
337
  )
338
  @property
339
+
340
  def tracker(self):
341
  return self.writer
342
  @on_main_process
343
+
344
  def store_init_configuration(self, values: dict):
345
  """
346
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
 
352
  self.writer.log_parameters(values)
353
  logger.debug("Stored initial configuration hyperparameters to CometML")
354
  @on_main_process
355
+
356
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
357
  """
358
  Logs `values` to the current run.
 
377
  self.writer.log_metrics(v, step=step, **kwargs)
378
  logger.debug("Successfully logged to CometML")
379
  @on_main_process
380
+
381
  def finish(self):
382
  """
383
  Closes `comet-ml` writer
 
396
  name = "aim"
397
  requires_logging_directory = True
398
  @on_main_process
399
+
400
  def __init__(self, run_name: str, logging_dir: Optional[Union[str, os.PathLike]] = ".", **kwargs):
401
  self.run_name = run_name
402
  from aim import Run
 
407
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
408
  )
409
  @property
410
+
411
  def tracker(self):
412
  return self.writer
413
  @on_main_process
414
+
415
  def store_init_configuration(self, values: dict):
416
  """
417
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
 
421
  """
422
  self.writer["hparams"] = values
423
  @on_main_process
424
+
425
  def log(self, values: dict, step: Optional[int], **kwargs):
426
  """
427
  Logs `values` to the current run.
 
437
  for key, value in values.items():
438
  self.writer.track(value, name=key, step=step, **kwargs)
439
  @on_main_process
440
+
441
  def log_images(self, values: dict, step: Optional[int] = None, kwargs: Optional[Dict[str, dict]] = None):
442
  """
443
  Logs `images` to the current run.
 
465
  aim_image = aim.Image(img, caption=caption, **aim_image_kw)
466
  self.writer.track(aim_image, name=key, step=step, **track_kw)
467
  @on_main_process
468
+
469
  def finish(self):
470
  """
471
  Closes `aim` writer
 
500
  name = "mlflow"
501
  requires_logging_directory = False
502
  @on_main_process
503
+
504
  def __init__(
505
  self,
506
  experiment_name: str = None,
 
542
  "Make sure to log any initial configurations with `self.store_init_configuration` before training!"
543
  )
544
  @property
545
+
546
  def tracker(self):
547
  return self.active_run
548
  @on_main_process
549
+
550
  def store_init_configuration(self, values: dict):
551
  """
552
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
 
569
  mlflow.log_params(dict(values_list[i : i + mlflow.utils.validation.MAX_PARAMS_TAGS_PER_BATCH]))
570
  logger.debug("Stored initial configuration hyperparameters to MLflow")
571
  @on_main_process
572
+
573
  def log(self, values: dict, step: Optional[int]):
574
  """
575
  Logs `values` to the current run.
 
592
  mlflow.log_metrics(metrics, step=step)
593
  logger.debug("Successfully logged to mlflow")
594
  @on_main_process
595
+
596
  def finish(self):
597
  """
598
  End the active MLflow run.
 
612
  name = "clearml"
613
  requires_logging_directory = False
614
  @on_main_process
615
+
616
  def __init__(self, run_name: str = None, **kwargs):
617
  from clearml import Task
618
  current_task = Task.current_task()
 
625
  kwargs.setdefault("task_name", os.environ.get("CLEARML_TASK", run_name))
626
  self.task = Task.init(**kwargs)
627
  @property
628
+
629
  def tracker(self):
630
  return self.task
631
  @on_main_process
632
+
633
  def store_init_configuration(self, values: dict):
634
  """
635
  Connect configuration dictionary to the Task object. Should be run at the beginning of your experiment.
 
639
  """
640
  return self.task.connect_configuration(values)
641
  @on_main_process
642
+
643
  def log(self, values: Dict[str, Union[int, float]], step: Optional[int] = None, **kwargs):
644
  """
645
  Logs `values` dictionary to the current run. The dictionary keys must be strings. The dictionary values must be
 
672
  title, series = ClearMLTracker._get_title_series(k)
673
  clearml_logger.report_scalar(title=title, series=series, value=v, iteration=step, **kwargs)
674
  @on_main_process
675
+
676
  def log_images(self, values: dict, step: Optional[int] = None, **kwargs):
677
  """
678
  Logs `images` to the current run.
 
689
  title, series = ClearMLTracker._get_title_series(k)
690
  clearml_logger.report_image(title=title, series=series, iteration=step, image=v, **kwargs)
691
  @on_main_process
692
+
693
  def log_table(
694
  self,
695
  table_name: str,
 
726
  title, series = ClearMLTracker._get_title_series(table_name)
727
  self.task.get_logger().report_table(title=title, series=series, table_plot=to_report, iteration=step, **kwargs)
728
  @on_main_process
729
+
730
  def finish(self):
731
  """
732
  Close the ClearML task. If the task was initialized externally (e.g. by manually calling `Task.init`), this
 
735
  if self.task and not self._initialized_externally:
736
  self.task.close()
737
  @staticmethod
738
+
739
  def _get_title_series(name):
740
  for prefix in ["eval", "test", "train"]:
741
  if name.startswith(prefix + "_"):
 
759
  name = "dvclive"
760
  requires_logging_directory = False
761
  @on_main_process
762
+
763
  def __init__(self, run_name: Optional[str] = None, live: Optional[Any] = None, **kwargs):
764
  from dvclive import Live
765
  super().__init__()
766
  self.live = live if live is not None else Live(**kwargs)
767
  @property
768
+
769
  def tracker(self):
770
  return self.live
771
  @on_main_process
772
+
773
  def store_init_configuration(self, values: dict):
774
  """
775
  Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment. Stores the
 
781
  """
782
  self.live.log_params(values)
783
  @on_main_process
784
+
785
  def log(self, values: dict, step: Optional[int] = None, **kwargs):
786
  """
787
  Logs `values` to the current run.
 
807
  "is incorrect so we dropped this attribute."
808
  )
809
  @on_main_process
810
+
811
  def finish(self):
812
  """
813
  Closes `dvclive.Live()`.
src/utils/dataclasses.py CHANGED
@@ -5,8 +5,10 @@ class KwargsHandler:
5
  """
6
  Internal mixin that implements a `to_kwargs()` method for a dataclass.
7
  """
 
8
  def to_dict(self):
9
  return copy.deepcopy(self.__dict__)
 
10
  def to_kwargs(self):
11
  """
12
  Returns a dictionary containing the attributes with values different from the default of this class.
@@ -159,6 +161,7 @@ class FP8RecipeKwargs(KwargsHandler):
159
  amax_history_len: int = 1
160
  amax_compute_algo: AmaxComputeAlgorithm = "most_recent"
161
  override_linear_precision: Tuple[bool, bool, bool] = (False, False, False)
 
162
  def __post_init__(self):
163
  self.backend = self.backend.upper()
164
  if self.backend not in get_args(Backend):
@@ -175,6 +178,7 @@ class FP8RecipeKwargs(KwargsHandler):
175
  raise ValueError(f"`optimization_level` must be one of {' or '.join(get_args(OptLevel))}")
176
  class EnumWithContains(enum.EnumMeta):
177
  "A metaclass that adds the ability to check if `self` contains an item with the `in` operator"
 
178
  def __contains__(cls, item):
179
  try:
180
  cls(item)
@@ -183,9 +187,11 @@ class EnumWithContains(enum.EnumMeta):
183
  return True
184
  class BaseEnum(enum.Enum, metaclass=EnumWithContains):
185
  "An enum class that can get the value of an item with `str(Enum.key)`"
 
186
  def __str__(self):
187
  return self.value
188
  @classmethod
 
189
  def list(cls):
190
  "Method to list all the possible items in `cls`"
191
  return list(map(str, cls))
@@ -354,11 +360,13 @@ class ProjectConfiguration:
354
  )
355
  },
356
  )
 
357
  def set_directories(self, project_dir: str = None):
358
  "Sets `self.project_dir` and `self.logging_dir` to the appropriate values."
359
  self.project_dir = project_dir
360
  if self.logging_dir is None:
361
  self.logging_dir = project_dir
 
362
  def __post_init__(self):
363
  self.set_directories(self.project_dir)
364
  @dataclass
@@ -395,6 +403,7 @@ class TorchDynamoPlugin(KwargsHandler):
395
  dynamic: bool = field(default=None, metadata={"help": "Whether to use dynamic shape for tracing"})
396
  options: Any = field(default=None, metadata={"help": "A dictionary of options to pass to the backend."})
397
  disable: bool = field(default=False, metadata={"help": "Turn torch.compile() into a no-op for testing"})
 
398
  def __post_init__(self):
399
  prefix = "ACCELERATE_DYNAMO_"
400
  if self.backend is None:
@@ -406,6 +415,7 @@ class TorchDynamoPlugin(KwargsHandler):
406
  self.fullgraph = str_to_bool(os.environ.get(prefix + "USE_FULLGRAPH", "False")) == 1
407
  if self.dynamic is None:
408
  self.dynamic = str_to_bool(os.environ.get(prefix + "USE_DYNAMIC", "False")) == 1
 
409
  def to_dict(self):
410
  dynamo_config = copy.deepcopy(self.__dict__)
411
  dynamo_config["backend"] = dynamo_config["backend"].value.lower()
@@ -463,6 +473,7 @@ class DeepSpeedPlugin:
463
  default=None,
464
  metadata={"help": "Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3."},
465
  )
 
466
  def __post_init__(self):
467
  from .deepspeed import HfDeepSpeedConfig
468
  if self.gradient_accumulation_steps is None:
@@ -554,6 +565,7 @@ class DeepSpeedPlugin:
554
  if self.zero3_init_flag and not self.hf_ds_config.is_zero3():
555
  warnings.warn("DeepSpeed Zero3 Init flag is only applicable for ZeRO Stage 3. Setting it to False.")
556
  self.zero3_init_flag = False
 
557
  def fill_match(self, ds_key_long, mismatches=None, must_match=True, **kwargs):
558
  mismatches = [] if mismatches is None else mismatches
559
  config, ds_key = self.hf_ds_config.find_config_node(ds_key_long)
@@ -575,6 +587,7 @@ class DeepSpeedPlugin:
575
  if ds_val is not None and ds_key_long in kwargs:
576
  if ds_val != kwargs[ds_key_long]:
577
  mismatches.append(f"- ds {ds_key_long}={ds_val} vs arg {ds_key_long}={kwargs[ds_key_long]}")
 
