Remove whitespace
Browse files- README.md +5 -0
- src/accelerator.py +0 -453
- src/big_modeling.py +0 -64
- src/checkpointing.py +0 -24
- src/commands/accelerate_cli.py +0 -6
- src/commands/config/cluster.py +0 -27
- src/commands/config/config.py +0 -13
- src/commands/config/config_args.py +0 -20
- src/commands/config/config_utils.py +0 -18
- src/commands/config/default.py +0 -8
- src/commands/config/sagemaker.py +0 -18
- src/commands/config/update.py +0 -8
- src/commands/env.py +0 -14
- src/commands/estimate.py +0 -34
- src/commands/launch.py +0 -71
- src/commands/test.py +0 -10
- src/commands/tpu.py +0 -13
- src/data_loader.py +0 -132
- src/hooks.py +0 -79
- src/launchers.py +0 -29
- src/local_sgd.py +0 -14
- src/logging.py +0 -20
- src/optimizer.py +0 -27
- src/scheduler.py +0 -13
- src/state.py +0 -153
- src/tracking.py +0 -151
- src/utils/bnb.py +0 -51
- src/utils/constants.py +0 -3
- src/utils/dataclasses.py +0 -190
- src/utils/deepspeed.py +0 -43
- src/utils/environment.py +0 -18
- src/utils/fsdp_utils.py +0 -11
- src/utils/imports.py +0 -75
- src/utils/launch.py +0 -55
- src/utils/megatron_lm.py +0 -193
- src/utils/memory.py +0 -20
- src/utils/modeling.py +0 -199
- src/utils/offload.py +0 -33
- src/utils/operations.py +0 -123
- src/utils/other.py +0 -61
- src/utils/random.py +0 -7
- src/utils/rich.py +0 -2
- src/utils/torch_xla.py +0 -5
- src/utils/tqdm.py +0 -1
- src/utils/transformer_engine.py +0 -6
- src/utils/versions.py +0 -6
README.md
CHANGED
@@ -50,4 +50,9 @@ To:
|
|
50 |
```
|
51 |
Then remove all import statements (as we only care about the content).
|
52 |
|
|
|
|
|
|
|
|
|
|
|
53 |
**WARNING**: It is known that this will seperate out the `_inner()` in the source code and use it as a seperate function losing the context. Trying out with this issue for now.
|
|
|
50 |
```
|
51 |
Then remove all import statements (as we only care about the content).
|
52 |
|
53 |
+
Strip all blank spaces/whitespace:
|
54 |
+
```regex
|
55 |
+
^(?:[\t ]*(?:\r?\n|\r))+
|
56 |
+
```
|
57 |
+
|
58 |
**WARNING**: It is known that this will seperate out the `_inner()` in the source code and use it as a seperate function losing the context. Trying out with this issue for now.
|
src/accelerator.py
CHANGED
@@ -1,11 +1,7 @@
|
|
1 |
-
|
2 |
logger = get_logger(__name__)
|
3 |
-
|
4 |
-
|
5 |
class Accelerator:
|
6 |
"""
|
7 |
Creates an instance of an accelerator for distributed training (on multi-GPU, TPU) or mixed precision training.
|
8 |
-
|
9 |
Args:
|
10 |
device_placement (`bool`, *optional*, defaults to `True`):
|
11 |
Whether or not the accelerator should put objects on device (tensors yielded by the dataloader, model,
|
@@ -39,17 +35,14 @@ class Accelerator:
|
|
39 |
rng_types (list of `str` or [`~utils.RNGType`]):
|
40 |
The list of random number generators to synchronize at the beginning of each iteration in your prepared
|
41 |
dataloaders. Should be one or several of:
|
42 |
-
|
43 |
- `"torch"`: the base torch random number generator
|
44 |
- `"cuda"`: the CUDA random number generator (GPU only)
|
45 |
- `"xla"`: the XLA random number generator (TPU only)
|
46 |
- `"generator"`: the `torch.Generator` of the sampler (or batch sampler if there is no sampler in your
|
47 |
dataloader) or of the iterable dataset (if it exists) if the underlying dataset is of that type.
|
48 |
-
|
49 |
Will default to `["torch"]` for PyTorch versions <=1.5.1 and `["generator"]` for PyTorch versions >= 1.6.
|
50 |
log_with (list of `str`, [`~utils.LoggerType`] or [`~tracking.GeneralTracker`], *optional*):
|
51 |
A list of loggers to be setup for experiment tracking. Should be one or several of:
|
52 |
-
|
53 |
- `"all"`
|
54 |
- `"tensorboard"`
|
55 |
- `"wandb"`
|
@@ -80,9 +73,7 @@ class Accelerator:
|
|
80 |
gradient_accumulation_plugin (`GradientAccumulationPlugin`, *optional*):
|
81 |
A configuration for how gradient accumulation should be handled, if more tweaking than just the
|
82 |
`gradient_accumulation_steps` is needed.
|
83 |
-
|
84 |
**Available attributes:**
|
85 |
-
|
86 |
- **device** (`torch.device`) -- The device to use.
|
87 |
- **distributed_type** ([`~utils.DistributedType`]) -- The distributed training configuration.
|
88 |
- **local_process_index** (`int`) -- The process index on the current machine.
|
@@ -130,9 +121,7 @@ class Accelerator:
|
|
130 |
raise ValueError(
|
131 |
f"Unknown mixed_precision mode: {mixed_precision}. Choose between {PrecisionType.list()}"
|
132 |
)
|
133 |
-
|
134 |
dynamo_plugin = TorchDynamoPlugin() if dynamo_backend is None else TorchDynamoPlugin(backend=dynamo_backend)
|
135 |
-
|
136 |
if deepspeed_plugin is None: # init from env variables
|
137 |
deepspeed_plugin = (
|
138 |
DeepSpeedPlugin() if os.environ.get("ACCELERATE_USE_DEEPSPEED", "false") == "true" else None
|
@@ -147,19 +136,16 @@ class Accelerator:
|
|
147 |
raise ImportError("DeepSpeed is not installed => run `pip install deepspeed` or build it from source.")
|
148 |
if compare_versions("deepspeed", "<", "0.9.3"):
|
149 |
raise ImportError("DeepSpeed version must be >= 0.9.3. Please update DeepSpeed.")
|
150 |
-
|
151 |
mixed_precision = (
|
152 |
os.environ.get("ACCELERATE_MIXED_PRECISION", "no") if mixed_precision is None else mixed_precision
|
153 |
)
|
154 |
deepspeed_plugin.set_mixed_precision(mixed_precision)
|
155 |
deepspeed_plugin.set_deepspeed_weakref()
|
156 |
-
|
157 |
if os.environ.get("ACCELERATE_USE_FSDP", "false") == "true" or isinstance(
|
158 |
fsdp_plugin, FullyShardedDataParallelPlugin
|
159 |
):
|
160 |
if is_torch_version("<", FSDP_PYTORCH_VERSION):
|
161 |
raise ValueError(f"FSDP requires PyTorch >= {FSDP_PYTORCH_VERSION}")
|
162 |
-
|
163 |
if fsdp_plugin is None: # init from env variables
|
164 |
fsdp_plugin = (
|
165 |
FullyShardedDataParallelPlugin() if os.environ.get("ACCELERATE_USE_FSDP", "false") == "true" else None
|
@@ -168,7 +154,6 @@ class Accelerator:
|
|
168 |
if not isinstance(fsdp_plugin, FullyShardedDataParallelPlugin):
|
169 |
raise TypeError("`fsdp_plugin` must be a FullyShardedDataParallelPlugin object.")
|
170 |
os.environ["ACCELERATE_USE_FSDP"] = "true" # use FSDP if plugin is provided
|
171 |
-
|
172 |
if megatron_lm_plugin is None: # init from env variables
|
173 |
megatron_lm_plugin = (
|
174 |
MegatronLMPlugin() if os.environ.get("ACCELERATE_USE_MEGATRON_LM", "false") == "true" else None
|
@@ -177,11 +162,9 @@ class Accelerator:
|
|
177 |
if not isinstance(megatron_lm_plugin, MegatronLMPlugin):
|
178 |
raise TypeError("`megatron_lm_plugin` must be a MegatronLMPlugin object.")
|
179 |
os.environ["ACCELERATE_USE_MEGATRON_LM"] = "true" # use MegatronLM if plugin is provided
|
180 |
-
|
181 |
if megatron_lm_plugin:
|
182 |
if not is_megatron_lm_available():
|
183 |
raise ImportError("Megatron is not installed. please build it from source.")
|
184 |
-
|
185 |
# Kwargs handlers
|
186 |
self.ddp_handler = None
|
187 |
self.scaler_handler = None
|
@@ -220,7 +203,6 @@ class Accelerator:
|
|
220 |
self.autocast_handler = handler
|
221 |
if self.fp8_recipe_handler is None and mixed_precision == "fp8":
|
222 |
self.fp8_recipe_handler = FP8RecipeKwargs()
|
223 |
-
|
224 |
kwargs = self.init_handler.to_kwargs() if self.init_handler is not None else {}
|
225 |
self.state = AcceleratorState(
|
226 |
mixed_precision=mixed_precision,
|
@@ -232,19 +214,16 @@ class Accelerator:
|
|
232 |
_from_accelerator=True,
|
233 |
**kwargs,
|
234 |
)
|
235 |
-
|
236 |
trackers = filter_trackers(log_with, self.logging_dir)
|
237 |
if len(trackers) < 1 and log_with is not None:
|
238 |
warnings.warn(f"`log_with={log_with}` was passed but no supported trackers are currently installed.")
|
239 |
self.log_with = trackers
|
240 |
-
|
241 |
if (
|
242 |
(mixed_precision != "bf16")
|
243 |
and getattr(self.state, "downcast_bfloat", False)
|
244 |
and (self.state.distributedType != DistributedType.TPU)
|
245 |
):
|
246 |
raise ValueError("Can only use `downcast_bf16` when using `mixed_precision='bf16'` and on a TPU")
|
247 |
-
|
248 |
if gradient_accumulation_plugin is not None:
|
249 |
if gradient_accumulation_steps != 1:
|
250 |
raise ValueError(
|
@@ -263,13 +242,11 @@ class Accelerator:
|
|
263 |
raise ValueError(
|
264 |
"Gradient accumulation is not supported on TPU. Please set `gradient_accumulation_steps` to 1 and don't pass in a `GradientAccumulationPlugin` object."
|
265 |
)
|
266 |
-
|
267 |
self.device_placement = device_placement
|
268 |
self.split_batches = split_batches
|
269 |
self.dispatch_batches = dispatch_batches
|
270 |
self.even_batches = even_batches
|
271 |
self.step_scheduler_with_optimizer = step_scheduler_with_optimizer
|
272 |
-
|
273 |
# Mixed precision attributes
|
274 |
self.scaler = None
|
275 |
self.native_amp = False
|
@@ -285,13 +262,11 @@ class Accelerator:
|
|
285 |
kwargs = self.scaler_handler.to_kwargs() if self.scaler_handler is not None else {}
|
286 |
if self.distributed_type == DistributedType.FSDP:
|
287 |
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
|
288 |
-
|
289 |
self.scaler = ShardedGradScaler(**kwargs)
|
290 |
elif is_npu_available():
|
291 |
self.scaler = torch.npu.amp.GradScaler(**kwargs)
|
292 |
else:
|
293 |
self.scaler = torch.cuda.amp.GradScaler(**kwargs)
|
294 |
-
|
295 |
elif self.state.mixed_precision == "bf16" and self.distributed_type not in (
|
296 |
DistributedType.DEEPSPEED,
|
297 |
DistributedType.MEGATRON_LM,
|
@@ -302,80 +277,62 @@ class Accelerator:
|
|
302 |
self.native_amp = is_bf16_available(True)
|
303 |
if mixed_precision == "bf16" and not self.native_amp and not is_tpu_available():
|
304 |
raise ValueError(err.format(mode="bf16", requirement="PyTorch >= 1.10 and a supported device."))
|
305 |
-
|
306 |
# Start of internal step tracking
|
307 |
self.step = 0
|
308 |
-
|
309 |
# Internal references to the training objects
|
310 |
self._optimizers = []
|
311 |
self._models = []
|
312 |
self._schedulers = []
|
313 |
self._dataloaders = []
|
314 |
self._custom_objects = []
|
315 |
-
|
316 |
# Hooks
|
317 |
self._load_model_state_pre_hook = OrderedDict()
|
318 |
self._save_model_state_pre_hook = OrderedDict()
|
319 |
-
|
320 |
# RNG Types
|
321 |
self.rng_types = rng_types
|
322 |
if self.rng_types is None:
|
323 |
self.rng_types = ["generator"]
|
324 |
-
|
325 |
# Set a flag tensor for early stopping and other breakpoints
|
326 |
self.flag_tensor = None
|
327 |
-
|
328 |
check_os_kernel()
|
329 |
-
|
330 |
@property
|
331 |
def use_distributed(self):
|
332 |
"""
|
333 |
Whether the Accelerator is configured for distributed training
|
334 |
"""
|
335 |
return self.state.use_distributed
|
336 |
-
|
337 |
@property
|
338 |
def distributed_type(self):
|
339 |
return self.state.distributed_type
|
340 |
-
|
341 |
@property
|
342 |
def num_processes(self):
|
343 |
return self.state.num_processes
|
344 |
-
|
345 |
@property
|
346 |
def process_index(self):
|
347 |
return self.state.process_index
|
348 |
-
|
349 |
@property
|
350 |
def local_process_index(self):
|
351 |
return self.state.local_process_index
|
352 |
-
|
353 |
@property
|
354 |
def device(self):
|
355 |
return self.state.device
|
356 |
-
|
357 |
@property
|
358 |
def project_dir(self):
|
359 |
return self.project_configuration.project_dir
|
360 |
-
|
361 |
@property
|
362 |
def logging_dir(self):
|
363 |
return self.project_configuration.logging_dir
|
364 |
-
|
365 |
@property
|
366 |
def save_iteration(self):
|
367 |
return self.project_configuration.iteration
|
368 |
-
|
369 |
@property
|
370 |
def is_main_process(self):
|
371 |
"""True for one process only."""
|
372 |
return self.state.is_main_process
|
373 |
-
|
374 |
@property
|
375 |
def is_local_main_process(self):
|
376 |
"""True for one process per server."""
|
377 |
return self.state.is_local_main_process
|
378 |
-
|
379 |
@property
|
380 |
def use_fp16(self):
|
381 |
warnings.warn(
|
@@ -384,23 +341,18 @@ class Accelerator:
|
|
384 |
FutureWarning,
|
385 |
)
|
386 |
return self.mixed_precision != "no"
|
387 |
-
|
388 |
@property
|
389 |
def is_last_process(self):
|
390 |
return self.process_index == self.num_processes - 1
|
391 |
-
|
392 |
@property
|
393 |
def mixed_precision(self):
|
394 |
return self.state.mixed_precision
|
395 |
-
|
396 |
@contextmanager
|
397 |
def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
|
398 |
"""
|
399 |
Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
|
400 |
distributed inference, such as with different prompts.
|
401 |
-
|
402 |
Note that when using a `dict`, all keys need to have the same number of elements.
|
403 |
-
|
404 |
Args:
|
405 |
inputs (`list`, `tuple`, `torch.Tensor`, or `dict` of `list`/`tuple`/`torch.Tensor`):
|
406 |
The input to split between processes.
|
@@ -409,13 +361,10 @@ class Accelerator:
|
|
409 |
number of elements. Useful when trying to perform actions such as `Accelerator.gather()` on the outputs
|
410 |
or passing in less inputs than there are processes. If so, just remember to drop the padded elements
|
411 |
afterwards.
|
412 |
-
|
413 |
Example:
|
414 |
-
|
415 |
```python
|
416 |
# Assume there are two processes
|
417 |
from accelerate import Accelerator
|
418 |
-
|
419 |
accelerator = Accelerator()
|
420 |
with accelerator.split_between_processes(["A", "B", "C"]) as inputs:
|
421 |
print(inputs)
|
@@ -423,7 +372,6 @@ class Accelerator:
|
|
423 |
["A", "B"]
|
424 |
# Process 1
|
425 |
["C"]
|
426 |
-
|
427 |
with accelerator.split_between_processes(["A", "B", "C"], apply_padding=True) as inputs:
|
428 |
print(inputs)
|
429 |
# Process 0
|
@@ -434,28 +382,19 @@ class Accelerator:
|
|
434 |
"""
|
435 |
with PartialState().split_between_processes(inputs, apply_padding=apply_padding) as inputs:
|
436 |
yield inputs
|
437 |
-
|
438 |
def on_main_process(self, function: Callable[..., Any] = None):
|
439 |
"""
|
440 |
A decorator that will run the decorated function on the main process only. Can also be called using the
|
441 |
`PartialState` class.
|
442 |
-
|
443 |
Args:
|
444 |
function (`Callable`): The function to decorate.
|
445 |
-
|
446 |
Example:
|
447 |
-
|
448 |
```python
|
449 |
>>> from accelerate import Accelerator
|
450 |
-
|
451 |
>>> accelerator = Accelerator()
|
452 |
-
|
453 |
-
|
454 |
>>> @accelerator.on_main_process
|
455 |
... def print_something():
|
456 |
... print("This will be printed by process 0 only.")
|
457 |
-
|
458 |
-
|
459 |
>>> print_something()
|
460 |
"This will be printed by process 0 only"
|
461 |
```
|
@@ -468,33 +407,23 @@ class Accelerator:
|
|
468 |
raise ValueError(
|
469 |
"The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
|
470 |
)
|
471 |
-
|
472 |
def _inner(*args, **kwargs):
|
473 |
return PartialState().on_main_process(function)(*args, **kwargs)
|
474 |
-
|
475 |
return _inner
|
476 |
-
|
477 |
def on_local_main_process(self, function: Callable[..., Any] = None):
|
478 |
"""
|
479 |
A decorator that will run the decorated function on the local main process only. Can also be called using the
|
480 |
`PartialState` class.
|
481 |
-
|
482 |
Args:
|
483 |
function (`Callable`): The function to decorate.
|
484 |
-
|
485 |
Example:
|
486 |
```python
|
487 |
# Assume we have 2 servers with 4 processes each.
|
488 |
from accelerate import Accelerator
|
489 |
-
|
490 |
accelerator = Accelerator()
|
491 |
-
|
492 |
-
|
493 |
@accelerator.on_local_main_process
|
494 |
def print_something():
|
495 |
print("This will be printed by process 0 only on each server.")
|
496 |
-
|
497 |
-
|
498 |
print_something()
|
499 |
# On server 1:
|
500 |
"This will be printed by process 0 only"
|
@@ -510,33 +439,23 @@ class Accelerator:
|
|
510 |
raise ValueError(
|
511 |
"The `on_local_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
|
512 |
)
|
513 |
-
|
514 |
def _inner(*args, **kwargs):
|
515 |
return PartialState().on_local_main_process(function)(*args, **kwargs)
|
516 |
-
|
517 |
return _inner
|
518 |
-
|
519 |
def on_last_process(self, function: Callable[..., Any]):
|
520 |
"""
|
521 |
A decorator that will run the decorated function on the last process only. Can also be called using the
|
522 |
`PartialState` class.
|
523 |
-
|
524 |
Args:
|
525 |
function (`Callable`): The function to decorate.
|
526 |
-
|
527 |
Example:
|
528 |
```python
|
529 |
# Assume we have 4 processes.
|
530 |
from accelerate import Accelerator
|
531 |
-
|
532 |
accelerator = Accelerator()
|
533 |
-
|
534 |
-
|
535 |
@accelerator.on_last_process
|
536 |
def print_something():
|
537 |
print(f"Printed on process {accelerator.process_index}")
|
538 |
-
|
539 |
-
|
540 |
print_something()
|
541 |
"Printed on process 3"
|
542 |
```
|
@@ -549,36 +468,26 @@ class Accelerator:
|
|
549 |
raise ValueError(
|
550 |
"The `on_last_process` decorator must be called with a function on an instantiated `Accelerator` object."