578
  def deepspeed_config_process(self, prefix="", mismatches=None, config=None, must_match=True, **kwargs):
579
  """Process the DeepSpeed config with the values from the kwargs."""
580
  mismatches = [] if mismatches is None else mismatches
@@ -593,6 +606,7 @@ class DeepSpeedPlugin:
593
  "Please correct the following DeepSpeed config values that mismatch kwargs "
594
  f" values:\n{mismatches_msg}\nThe easiest method is to set these DeepSpeed config values to 'auto'."
595
  )
 
596
  def set_mixed_precision(self, mixed_precision):
597
  ds_config = self.deepspeed_config
598
  kwargs = {
@@ -616,6 +630,7 @@ class DeepSpeedPlugin:
616
  ds_config[dtype] = {"enabled": False}
617
  self.fill_match("fp16.enabled", must_match=False, **kwargs)
618
  self.fill_match("bf16.enabled", must_match=False, **kwargs)
 
619
  def set_deepspeed_weakref(self):
620
  from .imports import is_transformers_available
621
  if self.zero3_init_flag:
@@ -639,9 +654,11 @@ class DeepSpeedPlugin:
639
  else:
640
  from transformers.integrations import HfDeepSpeedConfig
641
  self.dschf = HfDeepSpeedConfig(ds_config) # keep this object alive # noqa
 
642
  def is_zero3_init_enabled(self):
643
  return self.zero3_init_flag
644
  @contextmanager
 
645
  def zero3_init_context_manager(self, enable=False):
646
  old = self.zero3_init_flag
647
  if old == enable:
@@ -654,6 +671,7 @@ class DeepSpeedPlugin:
654
  self.zero3_init_flag = old
655
  self.dschf = None
656
  self.set_deepspeed_weakref()
 
657
  def _deepspeed_config_checks(self):
658
  env_variable_names_to_ignore = [
659
  "ACCELERATE_GRADIENT_ACCUMULATION_STEPS",
@@ -784,6 +802,7 @@ class FullyShardedDataParallelPlugin:
784
  "for reduced memory usage."
785
  },
786
  )
 
787
  def __post_init__(self):
788
  from torch.distributed.fsdp.fully_sharded_data_parallel import BackwardPrefetch, CPUOffload, ShardingStrategy
789
  prefix = "FSDP_"
@@ -824,6 +843,7 @@ class FullyShardedDataParallelPlugin:
824
  )
825
  self.param_init_fn = lambda x: x.to_empty(device=device, recurse=False)
826
  @staticmethod
 
827
  def get_module_class_from_name(module, name):
828
  """
829
  Gets a class from a module by its name.
@@ -841,6 +861,7 @@ class FullyShardedDataParallelPlugin:
841
  module_class = FullyShardedDataParallelPlugin.get_module_class_from_name(child_module, name)
842
  if module_class is not None:
843
  return module_class
 
844
  def set_auto_wrap_policy(self, model):
845
  from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy, transformer_auto_wrap_policy
846
  default_transformer_cls_names_to_wrap = (
@@ -870,6 +891,7 @@ class FullyShardedDataParallelPlugin:
870
  self.auto_wrap_policy = functools.partial(
871
  size_based_auto_wrap_policy, min_num_params=min_num_params
872
  )
 
873
  def set_mixed_precision(self, mixed_precision):
874
  if mixed_precision == "fp16":
875
  dtype = torch.float16
@@ -880,6 +902,7 @@ class FullyShardedDataParallelPlugin:
880
  from torch.distributed.fsdp.fully_sharded_data_parallel import MixedPrecision
881
  if self.mixed_precision_policy is None:
882
  self.mixed_precision_policy = MixedPrecision(param_dtype=dtype, reduce_dtype=dtype, buffer_dtype=dtype)
 
883
  def set_state_dict_type(self, state_dict_type_policy):
884
  from torch.distributed.fsdp.fully_sharded_data_parallel import (
885
  FullOptimStateDictConfig,
@@ -1047,6 +1070,7 @@ class MegatronLMPlugin:
1047
  default=None,
1048
  metadata={"help": "Other Megatron-LM arguments. Please refer Megatron-LM"},
1049
  )
 
1050
  def __post_init__(self):
1051
  prefix = "MEGATRON_LM_"
1052
  if self.tp_degree is None:
@@ -1100,6 +1124,7 @@ class MegatronLMPlugin:
1100
  self.set_tensorboard_logging_options()
1101
  if self.other_megatron_args is not None:
1102
  self.megatron_lm_default_args.update(self.other_megatron_args)
 
1103
  def set_network_size_args(self, model, batch_data=None):
1104
  # Check if the model is either BERT, GPT or T5 else raise error
1105
  # set 'num_layers', 'hidden_size', 'num_attention_heads', 'max_position_embeddings'
@@ -1180,6 +1205,7 @@ class MegatronLMPlugin:
1180
  self.megatron_lm_default_args["model_return_dict"] = model.config.return_dict
1181
  if model_type_name == "bert":
1182
  self.megatron_lm_default_args["num_labels"] = num_labels
 
1183
  def set_mixed_precision(self, mixed_precision):
1184
  if mixed_precision == "fp16":
1185
  self.megatron_lm_default_args["fp16"] = True
@@ -1187,6 +1213,7 @@ class MegatronLMPlugin:
1187
  self.megatron_lm_default_args["bf16"] = True
1188
  self.DDP_impl = "local"
1189
  self.megatron_lm_default_args["DDP_impl"] = self.DDP_impl
 
1190
  def set_training_args(self, micro_batch_size, dp_degree):
1191
  self.data_parallel_size = dp_degree
1192
  self.micro_batch_size = micro_batch_size
@@ -1194,6 +1221,7 @@ class MegatronLMPlugin:
1194
  self.megatron_lm_default_args["data_parallel_size"] = self.data_parallel_size
1195
  self.megatron_lm_default_args["micro_batch_size"] = self.micro_batch_size
1196
  self.megatron_lm_default_args["global_batch_size"] = self.global_batch_size
 
1197
  def set_optimizer_type(self, optimizer):
1198
  optimizer_name = optimizer.__class__.__name__.lower()
1199
  if "adam" in optimizer_name:
@@ -1208,6 +1236,7 @@ class MegatronLMPlugin:
1208
  raise ValueError(f"Optimizer {optimizer_name} is not supported by Megatron-LM")
1209
  self.megatron_lm_default_args["lr"] = optimizer.defaults["lr"]
1210
  self.megatron_lm_default_args["weight_decay"] = optimizer.defaults["weight_decay"]
 
1211
  def set_scheduler_args(self, scheduler):
1212
  if self.train_iters is None:
1213
  self.train_iters = scheduler.total_num_steps // self.megatron_lm_default_args["data_parallel_size"]
@@ -1235,6 +1264,7 @@ class MegatronLMPlugin:
1235
  self.megatron_lm_default_args["start_weight_decay"] = self.start_weight_decay
1236
  self.megatron_lm_default_args["end_weight_decay"] = self.end_weight_decay
1237
  self.megatron_lm_default_args["min_lr"] = self.min_lr
 
1238
  def set_tensorboard_logging_options(self):
1239
  from megatron.arguments import _add_logging_args
1240
  parser = argparse.ArgumentParser()
@@ -1292,6 +1322,7 @@ class BnbQuantizationConfig:
1292
  default=None,
1293
  metadata={"help": "an explicit list of the modules that we don't quantize. We keep them in `torch.float32`."},
1294
  )
 
1295
  def __post_init__(self):
1296
  """
1297
  Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
 
5
  """
6
  Internal mixin that implements a `to_kwargs()` method for a dataclass.
7
  """
8
+
9
  def to_dict(self):
10
  return copy.deepcopy(self.__dict__)
11
+
12
  def to_kwargs(self):
13
  """
14
  Returns a dictionary containing the attributes with values different from the default of this class.
 
161
  amax_history_len: int = 1
162
  amax_compute_algo: AmaxComputeAlgorithm = "most_recent"
163
  override_linear_precision: Tuple[bool, bool, bool] = (False, False, False)
164
+
165
  def __post_init__(self):
166
  self.backend = self.backend.upper()
167
  if self.backend not in get_args(Backend):
 
178
  raise ValueError(f"`optimization_level` must be one of {' or '.join(get_args(OptLevel))}")
179
  class EnumWithContains(enum.EnumMeta):
180
  "A metaclass that adds the ability to check if `self` contains an item with the `in` operator"
181
+
182
  def __contains__(cls, item):
183
  try:
184
  cls(item)
 
187
  return True
188
  class BaseEnum(enum.Enum, metaclass=EnumWithContains):
189
  "An enum class that can get the value of an item with `str(Enum.key)`"
190
+
191
  def __str__(self):
192
  return self.value
193
  @classmethod
194
+
195
  def list(cls):
196
  "Method to list all the possible items in `cls`"
197
  return list(map(str, cls))
 
360
  )
361
  },
362
  )
363
+
364
  def set_directories(self, project_dir: str = None):
365
  "Sets `self.project_dir` and `self.logging_dir` to the appropriate values."
366
  self.project_dir = project_dir
367
  if self.logging_dir is None:
368
  self.logging_dir = project_dir
369
+
370
  def __post_init__(self):
371
  self.set_directories(self.project_dir)
372
  @dataclass
 
403
  dynamic: bool = field(default=None, metadata={"help": "Whether to use dynamic shape for tracing"})
404
  options: Any = field(default=None, metadata={"help": "A dictionary of options to pass to the backend."})
405
  disable: bool = field(default=False, metadata={"help": "Turn torch.compile() into a no-op for testing"})
406
+
407
  def __post_init__(self):
408
  prefix = "ACCELERATE_DYNAMO_"
409
  if self.backend is None:
 
415
  self.fullgraph = str_to_bool(os.environ.get(prefix + "USE_FULLGRAPH", "False")) == 1
416
  if self.dynamic is None:
417
  self.dynamic = str_to_bool(os.environ.get(prefix + "USE_DYNAMIC", "False")) == 1
418
+
419
  def to_dict(self):
420
  dynamo_config = copy.deepcopy(self.__dict__)
421
  dynamo_config["backend"] = dynamo_config["backend"].value.lower()
 
473
  default=None,
474
  metadata={"help": "Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3."},
475
  )
476
+
477
  def __post_init__(self):
478
  from .deepspeed import HfDeepSpeedConfig
479
  if self.gradient_accumulation_steps is None:
 
565
  if self.zero3_init_flag and not self.hf_ds_config.is_zero3():
566
  warnings.warn("DeepSpeed Zero3 Init flag is only applicable for ZeRO Stage 3. Setting it to False.")
567
  self.zero3_init_flag = False
568
+
569
  def fill_match(self, ds_key_long, mismatches=None, must_match=True, **kwargs):
570
  mismatches = [] if mismatches is None else mismatches
571
  config, ds_key = self.hf_ds_config.find_config_node(ds_key_long)
 
587
  if ds_val is not None and ds_key_long in kwargs:
588
  if ds_val != kwargs[ds_key_long]:
589
  mismatches.append(f"- ds {ds_key_long}={ds_val} vs arg {ds_key_long}={kwargs[ds_key_long]}")
590
+
591
  def deepspeed_config_process(self, prefix="", mismatches=None, config=None, must_match=True, **kwargs):
592
  """Process the DeepSpeed config with the values from the kwargs."""
593
  mismatches = [] if mismatches is None else mismatches
 
606
  "Please correct the following DeepSpeed config values that mismatch kwargs "
607
  f" values:\n{mismatches_msg}\nThe easiest method is to set these DeepSpeed config values to 'auto'."
608
  )
609
+
610
  def set_mixed_precision(self, mixed_precision):
611
  ds_config = self.deepspeed_config
612
  kwargs = {
 
630
  ds_config[dtype] = {"enabled": False}
631
  self.fill_match("fp16.enabled", must_match=False, **kwargs)
632
  self.fill_match("bf16.enabled", must_match=False, **kwargs)
633
+
634
  def set_deepspeed_weakref(self):
635
  from .imports import is_transformers_available
636
  if self.zero3_init_flag:
 
654
  else:
655
  from transformers.integrations import HfDeepSpeedConfig
656
  self.dschf = HfDeepSpeedConfig(ds_config) # keep this object alive # noqa
657
+
658
  def is_zero3_init_enabled(self):
659
  return self.zero3_init_flag
660
  @contextmanager
661
+
662
  def zero3_init_context_manager(self, enable=False):
663
  old = self.zero3_init_flag
664
  if old == enable:
 
671
  self.zero3_init_flag = old
672
  self.dschf = None
673
  self.set_deepspeed_weakref()
674
+
675
  def _deepspeed_config_checks(self):
676
  env_variable_names_to_ignore = [
677
  "ACCELERATE_GRADIENT_ACCUMULATION_STEPS",
 
802
  "for reduced memory usage."
803
  },
804
  )
805
+
806
  def __post_init__(self):
807
  from torch.distributed.fsdp.fully_sharded_data_parallel import BackwardPrefetch, CPUOffload, ShardingStrategy
808
  prefix = "FSDP_"
 
843
  )
844
  self.param_init_fn = lambda x: x.to_empty(device=device, recurse=False)
845
  @staticmethod
846
+
847
  def get_module_class_from_name(module, name):
848
  """
849
  Gets a class from a module by its name.
 
861
  module_class = FullyShardedDataParallelPlugin.get_module_class_from_name(child_module, name)
862
  if module_class is not None:
863
  return module_class
864
+
865
  def set_auto_wrap_policy(self, model):
866
  from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy, transformer_auto_wrap_policy
867
  default_transformer_cls_names_to_wrap = (
 
891
  self.auto_wrap_policy = functools.partial(
892
  size_based_auto_wrap_policy, min_num_params=min_num_params
893
  )
894
+
895
  def set_mixed_precision(self, mixed_precision):
896
  if mixed_precision == "fp16":
897
  dtype = torch.float16
 
902
  from torch.distributed.fsdp.fully_sharded_data_parallel import MixedPrecision
903
  if self.mixed_precision_policy is None:
904
  self.mixed_precision_policy = MixedPrecision(param_dtype=dtype, reduce_dtype=dtype, buffer_dtype=dtype)
905
+
906
  def set_state_dict_type(self, state_dict_type_policy):
907
  from torch.distributed.fsdp.fully_sharded_data_parallel import (
908
  FullOptimStateDictConfig,
 
1070
  default=None,
1071
  metadata={"help": "Other Megatron-LM arguments. Please refer Megatron-LM"},
1072
  )
1073
+
1074
  def __post_init__(self):
1075
  prefix = "MEGATRON_LM_"
1076
  if self.tp_degree is None:
 
1124
  self.set_tensorboard_logging_options()
1125
  if self.other_megatron_args is not None:
1126
  self.megatron_lm_default_args.update(self.other_megatron_args)
1127
+
1128
  def set_network_size_args(self, model, batch_data=None):
1129
  # Check if the model is either BERT, GPT or T5 else raise error
1130
  # set 'num_layers', 'hidden_size', 'num_attention_heads', 'max_position_embeddings'
 
1205
  self.megatron_lm_default_args["model_return_dict"] = model.config.return_dict
1206
  if model_type_name == "bert":
1207
  self.megatron_lm_default_args["num_labels"] = num_labels
1208
+
1209
  def set_mixed_precision(self, mixed_precision):
1210
  if mixed_precision == "fp16":
1211
  self.megatron_lm_default_args["fp16"] = True
 
1213
  self.megatron_lm_default_args["bf16"] = True
1214
  self.DDP_impl = "local"
1215
  self.megatron_lm_default_args["DDP_impl"] = self.DDP_impl
1216
+
1217
  def set_training_args(self, micro_batch_size, dp_degree):
1218
  self.data_parallel_size = dp_degree
1219
  self.micro_batch_size = micro_batch_size
 
1221
  self.megatron_lm_default_args["data_parallel_size"] = self.data_parallel_size
1222
  self.megatron_lm_default_args["micro_batch_size"] = self.micro_batch_size
1223
  self.megatron_lm_default_args["global_batch_size"] = self.global_batch_size
1224
+
1225
  def set_optimizer_type(self, optimizer):
1226
  optimizer_name = optimizer.__class__.__name__.lower()
1227
  if "adam" in optimizer_name:
 
1236
  raise ValueError(f"Optimizer {optimizer_name} is not supported by Megatron-LM")
1237
  self.megatron_lm_default_args["lr"] = optimizer.defaults["lr"]
1238
  self.megatron_lm_default_args["weight_decay"] = optimizer.defaults["weight_decay"]
1239
+
1240
  def set_scheduler_args(self, scheduler):
1241
  if self.train_iters is None:
1242
  self.train_iters = scheduler.total_num_steps // self.megatron_lm_default_args["data_parallel_size"]
 
1264
  self.megatron_lm_default_args["start_weight_decay"] = self.start_weight_decay
1265
  self.megatron_lm_default_args["end_weight_decay"] = self.end_weight_decay
1266
  self.megatron_lm_default_args["min_lr"] = self.min_lr
1267
+
1268
  def set_tensorboard_logging_options(self):
1269
  from megatron.arguments import _add_logging_args
1270
  parser = argparse.ArgumentParser()
 
1322
  default=None,
1323
  metadata={"help": "an explicit list of the modules that we don't quantize. We keep them in `torch.float32`."},
1324
  )
1325
+
1326
  def __post_init__(self):
1327
  """
1328
  Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
src/utils/deepspeed.py CHANGED
@@ -10,6 +10,7 @@ class HfDeepSpeedConfig:
10
  Args:
11
  config_file_or_dict (`Union[str, Dict]`): path to DeepSpeed config file or dict.
12
  """
 
13
  def __init__(self, config_file_or_dict):
14
  if isinstance(config_file_or_dict, dict):
15
  # Don't modify user's data should they want to reuse it (e.g. in tests), because once we
@@ -28,6 +29,7 @@ class HfDeepSpeedConfig:
28
  )
29
  self.config = config
30
  self.set_stage_and_offload()
 
31
  def set_stage_and_offload(self):
32
  # zero stage - this is done as early as possible, before model is created, to allow
33
  # ``is_deepspeed_zero3_enabled`` query and getting to the early deepspeed config object
@@ -45,6 +47,7 @@ class HfDeepSpeedConfig:
45
  )
46
  if len(offload_devices & offload_devices_valid) > 0:
47
  self._offload = True
 
48
  def find_config_node(self, ds_key_long):
49
  config = self.config
50
  # find the config node of interest if it exists
@@ -55,6 +58,7 @@ class HfDeepSpeedConfig:
55
  if config is None:
56
  return None, ds_key
57
  return config, ds_key
 
58
  def get_value(self, ds_key_long, default=None):
59
  """
60
  Returns the set value or `default` if no value is set
@@ -63,6 +67,7 @@ class HfDeepSpeedConfig:
63
  if config is None:
64
  return default
65
  return config.get(ds_key, default)
 
66
  def del_config_sub_tree(self, ds_key_long, must_exist=False):
67
  """
68
  Deletes a sub-section of the config file if it's found.
@@ -82,6 +87,7 @@ class HfDeepSpeedConfig:
82
  # if found remove it
83
  if parent_config is not None:
84
  parent_config.pop(node)
 
85
  def is_true(self, ds_key_long):
86
  """
87
  Returns `True`/``False` only if the value is set, always `False` otherwise. So use this method to ask the very
@@ -89,6 +95,7 @@ class HfDeepSpeedConfig:
89
  """
90
  value = self.get_value(ds_key_long)
91
  return False if value is None else bool(value)
 
92
  def is_false(self, ds_key_long):
93
  """
94
  Returns `True`/``False` only if the value is set, always `False` otherwise. So use this method to ask the very
@@ -96,10 +103,13 @@ class HfDeepSpeedConfig:
96
  """
97
  value = self.get_value(ds_key_long)
98
  return False if value is None else not bool(value)
 