|
551 |
)
|
552 |
-
|
553 |
def _inner(*args, **kwargs):
|
554 |
return PartialState().on_last_process(function)(*args, **kwargs)
|
555 |
-
|
556 |
return _inner
|
557 |
-
|
558 |
def on_process(self, function: Callable[..., Any] = None, process_index: int = None):
|
559 |
"""
|
560 |
A decorator that will run the decorated function on a given process index only. Can also be called using the
|
561 |
`PartialState` class.
|
562 |
-
|
563 |
Args:
|
564 |
function (`Callable`, `optional`):
|
565 |
The function to decorate.
|
566 |
process_index (`int`, `optional`):
|
567 |
The index of the process on which to run the function.
|
568 |
-
|
569 |
Example:
|
570 |
```python
|
571 |
# Assume we have 4 processes.
|
572 |
from accelerate import Accelerator
|
573 |
-
|
574 |
accelerator = Accelerator()
|
575 |
-
|
576 |
-
|
577 |
@accelerator.on_process(process_index=2)
|
578 |
def print_something():
|
579 |
print(f"Printed on process {accelerator.process_index}")
|
580 |
-
|
581 |
-
|
582 |
print_something()
|
583 |
"Printed on process 2"
|
584 |
```
|
@@ -594,36 +503,26 @@ class Accelerator:
|
|
594 |
raise ValueError(
|
595 |
"The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
|
596 |
)
|
597 |
-
|
598 |
def _inner(*args, **kwargs):
|
599 |
return PartialState().on_process(function, process_index)(*args, **kwargs)
|
600 |
-
|
601 |
return _inner
|
602 |
-
|
603 |
def on_local_process(self, function: Callable[..., Any] = None, local_process_index: int = None):
|
604 |
"""
|
605 |
A decorator that will run the decorated function on a given local process index only. Can also be called using
|
606 |
the `PartialState` class.
|
607 |
-
|
608 |
Args:
|
609 |
function (`Callable`, *optional*):
|
610 |
The function to decorate.
|
611 |
local_process_index (`int`, *optional*):
|
612 |
The index of the local process on which to run the function.
|
613 |
-
|
614 |
Example:
|
615 |
```python
|
616 |
# Assume we have 2 servers with 4 processes each.
|
617 |
from accelerate import Accelerator
|
618 |
-
|
619 |
accelerator = Accelerator()
|
620 |
-
|
621 |
-
|
622 |
@accelerator.on_local_process(local_process_index=2)
|
623 |
def print_something():
|
624 |
print(f"Printed on process {accelerator.local_process_index}")
|
625 |
-
|
626 |
-
|
627 |
print_something()
|
628 |
# On server 1:
|
629 |
"Printed on process 2"
|
@@ -642,24 +541,17 @@ class Accelerator:
|
|
642 |
raise ValueError(
|
643 |
"The `on_main_process` decorator must be called with a function on an instantiated `Accelerator` object."
|
644 |
)
|
645 |
-
|
646 |
def _inner(*args, **kwargs):
|
647 |
return PartialState().on_local_process(function, local_process_index)(*args, **kwargs)
|
648 |
-
|
649 |
return _inner
|
650 |
-
|
651 |
@contextmanager
|
652 |
def main_process_first(self):
|
653 |
"""
|
654 |
Lets the main process go first inside a with block.
|
655 |
-
|
656 |
The other processes will enter the with block after the main process exits.
|
657 |
-
|
658 |
Example:
|
659 |
-
|
660 |
```python
|
661 |
>>> from accelerate import Accelerator
|
662 |
-
|
663 |
>>> accelerator = Accelerator()
|
664 |
>>> with accelerator.main_process_first():
|
665 |
... # This will be printed first by process 0 then in a seemingly
|
@@ -669,19 +561,14 @@ class Accelerator:
|
|
669 |
"""
|
670 |
with self.state.main_process_first():
|
671 |
yield
|
672 |
-
|
673 |
@contextmanager
|
674 |
def local_main_process_first(self):
|
675 |
"""
|
676 |
Lets the local main process go inside a with block.
|
677 |
-
|
678 |
The other processes will enter the with block after the main process exits.
|
679 |
-
|
680 |
Example:
|
681 |
-
|
682 |
```python
|
683 |
>>> from accelerate import Accelerator
|
684 |
-
|
685 |
>>> accelerator = Accelerator()
|
686 |
>>> with accelerator.local_main_process_first():
|
687 |
... # This will be printed first by local process 0 then in a seemingly
|
@@ -691,29 +578,22 @@ class Accelerator:
|
|
691 |
"""
|
692 |
with self.state.local_main_process_first():
|
693 |
yield
|
694 |
-
|
695 |
@contextmanager
|
696 |
def no_sync(self, model):
|
697 |
"""
|
698 |
A context manager to disable gradient synchronizations across DDP processes by calling
|
699 |
`torch.nn.parallel.DistributedDataParallel.no_sync`.
|
700 |
-
|
701 |
If `model` is not in DDP, this context manager does nothing
|
702 |
-
|
703 |
Args:
|
704 |
model (`torch.nn.Module`):
|
705 |
PyTorch Module that was prepared with `Accelerator.prepare`
|
706 |
-
|
707 |
Example:
|
708 |
-
|
709 |
```python
|
710 |
>>> from accelerate import Accelerator
|
711 |
-
|
712 |
>>> accelerator = Accelerator()
|
713 |
>>> dataloader, model, optimizer = accelerator.prepare(dataloader, model, optimizer)
|
714 |
>>> input_a = next(iter(dataloader))
|
715 |
>>> input_b = next(iter(dataloader))
|
716 |
-
|
717 |
>>> with accelerator.no_sync():
|
718 |
... outputs = model(input_a)
|
719 |
... loss = loss_func(outputs)
|
@@ -729,30 +609,22 @@ class Accelerator:
|
|
729 |
context = contextlib.nullcontext
|
730 |
if self.use_distributed:
|
731 |
context = getattr(model, "no_sync", context)
|
732 |
-
|
733 |
with context():
|
734 |
yield
|
735 |
-
|
736 |
@staticmethod
|
737 |
@contextmanager
|
738 |
def trigger_sync_in_backward(model):
|
739 |
"""Trigger the sync of the gradients in the next backward pass of the model after multiple forward passes under
|
740 |
`Accelerator.no_sync` (only applicable in multi-GPU scenarios).
|
741 |
-
|
742 |
If the script is not launched in distributed mode, this context manager does nothing.
|
743 |
-
|
744 |
Args:
|
745 |
model (`torch.nn.Module`):
|
746 |
The model for which to trigger the gradient synchronization.
|
747 |
-
|
748 |
Example:
|
749 |
-
|
750 |
```python
|
751 |
>>> from accelerate import Accelerator
|
752 |
-
|
753 |
>>> accelerator = Accelerator()
|
754 |
>>> dataloader, model, optimizer = accelerator.prepare(dataloader, model, optimizer)
|
755 |
-
|
756 |
>>> with accelerator.no_sync():
|
757 |
... loss_a = loss_func(model(input_a)) # first forward pass
|
758 |
... loss_b = loss_func(model(input_b)) # second forward pass
|
@@ -766,10 +638,8 @@ class Accelerator:
|
|
766 |
if not isinstance(model, torch.nn.parallel.DistributedDataParallel):
|
767 |
yield
|
768 |
return
|
769 |
-
|
770 |
old_require_backward_grad_sync = model.require_backward_grad_sync
|
771 |
old_require_forward_param_sync = model.require_forward_param_sync
|
772 |
-
|
773 |
# EXPERIMENTAL: This will force grad sync during `backward()`, but it is unknown if it breaks other DDP features.
|
774 |
# https://github.com/pytorch/pytorch/blob/e1502c0cdbfd17548c612f25d5a65b1e4b86224d/torch/nn/parallel/distributed.py#L1453-L1466
|
775 |
model.require_backward_grad_sync = True
|
@@ -781,7 +651,6 @@ class Accelerator:
|
|
781 |
finally:
|
782 |
model.require_backward_grad_sync = old_require_backward_grad_sync
|
783 |
model.require_forward_param_sync = old_require_forward_param_sync
|
784 |
-
|
785 |
def _do_sync(self):
|
786 |
"Sets the right `sync_gradients` context and either resets or increases `self.step`"
|
787 |
if self.gradient_state.sync_with_dataloader and self.gradient_state.end_of_dataloader:
|
@@ -790,41 +659,31 @@ class Accelerator:
|
|
790 |
else:
|
791 |
self.step += 1
|
792 |
self.gradient_state._set_sync_gradients((self.step % self.gradient_state.num_steps) == 0)
|
793 |
-
|
794 |
@property
|
795 |
def sync_gradients(self):
|
796 |
return self.gradient_state.sync_gradients
|
797 |
-
|
798 |
@sync_gradients.setter
|
799 |
def sync_gradients(self, sync_gradients):
|
800 |
self.gradient_state.sync_gradients = sync_gradients
|
801 |
-
|
802 |
@property
|
803 |
def gradient_accumulation_steps(self):
|
804 |
return self.gradient_state.num_steps
|
805 |
-
|
806 |
@gradient_accumulation_steps.setter
|
807 |
def gradient_accumulation_steps(self, gradient_accumulation_steps):
|
808 |
self.gradient_state.plugin_kwargs.update({"num_steps": gradient_accumulation_steps})
|
809 |
-
|
810 |
@contextmanager
|
811 |
def accumulate(self, *models):
|
812 |
"""
|
813 |
A context manager that will lightly wrap around and perform gradient accumulation automatically
|
814 |
-
|
815 |
Args:
|
816 |
*models (list of `torch.nn.Module`):
|
817 |
PyTorch Modules that were prepared with `Accelerator.prepare`. Models passed to `accumulate()` will
|
818 |
skip gradient syncing during backward pass in distributed training
|
819 |
-
|
820 |
Example:
|
821 |
-
|
822 |
```python
|
823 |
>>> from accelerate import Accelerator
|
824 |
-
|
825 |
>>> accelerator = Accelerator(gradient_accumulation_steps=1)
|
826 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
827 |
-
|
828 |
>>> for input, output in dataloader:
|
829 |
... with accelerator.accumulate(model):
|
830 |
... outputs = model(input)
|
@@ -840,14 +699,12 @@ class Accelerator:
|
|
840 |
for m in models:
|
841 |
cm_stack.enter_context(contextlib.nullcontext() if self.sync_gradients else self.no_sync(m))
|
842 |
yield
|
843 |
-
|
844 |
@contextmanager
|
845 |
def join_uneven_inputs(self, joinables, even_batches=None):
|
846 |
"""
|
847 |
A context manager that facilitates distributed training or evaluation on uneven inputs, which acts as a wrapper
|
848 |
around `torch.distributed.algorithms.join`. This is useful when the total batch size does not evenly divide the
|
849 |
length of the dataset.
|
850 |
-
|
851 |
Args:
|
852 |
joinables (`list[torch.distributed.algorithms.Joinable]`):
|
853 |
A list of models or optimizers that subclass `torch.distributed.algorithms.Joinable`. Most commonly, a
|
@@ -855,28 +712,18 @@ class Accelerator:
|
|
855 |
even_batches (`bool`, *optional*)
|
856 |
If set, this will override the value of `even_batches` set in the `Accelerator`. If it is not provided,
|
857 |
the default `Accelerator` value wil be used.
|
858 |
-
|
859 |
<Tip warning={true}>
|
860 |
-
|
861 |
`join_uneven_inputs` is only supported for Distributed Data Parallel training on multiple GPUs. For any other
|
862 |
configuration, this method will have no effect.
|
863 |
-
|
864 |
</Tip>
|
865 |
-
|
866 |
<Tip warning={true}>
|
867 |
-
|
868 |
Overidding `even_batches` will not affect iterable-style data loaders.
|
869 |
-
|
870 |
</Tip>
|
871 |
-
|
872 |
Example:
|
873 |
-
|
874 |
```python
|
875 |
>>> from accelerate import Accelerator
|
876 |
-
|
877 |
>>> accelerator = Accelerator(even_batches=True)
|
878 |
>>> ddp_model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
|
879 |
-
|
880 |
>>> with accelerator.join_uneven_inputs([ddp_model], even_batches=False):
|
881 |
... for input, output in dataloader:
|
882 |
... outputs = model(input)
|
@@ -888,7 +735,6 @@ class Accelerator:
|
|
888 |
"""
|
889 |
if self.distributed_type in (DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.MULTI_XPU):
|
890 |
dl_even_batches_values = []
|
891 |
-
|
892 |
if even_batches is not None:
|
893 |
iterable_dl_seen = False
|
894 |
# override value in batch sampler for map-style datasets
|
@@ -898,14 +744,12 @@ class Accelerator:
|
|
898 |
continue
|
899 |
dl_even_batches_values.append((dl_idx, dl.batch_sampler.even_batches))
|
900 |
dl.batch_sampler.even_batches = even_batches
|
901 |
-
|
902 |
if iterable_dl_seen:
|
903 |
warnings.warn(
|
904 |
"Overridding even_batches is only supported for map-style datasets, yet some dataloaders given were iterable"
|
905 |
)
|
906 |
else:
|
907 |
even_batches = self.even_batches
|
908 |
-
|
909 |
enable_join = False if even_batches else True
|
910 |
try:
|
911 |
with Join(joinables, enable=enable_join, throw_on_early_termination=False):
|
@@ -920,25 +764,19 @@ class Accelerator:
|
|
920 |
warnings.warn(
|
921 |
"Joining uneven inputs is only supported for multi-GPU training, as a result `join_uneven_inputs` will have no effect."
|
922 |
)
|
923 |
-
|
924 |
with contextlib.nullcontext(joinables):
|
925 |
yield
|
926 |
-
|
927 |
def print(self, *args, **kwargs):
|
928 |
"""
|
929 |
Drop in replacement of `print()` to only print once per server.
|
930 |
-
|
931 |
Example:
|
932 |
-
|
933 |
```python
|
934 |
>>> from accelerate import Accelerator
|
935 |
-
|
936 |
>>> accelerator = Accelerator()
|
937 |
>>> accelerator.print("Hello world!")
|
938 |
```
|
939 |
"""
|
940 |
self.state.print(*args, **kwargs)
|
941 |
-
|
942 |
def _prepare_one(self, obj, first_pass=False, device_placement=None):
|
943 |
# First pass of preparation: DataLoader, model, optimizer
|
944 |
if first_pass:
|
@@ -955,44 +793,32 @@ class Accelerator:
|
|
955 |
return scheduler
|
956 |
# Return the unprocessed object if previous criteria was not met
|
957 |
return obj
|
958 |
-
|
959 |
def prepare(self, *args, device_placement=None):
|
960 |
"""
|
961 |
Prepare all objects passed in `args` for distributed training and mixed precision, then return them in the same
|
962 |
order.
|
963 |
-
|
964 |
Args:
|
965 |
*args (list of objects):
|
966 |
Any of the following type of objects:
|
967 |
-
|
968 |
- `torch.utils.data.DataLoader`: PyTorch Dataloader
|
969 |
- `torch.nn.Module`: PyTorch Module
|
970 |
- `torch.optim.Optimizer`: PyTorch Optimizer
|
971 |
- `torch.optim.lr_scheduler.LRScheduler`: PyTorch LR Scheduler
|
972 |
-
|
973 |
device_placement (`list[bool]`, *optional*):
|
974 |
Used to customize whether automatic device placement should be performed for each object passed. Needs
|
975 |
to be a list of the same length as `args`. Not compatible with DeepSpeed or FSDP.
|
976 |
-
|
977 |
<Tip>
|
978 |
-
|
979 |
You don't need to prepare a model if you only use it for inference without any kind of mixed precision
|
980 |
-
|
981 |
</Tip>
|
982 |
-
|
983 |
Examples:
|
984 |
-
|
985 |
```python
|
986 |
>>> from accelerate import Accelerator
|
987 |
-
|
988 |
>>> accelerator = Accelerator()
|
989 |
>>> # Assume a model, optimizer, data_loader and scheduler are defined
|
990 |
>>> model, optimizer, data_loader, scheduler = accelerator.prepare(model, optimizer, data_loader, scheduler)
|
991 |
```
|
992 |
-
|
993 |
```python
|
994 |
>>> from accelerate import Accelerator
|
995 |
-
|
996 |
>>> accelerator = Accelerator()
|
997 |
>>> # Assume a model, optimizer, data_loader and scheduler are defined
|
998 |
>>> device_placement = [True, True, False, False]
|
@@ -1010,7 +836,6 @@ class Accelerator:
|
|
1010 |
raise ValueError(
|
1011 |
f"`device_placement` should be a list with {len(args)} elements (the number of objects passed)."
|
1012 |
)
|
1013 |
-
|
1014 |
for obj in args:
|
1015 |
# TODO: Look at enabling native TP training directly with a proper config
|
1016 |
if (
|
@@ -1023,7 +848,6 @@ class Accelerator:
|
|
1023 |
"You can't train a model that has been loaded with `device_map='auto'` in any distributed mode."
|
1024 |
" Please rerun your script specifying `--num_processes=1` or by launching with `python {{myscript.py}}`."
|
1025 |
)
|
1026 |
-
|
1027 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
1028 |
model_count = 0
|
1029 |
for obj in args:
|
@@ -1033,7 +857,6 @@ class Accelerator:
|
|
1033 |
raise AssertionError(
|
1034 |
"You can't use same `Accelerator()` instance with multiple models when using DeepSpeed"
|
1035 |
)
|
1036 |
-
|
1037 |
# On TPUs, putting the model on the XLA device will create new parameters, so the corresponding optimizer will
|
1038 |
# have parameters disconnected from the model (so no training :-( ).
|
1039 |
# If the model and optimizer have parameters on different devices we raise an error.
|
@@ -1047,13 +870,11 @@ class Accelerator:
|
|
1047 |
"the flag default value for `device_placement` in your `Accelerator` to let it handle that "
|
1048 |
"part for you."
|
1049 |
)
|
1050 |
-
|
1051 |
# If we're dealing with device placement, this deals with that by...
|
1052 |
tpu_should_fix_optimizer = self.device_placement and self.distributed_type == DistributedType.TPU
|
1053 |
if tpu_should_fix_optimizer or (self.mixed_precision == "fp8" and self.fp8_recipe_handler.backend == "TE"):
|
1054 |
# 1. grabbing old model parameters
|
1055 |
old_named_params = self._get_named_parameters(*args)
|
1056 |
-
|
1057 |
if self.distributed_type in [DistributedType.MULTI_CPU, DistributedType.MULTI_XPU, DistributedType.NO]:
|
1058 |
if self.device.type == "cpu" and self.state.use_ipex:
|
1059 |
args = self._prepare_ipex(*args)
|
@@ -1072,7 +893,6 @@ class Accelerator:
|
|
1072 |
self._prepare_one(obj, first_pass=True, device_placement=d) for obj, d in zip(args, device_placement)
|
1073 |
)
|
1074 |
result = tuple(self._prepare_one(obj, device_placement=d) for obj, d in zip(result, device_placement))
|
1075 |
-
|
1076 |
if tpu_should_fix_optimizer or (self.mixed_precision == "fp8" and self.fp8_recipe_handler.backend == "TE"):
|
1077 |
# 2. grabbing new model parameters
|
1078 |
new_named_params = self._get_named_parameters(*result)
|
@@ -1082,21 +902,17 @@ class Accelerator:
|
|
1082 |
for obj in result:
|
1083 |
if isinstance(obj, torch.optim.Optimizer):
|
1084 |
obj._switch_parameters(mapping)
|
1085 |
-
|
1086 |
for item in result:
|
1087 |
if any(
|
1088 |
item in container
|
1089 |
for container in (self._dataloaders, self._models, self._optimizers, self._schedulers)
|
1090 |
):
|
1091 |
setattr(item, "_is_accelerate_prepared", True)
|
1092 |
-
|
1093 |
return result if len(result) > 1 else result[0]
|
1094 |
-
|
1095 |
def prepare_model(self, model: torch.nn.Module, device_placement: bool = None, evaluation_mode: bool = False):
|
1096 |
"""
|
1097 |
Prepares a PyTorch model for training in any distributed setup. It is recommended to use
|
1098 |
[`Accelerator.prepare`] instead.
|
1099 |
-
|
1100 |
Args:
|
1101 |
model (`torch.nn.Module`):
|
1102 |
A PyTorch model to prepare. You don't need to prepare a model if it is used only for inference without
|
@@ -1106,12 +922,9 @@ class Accelerator:
|
|
1106 |
evaluation_mode (`bool`, *optional*, defaults to `False`):
|
1107 |
Whether or not to set the model for evaluation only, by just applying mixed precision and
|
1108 |
`torch.compile` (if configured in the `Accelerator` object).
|
1109 |
-
|
1110 |
Example:
|
1111 |
-
|
1112 |
```python
|
1113 |
>>> from accelerate import Accelerator
|
1114 |
-
|
1115 |
>>> accelerator = Accelerator()
|
1116 |
>>> # Assume a model is defined
|
1117 |
>>> model = accelerator.prepare_model(model)
|
@@ -1120,7 +933,6 @@ class Accelerator:
|
|
1120 |
if device_placement is None:
|
1121 |
device_placement = self.device_placement and self.distributed_type != DistributedType.FSDP
|
1122 |
self._models.append(model)
|
1123 |
-
|
1124 |
# TODO: Look at enabling native TP training directly with a proper config
|
1125 |
if (
|
1126 |
self.verify_device_map(model)
|
@@ -1131,7 +943,6 @@ class Accelerator:
|
|
1131 |
"You can't train a model that has been loaded with `device_map='auto'` in any distributed mode."
|
1132 |
" Please rerun your script specifying `--num_processes=1` or by launching with `python {{myscript.py}}`."
|
1133 |
)
|
1134 |
-
|
1135 |
if self.native_amp:
|
1136 |
model._original_forward = model.forward
|
1137 |
model_forward_func = model.forward.__func__ if hasattr(model.forward, "__func__") else model.forward
|
@@ -1148,13 +959,11 @@ class Accelerator:
|
|
1148 |
convert_model(model)
|
1149 |
model._converted_to_transformer_engine = True
|
1150 |
model._original_forward = model.forward
|
1151 |
-
|
1152 |
kwargs = self.fp8_recipe_handler.to_kwargs() if self.fp8_recipe_handler is not None else {}
|
1153 |
if "fp8_format" in kwargs:
|
1154 |
kwargs["fp8_format"] = getattr(te_recipe.Format, kwargs["fp8_format"])
|
1155 |
fp8_recipe = te_recipe.DelayedScaling(**kwargs)
|
1156 |
model.forward = fp8_autocast(enabled=True, fp8_recipe=fp8_recipe)(model.forward)
|
1157 |
-
|
1158 |
if (getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_loaded_in_4bit", False)) and getattr(
|
1159 |
model, "hf_device_map", False
|
1160 |
):
|
@@ -1167,7 +976,6 @@ class Accelerator:
|
|
1167 |
)
|
1168 |
current_device = list(model_devices)[0]
|
1169 |
current_device_index = current_device.index if isinstance(current_device, torch.device) else current_device
|
1170 |
-
|
1171 |
if torch.device(current_device_index) != self.device:
|
1172 |
# if on the first device (GPU 0) we don't care
|
1173 |
if (self.device.index is not None) or (current_device_index != 0):
|
@@ -1175,7 +983,6 @@ class Accelerator:
|
|
1175 |
"You can't train a model that has been loaded in 8-bit precision on a different device than the one "
|
1176 |
"you're training on. Make sure you loaded the model on the correct device using for example `device_map={'':torch.cuda.current_device() or device_map={'':torch.xpu.current_device()}"
|
1177 |
)
|
1178 |
-
|
1179 |
if "cpu" in model_devices or "disk" in model_devices:
|
1180 |
raise ValueError(
|
1181 |
"You can't train a model that has been loaded in 8-bit precision with CPU or disk offload."
|
@@ -1195,13 +1002,11 @@ class Accelerator:
|
|
1195 |
device_ids, output_device = [self.local_process_index], self.local_process_index
|
1196 |
else:
|
1197 |
device_ids, output_device = None, None
|
1198 |
-
|
1199 |
model = torch.nn.parallel.DistributedDataParallel(
|
1200 |
model, device_ids=device_ids, output_device=output_device, **kwargs
|
1201 |
)
|
1202 |
elif self.distributed_type == DistributedType.FSDP:
|
1203 |
from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel as FSDP
|
1204 |
-
|
1205 |
# Check if the model is already a FSDP model due to `Manual Wrapping` and if so,
|
1206 |
# don't wrap it again
|
1207 |
# In case the model is already compiled using PyTorch 2.0 and the wrapped model in it
|
@@ -1209,7 +1014,6 @@ class Accelerator:
|
|
1209 |
is_type_fsdp = isinstance(model, FSDP) or (
|
1210 |
is_compiled_module(model) and isinstance(model._orig_mod, FSDP)
|
1211 |
)
|
1212 |
-
|
1213 |
if not is_type_fsdp:
|
1214 |
self.state.fsdp_plugin.set_auto_wrap_policy(model)
|
1215 |
fsdp_plugin = self.state.fsdp_plugin
|
@@ -1234,7 +1038,6 @@ class Accelerator:
|
|
1234 |
apply_activation_checkpointing,
|
1235 |
checkpoint_wrapper,
|
1236 |
)
|
1237 |
-
|
1238 |
apply_activation_checkpointing(
|
1239 |
model,
|
1240 |
checkpoint_wrapper_fn=functools.partial(
|
@@ -1258,18 +1061,14 @@ class Accelerator:
|
|
1258 |
raise ValueError("Using `torch.compile` requires PyTorch 2.0 or higher.")
|
1259 |
model = torch.compile(model, **self.state.dynamo_plugin.to_kwargs())
|
1260 |
return model
|
1261 |
-
|
1262 |
def _prepare_deepspeed(self, *args):
|
1263 |
import deepspeed
|
1264 |
-
|
1265 |
deepspeed_plugin = self.state.deepspeed_plugin
|
1266 |
-
|
1267 |
is_dataloader_present = any(isinstance(obj, torch.utils.data.DataLoader) for obj in args)
|
1268 |
result = [
|
1269 |
self._prepare_one(obj, first_pass=True) if isinstance(obj, torch.utils.data.DataLoader) else obj
|
1270 |
for obj in args
|
1271 |
]
|
1272 |
-
|
1273 |
if deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] == "auto":
|
1274 |
if is_dataloader_present:
|
1275 |
batch_sizes = [obj.batch_size for obj in args if hasattr(obj, "batch_size")]
|
@@ -1281,7 +1080,6 @@ class Accelerator:
|
|
1281 |
)
|
1282 |
if self.split_batches:
|
1283 |
batch_sizes = [batch_size // self.num_processes for batch_size in batch_sizes]
|
1284 |
-
|
1285 |
batch_size_per_device = min(batch_sizes) if deepspeed_plugin.is_train_batch_min else max(batch_sizes)
|
1286 |
if len(batch_sizes) > 1:
|
1287 |
logger.info(
|
@@ -1297,14 +1095,12 @@ class Accelerator:
|
|
1297 |
)
|
1298 |
else:
|
1299 |
batch_size_per_device = deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"]
|
1300 |
-
|
1301 |
# handle `gradient_accumulation_steps` when the value is `auto`
|
1302 |
deepspeed_plugin.fill_match(
|
1303 |
"gradient_accumulation_steps",
|
1304 |
must_match=False,
|
1305 |
gradient_accumulation_steps=self.gradient_accumulation_steps,
|
1306 |
)
|
1307 |
-
|
1308 |
config_kwargs = {
|
1309 |
"train_micro_batch_size_per_gpu": batch_size_per_device,
|
1310 |
"train_batch_size": batch_size_per_device
|
@@ -1313,7 +1109,6 @@ class Accelerator:
|
|
1313 |
"gradient_clipping": 1.0,
|
1314 |
"zero_optimization.stage3_gather_16bit_weights_on_model_save": False,
|
1315 |
}
|
1316 |
-
|
1317 |
model = None
|
1318 |
optimizer = None
|
1319 |
scheduler = None
|
@@ -1326,7 +1121,6 @@ class Accelerator:
|
|
1326 |
type(obj).__name__ in deepspeed.runtime.lr_schedules.VALID_LR_SCHEDULES
|
1327 |
):
|
1328 |
scheduler = obj
|
1329 |
-
|
1330 |
if optimizer is not None:
|
1331 |
if "optimizer" in deepspeed_plugin.deepspeed_config and not isinstance(optimizer, (DummyOptim)):
|
1332 |
raise ValueError(
|
@@ -1338,10 +1132,8 @@ class Accelerator:
|
|
1338 |
raise ValueError(
|
1339 |
"You cannot create a `DummyOptim` without specifying an optimizer in the config file."
|
1340 |
)
|
1341 |
-
|
1342 |
if isinstance(optimizer, (torch.optim.Optimizer)):
|
1343 |
deepspeed_plugin.deepspeed_config["zero_allow_untested_optimizer"] = True
|
1344 |
-
|
1345 |
if scheduler is not None:
|
1346 |
if "scheduler" in deepspeed_plugin.deepspeed_config and not isinstance(scheduler, (DummyScheduler)):
|
1347 |
raise ValueError(
|
@@ -1358,14 +1150,12 @@ class Accelerator:
|
|
1358 |
"Either specify a scheduler in the config file or "
|
1359 |
"pass in the `lr_scheduler_callable` parameter when using `accelerate.utils.DummyScheduler`."
|
1360 |
)
|
1361 |
-
|
1362 |
if optimizer is not None and scheduler is not None:
|
1363 |
if isinstance(optimizer, (DummyOptim)) and not isinstance(scheduler, (DummyScheduler)):
|
1364 |
raise ValueError(
|
1365 |
"You can only specify `accelerate.utils.DummyScheduler` in the code when using "
|
1366 |
"`accelerate.utils.DummyOptim`."
|
1367 |
)
|
1368 |
-
|
1369 |
if model is not None:
|
1370 |
if hasattr(model, "config"):
|
1371 |
hidden_size = (
|
@@ -1381,7 +1171,6 @@ class Accelerator:
|
|
1381 |
"zero_optimization.stage3_param_persistence_threshold": 10 * hidden_size,
|
1382 |
}
|
1383 |
)
|
1384 |
-
|
1385 |
if isinstance(optimizer, (DummyOptim)):
|
1386 |
config_kwargs.update(
|
1387 |
{"optimizer.params.lr": optimizer.lr, "optimizer.params.weight_decay": optimizer.weight_decay}
|
@@ -1418,7 +1207,6 @@ class Accelerator:
|
|
1418 |
"device", "none"
|
1419 |
) != "none" and self.deepspeed_config.get("zero_force_ds_cpu_optimizer", True):
|
1420 |
from deepspeed.ops.adam import DeepSpeedCPUAdam
|
1421 |
-
|
1422 |
defaults = {k: v for k, v in optimizer.defaults.items() if k in ["lr", "weight_decay"]}
|
1423 |
optimizer = DeepSpeedCPUAdam(optimizer.param_groups, **defaults)
|
1424 |
kwargs["optimizer"] = optimizer
|
@@ -1428,7 +1216,6 @@ class Accelerator:
|
|
1428 |
or type(scheduler).__name__ in deepspeed.runtime.lr_schedules.VALID_LR_SCHEDULES
|
1429 |
):
|
1430 |
kwargs["lr_scheduler"] = scheduler
|
1431 |
-
|
1432 |
engine, optimizer, _, lr_scheduler = deepspeed.initialize(**kwargs)
|
1433 |
if optimizer is not None:
|
1434 |
optimizer = DeepSpeedOptimizerWrapper(optimizer)
|
@@ -1442,7 +1229,6 @@ class Accelerator:
|
|
1442 |
)
|
1443 |
else:
|
1444 |
scheduler = DeepSpeedSchedulerWrapper(lr_scheduler, optimizer)
|
1445 |
-
|
1446 |
for i in range(len(result)):
|
1447 |
if isinstance(result[i], torch.nn.Module):
|
1448 |
result[i] = engine
|
@@ -1464,7 +1250,6 @@ class Accelerator:
|
|
1464 |
"You can't use same `Accelerator()` instance with multiple models when using DeepSpeed"
|
1465 |
)
|
1466 |
return tuple(result)
|
1467 |
-
|
1468 |
def _prepare_megatron_lm(self, *args):
|
1469 |
megatron_lm_plugin = self.state.megatron_lm_plugin
|
1470 |
if not megatron_lm_plugin.megatron_dataset_flag:
|
@@ -1473,7 +1258,6 @@ class Accelerator:
|
|
1473 |
raise ValueError(
|
1474 |
"You must specify a training or evaluation dataloader in `accelerate.prepare()` when using Megatron-LM."
|
1475 |
)
|
1476 |
-
|
1477 |
micro_batch_size = min(batch_sizes) if megatron_lm_plugin.is_train_batch_min else max(batch_sizes)
|
1478 |
if len(batch_sizes) > 1:
|
1479 |
logger.info(
|
@@ -1485,10 +1269,8 @@ class Accelerator:
|
|
1485 |
if isinstance(obj, MegatronLMDummyDataLoader):
|
1486 |
micro_batch_size = obj.dataset_args["micro_batch_size"]
|
1487 |
break
|
1488 |
-
|
1489 |
dp_degree = self.num_processes // (megatron_lm_plugin.tp_degree * megatron_lm_plugin.pp_degree)
|
1490 |
megatron_lm_plugin.set_training_args(micro_batch_size, dp_degree)
|
1491 |
-
|
1492 |
model = None
|
1493 |
optimizer = None
|
1494 |
scheduler = None
|
@@ -1503,7 +1285,6 @@ class Accelerator:
|
|
1503 |
optimizer = obj
|
1504 |
elif isinstance(obj, (LRScheduler, MegatronLMDummyScheduler)):
|
1505 |
scheduler = obj
|
1506 |
-
|
1507 |
if model is not None:
|
1508 |
megatron_lm_plugin.set_network_size_args(model, batch_data)
|
1509 |
if optimizer is not None:
|
@@ -1515,7 +1296,6 @@ class Accelerator:
|
|
1515 |
"You can't use a custom scheduler with Megatron-LM. Please use the `accelerate.utils.MegatronLMDummyScheduler` instead."
|
1516 |
)
|
1517 |
megatron_lm_plugin.set_scheduler_args(scheduler)
|
1518 |
-
|
1519 |
# initialize megatron-lm
|
1520 |
megatron_lm_initialize(self, args_defaults=megatron_lm_plugin.megatron_lm_default_args)
|
1521 |
counter = 0
|
@@ -1532,21 +1312,18 @@ class Accelerator:
|
|
1532 |
counter += 1
|
1533 |
else:
|
1534 |
result.append(obj)
|
1535 |
-
|
1536 |
if model is not None:
|
1537 |
model = megatron_lm_prepare_model(self)
|
1538 |
if optimizer is not None:
|
1539 |
optimizer = megatron_lm_prepare_optimizer(self, model)
|
1540 |
if scheduler is not None:
|
1541 |
scheduler = megatron_lm_prepare_scheduler(self, optimizer, scheduler)
|
1542 |
-
|
1543 |
if model is not None:
|
1544 |
model = MegatronEngine(self, model, optimizer, scheduler)
|
1545 |
if optimizer is not None:
|
1546 |
optimizer = MegatronLMOptimizerWrapper(optimizer)
|
1547 |
if scheduler is not None:
|
1548 |
scheduler = MegatronLMSchedulerWrapper(scheduler, optimizer)
|
1549 |
-
|
1550 |
for i in range(len(result)):
|
1551 |
if isinstance(result[i], torch.nn.Module):
|
1552 |
result[i] = model
|
@@ -1565,7 +1342,6 @@ class Accelerator:
|
|
1565 |
"You can't use same `Accelerator()` instance with multiple models when using Megatron-LM"
|
1566 |
)
|
1567 |
return tuple(result)
|
1568 |
-
|
1569 |
def _prepare_ipex(self, *args):
|
1570 |
if not is_ipex_available():
|
1571 |
raise ImportError(
|
@@ -1574,7 +1350,6 @@ class Accelerator:
|
|
1574 |
)
|
1575 |
else:
|
1576 |
import intel_extension_for_pytorch as ipex
|
1577 |
-
|
1578 |
model = None
|
1579 |
optimizer = None
|
1580 |
result = [obj for obj in args]
|
@@ -1598,7 +1373,6 @@ class Accelerator:
|
|
1598 |
elif isinstance(result[i], (torch.optim.Optimizer)):
|
1599 |
result[i] = optimizer
|
1600 |
return tuple(result)
|
1601 |
-
|
1602 |
def _prepare_msamp(self, *args):
|
1603 |
if not is_msamp_available():
|
1604 |
raise ImportError(
|
@@ -1607,7 +1381,6 @@ class Accelerator:
|
|
1607 |
)
|
1608 |
else:
|
1609 |
import msamp
|
1610 |
-
|
1611 |
model, optimizer = None, None
|
1612 |
num_models, num_optimizers = 0, 0
|
1613 |
result = [obj for obj in args]
|
@@ -1634,14 +1407,12 @@ class Accelerator:
|
|
1634 |
elif isinstance(result[i], (torch.optim.Optimizer)):
|
1635 |
result[i] = optimizer
|
1636 |
return tuple(result)
|
1637 |
-
|
1638 |
def prepare_data_loader(
|
1639 |
self, data_loader: torch.utils.data.DataLoader, device_placement=None, slice_fn_for_dispatch=None
|
1640 |
):
|
1641 |
"""
|
1642 |
Prepares a PyTorch DataLoader for training in any distributed setup. It is recommended to use
|
1643 |
[`Accelerator.prepare`] instead.
|
1644 |
-
|
1645 |
Args:
|
1646 |
data_loader (`torch.utils.data.DataLoader`):
|
1647 |
A vanilla PyTorch DataLoader to prepare
|
@@ -1652,13 +1423,10 @@ class Accelerator:
|
|
1652 |
If passed, this function will be used to slice tensors across `num_processes`. Will default to
|
1653 |
[`~utils.slice_tensors`]. This argument is used only when `dispatch_batches` is set to `True` and will
|
1654 |
be ignored otherwise.
|
1655 |
-
|
1656 |
Example:
|
1657 |
-
|
1658 |
```python
|
1659 |
>>> import torch
|
1660 |
>>> from accelerate import Accelerator
|
1661 |
-
|
1662 |
>>> accelerator = Accelerator()
|
1663 |
>>> data_loader = torch.utils.data.DataLoader(...)
|
1664 |
>>> data_loader = accelerator.prepare_data_loader(data_loader, device_placement=True)
|
@@ -1685,24 +1453,19 @@ class Accelerator:
|
|
1685 |
)
|
1686 |
self._dataloaders.append(prepared_data_loader)
|
1687 |
return prepared_data_loader
|
1688 |
-
|
1689 |
def prepare_optimizer(self, optimizer: torch.optim.Optimizer, device_placement=None):
|
1690 |
"""
|
1691 |
Prepares a PyTorch Optimizer for training in any distributed setup. It is recommended to use
|
1692 |
[`Accelerator.prepare`] instead.