99
  def is_zero2(self):
100
  return self._stage == 2
 
101
  def is_zero3(self):
102
  return self._stage == 3
 
103
  def is_offload(self):
104
  return self._offload
105
  class DeepSpeedEngineWrapper:
@@ -108,8 +118,10 @@ class DeepSpeedEngineWrapper:
108
  Args:
109
  engine (deepspeed.runtime.engine.DeepSpeedEngine): deepspeed engine to wrap
110
  """
 
111
  def __init__(self, engine):
112
  self.engine = engine
 
113
  def backward(self, loss, **kwargs):
114
  # runs backpropagation and handles mixed precision
115
  self.engine.backward(loss, **kwargs)
@@ -131,14 +143,18 @@ class DeepSpeedOptimizerWrapper(AcceleratedOptimizer):
131
  optimizer (`torch.optim.optimizer.Optimizer`):
132
  The optimizer to wrap.
133
  """
 
134
  def __init__(self, optimizer):
135
  super().__init__(optimizer, device_placement=False, scaler=None)
136
  self.__has_overflow__ = hasattr(self.optimizer, "overflow")
 
137
  def zero_grad(self, set_to_none=None):
138
  pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
 
139
  def step(self):
140
  pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
141
  @property
 
142
  def step_was_skipped(self):
143
  """Whether or not the optimizer step was done, or skipped because of gradient overflow."""
144
  if self.__has_overflow__:
@@ -152,8 +168,10 @@ class DeepSpeedSchedulerWrapper(AcceleratedScheduler):
152
  The scheduler to wrap.
153
  optimizers (one or a list of `torch.optim.Optimizer`):
154
  """
 
155
  def __init__(self, scheduler, optimizers):
156
  super().__init__(scheduler, optimizers)
 
157
  def step(self):
158
  pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
159
  class DummyOptim:
@@ -170,6 +188,7 @@ class DummyOptim:
170
  **kwargs:
171
  Other arguments.
172
  """
 
173
  def __init__(self, params, lr=0.001, weight_decay=0, **kwargs):
174
  self.params = params
175
  self.lr = lr
@@ -191,6 +210,7 @@ class DummyScheduler:
191
  **kwargs:
192
  Other arguments.
193
  """
 
194
  def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, lr_scheduler_callable=None, **kwargs):
195
  self.optimizer = optimizer
196
  self.total_num_steps = total_num_steps
 
10
  Args:
11
  config_file_or_dict (`Union[str, Dict]`): path to DeepSpeed config file or dict.
12
  """
13
+
14
  def __init__(self, config_file_or_dict):
15
  if isinstance(config_file_or_dict, dict):
16
  # Don't modify user's data should they want to reuse it (e.g. in tests), because once we
 
29
  )
30
  self.config = config
31
  self.set_stage_and_offload()
32
+
33
  def set_stage_and_offload(self):
34
  # zero stage - this is done as early as possible, before model is created, to allow
35
  # ``is_deepspeed_zero3_enabled`` query and getting to the early deepspeed config object
 
47
  )
48
  if len(offload_devices & offload_devices_valid) > 0:
49
  self._offload = True
50
+
51
  def find_config_node(self, ds_key_long):
52
  config = self.config
53
  # find the config node of interest if it exists
 
58
  if config is None:
59
  return None, ds_key
60
  return config, ds_key
61
+
62
  def get_value(self, ds_key_long, default=None):
63
  """
64
  Returns the set value or `default` if no value is set
 
67
  if config is None:
68
  return default
69
  return config.get(ds_key, default)
70
+
71
  def del_config_sub_tree(self, ds_key_long, must_exist=False):
72
  """
73
  Deletes a sub-section of the config file if it's found.
 
87
  # if found remove it
88
  if parent_config is not None:
89
  parent_config.pop(node)
90
+
91
  def is_true(self, ds_key_long):
92
  """
93
  Returns `True`/``False` only if the value is set, always `False` otherwise. So use this method to ask the very
 
95
  """
96
  value = self.get_value(ds_key_long)
97
  return False if value is None else bool(value)
98
+
99
  def is_false(self, ds_key_long):
100
  """
101
  Returns `True`/``False` only if the value is set, always `False` otherwise. So use this method to ask the very
 
103
  """
104
  value = self.get_value(ds_key_long)
105
  return False if value is None else not bool(value)
106
+
107
  def is_zero2(self):
108
  return self._stage == 2
109
+
110
  def is_zero3(self):
111
  return self._stage == 3
112
+
113
  def is_offload(self):
114
  return self._offload
115
  class DeepSpeedEngineWrapper:
 
118
  Args:
119
  engine (deepspeed.runtime.engine.DeepSpeedEngine): deepspeed engine to wrap
120
  """
121
+
122
  def __init__(self, engine):
123
  self.engine = engine
124
+
125
  def backward(self, loss, **kwargs):
126
  # runs backpropagation and handles mixed precision
127
  self.engine.backward(loss, **kwargs)
 
143
  optimizer (`torch.optim.optimizer.Optimizer`):
144
  The optimizer to wrap.
145
  """
146
+
147
  def __init__(self, optimizer):
148
  super().__init__(optimizer, device_placement=False, scaler=None)
149
  self.__has_overflow__ = hasattr(self.optimizer, "overflow")
150
+
151
  def zero_grad(self, set_to_none=None):
152
  pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
153
+
154
  def step(self):
155
  pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
156
  @property
157
+
158
  def step_was_skipped(self):
159
  """Whether or not the optimizer step was done, or skipped because of gradient overflow."""
160
  if self.__has_overflow__:
 
168
  The scheduler to wrap.
169
  optimizers (one or a list of `torch.optim.Optimizer`):
170
  """
171
+
172
  def __init__(self, scheduler, optimizers):
173
  super().__init__(scheduler, optimizers)
174
+
175
  def step(self):
176
  pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
177
  class DummyOptim:
 
188
  **kwargs:
189
  Other arguments.
190
  """
191
+
192
  def __init__(self, params, lr=0.001, weight_decay=0, **kwargs):
193
  self.params = params
194
  self.lr = lr
 
210
  **kwargs:
211
  Other arguments.
212
  """
213
+
214
  def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, lr_scheduler_callable=None, **kwargs):
215
  self.optimizer = optimizer
216
  self.total_num_steps = total_num_steps
src/utils/imports.py CHANGED
@@ -149,6 +149,7 @@ def is_mlflow_available():
149
  def is_mps_available():
150
  return is_torch_version(">=", "1.12") and torch.backends.mps.is_available() and torch.backends.mps.is_built()
151
  def is_ipex_available():
 
152
  def get_major_and_minor_from_version(full_version):
153
  return str(version.parse(full_version).major) + "." + str(version.parse(full_version).minor)
154
  _torch_version = importlib.metadata.version("torch")
 
149
  def is_mps_available():
150
  return is_torch_version(">=", "1.12") and torch.backends.mps.is_available() and torch.backends.mps.is_built()
151
  def is_ipex_available():
152
+
153
  def get_major_and_minor_from_version(full_version):
154
  return str(version.parse(full_version).major) + "." + str(version.parse(full_version).minor)
155
  _torch_version = importlib.metadata.version("torch")
src/utils/launch.py CHANGED
@@ -297,6 +297,7 @@ def _convert_nargs_to_dict(nargs: List[str]) -> Dict[str, str]:
297
  if len(nargs) < 0:
298
  return {}
299
  # helper function to infer type for argsparser
 
300
  def _infer_type(s):
301
  try:
302
  s = float(s)
@@ -449,10 +450,12 @@ class PrepareForLaunch:
449
  debug (`bool`, *optional*, defaults to `False`):
450
  Whether or not this is a debug launch.
451
  """
 
452
  def __init__(self, launcher, distributed_type="NO", debug=False):
453
  self.launcher = launcher
454
  self.distributed_type = DistributedType(distributed_type)
455
  self.debug = debug
 
456
  def __call__(self, index, *args):
457
  if self.debug:
458
  world_size = int(os.environ.get("WORLD_SIZE"))
 
297
  if len(nargs) < 0:
298
  return {}
299
  # helper function to infer type for argsparser
300
+
301
  def _infer_type(s):
302
  try:
303
  s = float(s)
 
450
  debug (`bool`, *optional*, defaults to `False`):
451
  Whether or not this is a debug launch.
452
  """
453
+
454
  def __init__(self, launcher, distributed_type="NO", debug=False):
455
  self.launcher = launcher
456
  self.distributed_type = DistributedType(distributed_type)
457
  self.debug = debug
458
+
459
  def __call__(self, index, *args):
460
  if self.debug:
461
  world_size = int(os.environ.get("WORLD_SIZE"))
src/utils/megatron_lm.py CHANGED
@@ -63,6 +63,7 @@ class MegatronLMDummyDataLoader:
63
  Args:
64
  **dataset_kwargs: Megatron data arguments.
65
  """
 
66
  def __init__(self, **dataset_kwargs):
67
  parser = argparse.ArgumentParser()
68
  parser = _add_data_args(parser)
@@ -71,12 +72,15 @@ class MegatronLMDummyDataLoader:
71
  self.dataset_args = vars(data_args[0])
72
  self.dataset_args.update(dataset_kwargs)
73
  self.dataset_args["megatron_dataset_flag"] = True
 
74
  def set_megatron_data_args(self):
75
  args = get_args()
76
  for key, value in self.dataset_args.items():
77
  setattr(args, key, value)
 
78
  def get_train_valid_test_datasets_provider(self):
79
- def train_valid_test_datasets_provider(train_val_test_num_samples):
 
80
  """Build train, valid, and test datasets."""
81
  args = get_args()
82
  dataset_args = {
@@ -121,6 +125,7 @@ class MegatronLMDummyDataLoader:
121
  train_ds, valid_ds, test_ds = build_train_valid_test_datasets(**dataset_args)
122
  return train_ds, valid_ds, test_ds
123
  return train_valid_test_datasets_provider
 
124
  def build_pretraining_data_loader(self, dataset, consumed_samples):
125
  if dataset is None:
126
  return None
@@ -151,8 +156,10 @@ class MegatronLMDummyDataLoader:
151
  return torch.utils.data.DataLoader(
152
  dataset, batch_sampler=batch_sampler, num_workers=args.num_workers, pin_memory=True
153
  )
 
154
  def build_train_valid_test_data_iterators(self):
155
- def cyclic_iter(iter):
 
156
  while True:
157
  for x in iter:
158
  yield x
@@ -275,13 +282,17 @@ def prepare_data_loader(accelerator, dataloader):
275
  return train_data_iterator, valid_data_iterator, test_data_iterator
276
  # optimizer utilities
277
  class MegatronLMOptimizerWrapper(AcceleratedOptimizer):
 
278
  def __init__(self, optimizer):
279
  super().__init__(optimizer, device_placement=False, scaler=None)
 
280
  def zero_grad(self, set_to_none=None):
281
  pass # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
 
282
  def step(self):
283
  pass # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
284
  @property
 
285
  def step_was_skipped(self):
286
  """Whether or not the optimizer step was done, or skipped because of gradient overflow."""
287
  return self.optimizer.skipped_iter
@@ -305,14 +316,17 @@ class MegatronLMDummyScheduler:
305
  **kwargs:
306
  Other arguments.
307
  """
 
308
  def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, **kwargs):
309
  self.optimizer = optimizer
310
  self.total_num_steps = total_num_steps
311
  self.warmup_num_steps = warmup_num_steps
312
  self.kwargs = kwargs
313
  class MegatronLMSchedulerWrapper(AcceleratedScheduler):
 
314
  def __init__(self, scheduler, optimizers):
315
  super().__init__(scheduler, optimizers)
 
316
  def step(self, *args, **kwargs):
317
  return # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
318
  def prepare_scheduler(accelerator, optimizer, scheduler):
@@ -321,13 +335,17 @@ def prepare_scheduler(accelerator, optimizer, scheduler):
321
  return scheduler
322
  class AbstractTrainStep(ABC):
323
  """Abstract class for batching, forward pass and loss handler."""
 
324
  def __init__(self, name):
325
  super().__init__()
326
  self.name = name
 
327
  def get_batch_func(self):
328
  pass
 
329
  def get_forward_step_func(self):
330
  pass
 
331
  def get_loss_func(self):
332
  pass
333
  class BertTrainStep(AbstractTrainStep):
@@ -336,6 +354,7 @@ class BertTrainStep(AbstractTrainStep):
336
  Args:
337
  args (`argparse.Namespace`): Megatron-LM arguments.
338
  """
 
339
  def __init__(self, args):
340
  super().__init__("BertTrainStep")
341
  self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
@@ -345,8 +364,10 @@ class BertTrainStep(AbstractTrainStep):
345
  self.model_output_class = None
346
  else:
347
  self.model_output_class = SequenceClassifierOutput
 
348
  def get_batch_func(self, megatron_dataset_flag):
349
- def get_batch_megatron(data_iterator):
 
350
  """Build the batch."""
351
  # Items and their type.
352
  keys = ["text", "types", "labels", "is_random", "loss_mask", "padding_mask"]
@@ -365,7 +386,8 @@ class BertTrainStep(AbstractTrainStep):
365
  lm_labels = data_b["labels"].long()
366
  padding_mask = data_b["padding_mask"].long()
367
  return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
368
- def get_batch_transformer(data_iterator):
 
369
  """Build the batch."""
370
  data = next(data_iterator)
371
  data = send_to_device(data, torch.cuda.current_device())
@@ -391,8 +413,10 @@ class BertTrainStep(AbstractTrainStep):
391
  return get_batch_megatron
392
  else:
393
  return get_batch_transformer
 
394
  def get_loss_func(self, pretraining_flag, num_labels):
395
- def loss_func_pretrain(loss_mask, sentence_order, output_tensor):
 
396
  lm_loss_, sop_logits = output_tensor
397
  lm_loss_ = lm_loss_.float()
398
  loss_mask = loss_mask.float()
@@ -407,7 +431,8 @@ class BertTrainStep(AbstractTrainStep):
407
  loss = lm_loss
408
  averaged_losses = average_losses_across_data_parallel_group([lm_loss])
409
  return loss, {"lm loss": averaged_losses[0]}
410
- def loss_func_finetune(labels, logits):
 
411
  if num_labels == 1:
412
  # We are doing regression
413
  loss_fct = MSELoss()
@@ -424,8 +449,10 @@ class BertTrainStep(AbstractTrainStep):
424
  return loss_func_pretrain
425
  else:
426
  return loss_func_finetune
 
427
  def get_forward_step_func(self, pretraining_flag, bert_binary_head):
428
- def forward_step(data_iterator, model):
 
429
  """Forward step."""
430
  tokens, types, sentence_order, loss_mask, labels, padding_mask = self.get_batch(data_iterator)
431
  if not bert_binary_head:
@@ -444,6 +471,7 @@ class GPTTrainStep(AbstractTrainStep):
444
  Args:
445
  args (`argparse.Namespace`): Megatron-LM arguments.
446
  """
 
447
  def __init__(self, args):
448
  super().__init__("GPTTrainStep")
449
  self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
@@ -460,8 +488,10 @@ class GPTTrainStep(AbstractTrainStep):
460
  self.model_output_class = None
461
  else:
462
  self.model_output_class = CausalLMOutputWithCrossAttentions
 
463
  def get_batch_func(self, megatron_dataset_flag):
464
- def get_batch_megatron(data_iterator):
 
465
  """Generate a batch"""
466
  # Items and their type.
467
  keys = ["text"]
@@ -481,7 +511,8 @@ class GPTTrainStep(AbstractTrainStep):
481
  tokens, self.eod_token, self.reset_position_ids, self.reset_attention_mask, self.eod_mask_loss
482
  )
483
  return tokens, labels, loss_mask, attention_mask, position_ids
484
- def get_batch_transformer(data_iterator):
 
485
  data = next(data_iterator)
486
  data = {"input_ids": data["input_ids"]}
487
  data = send_to_device(data, torch.cuda.current_device())
@@ -499,9 +530,11 @@ class GPTTrainStep(AbstractTrainStep):
499
  return get_batch_megatron
500
  else:
501
  return get_batch_transformer
 
502
  def get_loss_func(self):
503
  args = get_args()
504
- def loss_func(loss_mask, output_tensor):
 
505
  if args.return_logits:
506
  losses, logits = output_tensor
507
  else:
@@ -516,8 +549,10 @@ class GPTTrainStep(AbstractTrainStep):
516
  output_dict.update({"logits": logits})
517
  return loss, output_dict
518
  return loss_func
 
519
  def get_forward_step_func(self):
520
- def forward_step(data_iterator, model):
 
521
  """Forward step."""
522
  # Get the batch.
523
  tokens, labels, loss_mask, attention_mask, position_ids = self.get_batch(data_iterator)
@@ -530,6 +565,7 @@ class T5TrainStep(AbstractTrainStep):
530
  Args:
531
  args (`argparse.Namespace`): Megatron-LM arguments.
532
  """
 
533
  def __init__(self, args):
534
  super().__init__("T5TrainStep")
535
  self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
@@ -540,6 +576,7 @@ class T5TrainStep(AbstractTrainStep):
540
  else:
541
  self.model_output_class = Seq2SeqLMOutput
542
  @staticmethod
 
543
  def attn_mask_postprocess(attention_mask):
544
  # We create a 3D attention mask from a 2D tensor mask.
545
  # [b, 1, s]
@@ -552,11 +589,13 @@ class T5TrainStep(AbstractTrainStep):
552
  extended_attention_mask = attention_mask_bss < 0.5
553
  return extended_attention_mask
554
  @staticmethod
 
555
  def get_decoder_mask(seq_length, device):
556
  attention_mask = torch.tril(torch.ones((1, seq_length, seq_length), device=device))
557
  attention_mask = attention_mask < 0.5
558
  return attention_mask
559
  @staticmethod
 
560
  def get_enc_dec_mask(attention_mask, dec_seq_length, device):
561
  batch_size, _ = attention_mask.shape
562
  # We create a 3D attention mask from a 2D tensor mask.
@@ -567,8 +606,10 @@ class T5TrainStep(AbstractTrainStep):
567
  attention_mask_bss = attention_mask_bs1 * attention_mask_b1s
568
  extended_attention_mask = attention_mask_bss < 0.5
569
  return extended_attention_mask
 
570
  def get_batch_func(self, megatron_dataset_flag):
571
- def get_batch_megatron(data_iterator):
 
572
  """Build the batch."""
573
  keys = ["text_enc", "text_dec", "labels", "loss_mask", "enc_mask", "dec_mask", "enc_dec_mask"]
574
  datatype = torch.int64
@@ -587,7 +628,8 @@ class T5TrainStep(AbstractTrainStep):
587
  dec_mask = data_b["dec_mask"] < 0.5
588
  enc_dec_mask = data_b["enc_dec_mask"] < 0.5
589
  return tokens_enc, tokens_dec, loss_mask, labels, enc_mask, dec_mask, enc_dec_mask
590
- def get_batch_transformer(data_iterator):
 
591
  """Build the batch."""
592
  data = next(data_iterator)
593
  data = send_to_device(data, torch.cuda.current_device())
@@ -611,16 +653,20 @@ class T5TrainStep(AbstractTrainStep):
611
  return get_batch_megatron
612
  else:
613
  return get_batch_transformer
 
614
  def get_loss_func(self):
615
- def loss_func(loss_mask, output_tensor):
 
616
  lm_loss_ = output_tensor.float()
617
  lm_loss = torch.sum(lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
618
  loss = lm_loss
619
  averaged_losses = average_losses_across_data_parallel_group([lm_loss])
620
  return loss, {"lm loss": averaged_losses[0]}
621
  return loss_func
 
622
  def get_forward_step_func(self):
623
- def forward_step(data_iterator, model):
 
624
  """Forward step."""
625
  # Get the batch.
626
  tokens_enc, tokens_dec, loss_mask, lm_labels, enc_mask, dec_mask, enc_dec_mask = self.get_batch(
@@ -658,6 +704,7 @@ def initialize(accelerator, extra_args_provider=None, args_defaults={}):
658
  # tensorboard-writer, and timers.
659
  set_global_variables(args)
660
  # torch.distributed initialization
 
661
  def finish_mpu_init():
662
  args = get_args()
663
  # Pytorch distributed.
@@ -710,6 +757,7 @@ class MegatronEngine(torch.nn.Module):
710
  optimizer: Megatron-LM optimizer
711
  lr_scheduler: Megatron-LM lr scheduler
712
  """
 