|
1693 |
-
|
1694 |
Args:
|
1695 |
optimizer (`torch.optim.Optimizer`):
|
1696 |
A vanilla PyTorch optimizer to prepare
|
1697 |
device_placement (`bool`, *optional*):
|
1698 |
Whether or not to place the optimizer on the proper device. Will default to `self.device_placement`.
|
1699 |
-
|
1700 |
Example:
|
1701 |
-
|
1702 |
```python
|
1703 |
>>> import torch
|
1704 |
>>> from accelerate import Accelerator
|
1705 |
-
|
1706 |
>>> accelerator = Accelerator()
|
1707 |
>>> optimizer = torch.optim.Adam(...)
|
1708 |
>>> optimizer = accelerator.prepare_optimizer(optimizer, device_placement=True)
|
@@ -1718,22 +1481,17 @@ class Accelerator:
|
|
1718 |
optimizer = AcceleratedOptimizer(optimizer, device_placement=device_placement, scaler=self.scaler)
|
1719 |
self._optimizers.append(optimizer)
|
1720 |
return optimizer
|
1721 |
-
|
1722 |
def prepare_scheduler(self, scheduler: LRScheduler):
|
1723 |
"""
|
1724 |
Prepares a PyTorch Scheduler for training in any distributed setup. It is recommended to use
|
1725 |
[`Accelerator.prepare`] instead.
|
1726 |
-
|
1727 |
Args:
|
1728 |
scheduler (`torch.optim.lr_scheduler.LRScheduler`):
|
1729 |
A vanilla PyTorch scheduler to prepare
|
1730 |
-
|
1731 |
Example:
|
1732 |
-
|
1733 |
```python
|
1734 |
>>> import torch
|
1735 |
>>> from accelerate import Accelerator
|
1736 |
-
|
1737 |
>>> accelerator = Accelerator()
|
1738 |
>>> optimizer = torch.optim.Adam(...)
|
1739 |
>>> scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, ...)
|
@@ -1759,19 +1517,14 @@ class Accelerator:
|
|
1759 |
)
|
1760 |
self._schedulers.append(scheduler)
|
1761 |
return scheduler
|
1762 |
-
|
1763 |
def backward(self, loss, **kwargs):
|
1764 |
"""
|
1765 |
Scales the gradients in accordance to the `GradientAccumulationPlugin` and calls the correct `backward()` based
|
1766 |
on the configuration.
|
1767 |
-
|
1768 |
Should be used in lieu of `loss.backward()`.
|
1769 |
-
|
1770 |
Example:
|
1771 |
-
|
1772 |
```python
|
1773 |
>>> from accelerate import Accelerator
|
1774 |
-
|
1775 |
>>> accelerator = Accelerator(gradient_accumulation_steps=2)
|
1776 |
>>> outputs = model(inputs)
|
1777 |
>>> loss = loss_fn(outputs, labels)
|
@@ -1789,20 +1542,15 @@ class Accelerator:
|
|
1789 |
self.scaler.scale(loss).backward(**kwargs)
|
1790 |
else:
|
1791 |
loss.backward(**kwargs)
|
1792 |
-
|
1793 |
def set_trigger(self):
|
1794 |
"""
|
1795 |
Sets the internal trigger tensor to 1 on the current process. A latter check should follow using this which
|
1796 |
will check across all processes.
|
1797 |
-
|
1798 |
Note:
|
1799 |
Does not require `wait_for_everyone()`
|
1800 |
-
|
1801 |
Example:
|
1802 |
-
|
1803 |
```python
|
1804 |
>>> from accelerate import Accelerator
|
1805 |
-
|
1806 |
>>> accelerator = Accelerator()
|
1807 |
>>> # Assume later in the training script
|
1808 |
>>> # `should_do_breakpoint` is a custom function to monitor when to break,
|
@@ -1815,20 +1563,15 @@ class Accelerator:
|
|
1815 |
```
|
1816 |
"""
|
1817 |
self.flag_tensor = torch.tensor(1, device=self.device)
|
1818 |
-
|
1819 |
def check_trigger(self):
|
1820 |
"""
|
1821 |
Checks if the internal trigger tensor has been set to 1 in any of the processes. If so, will return `True` and
|
1822 |
reset the trigger tensor to 0.
|
1823 |
-
|
1824 |
Note:
|
1825 |
Does not require `wait_for_everyone()`
|
1826 |
-
|
1827 |
Example:
|
1828 |
-
|
1829 |
```python
|
1830 |
>>> from accelerate import Accelerator
|
1831 |
-
|
1832 |
>>> accelerator = Accelerator()
|
1833 |
>>> # Assume later in the training script
|
1834 |
>>> # `should_do_breakpoint` is a custom function to monitor when to break,
|
@@ -1848,23 +1591,17 @@ class Accelerator:
|
|
1848 |
self.flag_tensor = torch.tensor(0, device=self.device)
|
1849 |
return True
|
1850 |
return False
|
1851 |
-
|
1852 |
def unscale_gradients(self, optimizer=None):
|
1853 |
"""
|
1854 |
Unscale the gradients in mixed precision training with AMP. This is a noop in all other settings.
|
1855 |
-
|
1856 |
Likely should be called through [`Accelerator.clip_grad_norm_`] or [`Accelerator.clip_grad_value_`]
|
1857 |
-
|
1858 |
Args:
|
1859 |
optimizer (`torch.optim.Optimizer` or `list[torch.optim.Optimizer]`, *optional*):
|
1860 |
The optimizer(s) for which to unscale gradients. If not set, will unscale gradients on all optimizers
|
1861 |
that were passed to [`~Accelerator.prepare`].
|
1862 |
-
|
1863 |
Example:
|
1864 |
-
|
1865 |
```python
|
1866 |
>>> from accelerate import Accelerator
|
1867 |
-
|
1868 |
>>> accelerator = Accelerator()
|
1869 |
>>> model, optimizer = accelerator.prepare(model, optimizer)
|
1870 |
>>> outputs = model(inputs)
|
@@ -1887,22 +1624,16 @@ class Accelerator:
|
|
1887 |
gradients = xm._fetch_gradients(opt)
|
1888 |
self.reduce(gradients, scale=1.0 / self.num_processes)
|
1889 |
self.scaler.unscale_(opt)
|
1890 |
-
|
1891 |
def clip_grad_norm_(self, parameters, max_norm, norm_type=2):
|
1892 |
"""
|
1893 |
Should be used in place of `torch.nn.utils.clip_grad_norm_`.
|
1894 |
-
|
1895 |
Returns:
|
1896 |
`torch.Tensor`: Total norm of the parameter gradients (viewed as a single vector).
|
1897 |
-
|
1898 |
Example:
|
1899 |
-
|
1900 |
```python
|
1901 |
>>> from accelerate import Accelerator
|
1902 |
-
|
1903 |
>>> accelerator = Accelerator(gradient_accumulation_steps=2)
|
1904 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
1905 |
-
|
1906 |
>>> for input, target in dataloader:
|
1907 |
... optimizer.zero_grad()
|
1908 |
... output = model(input)
|
@@ -1925,19 +1656,14 @@ class Accelerator:
|
|
1925 |
return None
|
1926 |
self.unscale_gradients()
|
1927 |
return torch.nn.utils.clip_grad_norm_(parameters, max_norm, norm_type=norm_type)
|
1928 |
-
|
1929 |
def clip_grad_value_(self, parameters, clip_value):
|
1930 |
"""
|
1931 |
Should be used in place of `torch.nn.utils.clip_grad_value_`.
|
1932 |
-
|
1933 |
Example:
|
1934 |
-
|
1935 |
```python
|
1936 |
>>> from accelerate import Accelerator
|
1937 |
-
|
1938 |
>>> accelerator = Accelerator(gradient_accumulation_steps=2)
|
1939 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
1940 |
-
|
1941 |
>>> for input, target in dataloader:
|
1942 |
... optimizer.zero_grad()
|
1943 |
... output = model(input)
|
@@ -1952,30 +1678,23 @@ class Accelerator:
|
|
1952 |
raise Exception("DeepSpeed and FSDP do not support `clip_grad_value_`. Use `clip_grad_norm_` instead.")
|
1953 |
self.unscale_gradients()
|
1954 |
torch.nn.utils.clip_grad_value_(parameters, clip_value)
|
1955 |
-
|
1956 |
def gather(self, tensor):
|
1957 |
"""
|
1958 |
Gather the values in *tensor* across all processes and concatenate them on the first dimension. Useful to
|
1959 |
regroup the predictions from all processes when doing evaluation.
|
1960 |
-
|
1961 |
Note:
|
1962 |
This gather happens in all processes.
|
1963 |
-
|
1964 |
Args:
|
1965 |
tensor (`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`):
|
1966 |
The tensors to gather across all processes.
|
1967 |
-
|
1968 |
Returns:
|
1969 |
`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`: The gathered tensor(s). Note that the
|
1970 |
first dimension of the result is *num_processes* multiplied by the first dimension of the input tensors.
|
1971 |
-
|
1972 |
Example:
|
1973 |
-
|
1974 |
```python
|
1975 |
>>> # Assuming four processes
|
1976 |
>>> import torch
|
1977 |
>>> from accelerate import Accelerator
|
1978 |
-
|
1979 |
>>> accelerator = Accelerator()
|
1980 |
>>> process_tensor = torch.tensor([accelerator.process_index])
|
1981 |
>>> gathered_tensor = accelerator.gather(process_tensor)
|
@@ -1984,23 +1703,18 @@ class Accelerator:
|
|
1984 |
```
|
1985 |
"""
|
1986 |
return gather(tensor)
|
1987 |
-
|
1988 |
def gather_for_metrics(self, input_data):
|
1989 |
"""
|
1990 |
Gathers `input_data` and potentially drops duplicates in the last batch if on a distributed system. Should be
|
1991 |
used for gathering the inputs and targets for metric calculation.
|
1992 |
-
|
1993 |
Args:
|
1994 |
input (`torch.Tensor`, `object`, a nested tuple/list/dictionary of `torch.Tensor`, or a nested tuple/list/dictionary of `object`):
|
1995 |
The tensors or objects for calculating metrics across all processes
|
1996 |
-
|
1997 |
Example:
|
1998 |
-
|
1999 |
```python
|
2000 |
>>> # Assuming two processes, with a batch size of 5 on a dataset with 9 samples
|
2001 |
>>> import torch
|
2002 |
>>> from accelerate import Accelerator
|
2003 |
-
|
2004 |
>>> accelerator = Accelerator()
|
2005 |
>>> dataloader = torch.utils.data.DataLoader(range(9), batch_size=5)
|
2006 |
>>> dataloader = accelerator.prepare(dataloader)
|
@@ -2015,12 +1729,10 @@ class Accelerator:
|
|
2015 |
all_tensors = True
|
2016 |
except TypeError:
|
2017 |
all_tensors = False
|
2018 |
-
|
2019 |
if not all_tensors:
|
2020 |
data = gather_object(input_data)
|
2021 |
else:
|
2022 |
data = self.gather(input_data)
|
2023 |
-
|
2024 |
try:
|
2025 |
if self.gradient_state.end_of_dataloader:
|
2026 |
# at the end of a dataloader, `gather_for_metrics` regresses to
|
@@ -2034,7 +1746,6 @@ class Accelerator:
|
|
2034 |
# Last batch needs to be truncated on distributed systems as it contains additional samples
|
2035 |
def _adjust_samples(tensor):
|
2036 |
return tensor[: self.gradient_state.remainder]
|
2037 |
-
|
2038 |
return recursively_apply(_adjust_samples, data)
|
2039 |
else: # remainder is 0
|
2040 |
# no remainder even though at end of dataloader, so nothing to do.
|
@@ -2045,14 +1756,11 @@ class Accelerator:
|
|
2045 |
except Exception:
|
2046 |
# Dataset had no length or raised an error
|
2047 |
return data
|
2048 |
-
|
2049 |
def reduce(self, tensor, reduction="sum", scale=1.0):
|
2050 |
"""
|
2051 |
Reduce the values in *tensor* across all processes based on *reduction*.
|
2052 |
-
|
2053 |
Note:
|
2054 |
All processes get the reduced value.
|
2055 |
-
|
2056 |
Args:
|
2057 |
tensor (`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`):
|
2058 |
The tensors to reduce across all processes.
|
@@ -2060,18 +1768,14 @@ class Accelerator:
|
|
2060 |
A reduction type, can be one of 'sum', 'mean', or 'none'. If 'none', will not perform any operation.
|
2061 |
scale (`float`, *optional*, defaults to 1.0):
|
2062 |
A default scaling value to be applied after the reduce, only valied on XLA.
|
2063 |
-
|
2064 |
Returns:
|
2065 |
`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`:
|
2066 |
The reduced tensor(s).
|
2067 |
-
|
2068 |
Example:
|
2069 |
-
|
2070 |
```python
|
2071 |
>>> # Assuming two processes
|
2072 |
>>> import torch
|
2073 |
>>> from accelerate import Accelerator
|
2074 |
-
|
2075 |
>>> accelerator = Accelerator()
|
2076 |
>>> process_tensor = torch.arange(accelerator.num_processes) + 1 + (2 * accelerator.process_index)
|
2077 |
>>> process_tensor = process_tensor.to(accelerator.device)
|
@@ -2081,12 +1785,10 @@ class Accelerator:
|
|
2081 |
```
|
2082 |
"""
|
2083 |
return reduce(tensor, reduction, scale)
|
2084 |
-
|
2085 |
def pad_across_processes(self, tensor, dim=0, pad_index=0, pad_first=False):
|
2086 |
"""
|
2087 |
Recursively pad the tensors in a nested list/tuple/dictionary of tensors from all devices to the same size so
|
2088 |
they can safely be gathered.
|
2089 |
-
|
2090 |
Args:
|
2091 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
2092 |
The data to gather.
|
@@ -2096,18 +1798,14 @@ class Accelerator:
|
|
2096 |
The value with which to pad.
|
2097 |
pad_first (`bool`, *optional*, defaults to `False`):
|
2098 |
Whether to pad at the beginning or the end.
|
2099 |
-
|
2100 |
Returns:
|
2101 |
`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`:
|
2102 |
The padded tensor(s).
|
2103 |
-
|
2104 |
Example:
|
2105 |
-
|
2106 |
```python
|
2107 |
>>> # Assuming two processes, with the first processes having a tensor of size 1 and the second of size 2
|
2108 |
>>> import torch
|
2109 |
>>> from accelerate import Accelerator
|
2110 |
-
|
2111 |
>>> accelerator = Accelerator()
|
2112 |
>>> process_tensor = torch.arange(accelerator.process_index + 1).to(accelerator.device)
|
2113 |
>>> padded_tensor = accelerator.pad_across_processes(process_tensor)
|
@@ -2116,52 +1814,41 @@ class Accelerator:
|
|
2116 |
```
|
2117 |
"""
|
2118 |
return pad_across_processes(tensor, dim=dim, pad_index=pad_index, pad_first=pad_first)
|
2119 |
-
|
2120 |
def unwrap_model(self, model, keep_fp32_wrapper: bool = True):
|
2121 |
"""
|
2122 |
Unwraps the `model` from the additional layer possible added by [`~Accelerator.prepare`]. Useful before saving
|
2123 |
the model.
|
2124 |
-
|
2125 |
Args:
|
2126 |
model (`torch.nn.Module`):
|
2127 |
The model to unwrap.
|
2128 |
keep_fp32_wrapper (`bool`, *optional*, defaults to `True`):
|
2129 |
Whether to not remove the mixed precision hook if it was added.
|
2130 |
-
|
2131 |
Returns:
|
2132 |
`torch.nn.Module`: The unwrapped model.
|
2133 |
-
|
2134 |
Example:
|
2135 |
-
|
2136 |
```python
|
2137 |
>>> # Assuming two GPU processes
|
2138 |
>>> from torch.nn.parallel import DistributedDataParallel
|
2139 |
>>> from accelerate import Accelerator
|
2140 |
-
|
2141 |
>>> accelerator = Accelerator()
|
2142 |
>>> model = accelerator.prepare(MyModel())
|
2143 |
>>> print(model.__class__.__name__)
|
2144 |
DistributedDataParallel
|
2145 |
-
|
2146 |
>>> model = accelerator.unwrap_model(model)
|
2147 |
>>> print(model.__class__.__name__)
|
2148 |
MyModel
|
2149 |
```
|
2150 |
"""
|
2151 |
return extract_model_from_parallel(model, keep_fp32_wrapper)
|
2152 |
-
|
2153 |
def wait_for_everyone(self):
|
2154 |
"""
|
2155 |
Will stop the execution of the current process until every other process has reached that point (so this does
|
2156 |
nothing when the script is only run in one process). Useful to do before saving a model.
|
2157 |
-
|
2158 |
Example:
|
2159 |
-
|
2160 |
```python
|
2161 |
>>> # Assuming two GPU processes
|
2162 |
>>> import time
|
2163 |
>>> from accelerate import Accelerator
|
2164 |
-
|
2165 |
>>> accelerator = Accelerator()
|
2166 |
>>> if accelerator.is_main_process:
|
2167 |
... time.sleep(2)
|
@@ -2173,12 +1860,10 @@ class Accelerator:
|
|
2173 |
```
|
2174 |
"""
|
2175 |
wait_for_everyone()
|
2176 |
-
|
2177 |
@on_main_process
|
2178 |
def init_trackers(self, project_name: str, config: dict | None = None, init_kwargs: dict | None = {}):
|
2179 |
"""
|
2180 |
Initializes a run for all trackers stored in `self.log_with`, potentially with starting configurations
|
2181 |
-
|
2182 |
Args:
|
2183 |
project_name (`str`):
|
2184 |
The name of the project. All trackers will save their data based on this
|
@@ -2190,12 +1875,9 @@ class Accelerator:
|
|
2190 |
```python
|
2191 |
{"wandb": {"tags": ["tag_a", "tag_b"]}}
|
2192 |
```
|
2193 |
-
|
2194 |
Example:
|
2195 |
-
|
2196 |
```python
|
2197 |
>>> from accelerate import Accelerator
|
2198 |
-
|
2199 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
2200 |
>>> accelerator.init_trackers(
|
2201 |
... project_name="my_project",
|
@@ -2220,26 +1902,20 @@ class Accelerator:
|
|
2220 |
if config is not None:
|
2221 |
for tracker in self.trackers:
|
2222 |
tracker.store_init_configuration(config)
|
2223 |
-
|
2224 |
def get_tracker(self, name: str, unwrap: bool = False):
|
2225 |
"""
|
2226 |
Returns a `tracker` from `self.trackers` based on `name` on the main process only.
|
2227 |
-
|
2228 |
Args:
|
2229 |
name (`str`):
|
2230 |
The name of a tracker, corresponding to the `.name` property.
|
2231 |
unwrap (`bool`):
|
2232 |
Whether to return the internal tracking mechanism or to return the wrapped tracker instead
|
2233 |
(recommended).
|
2234 |
-
|
2235 |
Returns:
|
2236 |
`GeneralTracker`: The tracker corresponding to `name` if it exists.
|
2237 |
-
|
2238 |
Example:
|
2239 |
-
|
2240 |
```python
|
2241 |
>>> from accelerate import Accelerator
|
2242 |
-
|
2243 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
2244 |
>>> accelerator.init_trackers("my_project")
|
2245 |
>>> tensorboard_tracker = accelerator.get_tracker("tensorboard")
|
@@ -2252,12 +1928,10 @@ class Accelerator:
|
|
2252 |
raise ValueError(f"{name} is not an available tracker stored inside the `Accelerator`.")