713
  def __init__(self, accelerator, model, optimizer, scheduler):
714
  super(MegatronEngine, self).__init__()
715
  self.module = model
@@ -737,13 +785,16 @@ class MegatronEngine(torch.nn.Module):
737
  self.report_memory_flag = True
738
  if args.tensorboard_dir is not None:
739
  write_args_to_tensorboard()
 
740
  def train(self):
741
  for model_module in self.module:
742
  model_module.train()
743
  self.log_eval_results()
 
744
  def eval(self):
745
  for model_module in self.module:
746
  model_module.eval()
 
747
  def train_step(self, **batch_data):
748
  """
749
  Training step for Megatron-LM
@@ -829,6 +880,7 @@ class MegatronEngine(torch.nn.Module):
829
  loss_reduced[key] = torch.concat(losses_reduced_for_key)
830
  return loss_reduced, skipped_iter, grad_norm, num_zeros_in_grad
831
  return {}, skipped_iter, grad_norm, num_zeros_in_grad
 
832
  def eval_step(self, **batch_data):
833
  """
834
  Evaluation step for Megatron-LM
@@ -875,6 +927,7 @@ class MegatronEngine(torch.nn.Module):
875
  return loss_reduced
876
  else:
877
  return {}
 
878
  def forward(self, **batch_data):
879
  # During training, we use train_step()
880
  # model(**batch_data) performs following operations by delegating it to `self.train_step`:
@@ -933,6 +986,7 @@ class MegatronEngine(torch.nn.Module):
933
  if self.train_step_handler.model_output_class is not None:
934
  return self.train_step_handler.model_output_class(loss=loss, logits=logits)
935
  return loss
 
936
  def log_eval_results(self):
937
  args = get_args()
938
  if args.tensorboard_dir is None or self.iteration == 0:
@@ -957,6 +1011,7 @@ class MegatronEngine(torch.nn.Module):
957
  print_rank_last(string)
958
  print_rank_last("-" * length)
959
  self.eval_total_loss_dict = {}
 
960
  def save_checkpoint(self, output_dir):
961
  self.log_eval_results()
962
  args = get_args()
@@ -964,6 +1019,7 @@ class MegatronEngine(torch.nn.Module):
964
  torch.distributed.barrier()
965
  save_checkpoint(self.iteration, self.module, self.optimizer, self.scheduler)
966
  torch.distributed.barrier()
 
967
  def load_checkpoint(self, input_dir):
968
  args = get_args()
969
  args.load = input_dir
@@ -975,6 +1031,7 @@ class MegatronEngine(torch.nn.Module):
975
  self.iteration = iteration
976
  if args.fp16 and self.iteration == 0:
977
  self.optimizer.reload_model_params()
 
978
  def megatron_generate(
979
  self,
980
  inputs,
@@ -1143,6 +1200,7 @@ def gather_across_data_parallel_groups(tensor):
1143
  tensor (nested list/tuple/dictionary of `torch.Tensor`):
1144
  The data to gather across data parallel ranks.
1145
  """
 
1146
  def _gpu_gather_one(tensor):
1147
  if tensor.ndim == 0:
1148
  tensor = tensor.clone()[None]
 
63
  Args:
64
  **dataset_kwargs: Megatron data arguments.
65
  """
66
+
67
  def __init__(self, **dataset_kwargs):
68
  parser = argparse.ArgumentParser()
69
  parser = _add_data_args(parser)
 
72
  self.dataset_args = vars(data_args[0])
73
  self.dataset_args.update(dataset_kwargs)
74
  self.dataset_args["megatron_dataset_flag"] = True
75
+
76
  def set_megatron_data_args(self):
77
  args = get_args()
78
  for key, value in self.dataset_args.items():
79
  setattr(args, key, value)
80
+
81
  def get_train_valid_test_datasets_provider(self):
82
+
83
+ def train_valid_test_datasets_provider(train_val_test_num_samples):
84
  """Build train, valid, and test datasets."""
85
  args = get_args()
86
  dataset_args = {
 
125
  train_ds, valid_ds, test_ds = build_train_valid_test_datasets(**dataset_args)
126
  return train_ds, valid_ds, test_ds
127
  return train_valid_test_datasets_provider
128
+
129
  def build_pretraining_data_loader(self, dataset, consumed_samples):
130
  if dataset is None:
131
  return None
 
156
  return torch.utils.data.DataLoader(
157
  dataset, batch_sampler=batch_sampler, num_workers=args.num_workers, pin_memory=True
158
  )
159
+
160
  def build_train_valid_test_data_iterators(self):
161
+
162
+ def cyclic_iter(iter):
163
  while True:
164
  for x in iter:
165
  yield x
 
282
  return train_data_iterator, valid_data_iterator, test_data_iterator
283
  # optimizer utilities
284
  class MegatronLMOptimizerWrapper(AcceleratedOptimizer):
285
+
286
  def __init__(self, optimizer):
287
  super().__init__(optimizer, device_placement=False, scaler=None)
288
+
289
  def zero_grad(self, set_to_none=None):
290
  pass # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
291
+
292
  def step(self):
293
  pass # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
294
  @property
295
+
296
  def step_was_skipped(self):
297
  """Whether or not the optimizer step was done, or skipped because of gradient overflow."""
298
  return self.optimizer.skipped_iter
 
316
  **kwargs:
317
  Other arguments.
318
  """
319
+
320
  def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, **kwargs):
321
  self.optimizer = optimizer
322
  self.total_num_steps = total_num_steps
323
  self.warmup_num_steps = warmup_num_steps
324
  self.kwargs = kwargs
325
  class MegatronLMSchedulerWrapper(AcceleratedScheduler):
326
+
327
  def __init__(self, scheduler, optimizers):
328
  super().__init__(scheduler, optimizers)
329
+
330
  def step(self, *args, **kwargs):
331
  return # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
332
  def prepare_scheduler(accelerator, optimizer, scheduler):
 
335
  return scheduler
336
  class AbstractTrainStep(ABC):
337
  """Abstract class for batching, forward pass and loss handler."""
338
+
339
  def __init__(self, name):
340
  super().__init__()
341
  self.name = name
342
+
343
  def get_batch_func(self):
344
  pass
345
+
346
  def get_forward_step_func(self):
347
  pass
348
+
349
  def get_loss_func(self):
350
  pass
351
  class BertTrainStep(AbstractTrainStep):
 
354
  Args:
355
  args (`argparse.Namespace`): Megatron-LM arguments.
356
  """
357
+
358
  def __init__(self, args):
359
  super().__init__("BertTrainStep")
360
  self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
 
364
  self.model_output_class = None
365
  else:
366
  self.model_output_class = SequenceClassifierOutput
367
+
368
  def get_batch_func(self, megatron_dataset_flag):
369
+
370
+ def get_batch_megatron(data_iterator):
371
  """Build the batch."""
372
  # Items and their type.
373
  keys = ["text", "types", "labels", "is_random", "loss_mask", "padding_mask"]
 
386
  lm_labels = data_b["labels"].long()
387
  padding_mask = data_b["padding_mask"].long()
388
  return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
389
+
390
+ def get_batch_transformer(data_iterator):
391
  """Build the batch."""
392
  data = next(data_iterator)
393
  data = send_to_device(data, torch.cuda.current_device())
 
413
  return get_batch_megatron
414
  else:
415
  return get_batch_transformer
416
+
417
  def get_loss_func(self, pretraining_flag, num_labels):
418
+
419
+ def loss_func_pretrain(loss_mask, sentence_order, output_tensor):
420
  lm_loss_, sop_logits = output_tensor
421
  lm_loss_ = lm_loss_.float()
422
  loss_mask = loss_mask.float()
 
431
  loss = lm_loss
432
  averaged_losses = average_losses_across_data_parallel_group([lm_loss])
433
  return loss, {"lm loss": averaged_losses[0]}
434
+
435
+ def loss_func_finetune(labels, logits):
436
  if num_labels == 1:
437
  # We are doing regression
438
  loss_fct = MSELoss()
 
449
  return loss_func_pretrain
450
  else:
451
  return loss_func_finetune
452
+
453
  def get_forward_step_func(self, pretraining_flag, bert_binary_head):
454
+
455
+ def forward_step(data_iterator, model):
456
  """Forward step."""
457
  tokens, types, sentence_order, loss_mask, labels, padding_mask = self.get_batch(data_iterator)
458
  if not bert_binary_head:
 
471
  Args:
472
  args (`argparse.Namespace`): Megatron-LM arguments.
473
  """
474
+
475
  def __init__(self, args):
476
  super().__init__("GPTTrainStep")
477
  self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
 
488
  self.model_output_class = None
489
  else:
490
  self.model_output_class = CausalLMOutputWithCrossAttentions
491
+
492
  def get_batch_func(self, megatron_dataset_flag):
493
+
494
+ def get_batch_megatron(data_iterator):
495
  """Generate a batch"""
496
  # Items and their type.
497
  keys = ["text"]
 
511
  tokens, self.eod_token, self.reset_position_ids, self.reset_attention_mask, self.eod_mask_loss
512
  )
513
  return tokens, labels, loss_mask, attention_mask, position_ids
514
+
515
+ def get_batch_transformer(data_iterator):
516
  data = next(data_iterator)
517
  data = {"input_ids": data["input_ids"]}
518
  data = send_to_device(data, torch.cuda.current_device())
 
530
  return get_batch_megatron
531
  else:
532
  return get_batch_transformer
533
+
534
  def get_loss_func(self):
535
  args = get_args()
536
+
537
+ def loss_func(loss_mask, output_tensor):
538
  if args.return_logits:
539
  losses, logits = output_tensor
540
  else:
 
549
  output_dict.update({"logits": logits})
550
  return loss, output_dict
551
  return loss_func
552
+
553
  def get_forward_step_func(self):
554
+
555
+ def forward_step(data_iterator, model):
556
  """Forward step."""
557
  # Get the batch.
558
  tokens, labels, loss_mask, attention_mask, position_ids = self.get_batch(data_iterator)
 
565
  Args:
566
  args (`argparse.Namespace`): Megatron-LM arguments.
567
  """
568
+
569
  def __init__(self, args):
570
  super().__init__("T5TrainStep")
571
  self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
 
576
  else:
577
  self.model_output_class = Seq2SeqLMOutput
578
  @staticmethod
579
+
580
  def attn_mask_postprocess(attention_mask):
581
  # We create a 3D attention mask from a 2D tensor mask.
582
  # [b, 1, s]
 
589
  extended_attention_mask = attention_mask_bss < 0.5
590
  return extended_attention_mask
591
  @staticmethod
592
+
593
  def get_decoder_mask(seq_length, device):
594
  attention_mask = torch.tril(torch.ones((1, seq_length, seq_length), device=device))
595
  attention_mask = attention_mask < 0.5
596
  return attention_mask
597
  @staticmethod
598
+
599
  def get_enc_dec_mask(attention_mask, dec_seq_length, device):
600
  batch_size, _ = attention_mask.shape
601
  # We create a 3D attention mask from a 2D tensor mask.
 