|
2253 |
# Handle tracker only made on main process
|
2254 |
return GeneralTracker(_blank=True)
|
2255 |
-
|
2256 |
@on_main_process
|
2257 |
def log(self, values: dict, step: int | None = None, log_kwargs: dict | None = {}):
|
2258 |
"""
|
2259 |
Logs `values` to all stored trackers in `self.trackers` on the main process only.
|
2260 |
-
|
2261 |
Args:
|
2262 |
values (`dict`):
|
2263 |
Values should be a dictionary-like object containing only types `int`, `float`, or `str`.
|
@@ -2269,12 +1943,9 @@ class Accelerator:
|
|
2269 |
```python
|
2270 |
{"wandb": {"tags": ["tag_a", "tag_b"]}}
|
2271 |
```
|
2272 |
-
|
2273 |
Example:
|
2274 |
-
|
2275 |
```python
|
2276 |
>>> from accelerate import Accelerator
|
2277 |
-
|
2278 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
2279 |
>>> accelerator.init_trackers("my_project")
|
2280 |
>>> accelerator.log({"loss": 0.5, "accuracy": 0.9})
|
@@ -2282,18 +1953,14 @@ class Accelerator:
|
|
2282 |
"""
|
2283 |
for tracker in self.trackers:
|
2284 |
tracker.log(values, step=step, **log_kwargs.get(tracker.name, {}))
|
2285 |
-
|
2286 |
@on_main_process
|
2287 |
def end_training(self):
|
2288 |
"""
|
2289 |
Runs any special end training behaviors, such as stopping trackers on the main process only. Should always be
|
2290 |
called at the end of your script if using experiment tracking.
|
2291 |
-
|
2292 |
Example:
|
2293 |
-
|
2294 |
```python
|
2295 |
>>> from accelerate import Accelerator
|
2296 |
-
|
2297 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
2298 |
>>> accelerator.init_trackers("my_project")
|
2299 |
>>> # Do training
|
@@ -2302,25 +1969,19 @@ class Accelerator:
|
|
2302 |
"""
|
2303 |
for tracker in self.trackers:
|
2304 |
tracker.finish()
|
2305 |
-
|
2306 |
def save(self, obj, f, safe_serialization=False):
|
2307 |
"""
|
2308 |
Save the object passed to disk once per machine. Use in place of `torch.save`.
|
2309 |
-
|
2310 |
Args:
|
2311 |
obj (`object`): The object to save.
|
2312 |
f (`str` or `os.PathLike`): Where to save the content of `obj`.
|
2313 |
safe_serialization (`bool`, *optional*, defaults to `False`): Whether to save `obj` using `safetensors`
|
2314 |
-
|
2315 |
Note:
|
2316 |
If `save_on_each_node` was passed in as a `ProjectConfiguration`, will save the object once per node,
|
2317 |
rather than only once on the main node.
|
2318 |
-
|
2319 |
Example:
|
2320 |
-
|
2321 |
```python
|
2322 |
>>> from accelerate import Accelerator
|
2323 |
-
|
2324 |
>>> accelerator = Accelerator()
|
2325 |
>>> arr = [0, 1, 2, 3]
|
2326 |
>>> accelerator.save(arr, "array.pkl")
|
@@ -2332,7 +1993,6 @@ class Accelerator:
|
|
2332 |
save_on_each_node=self.project_configuration.save_on_each_node,
|
2333 |
safe_serialization=safe_serialization,
|
2334 |
)
|
2335 |
-
|
2336 |
def save_model(
|
2337 |
self,
|
2338 |
model: torch.nn.Module,
|
@@ -2342,7 +2002,6 @@ class Accelerator:
|
|
2342 |
):
|
2343 |
"""
|
2344 |
Save a model so that it can be re-loaded using load_checkpoint_in_model
|
2345 |
-
|
2346 |
Arguments:
|
2347 |
model: (`torch.nn.Module`):
|
2348 |
Model to be saved. The model can be wrapped or unwraped.
|
@@ -2351,22 +2010,15 @@ class Accelerator:
|
|
2351 |
max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
|
2352 |
The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
|
2353 |
lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5MB"`).
|
2354 |
-
|
2355 |
<Tip warning={true}>
|
2356 |
-
|
2357 |
If a single weight of the model is bigger than `max_shard_size`, it will be in its own checkpoint shard
|
2358 |
which will be bigger than `max_shard_size`.
|
2359 |
-
|
2360 |
</Tip>
|
2361 |
-
|
2362 |
safe_serialization (`bool`, *optional*, defaults to `True`):
|
2363 |
Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
|
2364 |
-
|
2365 |
Example:
|
2366 |
-
|
2367 |
```python
|
2368 |
>>> from accelerate import Accelerator
|
2369 |
-
|
2370 |
>>> accelerator = Accelerator()
|
2371 |
>>> model = ...
|
2372 |
>>> accelerator.save_model(model, save_directory)
|
@@ -2375,9 +2027,7 @@ class Accelerator:
|
|
2375 |
if os.path.isfile(save_directory):
|
2376 |
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
|
2377 |
return
|
2378 |
-
|
2379 |
os.makedirs(save_directory, exist_ok=True)
|
2380 |
-
|
2381 |
# get the state_dict of the model
|
2382 |
if any(
|
2383 |
[
|
@@ -2391,25 +2041,20 @@ class Accelerator:
|
|
2391 |
if any(param.device == torch.device("meta") for param in model.parameters()):
|
2392 |
raise RuntimeError("You can't save the model since some parameters are on the meta device.")
|
2393 |
state_dict = self.get_state_dict(model)
|
2394 |
-
|
2395 |
if safe_serialization:
|
2396 |
state_dict = clean_state_dict_for_safetensors(state_dict)
|
2397 |
weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
|
2398 |
-
|
2399 |
# Shard the model if it is too big.
|
2400 |
shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size, weights_name=weights_name)
|
2401 |
-
|
2402 |
# Clean the folder from a previous save
|
2403 |
for filename in os.listdir(save_directory):
|
2404 |
full_filename = os.path.join(save_directory, filename)
|
2405 |
# If we have a shard file that is not going to be replaced, we delete it, but only from the main process
|
2406 |
# in distributed settings to avoid race conditions.
|
2407 |
weights_no_suffix = weights_name.replace(".bin", "")
|
2408 |
-
|
2409 |
# make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
|
2410 |
filename_no_suffix = filename.replace(".bin", "")
|
2411 |
reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
|
2412 |
-
|
2413 |
if (
|
2414 |
filename.startswith(weights_no_suffix)
|
2415 |
and os.path.isfile(full_filename)
|
@@ -2418,11 +2063,9 @@ class Accelerator:
|
|
2418 |
and PartialState().is_main_process
|
2419 |
):
|
2420 |
os.remove(full_filename)
|
2421 |
-
|
2422 |
# Save the model
|
2423 |
for shard_file, shard in shards.items():
|
2424 |
self.save(shard, os.path.join(save_directory, shard_file), safe_serialization=safe_serialization)
|
2425 |
-
|
2426 |
if index is None:
|
2427 |
path_to_weights = os.path.join(save_directory, WEIGHTS_NAME)
|
2428 |
logger.info(f"Model weights saved in {path_to_weights}")
|
@@ -2438,31 +2081,22 @@ class Accelerator:
|
|
2438 |
f"split in {len(shards)} checkpoint shards. You can find where each parameters has been saved in the "
|
2439 |
f"index located at {save_index_file}."
|
2440 |
)
|
2441 |
-
|
2442 |
def register_save_state_pre_hook(self, hook: Callable[..., None]) -> hooks.RemovableHandle:
|
2443 |
"""
|
2444 |
Registers a pre hook to be run before `save_checkpoint` is called in [`Accelerator.save_state`].
|
2445 |
-
|
2446 |
Args:
|
2447 |
hook (`Callable`):
|
2448 |
A function to be called in [`Accelerator.save_state`] before `save_checkpoint`.
|
2449 |
-
|
2450 |
The hook should have the following signature:
|
2451 |
-
|
2452 |
`hook(models: list[torch.nn.Module], weights: list[dict[str, torch.Tensor]], input_dir: str) -> None`
|
2453 |
-
|
2454 |
The `models` argument are the models as saved in the accelerator state under `accelerator._models`, `weigths`
|
2455 |
argument are the state dicts of the `models`, and the `input_dir` argument is the `input_dir` argument passed
|
2456 |
to [`Accelerator.load_state`].
|
2457 |
-
|
2458 |
<Tip>
|
2459 |
-
|
2460 |
Should only be used in conjunction with [`Accelerator.register_load_state_pre_hook`]. Can be useful to save
|
2461 |
configurations in addition to model weights. Can also be used to overwrite model saving with a customized
|
2462 |
method. In this case, make sure to remove already loaded weights from the weights list.
|
2463 |
-
|
2464 |
</Tip>
|
2465 |
-
|
2466 |
Returns:
|
2467 |
`torch.utils.hooks.RemovableHandle`: a handle that can be used to remove the added hook by calling
|
2468 |
`handle.remove()`
|
@@ -2470,25 +2104,18 @@ class Accelerator:
|
|
2470 |
handle = hooks.RemovableHandle(self._save_model_state_pre_hook)
|
2471 |
self._save_model_state_pre_hook[handle.id] = hook
|
2472 |
return handle
|
2473 |
-
|
2474 |
def save_state(self, output_dir: str = None, safe_serialization: bool = True, **save_model_func_kwargs):
|
2475 |
"""
|
2476 |
Saves the current states of the model, optimizer, scaler, RNG generators, and registered objects to a folder.
|
2477 |
-
|
2478 |
If a `ProjectConfiguration` was passed to the `Accelerator` object with `automatic_checkpoint_naming` enabled
|
2479 |
then checkpoints will be saved to `self.project_dir/checkpoints`. If the number of current saves is greater
|
2480 |
than `total_limit` then the oldest save is deleted. Each checkpoint is saved in seperate folders named
|
2481 |
`checkpoint_<iteration>`.
|
2482 |
-
|
2483 |
Otherwise they are just saved to `output_dir`.
|
2484 |
-
|
2485 |
<Tip>
|
2486 |
-
|
2487 |
Should only be used when wanting to save a checkpoint during training and restoring the state in the same
|
2488 |
environment.
|
2489 |
-
|
2490 |
</Tip>
|
2491 |
-
|
2492 |
Args:
|
2493 |
output_dir (`str` or `os.PathLike`):
|
2494 |
The name of the folder to save all relevant weights and states.
|
@@ -2497,12 +2124,9 @@ class Accelerator:
|
|
2497 |
save_model_func_kwargs (`dict`, *optional*):
|
2498 |
Additional keyword arguments for saving model which can be passed to the underlying save function, such
|
2499 |
as optional arguments for DeepSpeed's `save_checkpoint` function.
|
2500 |
-
|
2501 |
Example:
|
2502 |
-
|
2503 |
```python
|
2504 |
>>> from accelerate import Accelerator
|
2505 |
-
|
2506 |
>>> accelerator = Accelerator()
|
2507 |
>>> model, optimizer, lr_scheduler = ...
|
2508 |
>>> model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
|
@@ -2519,10 +2143,8 @@ class Accelerator:
|
|
2519 |
and (len(folders) + 1 > self.project_configuration.total_limit)
|
2520 |
and self.is_main_process
|
2521 |
):
|
2522 |
-
|
2523 |
def _inner(folder):
|
2524 |
return list(map(int, re.findall(r"[\/]?([0-9]+)(?=[^\/]*$)", folder)))[0]
|
2525 |
-
|
2526 |
folders.sort(key=_inner)
|
2527 |
logger.warning(
|
2528 |
f"Deleting {len(folders) + 1 - self.project_configuration.total_limit} checkpoints to make room for new checkpoint."
|
@@ -2537,11 +2159,9 @@ class Accelerator:
|
|
2537 |
self.wait_for_everyone()
|
2538 |
os.makedirs(output_dir, exist_ok=True)
|
2539 |
logger.info(f"Saving current state to {output_dir}")
|
2540 |
-
|
2541 |
if self.distributed_type == DistributedType.TPU:
|
2542 |
# Finish running the previous step before checkpointing
|
2543 |
xm.mark_step()
|
2544 |
-
|
2545 |
# Save the models taking care of FSDP and DeepSpeed nuances
|
2546 |
weights = []
|
2547 |
for i, model in enumerate(self._models):
|
@@ -2560,7 +2180,6 @@ class Accelerator:
|
|
2560 |
logger.info(f"Megatron-LM Model , Optimizer and Scheduler saved to output dir {output_dir}")
|
2561 |
else:
|
2562 |
weights.append(self.get_state_dict(model, unwrap=False))
|
2563 |
-
|
2564 |
# Save the optimizers taking care of FSDP and DeepSpeed nuances
|
2565 |
optimizers = []
|
2566 |
if self.distributed_type == DistributedType.FSDP:
|
@@ -2570,7 +2189,6 @@ class Accelerator:
|
|
2570 |
logger.info(f"FSDP Optimizer saved to output dir {output_dir}")
|
2571 |
elif self.distributed_type not in [DistributedType.DEEPSPEED, DistributedType.MEGATRON_LM]:
|
2572 |
optimizers = self._optimizers
|
2573 |
-
|
2574 |
# Save the lr schedulers taking care of DeepSpeed nuances
|
2575 |
schedulers = []
|
2576 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
@@ -2580,15 +2198,12 @@ class Accelerator:
|
|
2580 |
schedulers.append(scheduler)
|
2581 |
elif self.distributed_type not in [DistributedType.MEGATRON_LM]:
|
2582 |
schedulers = self._schedulers
|
2583 |
-
|
2584 |
# Save the samplers of the dataloaders
|
2585 |
dataloaders = self._dataloaders
|
2586 |
-
|
2587 |
# Call model loading hooks that might have been registered with
|
2588 |
# accelerator.register_model_state_hook
|
2589 |
for hook in self._save_model_state_pre_hook.values():
|
2590 |
hook(self._models, weights, output_dir)
|
2591 |
-
|
2592 |
save_location = save_accelerator_state(
|
2593 |
output_dir,
|
2594 |
weights,
|
@@ -2604,30 +2219,21 @@ class Accelerator:
|
|
2604 |
save_custom_state(obj, output_dir, i, save_on_each_node=self.project_configuration.save_on_each_node)
|
2605 |
self.project_configuration.iteration += 1
|
2606 |
return save_location
|
2607 |
-
|
2608 |
def register_load_state_pre_hook(self, hook: Callable[..., None]) -> hooks.RemovableHandle:
|
2609 |
"""
|
2610 |
Registers a pre hook to be run before [`load_checkpoint`] is called in [`Accelerator.load_state`].
|
2611 |
-
|
2612 |
Args:
|
2613 |
hook (`Callable`):
|
2614 |
A function to be called in [`Accelerator.load_state`] before `load_checkpoint`.
|
2615 |
-
|
2616 |
The hook should have the following signature:
|
2617 |
-
|
2618 |
`hook(models: list[torch.nn.Module], input_dir: str) -> None`
|
2619 |
-
|
2620 |
The `models` argument are the models as saved in the accelerator state under `accelerator._models`, and the
|
2621 |
`input_dir` argument is the `input_dir` argument passed to [`Accelerator.load_state`].
|
2622 |
-
|
2623 |
<Tip>
|
2624 |
-
|
2625 |
Should only be used in conjunction with [`Accelerator.register_save_state_pre_hook`]. Can be useful to load
|
2626 |
configurations in addition to model weights. Can also be used to overwrite model loading with a customized
|
2627 |
method. In this case, make sure to remove already loaded models from the models list.
|
2628 |
-
|
2629 |
</Tip>
|
2630 |
-
|
2631 |
Returns:
|
2632 |
`torch.utils.hooks.RemovableHandle`: a handle that can be used to remove the added hook by calling
|
2633 |
`handle.remove()`
|
@@ -2635,18 +2241,13 @@ class Accelerator:
|
|
2635 |
handle = hooks.RemovableHandle(self._load_model_state_pre_hook)
|
2636 |
self._load_model_state_pre_hook[handle.id] = hook
|
2637 |
return handle
|
2638 |
-
|
2639 |
def load_state(self, input_dir: str = None, **load_model_func_kwargs):
|
2640 |
"""
|
2641 |
Loads the current states of the model, optimizer, scaler, RNG generators, and registered objects.
|
2642 |
-
|
2643 |
<Tip>
|
2644 |
-
|
2645 |
Should only be used in conjunction with [`Accelerator.save_state`]. If a file is not registered for
|
2646 |
checkpointing, it will not be loaded if stored in the directory.
|
2647 |
-
|
2648 |
</Tip>
|
2649 |
-
|
2650 |
Args:
|
2651 |
input_dir (`str` or `os.PathLike`):
|
2652 |
The name of the folder all relevant weights and states were saved in. Can be `None` if
|
@@ -2655,12 +2256,9 @@ class Accelerator:
|
|
2655 |
Additional keyword arguments for loading model which can be passed to the underlying load function,
|
2656 |
such as optional arguments for DeepSpeed's `load_checkpoint` function or a `map_location` to load the
|
2657 |
model and optimizer on.
|
2658 |
-
|
2659 |
Example:
|
2660 |
-
|
2661 |
```python
|
2662 |
>>> from accelerate import Accelerator
|
2663 |
-
|
2664 |
>>> accelerator = Accelerator()
|
2665 |
>>> model, optimizer, lr_scheduler = ...
|
2666 |
>>> model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
|
@@ -2676,16 +2274,13 @@ class Accelerator:
|
|
2676 |
# Pick up from automatic checkpoint naming
|
2677 |
input_dir = os.path.join(self.project_dir, "checkpoints")
|
2678 |
folders = [os.path.join(input_dir, folder) for folder in os.listdir(input_dir)]
|
2679 |
-
|
2680 |
def _inner(folder):
|
2681 |
return list(map(int, re.findall(r"[\/]?([0-9]+)(?=[^\/]*$)", folder)))[0]
|
2682 |
-
|
2683 |
folders.sort(key=_inner)
|
2684 |
input_dir = folders[-1]
|
2685 |
else:
|
2686 |
raise ValueError("No input_dir provided and automatic checkpoint naming is disabled.")
|
2687 |
logger.info(f"Loading states from {input_dir}")
|
2688 |
-
|
2689 |
# Load the models taking care of FSDP and DeepSpeed nuances
|
2690 |
models = []
|
2691 |
for i, model in enumerate(self._models):
|
@@ -2704,7 +2299,6 @@ class Accelerator:
|
|
2704 |
logger.info(f"Megatron-LM Model , Optimizer and Scheduler loaded from input dir {input_dir}")
|
2705 |
else:
|
2706 |
models.append(model)
|
2707 |
-
|
2708 |
# Load the optimizers taking care of FSDP and DeepSpeed nuances
|
2709 |
optimizers = []
|
2710 |
if self.distributed_type == DistributedType.FSDP:
|
@@ -2714,7 +2308,6 @@ class Accelerator:
|
|
2714 |
logger.info(f"FSDP Optimizer loaded from input dir {input_dir}")
|
2715 |
elif self.distributed_type not in [DistributedType.DEEPSPEED, DistributedType.MEGATRON_LM]:
|
2716 |
optimizers = self._optimizers
|
2717 |
-
|
2718 |
# Load the lr schedulers taking care of DeepSpeed nuances
|
2719 |
schedulers = []
|
2720 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
@@ -2724,14 +2317,11 @@ class Accelerator:
|
|
2724 |
schedulers.append(scheduler)
|
2725 |
elif self.distributed_type not in [DistributedType.MEGATRON_LM]:
|
2726 |
schedulers = self._schedulers
|
2727 |
-
|
2728 |
dataloaders = self._dataloaders
|
2729 |
-
|
2730 |
# Call model loading hooks that might have been registered with
|
2731 |
# accelerator.register_model_state_hook
|
2732 |
for hook in self._load_model_state_pre_hook.values():
|
2733 |
hook(models, input_dir)
|
2734 |
-
|
2735 |
map_location = load_model_func_kwargs.pop("map_location", None)
|
2736 |
if map_location is None:
|
2737 |
if self.num_processes > 1 and self.distributed_type in (
|
@@ -2741,7 +2331,6 @@ class Accelerator:
|
|
2741 |
map_location = "on_device"
|
2742 |
else:
|
2743 |
map_location = "cpu"
|
2744 |
-
|
2745 |
load_accelerator_state(
|
2746 |
input_dir,
|
2747 |
models,
|
@@ -2767,17 +2356,13 @@ class Accelerator:
|
|
2767 |
logger.info(f"Loading in {len(custom_checkpoints)} custom states")
|
2768 |
for index, obj in enumerate(self._custom_objects):
|
2769 |
load_custom_state(obj, input_dir, index)
|
2770 |
-
|
2771 |
def free_memory(self):
|
2772 |
"""
|
2773 |
Will release all references to the internal objects stored and call the garbage collector. You should call this
|
2774 |
method between two trainings with different models/optimizers. Also will reset `Accelerator.step` to 0.