606
  attention_mask_bss = attention_mask_bs1 * attention_mask_b1s
607
  extended_attention_mask = attention_mask_bss < 0.5
608
  return extended_attention_mask
609
+
610
  def get_batch_func(self, megatron_dataset_flag):
611
+
612
+ def get_batch_megatron(data_iterator):
613
  """Build the batch."""
614
  keys = ["text_enc", "text_dec", "labels", "loss_mask", "enc_mask", "dec_mask", "enc_dec_mask"]
615
  datatype = torch.int64
 
628
  dec_mask = data_b["dec_mask"] < 0.5
629
  enc_dec_mask = data_b["enc_dec_mask"] < 0.5
630
  return tokens_enc, tokens_dec, loss_mask, labels, enc_mask, dec_mask, enc_dec_mask
631
+
632
+ def get_batch_transformer(data_iterator):
633
  """Build the batch."""
634
  data = next(data_iterator)
635
  data = send_to_device(data, torch.cuda.current_device())
 
653
  return get_batch_megatron
654
  else:
655
  return get_batch_transformer
656
+
657
  def get_loss_func(self):
658
+
659
+ def loss_func(loss_mask, output_tensor):
660
  lm_loss_ = output_tensor.float()
661
  lm_loss = torch.sum(lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
662
  loss = lm_loss
663
  averaged_losses = average_losses_across_data_parallel_group([lm_loss])
664
  return loss, {"lm loss": averaged_losses[0]}
665
  return loss_func
666
+
667
  def get_forward_step_func(self):
668
+
669
+ def forward_step(data_iterator, model):
670
  """Forward step."""
671
  # Get the batch.
672
  tokens_enc, tokens_dec, loss_mask, lm_labels, enc_mask, dec_mask, enc_dec_mask = self.get_batch(
 
704
  # tensorboard-writer, and timers.
705
  set_global_variables(args)
706
  # torch.distributed initialization
707
+
708
  def finish_mpu_init():
709
  args = get_args()
710
  # Pytorch distributed.
 
757
  optimizer: Megatron-LM optimizer
758
  lr_scheduler: Megatron-LM lr scheduler
759
  """
760
+
761
  def __init__(self, accelerator, model, optimizer, scheduler):
762
  super(MegatronEngine, self).__init__()
763
  self.module = model
 
785
  self.report_memory_flag = True
786
  if args.tensorboard_dir is not None:
787
  write_args_to_tensorboard()
788
+
789
  def train(self):
790
  for model_module in self.module:
791
  model_module.train()
792
  self.log_eval_results()
793
+
794
  def eval(self):
795
  for model_module in self.module:
796
  model_module.eval()
797
+
798
  def train_step(self, **batch_data):
799
  """
800
  Training step for Megatron-LM
 
880
  loss_reduced[key] = torch.concat(losses_reduced_for_key)
881
  return loss_reduced, skipped_iter, grad_norm, num_zeros_in_grad
882
  return {}, skipped_iter, grad_norm, num_zeros_in_grad
883
+
884
  def eval_step(self, **batch_data):
885
  """
886
  Evaluation step for Megatron-LM
 
927
  return loss_reduced
928
  else:
929
  return {}
930
+
931
  def forward(self, **batch_data):
932
  # During training, we use train_step()
933
  # model(**batch_data) performs following operations by delegating it to `self.train_step`:
 
986
  if self.train_step_handler.model_output_class is not None:
987
  return self.train_step_handler.model_output_class(loss=loss, logits=logits)
988
  return loss
989
+
990
  def log_eval_results(self):
991
  args = get_args()
992
  if args.tensorboard_dir is None or self.iteration == 0:
 
1011
  print_rank_last(string)
1012
  print_rank_last("-" * length)
1013
  self.eval_total_loss_dict = {}
1014
+
1015
  def save_checkpoint(self, output_dir):
1016
  self.log_eval_results()
1017
  args = get_args()
 
1019
  torch.distributed.barrier()
1020
  save_checkpoint(self.iteration, self.module, self.optimizer, self.scheduler)
1021
  torch.distributed.barrier()
1022
+
1023
  def load_checkpoint(self, input_dir):
1024
  args = get_args()
1025
  args.load = input_dir
 
1031
  self.iteration = iteration
1032
  if args.fp16 and self.iteration == 0:
1033
  self.optimizer.reload_model_params()
1034
+
1035
  def megatron_generate(
1036
  self,
1037
  inputs,
 
1200
  tensor (nested list/tuple/dictionary of `torch.Tensor`):
1201
  The data to gather across data parallel ranks.
1202
  """
1203
+
1204
  def _gpu_gather_one(tensor):
1205
  if tensor.ndim == 0:
1206
  tensor = tensor.clone()[None]
src/utils/memory.py CHANGED
@@ -69,6 +69,7 @@ def find_executable_batch_size(function: callable = None, starting_batch_size: i
69
  if function is None:
70
  return functools.partial(find_executable_batch_size, starting_batch_size=starting_batch_size)
71
  batch_size = starting_batch_size
 
72
  def decorator(*args, **kwargs):
73
  nonlocal batch_size
74
  gc.collect()
 
69
  if function is None:
70
  return functools.partial(find_executable_batch_size, starting_batch_size=starting_batch_size)
71
  batch_size = starting_batch_size
72
+
73
  def decorator(*args, **kwargs):
74
  nonlocal batch_size
75
  gc.collect()
src/utils/modeling.py CHANGED
@@ -352,8 +352,10 @@ class FindTiedParametersResult(list):
352
  This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
353
  a list or on the `values` method as in the future this will be removed.
354
  """
 
355
  def __init__(self, *args, **kwargs):
356
  super().__init__(*args, **kwargs)
 
357
  def values(self):
358
  # TODO: at the next Transformers release (4.28.0) issue a deprecation warning here.
359
  return sum([x[1:] for x in self], [])
 
352
  This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
353
  a list or on the `values` method as in the future this will be removed.
354
  """
355
+
356
  def __init__(self, *args, **kwargs):
357
  super().__init__(*args, **kwargs)
358
+
359
  def values(self):
360
  # TODO: at the next Transformers release (4.28.0) issue a deprecation warning here.
361
  return sum([x[1:] for x in self], [])
src/utils/offload.py CHANGED
@@ -68,13 +68,17 @@ class PrefixedDataset(Mapping):
68
  dataset (`Mapping`): Any map with string keys.
69
  prefix (`str`): A prefix to add when trying to access any element in the underlying dataset.
70
  """
 
71
  def __init__(self, dataset: Mapping, prefix: str):
72
  self.dataset = dataset
73
  self.prefix = prefix
 
74
  def __getitem__(self, key):
75
  return self.dataset[f"{self.prefix}{key}"]
 
76
  def __iter__(self):
77
  return iter([key for key in self.dataset if key.startswith(self.prefix)])
 
78
  def __len__(self):
79
  return len(self.dataset)
80
  class OffloadedWeightsLoader(Mapping):
@@ -89,6 +93,7 @@ class OffloadedWeightsLoader(Mapping):
89
  A dictionary from weight name to their information (`dtype`/ `shape` or safetensors filename). Will default
90
  to the index saved in `save_folder`.
91
  """
 
92
  def __init__(
93
  self,
94
  state_dict: Dict[str, torch.Tensor] = None,
@@ -107,6 +112,7 @@ class OffloadedWeightsLoader(Mapping):
107
  self.all_keys = list(self.state_dict.keys())
108
  self.all_keys.extend([key for key in self.index if key not in self.all_keys])
109
  self.device = device
 
110
  def __getitem__(self, key: str):
111
  # State dict gets priority
112
  if key in self.state_dict:
@@ -129,8 +135,10 @@ class OffloadedWeightsLoader(Mapping):
129
  return tensor
130
  weight_file = os.path.join(self.save_folder, f"{key}.dat")
131
  return load_offloaded_weight(weight_file, weight_info)
 
132
  def __iter__(self):
133
  return iter(self.all_keys)
 
134
  def __len__(self):
135
  return len(self.all_keys)
136
  def extract_submodules_state_dict(state_dict: Dict[str, torch.Tensor], submodule_names: List[str]):
 
68
  dataset (`Mapping`): Any map with string keys.
69
  prefix (`str`): A prefix to add when trying to access any element in the underlying dataset.
70
  """
71
+
72
  def __init__(self, dataset: Mapping, prefix: str):
73
  self.dataset = dataset
74
  self.prefix = prefix
75
+
76
  def __getitem__(self, key):
77
  return self.dataset[f"{self.prefix}{key}"]
78
+
79
  def __iter__(self):
80
  return iter([key for key in self.dataset if key.startswith(self.prefix)])
81
+
82
  def __len__(self):
83
  return len(self.dataset)
84
  class OffloadedWeightsLoader(Mapping):
 
93
  A dictionary from weight name to their information (`dtype`/ `shape` or safetensors filename). Will default
94
  to the index saved in `save_folder`.
95
  """
96
+
97
  def __init__(
98
  self,
99
  state_dict: Dict[str, torch.Tensor] = None,
 
112
  self.all_keys = list(self.state_dict.keys())
113
  self.all_keys.extend([key for key in self.index if key not in self.all_keys])
114
  self.device = device
115
+
116
  def __getitem__(self, key: str):
117
  # State dict gets priority
118
  if key in self.state_dict:
 
135
  return tensor
136
  weight_file = os.path.join(self.save_folder, f"{key}.dat")
137
  return load_offloaded_weight(weight_file, weight_info)
138
+
139
  def __iter__(self):
140
  return iter(self.all_keys)
141
+
142
  def __len__(self):
143
  return len(self.all_keys)
144
  def extract_submodules_state_dict(state_dict: Dict[str, torch.Tensor], submodule_names: List[str]):
src/utils/operations.py CHANGED
@@ -130,6 +130,7 @@ def get_data_structure(data):
130
  Returns:
131
  The same data structure as `data` with [`~utils.TensorInformation`] instead of tensors.
132
  """
 
133
  def _get_data_structure(tensor):
134
  return TensorInformation(shape=tensor.shape, dtype=tensor.dtype)
135
  return recursively_apply(_get_data_structure, data)
@@ -142,6 +143,7 @@ def get_shape(data):
142
  Returns:
143
  The same data structure as `data` with lists of tensor shapes instead of tensors.
144
  """
 
145
  def _get_shape(tensor):
146
  return list(tensor.shape)
147
  return recursively_apply(_get_shape, data)
@@ -151,6 +153,7 @@ def initialize_tensors(data_structure):
151
  Returns:
152
  The same data structure as `data` with tensors instead of [`~utils.TensorInformation`].
153
  """
 
154
  def _initialize_tensor(tensor_info):
155
  return torch.empty(*tensor_info.shape, dtype=tensor_info.dtype)
156
  return recursively_apply(_initialize_tensor, data_structure, test_type=is_tensor_information)
@@ -180,6 +183,7 @@ def listify(data):
180
  Returns:
181
  The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
182
  """
 
183
  def _convert_to_list(tensor):
184
  tensor = tensor.detach().cpu()
185
  if tensor.dtype == torch.bfloat16:
@@ -190,6 +194,7 @@ def listify(data):
190
  return tensor.tolist()
191
  return recursively_apply(_convert_to_list, data)
192
  def _tpu_gather(tensor):
 
193
  def _tpu_gather_one(tensor):
194
  if tensor.ndim == 0:
195
  tensor = tensor.clone()[None]
@@ -206,6 +211,7 @@ def _gpu_gather(tensor):
206
  gather_op = torch.distributed.all_gather_into_tensor
207
  else:
208
  gather_op = torch.distributed._all_gather_base
 
209
  def _gpu_gather_one(tensor):
210
  if tensor.ndim == 0:
211
  tensor = tensor.clone()[None]
@@ -243,6 +249,7 @@ def verify_operation(function):
243
  Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
244
  """
245
  @wraps(function)
 
246
  def wrapper(*args, **kwargs):
247
  if PartialState().distributed_type == DistributedType.NO or not PartialState().debug:
248
  return function(*args, **kwargs)
@@ -275,6 +282,7 @@ def chained_operation(function):
275
  `DistributedOperationException`.
276
  """
277
  @wraps(function)
 
278
  def wrapper(*args, **kwargs):
279
  try:
280
  return function(*args, **kwargs)
@@ -321,6 +329,7 @@ def gather_object(object: Any):
321
  else:
322
  return object
323
  def _gpu_broadcast(data, src=0):
 
324
  def _gpu_broadcast_one(tensor, src=0):
325
  torch.distributed.broadcast(tensor, src=src)
326
  return tensor
@@ -377,6 +386,7 @@ def slice_tensors(data, tensor_slice, process_index=None, num_processes=None):
377
  Returns:
378
  The same data structure as `data` with all the tensors slices.
379
  """
 
380
  def _slice_tensor(tensor, tensor_slice):
381
  return tensor[tensor_slice]
382
  return recursively_apply(_slice_tensor, data, tensor_slice)
@@ -415,6 +425,7 @@ def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
415
  pad_first (`bool`, *optional*, defaults to `False`):
416
  Whether to pad at the beginning or the end.
417
  """
 
418
  def _pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
419
  if getattr(tensor, "is_nested", False):
420
  warnings.warn(
@@ -461,6 +472,7 @@ def reduce(tensor, reduction="mean", scale=1.0):
461
  Returns:
462
  The same data structure as `data` with all the tensors reduced.
463
  """
 
464
  def _reduce_across_processes(tensor, reduction="mean", scale=1.0):
465
  state = PartialState()
466
  cloned_tensor = tensor.clone()
@@ -485,8 +497,10 @@ def convert_to_fp32(tensor):
485
  Returns:
486
  The same data structure as `tensor` with all tensors that were in FP16/BF16 precision converted to FP32.
487
  """
 
488
  def _convert_to_fp32(tensor):
489
  return tensor.float()
 
490
  def _is_fp16_bf16_tensor(tensor):
491
  return hasattr(tensor, "dtype") and tensor.dtype in (torch.float16, torch.bfloat16)
492
  return recursively_apply(_convert_to_fp32, tensor, test_type=_is_fp16_bf16_tensor)
@@ -500,17 +514,21 @@ class ConvertOutputsToFp32:
500
  Returns:
501
  The same function as `model_forward` but with converted outputs.
502
  """
 
503
  def __init__(self, model_forward):
504
  self.model_forward = model_forward
505
  update_wrapper(self, model_forward)
 
506
  def __call__(self, *args, **kwargs):
507
  return convert_to_fp32(self.model_forward(*args, **kwargs))
 
508
  def __getstate__(self):
509
  raise pickle.PicklingError(
510
  "Cannot pickle a prepared model with automatic mixed precision, please unwrap the model with `Accelerator.unwrap_model(model)` before pickling it."
511
  )
512
  def convert_outputs_to_fp32(model_forward):
513
  model_forward = ConvertOutputsToFp32(model_forward)
 
514
  def forward(*args, **kwargs):
515
  return model_forward(*args, **kwargs)
516
  # To act like a decorator so that it can be popped when doing `extract_model_from_parallel`
 
130
  Returns:
131
  The same data structure as `data` with [`~utils.TensorInformation`] instead of tensors.
132
  """
133
+
134
  def _get_data_structure(tensor):
135
  return TensorInformation(shape=tensor.shape, dtype=tensor.dtype)
136
  return recursively_apply(_get_data_structure, data)
 
143
  Returns:
144
  The same data structure as `data` with lists of tensor shapes instead of tensors.
145
  """
146
+
147
  def _get_shape(tensor):
148
  return list(tensor.shape)
149
  return recursively_apply(_get_shape, data)
 
153
  Returns:
154
  The same data structure as `data` with tensors instead of [`~utils.TensorInformation`].
155
  """
156
+
157
  def _initialize_tensor(tensor_info):
158
  return torch.empty(*tensor_info.shape, dtype=tensor_info.dtype)
159
  return recursively_apply(_initialize_tensor, data_structure, test_type=is_tensor_information)
 
183
  Returns:
184
  The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
185
  """
186
+
187
  def _convert_to_list(tensor):
188
  tensor = tensor.detach().cpu()
189
  if tensor.dtype == torch.bfloat16:
 
194
  return tensor.tolist()
195
  return recursively_apply(_convert_to_list, data)
196
  def _tpu_gather(tensor):
197
+
198
  def _tpu_gather_one(tensor):
199
  if tensor.ndim == 0:
200
  tensor = tensor.clone()[None]
 
211
  gather_op = torch.distributed.all_gather_into_tensor
212
  else:
213
  gather_op = torch.distributed._all_gather_base
214
+
215
  def _gpu_gather_one(tensor):
216
  if tensor.ndim == 0:
217
  tensor = tensor.clone()[None]
 
249
  Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
250
  """
251
  @wraps(function)
252
+
253
  def wrapper(*args, **kwargs):
254
  if PartialState().distributed_type == DistributedType.NO or not PartialState().debug:
255
  return function(*args, **kwargs)
 
282
  `DistributedOperationException`.
283
  """
284
  @wraps(function)
285
+
286
  def wrapper(*args, **kwargs):
287
  try:
288
  return function(*args, **kwargs)
 
329
  else:
330
  return object
331
  def _gpu_broadcast(data, src=0):
332
+
333
  def _gpu_broadcast_one(tensor, src=0):
334
  torch.distributed.broadcast(tensor, src=src)
335
  return tensor
 
386
  Returns:
387
  The same data structure as `data` with all the tensors slices.
388
  """
389
+
390
  def _slice_tensor(tensor, tensor_slice):
391
  return tensor[tensor_slice]
392
  return recursively_apply(_slice_tensor, data, tensor_slice)
 
425
  pad_first (`bool`, *optional*, defaults to `False`):
426
  Whether to pad at the beginning or the end.
427
  """
428
+
429
  def _pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
430
  if getattr(tensor, "is_nested", False):
431
  warnings.warn(
 
472
  Returns:
473
  The same data structure as `data` with all the tensors reduced.
474
  """
475
+
476
  def _reduce_across_processes(tensor, reduction="mean", scale=1.0):
477
  state = PartialState()
478
  cloned_tensor = tensor.clone()
 
497
  Returns:
498
  The same data structure as `tensor` with all tensors that were in FP16/BF16 precision converted to FP32.
499
  """
500
+
501
  def _convert_to_fp32(tensor):
502
  return tensor.float()
503
+
504
  def _is_fp16_bf16_tensor(tensor):
505
  return hasattr(tensor, "dtype") and tensor.dtype in (torch.float16, torch.bfloat16)
506
  return recursively_apply(_convert_to_fp32, tensor, test_type=_is_fp16_bf16_tensor)
 
514
  Returns:
515
  The same function as `model_forward` but with converted outputs.
516
  """
517
+
518
  def __init__(self, model_forward):
519
  self.model_forward = model_forward
520
  update_wrapper(self, model_forward)
521
+
522
  def __call__(self, *args, **kwargs):
523
  return convert_to_fp32(self.model_forward(*args, **kwargs))
524
+
525
  def __getstate__(self):
526
  raise pickle.PicklingError(
527
  "Cannot pickle a prepared model with automatic mixed precision, please unwrap the model with `Accelerator.unwrap_model(model)` before pickling it."
528
  )
529
  def convert_outputs_to_fp32(model_forward):
530
  model_forward = ConvertOutputsToFp32(model_forward)
531
+
532
  def forward(*args, **kwargs):
533
  return model_forward(*args, **kwargs)
534
  # To act like a decorator so that it can be popped when doing `extract_model_from_parallel`