|
2775 |
-
|
2776 |
Example:
|
2777 |
-
|
2778 |
```python
|
2779 |
>>> from accelerate import Accelerator
|
2780 |
-
|
2781 |
>>> accelerator = Accelerator()
|
2782 |
>>> model, optimizer, scheduler = ...
|
2783 |
>>> model, optimizer, scheduler = accelerator.prepare(model, optimizer, scheduler)
|
@@ -2792,17 +2377,13 @@ class Accelerator:
|
|
2792 |
self.deepspeed_engine_wrapped = None
|
2793 |
self.step = 0
|
2794 |
release_memory()
|
2795 |
-
|
2796 |
def clear(self):
|
2797 |
"""
|
2798 |
Alias for [`Accelerate.free_memory`], releases all references to the internal objects stored and call the
|
2799 |
garbage collector. You should call this method between two trainings with different models/optimizers.
|
2800 |
-
|
2801 |
Example:
|
2802 |
-
|
2803 |
```python
|
2804 |
>>> from accelerate import Accelerator
|
2805 |
-
|
2806 |
>>> accelerator = Accelerator()
|
2807 |
>>> model, optimizer, scheduler = ...
|
2808 |
>>> model, optimizer, scheduler = accelerator.prepare(model, optimizer, scheduler)
|
@@ -2811,7 +2392,6 @@ class Accelerator:
|
|
2811 |
```
|
2812 |
"""
|
2813 |
self.free_memory()
|
2814 |
-
|
2815 |
def _get_named_parameters(self, *args):
|
2816 |
named_parameters = {}
|
2817 |
for obj in args:
|
@@ -2819,7 +2399,6 @@ class Accelerator:
|
|
2819 |
obj = extract_model_from_parallel(obj)
|
2820 |
named_parameters.update({n: p for n, p in obj.named_parameters()})
|
2821 |
return named_parameters
|
2822 |
-
|
2823 |
def _get_devices(self, *args):
|
2824 |
model_device = None
|
2825 |
optimizer_device = None
|
@@ -2836,27 +2415,21 @@ class Accelerator:
|
|
2836 |
optimizer_device = param_group["params"][0].device
|
2837 |
break
|
2838 |
return (model_device, optimizer_device)
|
2839 |
-
|
2840 |
def get_state_dict(self, model, unwrap=True):
|
2841 |
"""
|
2842 |
Returns the state dictionary of a model sent through [`Accelerator.prepare`] potentially without full
|
2843 |
precision.
|
2844 |
-
|
2845 |
Args:
|
2846 |
model (`torch.nn.Module`):
|
2847 |
A PyTorch model sent through [`Accelerator.prepare`]
|
2848 |
unwrap (`bool`, *optional*, defaults to `True`):
|
2849 |
Whether to return the original underlying state_dict of `model` or to return the wrapped state_dict
|
2850 |
-
|
2851 |
Returns:
|
2852 |
`dict`: The state dictionary of the model potentially without full precision.
|
2853 |
-
|
2854 |
Example:
|
2855 |
-
|
2856 |
```python
|
2857 |
>>> import torch
|
2858 |
>>> from accelerate import Accelerator
|
2859 |
-
|
2860 |
>>> accelerator = Accelerator()
|
2861 |
>>> net = torch.nn.Linear(2, 2)
|
2862 |
>>> net = accelerator.prepare(net)
|
@@ -2876,12 +2449,10 @@ class Accelerator:
|
|
2876 |
)
|
2877 |
else:
|
2878 |
from deepspeed.checkpoint.utils import clone_tensors_for_torch_save
|
2879 |
-
|
2880 |
state_dict = clone_tensors_for_torch_save(self.unwrap_model(model).state_dict())
|
2881 |
elif self.distributed_type == DistributedType.FSDP:
|
2882 |
from torch.distributed.fsdp import FullStateDictConfig, StateDictType
|
2883 |
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
|
2884 |
-
|
2885 |
full_state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
|
2886 |
with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, full_state_dict_config):
|
2887 |
state_dict = model.state_dict()
|
@@ -2889,27 +2460,18 @@ class Accelerator:
|
|
2889 |
if unwrap:
|
2890 |
model = self.unwrap_model(model)
|
2891 |
state_dict = model.state_dict()
|
2892 |
-
|
2893 |
return state_dict
|
2894 |
-
|
2895 |
def register_for_checkpointing(self, *objects):
|
2896 |
"""
|
2897 |
Makes note of `objects` and will save or load them in during `save_state` or `load_state`.
|
2898 |
-
|
2899 |
These should be utilized when the state is being loaded or saved in the same script. It is not designed to be
|
2900 |
used in different scripts.
|
2901 |
-
|
2902 |
<Tip>
|
2903 |
-
|
2904 |
Every `object` must have a `load_state_dict` and `state_dict` function to be stored.
|
2905 |
-
|
2906 |
</Tip>
|
2907 |
-
|
2908 |
Example:
|
2909 |
-
|
2910 |
```python
|
2911 |
>>> from accelerate import Accelerator
|
2912 |
-
|
2913 |
>>> accelerator = Accelerator()
|
2914 |
>>> # Assume `CustomObject` has a `state_dict` and `load_state_dict` function.
|
2915 |
>>> obj = CustomObject()
|
@@ -2927,21 +2489,16 @@ class Accelerator:
|
|
2927 |
err += f"\n\t- Item at index {index}, `{get_pretty_name(obj)}`"
|
2928 |
raise ValueError(err)
|
2929 |
self._custom_objects.extend(objects)
|
2930 |
-
|
2931 |
@contextmanager
|
2932 |
def autocast(self, cache_enabled: bool = False, autocast_handler: AutocastKwargs = None):
|
2933 |
"""
|
2934 |
Will apply automatic mixed-precision inside the block inside this context manager, if it is enabled. Nothing
|
2935 |
different will happen otherwise.
|
2936 |
-
|
2937 |
A different `autocast_handler` can be passed in to override the one set in the `Accelerator` object. This is
|
2938 |
useful in blocks under `autocast` where you want to revert to fp32.
|
2939 |
-
|
2940 |
Example:
|
2941 |
-
|
2942 |
```python
|
2943 |
>>> from accelerate import Accelerator
|
2944 |
-
|
2945 |
>>> accelerator = Accelerator(mixed_precision="fp16")
|
2946 |
>>> with accelerator.autocast():
|
2947 |
... train()
|
@@ -2963,7 +2520,6 @@ class Accelerator:
|
|
2963 |
autocast_context.__enter__()
|
2964 |
yield
|
2965 |
autocast_context.__exit__(*sys.exc_info())
|
2966 |
-
|
2967 |
@property
|
2968 |
def optimizer_step_was_skipped(self):
|
2969 |
"""
|
@@ -2974,20 +2530,15 @@ class Accelerator:
|
|
2974 |
if optimizer.step_was_skipped:
|
2975 |
return True
|
2976 |
return False
|
2977 |
-
|
2978 |
def skip_first_batches(self, dataloader, num_batches: int = 0):
|
2979 |
"""
|
2980 |
Creates a new `torch.utils.data.DataLoader` that will efficiently skip the first `num_batches`.
|
2981 |
-
|
2982 |
Args:
|
2983 |
dataloader (`torch.utils.data.DataLoader`): The data loader in which to skip batches.
|
2984 |
num_batches (`int`, *optional*, defaults to 0): The number of batches to skip
|
2985 |
-
|
2986 |
Example:
|
2987 |
-
|
2988 |
```python
|
2989 |
>>> from accelerate import Accelerator
|
2990 |
-
|
2991 |
>>> accelerator = Accelerator()
|
2992 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
2993 |
>>> skipped_dataloader = accelerator.skip_first_batches(dataloader, num_batches=2)
|
@@ -2998,7 +2549,6 @@ class Accelerator:
|
|
2998 |
... loss = loss_func(output, target)
|
2999 |
... accelerator.backward(loss)
|
3000 |
... optimizer.step()
|
3001 |
-
|
3002 |
>>> # subsequent epochs
|
3003 |
>>> for input, target in dataloader:
|
3004 |
... optimizer.zero_grad()
|
@@ -3006,11 +2556,9 @@ class Accelerator:
|
|
3006 |
```
|
3007 |
"""
|
3008 |
return skip_first_batches(dataloader, num_batches=num_batches)
|
3009 |
-
|
3010 |
def __deepcopy__(self, memo):
|
3011 |
logger.info("Deep copying the `Accelerator` object, note that this will point to the same original object.")
|
3012 |
return self
|
3013 |
-
|
3014 |
def verify_device_map(self, model: torch.nn.Module) -> bool:
|
3015 |
"""
|
3016 |
Verifies that `model` has not been prepared with big model inference with a device-map resembling `auto`.
|
@@ -3019,5 +2567,4 @@ class Accelerator:
|
|
3019 |
for m in model.modules():
|
3020 |
if hasattr(m, "hf_device_map") and len(m.hf_device_map) > 1:
|
3021 |
return True
|
3022 |
-
|
3023 |
return False
|
|
|
|
|
1 |
logger = get_logger(__name__)
|
|
|
|
|
2 |
class Accelerator:
|
3 |
"""
|
4 |
Creates an instance of an accelerator for distributed training (on multi-GPU, TPU) or mixed precision training.
|
|
|
5 |
Args:
|
6 |
device_placement (`bool`, *optional*, defaults to `True`):
|
7 |
Whether or not the accelerator should put objects on device (tensors yielded by the dataloader, model,
|
|
|
35 |
rng_types (list of `str` or [`~utils.RNGType`]):
|
36 |
The list of random number generators to synchronize at the beginning of each iteration in your prepared
|
37 |
dataloaders. Should be one or several of:
|
|
|
38 |
- `"torch"`: the base torch random number generator
|
39 |
- `"cuda"`: the CUDA random number generator (GPU only)
|
40 |
- `"xla"`: the XLA random number generator (TPU only)
|
41 |
- `"generator"`: the `torch.Generator` of the sampler (or batch sampler if there is no sampler in your
|
42 |
dataloader) or of the iterable dataset (if it exists) if the underlying dataset is of that type.
|
|
|
43 |
Will default to `["torch"]` for PyTorch versions <=1.5.1 and `["generator"]` for PyTorch versions >= 1.6.
|
44 |
log_with (list of `str`, [`~utils.LoggerType`] or [`~tracking.GeneralTracker`], *optional*):
|
45 |
A list of loggers to be setup for experiment tracking. Should be one or several of:
|
|
|
46 |
- `"all"`
|
47 |
- `"tensorboard"`
|
48 |
- `"wandb"`
|
|
|
73 |
gradient_accumulation_plugin (`GradientAccumulationPlugin`, *optional*):
|
74 |
A configuration for how gradient accumulation should be handled, if more tweaking than just the
|
75 |
`gradient_accumulation_steps` is needed.
|
|
|
76 |
**Available attributes:**
|
|
|
77 |
- **device** (`torch.device`) -- The device to use.
|
78 |
- **distributed_type** ([`~utils.DistributedType`]) -- The distributed training configuration.
|
79 |
- **local_process_index** (`int`) -- The process index on the current machine.
|
|
|
121 |
raise ValueError(
|
122 |
f"Unknown mixed_precision mode: {mixed_precision}. Choose between {PrecisionType.list()}"
|
123 |
)
|
|
|
124 |
dynamo_plugin = TorchDynamoPlugin() if dynamo_backend is None else TorchDynamoPlugin(backend=dynamo_backend)
|
|
|
125 |
if deepspeed_plugin is None: # init from env variables
|
126 |
deepspeed_plugin = (
|
127 |
DeepSpeedPlugin() if os.environ.get("ACCELERATE_USE_DEEPSPEED", "false") == "true" else None
|
|
|
136 |
raise ImportError("DeepSpeed is not installed => run `pip install deepspeed` or build it from source.")
|
137 |
if compare_versions("deepspeed", "<", "0.9.3"):
|
138 |
raise ImportError("DeepSpeed version must be >= 0.9.3. Please update DeepSpeed.")
|
|
|
139 |
mixed_precision = (
|
140 |
os.environ.get("ACCELERATE_MIXED_PRECISION", "no") if mixed_precision is None else mixed_precision
|
141 |
)
|
142 |
deepspeed_plugin.set_mixed_precision(mixed_precision)
|
143 |
deepspeed_plugin.set_deepspeed_weakref()
|
|
|
144 |
if os.environ.get("ACCELERATE_USE_FSDP", "false") == "true" or isinstance(
|
145 |
fsdp_plugin, FullyShardedDataParallelPlugin
|
146 |
):
|
147 |
if is_torch_version("<", FSDP_PYTORCH_VERSION):
|
148 |
raise ValueError(f"FSDP requires PyTorch >= {FSDP_PYTORCH_VERSION}")
|
|
|
149 |
if fsdp_plugin is None: # init from env variables
|
150 |
fsdp_plugin = (
|
151 |
FullyShardedDataParallelPlugin() if os.environ.get("ACCELERATE_USE_FSDP", "false") == "true" else None
|
|
|
154 |
if not isinstance(fsdp_plugin, FullyShardedDataParallelPlugin):
|
155 |
raise TypeError("`fsdp_plugin` must be a FullyShardedDataParallelPlugin object.")
|
156 |
os.environ["ACCELERATE_USE_FSDP"] = "true" # use FSDP if plugin is provided
|
|
|
157 |
if megatron_lm_plugin is None: # init from env variables
|
158 |
megatron_lm_plugin = (
|
159 |
MegatronLMPlugin() if os.environ.get("ACCELERATE_USE_MEGATRON_LM", "false") == "true" else None
|
|
|
162 |
if not isinstance(megatron_lm_plugin, MegatronLMPlugin):
|
163 |
raise TypeError("`megatron_lm_plugin` must be a MegatronLMPlugin object.")
|
164 |
os.environ["ACCELERATE_USE_MEGATRON_LM"] = "true" # use MegatronLM if plugin is provided
|
|
|
165 |
if megatron_lm_plugin:
|
166 |
if not is_megatron_lm_available():
|
167 |
raise ImportError("Megatron is not installed. please build it from source.")
|
|
|
168 |
# Kwargs handlers
|
169 |
self.ddp_handler = None
|
170 |
self.scaler_handler = None
|
|
|
203 |
self.autocast_handler = handler
|
204 |
if self.fp8_recipe_handler is None and mixed_precision == "fp8":
|
205 |
self.fp8_recipe_handler = FP8RecipeKwargs()
|
|
|
206 |
kwargs = self.init_handler.to_kwargs() if self.init_handler is not None else {}
|
207 |
self.state = AcceleratorState(
|
208 |
mixed_precision=mixed_precision,
|
|
|
214 |
_from_accelerator=True,
|
215 |
**kwargs,
|
216 |
)
|
|
|
217 |
trackers = filter_trackers(log_with, self.logging_dir)
|
218 |
if len(trackers) < 1 and log_with is not None:
|
219 |
warnings.warn(f"`log_with={log_with}` was passed but no supported trackers are currently installed.")
|
220 |
self.log_with = trackers
|
|
|
221 |
if (
|
222 |
(mixed_precision != "bf16")
|
223 |
and getattr(self.state, "downcast_bfloat", False)
|
224 |
and (self.state.distributedType != DistributedType.TPU)
|
225 |
):
|
226 |
raise ValueError("Can only use `downcast_bf16` when using `mixed_precision='bf16'` and on a TPU")
|
|
|
227 |
if gradient_accumulation_plugin is not None:
|
228 |
if gradient_accumulation_steps != 1:
|
229 |
raise ValueError(
|
|
|
242 |
raise ValueError(
|
243 |
"Gradient accumulation is not supported on TPU. Please set `gradient_accumulation_steps` to 1 and don't pass in a `GradientAccumulationPlugin` object."
|
244 |
)
|
|
|
245 |
self.device_placement = device_placement
|
246 |
self.split_batches = split_batches
|
247 |
self.dispatch_batches = dispatch_batches
|
248 |
self.even_batches = even_batches
|
249 |
self.step_scheduler_with_optimizer = step_scheduler_with_optimizer
|
|
|
250 |
# Mixed precision attributes
|
251 |
self.scaler = None
|
252 |
self.native_amp = False
|
|
|
262 |
kwargs = self.scaler_handler.to_kwargs() if self.scaler_handler is not None else {}
|
263 |
if self.distributed_type == DistributedType.FSDP:
|
264 |
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
|
|
|
265 |
self.scaler = ShardedGradScaler(**kwargs)
|
266 |
elif is_npu_available():
|
267 |
self.scaler = torch.npu.amp.GradScaler(**kwargs)
|
268 |
else:
|
269 |
self.scaler = torch.cuda.amp.GradScaler(**kwargs)
|
|
|
270 |
elif self.state.mixed_precision == "bf16" and self.distributed_type not in (
|
271 |
DistributedType.DEEPSPEED,
|
272 |
DistributedType.MEGATRON_LM,
|
|
|
277 |
self.native_amp = is_bf16_available(True)
|
278 |
if mixed_precision == "bf16" and not self.native_amp and not is_tpu_available():
|
279 |
raise ValueError(err.format(mode="bf16", requirement="PyTorch >= 1.10 and a supported device."))
|
|
|
280 |
# Start of internal step tracking
|
281 |
self.step = 0
|
|
|
282 |
# Internal references to the training objects
|
283 |
self._optimizers = []
|
284 |
self._models = []
|
285 |
self._schedulers = []
|
286 |
self._dataloaders = []
|
287 |
self._custom_objects = []
|
|
|
288 |
# Hooks
|
289 |
self._load_model_state_pre_hook = OrderedDict()
|
290 |
self._save_model_state_pre_hook = OrderedDict()
|
|
|
291 |
# RNG Types
|
292 |
self.rng_types = rng_types
|
293 |
if self.rng_types is None:
|
294 |
self.rng_types = ["generator"]
|
|
|
295 |
# Set a flag tensor for early stopping and other breakpoints
|
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(
|
|
|
341 |
FutureWarning,
|
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
|
354 |
distributed inference, such as with different prompts.
|
|
|
355 |
Note that when using a `dict`, all keys need to have the same number of elements.
|
|
|
356 |
Args:
|
357 |
inputs (`list`, `tuple`, `torch.Tensor`, or `dict` of `list`/`tuple`/`torch.Tensor`):
|
358 |
The input to split between processes.
|
|
|
361 |
number of elements. Useful when trying to perform actions such as `Accelerator.gather()` on the outputs
|
362 |
or passing in less inputs than there are processes. If so, just remember to drop the padded elements
|
363 |
afterwards.
|
|
|
364 |
Example:
|
|
|
365 |
```python
|
366 |
# Assume there are two processes
|
367 |
from accelerate import Accelerator
|
|
|
368 |
accelerator = Accelerator()
|
369 |
with accelerator.split_between_processes(["A", "B", "C"]) as inputs:
|
370 |
print(inputs)
|
|
|
372 |
["A", "B"]
|
373 |
# Process 1
|
374 |
["C"]
|
|
|
375 |
with accelerator.split_between_processes(["A", "B", "C"], apply_padding=True) as inputs:
|
376 |
print(inputs)
|
377 |
# Process 0
|
|
|
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
|
388 |
`PartialState` class.
|
|
|
389 |
Args:
|
390 |
function (`Callable`): The function to decorate.
|
|
|
391 |
Example:
|
|
|
392 |
```python
|
393 |
>>> from accelerate import Accelerator
|
|
|
394 |
>>> accelerator = Accelerator()
|
|
|
|
|
395 |
>>> @accelerator.on_main_process
|
396 |
... def print_something():
|
397 |
... print("This will be printed by process 0 only.")
|
|
|
|
|
398 |
>>> print_something()
|
399 |
"This will be printed by process 0 only"
|
400 |
```
|
|
|
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
|
416 |
`PartialState` class.
|
|
|
417 |
Args:
|
418 |
function (`Callable`): The function to decorate.
|
|
|
419 |
Example:
|
420 |
```python
|
421 |
# Assume we have 2 servers with 4 processes each.
|
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:
|
429 |
"This will be printed by process 0 only"
|
|
|
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
|
448 |
`PartialState` class.
|
|
|
449 |
Args:
|
450 |
function (`Callable`): The function to decorate.
|
|
|
451 |
Example:
|
452 |
```python
|
453 |
# Assume we have 4 processes.
|
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"
|
461 |
```
|
|
|
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
|
477 |
`PartialState` class.
|
|
|
478 |
Args:
|
479 |
function (`Callable`, `optional`):
|
480 |
The function to decorate.
|
481 |
process_index (`int`, `optional`):
|
482 |
The index of the process on which to run the function.
|
|
|
483 |
Example:
|
484 |
```python
|
485 |
# Assume we have 4 processes.
|
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"
|
493 |
```
|
|
|
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
|
512 |
the `PartialState` class.
|
|
|
513 |
Args:
|
514 |
function (`Callable`, *optional*):
|
515 |
The function to decorate.
|
516 |
local_process_index (`int`, *optional*):
|
517 |
The index of the local process on which to run the function.
|
|
|
518 |
Example:
|
519 |
```python
|
520 |
# Assume we have 2 servers with 4 processes each.
|
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:
|
528 |
"Printed on process 2"
|
|
|
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.
|
|
|
551 |
The other processes will enter the with block after the main process exits.
|
|
|
552 |
Example:
|
|
|
553 |
```python
|
554 |
>>> from accelerate import Accelerator
|
|
|
555 |
>>> accelerator = Accelerator()
|
556 |
>>> with accelerator.main_process_first():
|
557 |
... # This will be printed first by process 0 then in a seemingly
|
|
|
561 |
"""
|
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.
|
|
|
568 |
The other processes will enter the with block after the main process exits.
|
|
|
569 |
Example:
|
|
|
570 |
```python
|
571 |
>>> from accelerate import Accelerator
|
|
|
572 |
>>> accelerator = Accelerator()
|
573 |
>>> with accelerator.local_main_process_first():
|
574 |
... # This will be printed first by local process 0 then in a seemingly
|
|
|
578 |
"""
|
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
|
585 |
`torch.nn.parallel.DistributedDataParallel.no_sync`.
|
|
|
586 |
If `model` is not in DDP, this context manager does nothing
|
|
|
587 |
Args:
|
588 |
model (`torch.nn.Module`):
|
589 |
PyTorch Module that was prepared with `Accelerator.prepare`
|
|
|
590 |
Example:
|
|
|
591 |
```python
|
592 |
>>> from accelerate import Accelerator
|
|
|
593 |
>>> accelerator = Accelerator()
|
594 |
>>> dataloader, model, optimizer = accelerator.prepare(dataloader, model, optimizer)
|
595 |
>>> input_a = next(iter(dataloader))
|
596 |
>>> input_b = next(iter(dataloader))
|
|
|
597 |
>>> with accelerator.no_sync():
|
598 |
... outputs = model(input_a)
|
599 |
... loss = loss_func(outputs)
|
|
|
609 |
context = contextlib.nullcontext
|
610 |
if self.use_distributed:
|
611 |
context = getattr(model, "no_sync", context)
|
|
|
612 |
with context():
|
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).
|
|
|
619 |
If the script is not launched in distributed mode, this context manager does nothing.
|
|
|
620 |
Args:
|
621 |
model (`torch.nn.Module`):
|
622 |
The model for which to trigger the gradient synchronization.
|
|
|
623 |
Example:
|
|
|
624 |
```python
|
625 |
>>> from accelerate import Accelerator
|
|
|
626 |
>>> accelerator = Accelerator()
|
627 |
>>> dataloader, model, optimizer = accelerator.prepare(dataloader, model, optimizer)
|
|
|
628 |
>>> with accelerator.no_sync():
|
629 |
... loss_a = loss_func(model(input_a)) # first forward pass
|
630 |
... loss_b = loss_func(model(input_b)) # second forward pass
|
|
|
638 |
if not isinstance(model, torch.nn.parallel.DistributedDataParallel):
|
639 |
yield
|
640 |
return
|
|
|
641 |
old_require_backward_grad_sync = model.require_backward_grad_sync
|
642 |
old_require_forward_param_sync = model.require_forward_param_sync
|
|
|
643 |
# EXPERIMENTAL: This will force grad sync during `backward()`, but it is unknown if it breaks other DDP features.
|
644 |
# https://github.com/pytorch/pytorch/blob/e1502c0cdbfd17548c612f25d5a65b1e4b86224d/torch/nn/parallel/distributed.py#L1453-L1466
|
645 |
model.require_backward_grad_sync = True
|
|
|
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:
|
|
|
659 |
else:
|
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
|
|
|
678 |
Args:
|
679 |
*models (list of `torch.nn.Module`):
|
680 |
PyTorch Modules that were prepared with `Accelerator.prepare`. Models passed to `accumulate()` will
|
681 |
skip gradient syncing during backward pass in distributed training
|
|
|
682 |
Example:
|
|
|
683 |
```python
|
684 |
>>> from accelerate import Accelerator
|
|
|
685 |
>>> accelerator = Accelerator(gradient_accumulation_steps=1)
|
686 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
|
|
687 |
>>> for input, output in dataloader:
|
688 |
... with accelerator.accumulate(model):
|
689 |
... outputs = model(input)
|
|
|
699 |
for m in models:
|
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
|
706 |
around `torch.distributed.algorithms.join`. This is useful when the total batch size does not evenly divide the
|
707 |
length of the dataset.
|
|
|
708 |
Args:
|
709 |
joinables (`list[torch.distributed.algorithms.Joinable]`):
|
710 |
A list of models or optimizers that subclass `torch.distributed.algorithms.Joinable`. Most commonly, a
|
|
|
712 |
even_batches (`bool`, *optional*)
|
713 |
If set, this will override the value of `even_batches` set in the `Accelerator`. If it is not provided,
|
714 |
the default `Accelerator` value wil be used.
|
|
|
715 |
<Tip warning={true}>
|
|
|
716 |
`join_uneven_inputs` is only supported for Distributed Data Parallel training on multiple GPUs. For any other
|
717 |
configuration, this method will have no effect.
|
|
|
718 |
</Tip>
|
|
|
719 |
<Tip warning={true}>
|
|
|
720 |
Overidding `even_batches` will not affect iterable-style data loaders.
|
|
|
721 |
</Tip>
|
|
|
722 |
Example:
|
|
|
723 |
```python
|
724 |
>>> from accelerate import Accelerator
|
|
|
725 |
>>> accelerator = Accelerator(even_batches=True)
|
726 |
>>> ddp_model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
|
|
|
727 |
>>> with accelerator.join_uneven_inputs([ddp_model], even_batches=False):
|
728 |
... for input, output in dataloader:
|
729 |
... outputs = model(input)
|
|
|
735 |
"""
|
736 |
if self.distributed_type in (DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.MULTI_XPU):
|
737 |
dl_even_batches_values = []
|
|
|
738 |
if even_batches is not None:
|
739 |
iterable_dl_seen = False
|
740 |
# override value in batch sampler for map-style datasets
|
|
|
744 |
continue
|
745 |
dl_even_batches_values.append((dl_idx, dl.batch_sampler.even_batches))
|
746 |
dl.batch_sampler.even_batches = even_batches
|
|
|
747 |
if iterable_dl_seen:
|
748 |
warnings.warn(
|
749 |
"Overridding even_batches is only supported for map-style datasets, yet some dataloaders given were iterable"
|
750 |
)
|
751 |
else:
|
752 |
even_batches = self.even_batches
|
|
|
753 |
enable_join = False if even_batches else True
|
754 |
try:
|
755 |
with Join(joinables, enable=enable_join, throw_on_early_termination=False):
|
|
|
764 |
warnings.warn(
|
765 |
"Joining uneven inputs is only supported for multi-GPU training, as a result `join_uneven_inputs` will have no effect."
|
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.
|
|
|
772 |
Example:
|
|
|
773 |
```python
|
774 |
>>> from accelerate import Accelerator
|
|
|
775 |
>>> accelerator = Accelerator()
|
776 |
>>> accelerator.print("Hello world!")
|
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 |
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
|
799 |
order.
|
|
|
800 |
Args:
|
801 |
*args (list of objects):
|
802 |
Any of the following type of objects:
|
|
|
803 |
- `torch.utils.data.DataLoader`: PyTorch Dataloader
|
804 |
- `torch.nn.Module`: PyTorch Module
|
805 |
- `torch.optim.Optimizer`: PyTorch Optimizer
|
806 |
- `torch.optim.lr_scheduler.LRScheduler`: PyTorch LR Scheduler
|
|
|
807 |
device_placement (`list[bool]`, *optional*):
|
808 |
Used to customize whether automatic device placement should be performed for each object passed. Needs
|
809 |
to be a list of the same length as `args`. Not compatible with DeepSpeed or FSDP.
|
|
|
810 |
<Tip>
|
|
|
811 |
You don't need to prepare a model if you only use it for inference without any kind of mixed precision
|
|
|
812 |
</Tip>
|
|
|
813 |
Examples:
|
|
|
814 |
```python
|
815 |
>>> from accelerate import Accelerator
|
|
|
816 |
>>> accelerator = Accelerator()
|
817 |
>>> # Assume a model, optimizer, data_loader and scheduler are defined
|
818 |
>>> model, optimizer, data_loader, scheduler = accelerator.prepare(model, optimizer, data_loader, scheduler)
|
819 |
```
|
|
|
820 |
```python
|
821 |
>>> from accelerate import Accelerator
|
|
|
822 |
>>> accelerator = Accelerator()
|
823 |
>>> # Assume a model, optimizer, data_loader and scheduler are defined
|
824 |
>>> device_placement = [True, True, False, False]
|
|
|
836 |
raise ValueError(
|
837 |
f"`device_placement` should be a list with {len(args)} elements (the number of objects passed)."
|
838 |
)
|
|
|
839 |
for obj in args:
|
840 |
# TODO: Look at enabling native TP training directly with a proper config
|
841 |
if (
|
|
|
848 |
"You can't train a model that has been loaded with `device_map='auto'` in any distributed mode."
|
849 |
" Please rerun your script specifying `--num_processes=1` or by launching with `python {{myscript.py}}`."
|
850 |
)
|
|
|
851 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
852 |
model_count = 0
|
853 |
for obj in args:
|
|
|
857 |
raise AssertionError(
|
858 |
"You can't use same `Accelerator()` instance with multiple models when using DeepSpeed"
|
859 |
)
|
|
|
860 |
# On TPUs, putting the model on the XLA device will create new parameters, so the corresponding optimizer will
|
861 |
# have parameters disconnected from the model (so no training :-( ).
|
862 |
# If the model and optimizer have parameters on different devices we raise an error.
|
|
|
870 |
"the flag default value for `device_placement` in your `Accelerator` to let it handle that "
|
871 |
"part for you."
|
872 |
)
|
|
|
873 |
# If we're dealing with device placement, this deals with that by...
|
874 |
tpu_should_fix_optimizer = self.device_placement and self.distributed_type == DistributedType.TPU
|
875 |
if tpu_should_fix_optimizer or (self.mixed_precision == "fp8" and self.fp8_recipe_handler.backend == "TE"):
|
876 |
# 1. grabbing old model parameters
|
877 |
old_named_params = self._get_named_parameters(*args)
|
|
|
878 |
if self.distributed_type in [DistributedType.MULTI_CPU, DistributedType.MULTI_XPU, DistributedType.NO]:
|
879 |
if self.device.type == "cpu" and self.state.use_ipex:
|
880 |
args = self._prepare_ipex(*args)
|
|
|
893 |
self._prepare_one(obj, first_pass=True, device_placement=d) for obj, d in zip(args, device_placement)
|
894 |
)
|
895 |
result = tuple(self._prepare_one(obj, device_placement=d) for obj, d in zip(result, device_placement))
|
|
|
896 |
if tpu_should_fix_optimizer or (self.mixed_precision == "fp8" and self.fp8_recipe_handler.backend == "TE"):
|
897 |
# 2. grabbing new model parameters
|
898 |
new_named_params = self._get_named_parameters(*result)
|
|
|
902 |
for obj in result:
|
903 |
if isinstance(obj, torch.optim.Optimizer):
|
904 |
obj._switch_parameters(mapping)
|
|
|
905 |
for item in result:
|
906 |
if any(
|
907 |
item in container
|
908 |
for container in (self._dataloaders, self._models, self._optimizers, self._schedulers)
|
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
|
915 |
[`Accelerator.prepare`] instead.
|
|
|
916 |
Args:
|
917 |
model (`torch.nn.Module`):
|
918 |
A PyTorch model to prepare. You don't need to prepare a model if it is used only for inference without
|
|
|
922 |
evaluation_mode (`bool`, *optional*, defaults to `False`):
|
923 |
Whether or not to set the model for evaluation only, by just applying mixed precision and
|
924 |
`torch.compile` (if configured in the `Accelerator` object).
|
|
|
925 |
Example:
|
|
|
926 |
```python
|
927 |
>>> from accelerate import Accelerator
|
|
|
928 |
>>> accelerator = Accelerator()
|
929 |
>>> # Assume a model is defined
|
930 |
>>> model = accelerator.prepare_model(model)
|
|
|
933 |
if device_placement is None:
|
934 |
device_placement = self.device_placement and self.distributed_type != DistributedType.FSDP
|
935 |
self._models.append(model)
|
|
|
936 |
# TODO: Look at enabling native TP training directly with a proper config
|
937 |
if (
|
938 |
self.verify_device_map(model)
|
|
|
943 |
"You can't train a model that has been loaded with `device_map='auto'` in any distributed mode."
|
944 |
" Please rerun your script specifying `--num_processes=1` or by launching with `python {{myscript.py}}`."
|
945 |
)
|
|
|
946 |
if self.native_amp:
|
947 |
model._original_forward = model.forward
|
948 |
model_forward_func = model.forward.__func__ if hasattr(model.forward, "__func__") else model.forward
|
|
|
959 |
convert_model(model)
|
960 |
model._converted_to_transformer_engine = True
|
961 |
model._original_forward = model.forward
|
|
|
962 |
kwargs = self.fp8_recipe_handler.to_kwargs() if self.fp8_recipe_handler is not None else {}
|
963 |
if "fp8_format" in kwargs:
|
964 |
kwargs["fp8_format"] = getattr(te_recipe.Format, kwargs["fp8_format"])
|
965 |
fp8_recipe = te_recipe.DelayedScaling(**kwargs)
|
966 |
model.forward = fp8_autocast(enabled=True, fp8_recipe=fp8_recipe)(model.forward)
|
|
|
967 |
if (getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_loaded_in_4bit", False)) and getattr(
|
968 |
model, "hf_device_map", False
|
969 |
):
|
|
|
976 |
)
|
977 |
current_device = list(model_devices)[0]
|
978 |
current_device_index = current_device.index if isinstance(current_device, torch.device) else current_device
|
|
|
979 |
if torch.device(current_device_index) != self.device:
|
980 |
# if on the first device (GPU 0) we don't care
|
981 |
if (self.device.index is not None) or (current_device_index != 0):
|
|
|
983 |
"You can't train a model that has been loaded in 8-bit precision on a different device than the one "
|
984 |
"you're training on. Make sure you loaded the model on the correct device using for example `device_map={'':torch.cuda.current_device() or device_map={'':torch.xpu.current_device()}"
|
985 |
)
|
|
|
986 |
if "cpu" in model_devices or "disk" in model_devices:
|
987 |
raise ValueError(
|
988 |
"You can't train a model that has been loaded in 8-bit precision with CPU or disk offload."
|
|
|
1002 |
device_ids, output_device = [self.local_process_index], self.local_process_index
|
1003 |
else:
|
1004 |
device_ids, output_device = None, None
|
|
|
1005 |
model = torch.nn.parallel.DistributedDataParallel(
|
1006 |
model, device_ids=device_ids, output_device=output_device, **kwargs
|
1007 |
)
|
1008 |
elif self.distributed_type == DistributedType.FSDP:
|
1009 |
from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel as FSDP
|
|
|
1010 |
# Check if the model is already a FSDP model due to `Manual Wrapping` and if so,
|
1011 |
# don't wrap it again
|
1012 |
# In case the model is already compiled using PyTorch 2.0 and the wrapped model in it
|
|
|
1014 |
is_type_fsdp = isinstance(model, FSDP) or (
|
1015 |
is_compiled_module(model) and isinstance(model._orig_mod, FSDP)
|
1016 |
)
|
|
|
1017 |
if not is_type_fsdp:
|
1018 |
self.state.fsdp_plugin.set_auto_wrap_policy(model)
|
1019 |
fsdp_plugin = self.state.fsdp_plugin
|
|
|
1038 |
apply_activation_checkpointing,
|
1039 |
checkpoint_wrapper,
|
1040 |
)
|
|
|
1041 |
apply_activation_checkpointing(
|
1042 |
model,
|
1043 |
checkpoint_wrapper_fn=functools.partial(
|
|
|
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
|
|
|
1067 |
is_dataloader_present = any(isinstance(obj, torch.utils.data.DataLoader) for obj in args)
|
1068 |
result = [
|
1069 |
self._prepare_one(obj, first_pass=True) if isinstance(obj, torch.utils.data.DataLoader) else obj
|
1070 |
for obj in args
|
1071 |
]
|
|
|
1072 |
if deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] == "auto":
|
1073 |
if is_dataloader_present:
|
1074 |
batch_sizes = [obj.batch_size for obj in args if hasattr(obj, "batch_size")]
|
|
|
1080 |
)
|
1081 |
if self.split_batches:
|
1082 |
batch_sizes = [batch_size // self.num_processes for batch_size in batch_sizes]
|
|
|
1083 |
batch_size_per_device = min(batch_sizes) if deepspeed_plugin.is_train_batch_min else max(batch_sizes)
|
1084 |
if len(batch_sizes) > 1:
|
1085 |
logger.info(
|
|
|
1095 |
)
|
1096 |
else:
|
1097 |
batch_size_per_device = deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"]
|
|
|
1098 |
# handle `gradient_accumulation_steps` when the value is `auto`
|
1099 |
deepspeed_plugin.fill_match(
|
1100 |
"gradient_accumulation_steps",
|
1101 |
must_match=False,
|
1102 |
gradient_accumulation_steps=self.gradient_accumulation_steps,
|
1103 |
)
|
|
|
1104 |
config_kwargs = {
|
1105 |
"train_micro_batch_size_per_gpu": batch_size_per_device,
|
1106 |
"train_batch_size": batch_size_per_device
|
|
|
1109 |
"gradient_clipping": 1.0,
|
1110 |
"zero_optimization.stage3_gather_16bit_weights_on_model_save": False,
|
1111 |
}
|
|
|
1112 |
model = None
|
1113 |
optimizer = None
|
1114 |
scheduler = None
|
|
|
1121 |
type(obj).__name__ in deepspeed.runtime.lr_schedules.VALID_LR_SCHEDULES
|
1122 |
):
|
1123 |
scheduler = obj
|
|
|
1124 |
if optimizer is not None:
|
1125 |
if "optimizer" in deepspeed_plugin.deepspeed_config and not isinstance(optimizer, (DummyOptim)):
|
1126 |
raise ValueError(
|
|
|
1132 |
raise ValueError(
|
1133 |
"You cannot create a `DummyOptim` without specifying an optimizer in the config file."
|
1134 |
)
|
|
|
1135 |
if isinstance(optimizer, (torch.optim.Optimizer)):
|
1136 |
deepspeed_plugin.deepspeed_config["zero_allow_untested_optimizer"] = True
|
|
|
1137 |
if scheduler is not None:
|
1138 |
if "scheduler" in deepspeed_plugin.deepspeed_config and not isinstance(scheduler, (DummyScheduler)):
|
1139 |
raise ValueError(
|
|
|
1150 |
"Either specify a scheduler in the config file or "
|
1151 |
"pass in the `lr_scheduler_callable` parameter when using `accelerate.utils.DummyScheduler`."
|
1152 |
)
|
|
|
1153 |
if optimizer is not None and scheduler is not None:
|
1154 |
if isinstance(optimizer, (DummyOptim)) and not isinstance(scheduler, (DummyScheduler)):
|
1155 |
raise ValueError(
|
1156 |
"You can only specify `accelerate.utils.DummyScheduler` in the code when using "
|
1157 |
"`accelerate.utils.DummyOptim`."
|
1158 |
)
|
|
|
1159 |
if model is not None:
|
1160 |
if hasattr(model, "config"):
|
1161 |
hidden_size = (
|
|
|
1171 |
"zero_optimization.stage3_param_persistence_threshold": 10 * hidden_size,
|
1172 |
}
|
1173 |
)
|
|
|
1174 |
if isinstance(optimizer, (DummyOptim)):
|
1175 |
config_kwargs.update(
|
1176 |
{"optimizer.params.lr": optimizer.lr, "optimizer.params.weight_decay": optimizer.weight_decay}
|
|
|
1207 |
"device", "none"
|
1208 |
) != "none" and self.deepspeed_config.get("zero_force_ds_cpu_optimizer", True):
|
1209 |
from deepspeed.ops.adam import DeepSpeedCPUAdam
|
|
|
1210 |
defaults = {k: v for k, v in optimizer.defaults.items() if k in ["lr", "weight_decay"]}
|
1211 |
optimizer = DeepSpeedCPUAdam(optimizer.param_groups, **defaults)
|
1212 |
kwargs["optimizer"] = optimizer
|
|
|
1216 |
or type(scheduler).__name__ in deepspeed.runtime.lr_schedules.VALID_LR_SCHEDULES
|
1217 |
):
|
1218 |
kwargs["lr_scheduler"] = scheduler
|
|
|
1219 |
engine, optimizer, _, lr_scheduler = deepspeed.initialize(**kwargs)
|
1220 |
if optimizer is not None:
|
1221 |
optimizer = DeepSpeedOptimizerWrapper(optimizer)
|
|
|
1229 |
)
|
1230 |
else:
|
1231 |
scheduler = DeepSpeedSchedulerWrapper(lr_scheduler, optimizer)
|
|
|
1232 |
for i in range(len(result)):
|
1233 |
if isinstance(result[i], torch.nn.Module):
|
1234 |
result[i] = engine
|
|
|
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:
|
|
|
1258 |
raise ValueError(
|
1259 |
"You must specify a training or evaluation dataloader in `accelerate.prepare()` when using Megatron-LM."
|
1260 |
)
|
|
|
1261 |
micro_batch_size = min(batch_sizes) if megatron_lm_plugin.is_train_batch_min else max(batch_sizes)
|
1262 |
if len(batch_sizes) > 1:
|
1263 |
logger.info(
|
|
|
1269 |
if isinstance(obj, MegatronLMDummyDataLoader):
|
1270 |
micro_batch_size = obj.dataset_args["micro_batch_size"]
|
1271 |
break
|
|
|
1272 |
dp_degree = self.num_processes // (megatron_lm_plugin.tp_degree * megatron_lm_plugin.pp_degree)
|
1273 |
megatron_lm_plugin.set_training_args(micro_batch_size, dp_degree)
|
|
|
1274 |
model = None
|
1275 |
optimizer = None
|
1276 |
scheduler = None
|
|
|
1285 |
optimizer = obj
|
1286 |
elif isinstance(obj, (LRScheduler, MegatronLMDummyScheduler)):
|
1287 |
scheduler = obj
|
|
|
1288 |
if model is not None:
|
1289 |
megatron_lm_plugin.set_network_size_args(model, batch_data)
|
1290 |
if optimizer is not None:
|
|
|
1296 |
"You can't use a custom scheduler with Megatron-LM. Please use the `accelerate.utils.MegatronLMDummyScheduler` instead."
|
1297 |
)
|
1298 |
megatron_lm_plugin.set_scheduler_args(scheduler)
|
|
|
1299 |
# initialize megatron-lm
|
1300 |
megatron_lm_initialize(self, args_defaults=megatron_lm_plugin.megatron_lm_default_args)
|
1301 |
counter = 0
|
|
|
1312 |
counter += 1
|
1313 |
else:
|
1314 |
result.append(obj)
|
|
|
1315 |
if model is not None:
|
1316 |
model = megatron_lm_prepare_model(self)
|
1317 |
if optimizer is not None:
|
1318 |
optimizer = megatron_lm_prepare_optimizer(self, model)
|
1319 |
if scheduler is not None:
|
1320 |
scheduler = megatron_lm_prepare_scheduler(self, optimizer, scheduler)
|
|
|
1321 |
if model is not None:
|
1322 |
model = MegatronEngine(self, model, optimizer, scheduler)
|
1323 |
if optimizer is not None:
|
1324 |
optimizer = MegatronLMOptimizerWrapper(optimizer)
|
1325 |
if scheduler is not None:
|
1326 |
scheduler = MegatronLMSchedulerWrapper(scheduler, optimizer)
|
|
|
1327 |
for i in range(len(result)):
|
1328 |
if isinstance(result[i], torch.nn.Module):
|
1329 |
result[i] = model
|
|
|
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(
|
|
|
1350 |
)
|
1351 |
else:
|
1352 |
import intel_extension_for_pytorch as ipex
|
|
|
1353 |
model = None
|
1354 |
optimizer = None
|
1355 |
result = [obj for obj in args]
|
|
|
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(
|
|
|
1381 |
)
|
1382 |
else:
|
1383 |
import msamp
|
|
|
1384 |
model, optimizer = None, None
|
1385 |
num_models, num_optimizers = 0, 0
|
1386 |
result = [obj for obj in args]
|
|
|
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 |
):
|
1413 |
"""
|
1414 |
Prepares a PyTorch DataLoader for training in any distributed setup. It is recommended to use
|
1415 |
[`Accelerator.prepare`] instead.
|
|
|
1416 |
Args:
|
1417 |
data_loader (`torch.utils.data.DataLoader`):
|
1418 |
A vanilla PyTorch DataLoader to prepare
|
|
|
1423 |
If passed, this function will be used to slice tensors across `num_processes`. Will default to
|
1424 |
[`~utils.slice_tensors`]. This argument is used only when `dispatch_batches` is set to `True` and will
|
1425 |
be ignored otherwise.
|
|
|
1426 |
Example:
|
|
|
1427 |
```python
|
1428 |
>>> import torch
|
1429 |
>>> from accelerate import Accelerator
|
|
|
1430 |
>>> accelerator = Accelerator()
|
1431 |
>>> data_loader = torch.utils.data.DataLoader(...)
|
1432 |
>>> data_loader = accelerator.prepare_data_loader(data_loader, device_placement=True)
|
|
|
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
|
1459 |
[`Accelerator.prepare`] instead.
|
|
|
1460 |
Args:
|
1461 |
optimizer (`torch.optim.Optimizer`):
|
1462 |
A vanilla PyTorch optimizer to prepare
|
1463 |
device_placement (`bool`, *optional*):
|
1464 |
Whether or not to place the optimizer on the proper device. Will default to `self.device_placement`.
|
|
|
1465 |
Example:
|
|
|
1466 |
```python
|
1467 |
>>> import torch
|
1468 |
>>> from accelerate import Accelerator
|
|
|
1469 |
>>> accelerator = Accelerator()
|
1470 |
>>> optimizer = torch.optim.Adam(...)
|
1471 |
>>> optimizer = accelerator.prepare_optimizer(optimizer, device_placement=True)
|
|
|
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
|
1487 |
[`Accelerator.prepare`] instead.
|
|
|
1488 |
Args:
|
1489 |
scheduler (`torch.optim.lr_scheduler.LRScheduler`):
|
1490 |
A vanilla PyTorch scheduler to prepare
|
|
|
1491 |
Example:
|
|
|
1492 |
```python
|
1493 |
>>> import torch
|
1494 |
>>> from accelerate import Accelerator
|
|
|
1495 |
>>> accelerator = Accelerator()
|
1496 |
>>> optimizer = torch.optim.Adam(...)
|
1497 |
>>> scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, ...)
|
|
|
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
|
1523 |
on the configuration.
|
|
|
1524 |
Should be used in lieu of `loss.backward()`.
|
|
|
1525 |
Example:
|
|
|
1526 |
```python
|
1527 |
>>> from accelerate import Accelerator
|
|
|
1528 |
>>> accelerator = Accelerator(gradient_accumulation_steps=2)
|
1529 |
>>> outputs = model(inputs)
|
1530 |
>>> loss = loss_fn(outputs, labels)
|
|
|
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
|
1548 |
will check across all processes.
|
|
|
1549 |
Note:
|
1550 |
Does not require `wait_for_everyone()`
|
|
|
1551 |
Example:
|
|
|
1552 |
```python
|
1553 |
>>> from accelerate import Accelerator
|
|
|
1554 |
>>> accelerator = Accelerator()
|
1555 |
>>> # Assume later in the training script
|
1556 |
>>> # `should_do_breakpoint` is a custom function to monitor when to break,
|
|
|
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
|
1569 |
reset the trigger tensor to 0.
|
|
|
1570 |
Note:
|
1571 |
Does not require `wait_for_everyone()`
|
|
|
1572 |
Example:
|
|
|
1573 |
```python
|
1574 |
>>> from accelerate import Accelerator
|
|
|
1575 |
>>> accelerator = Accelerator()
|
1576 |
>>> # Assume later in the training script
|
1577 |
>>> # `should_do_breakpoint` is a custom function to monitor when to break,
|
|
|
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.
|
|
|
1597 |
Likely should be called through [`Accelerator.clip_grad_norm_`] or [`Accelerator.clip_grad_value_`]
|
|
|
1598 |
Args:
|
1599 |
optimizer (`torch.optim.Optimizer` or `list[torch.optim.Optimizer]`, *optional*):
|
1600 |
The optimizer(s) for which to unscale gradients. If not set, will unscale gradients on all optimizers
|
1601 |
that were passed to [`~Accelerator.prepare`].
|
|
|
1602 |
Example:
|
|
|
1603 |
```python
|
1604 |
>>> from accelerate import Accelerator
|
|
|
1605 |
>>> accelerator = Accelerator()
|
1606 |
>>> model, optimizer = accelerator.prepare(model, optimizer)
|
1607 |
>>> outputs = model(inputs)
|
|
|
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_`.
|
|
|
1630 |
Returns:
|
1631 |
`torch.Tensor`: Total norm of the parameter gradients (viewed as a single vector).
|
|
|
1632 |
Example:
|
|
|
1633 |
```python
|
1634 |
>>> from accelerate import Accelerator
|
|
|
1635 |
>>> accelerator = Accelerator(gradient_accumulation_steps=2)
|
1636 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
|
|
1637 |
>>> for input, target in dataloader:
|
1638 |
... optimizer.zero_grad()
|
1639 |
... output = model(input)
|
|
|
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_`.
|
|
|
1662 |
Example:
|
|
|
1663 |
```python
|
1664 |
>>> from accelerate import Accelerator
|
|
|
1665 |
>>> accelerator = Accelerator(gradient_accumulation_steps=2)
|
1666 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
|
|
1667 |
>>> for input, target in dataloader:
|
1668 |
... optimizer.zero_grad()
|
1669 |
... output = model(input)
|
|
|
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
|
1684 |
regroup the predictions from all processes when doing evaluation.
|
|
|
1685 |
Note:
|
1686 |
This gather happens in all processes.
|
|
|
1687 |
Args:
|
1688 |
tensor (`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`):
|
1689 |
The tensors to gather across all processes.
|
|
|
1690 |
Returns:
|
1691 |
`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`: The gathered tensor(s). Note that the
|
1692 |
first dimension of the result is *num_processes* multiplied by the first dimension of the input tensors.
|
|
|
1693 |
Example:
|
|
|
1694 |
```python
|
1695 |
>>> # Assuming four processes
|
1696 |
>>> import torch
|
1697 |
>>> from accelerate import Accelerator
|
|
|
1698 |
>>> accelerator = Accelerator()
|
1699 |
>>> process_tensor = torch.tensor([accelerator.process_index])
|
1700 |
>>> gathered_tensor = accelerator.gather(process_tensor)
|
|
|
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
|
1709 |
used for gathering the inputs and targets for metric calculation.
|
|
|
1710 |
Args:
|
1711 |
input (`torch.Tensor`, `object`, a nested tuple/list/dictionary of `torch.Tensor`, or a nested tuple/list/dictionary of `object`):
|
1712 |
The tensors or objects for calculating metrics across all processes
|
|
|
1713 |
Example:
|
|
|
1714 |
```python
|
1715 |
>>> # Assuming two processes, with a batch size of 5 on a dataset with 9 samples
|
1716 |
>>> import torch
|
1717 |
>>> from accelerate import Accelerator
|
|
|
1718 |
>>> accelerator = Accelerator()
|
1719 |
>>> dataloader = torch.utils.data.DataLoader(range(9), batch_size=5)
|
1720 |
>>> dataloader = accelerator.prepare(dataloader)
|
|
|
1729 |
all_tensors = True
|
1730 |
except TypeError:
|
1731 |
all_tensors = False
|
|
|
1732 |
if not all_tensors:
|
1733 |
data = gather_object(input_data)
|
1734 |
else:
|
1735 |
data = self.gather(input_data)
|
|
|
1736 |
try:
|
1737 |
if self.gradient_state.end_of_dataloader:
|
1738 |
# at the end of a dataloader, `gather_for_metrics` regresses to
|
|
|
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
|
1751 |
# no remainder even though at end of dataloader, so nothing to do.
|
|
|
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*.
|
|
|
1762 |
Note:
|
1763 |
All processes get the reduced value.
|
|
|
1764 |
Args:
|
1765 |
tensor (`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`):
|
1766 |
The tensors to reduce across all processes.
|
|
|
1768 |
A reduction type, can be one of 'sum', 'mean', or 'none'. If 'none', will not perform any operation.
|
1769 |
scale (`float`, *optional*, defaults to 1.0):
|
1770 |
A default scaling value to be applied after the reduce, only valied on XLA.
|
|
|
1771 |
Returns:
|
1772 |
`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`:
|
1773 |
The reduced tensor(s).
|
|
|
1774 |
Example:
|
|
|
1775 |
```python
|
1776 |
>>> # Assuming two processes
|
1777 |
>>> import torch
|
1778 |
>>> from accelerate import Accelerator
|
|
|
1779 |
>>> accelerator = Accelerator()
|
1780 |
>>> process_tensor = torch.arange(accelerator.num_processes) + 1 + (2 * accelerator.process_index)
|
1781 |
>>> process_tensor = process_tensor.to(accelerator.device)
|
|
|
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
|
1791 |
they can safely be gathered.
|
|
|
1792 |
Args:
|
1793 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
1794 |
The data to gather.
|
|
|
1798 |
The value with which to pad.
|
1799 |
pad_first (`bool`, *optional*, defaults to `False`):
|
1800 |
Whether to pad at the beginning or the end.
|
|
|
1801 |
Returns:
|
1802 |
`torch.Tensor`, or a nested tuple/list/dictionary of `torch.Tensor`:
|
1803 |
The padded tensor(s).
|
|
|
1804 |
Example:
|
|
|
1805 |
```python
|
1806 |
>>> # Assuming two processes, with the first processes having a tensor of size 1 and the second of size 2
|
1807 |
>>> import torch
|
1808 |
>>> from accelerate import Accelerator
|
|
|
1809 |
>>> accelerator = Accelerator()
|
1810 |
>>> process_tensor = torch.arange(accelerator.process_index + 1).to(accelerator.device)
|
1811 |
>>> padded_tensor = accelerator.pad_across_processes(process_tensor)
|
|
|
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
|
1820 |
the model.
|
|
|
1821 |
Args:
|
1822 |
model (`torch.nn.Module`):
|
1823 |
The model to unwrap.
|
1824 |
keep_fp32_wrapper (`bool`, *optional*, defaults to `True`):
|
1825 |
Whether to not remove the mixed precision hook if it was added.
|
|
|
1826 |
Returns:
|
1827 |
`torch.nn.Module`: The unwrapped model.
|
|
|
1828 |
Example:
|
|
|
1829 |
```python
|
1830 |
>>> # Assuming two GPU processes
|
1831 |
>>> from torch.nn.parallel import DistributedDataParallel
|
1832 |
>>> from accelerate import Accelerator
|
|
|
1833 |
>>> accelerator = Accelerator()
|
1834 |
>>> model = accelerator.prepare(MyModel())
|
1835 |
>>> print(model.__class__.__name__)
|
1836 |
DistributedDataParallel
|
|
|
1837 |
>>> model = accelerator.unwrap_model(model)
|
1838 |
>>> print(model.__class__.__name__)
|
1839 |
MyModel
|
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
|
1846 |
nothing when the script is only run in one process). Useful to do before saving a model.
|
|
|
1847 |
Example:
|
|
|
1848 |
```python
|
1849 |
>>> # Assuming two GPU processes
|
1850 |
>>> import time
|
1851 |
>>> from accelerate import Accelerator
|
|
|
1852 |
>>> accelerator = Accelerator()
|
1853 |
>>> if accelerator.is_main_process:
|
1854 |
... time.sleep(2)
|
|
|
1860 |
```
|
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
|
|
|
1867 |
Args:
|
1868 |
project_name (`str`):
|
1869 |
The name of the project. All trackers will save their data based on this
|
|
|
1875 |
```python
|
1876 |
{"wandb": {"tags": ["tag_a", "tag_b"]}}
|
1877 |
```
|
|
|
1878 |
Example:
|
|
|
1879 |
```python
|
1880 |
>>> from accelerate import Accelerator
|
|
|
1881 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
1882 |
>>> accelerator.init_trackers(
|
1883 |
... project_name="my_project",
|
|
|
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.
|
|
|
1908 |
Args:
|
1909 |
name (`str`):
|
1910 |
The name of a tracker, corresponding to the `.name` property.
|
1911 |
unwrap (`bool`):
|
1912 |
Whether to return the internal tracking mechanism or to return the wrapped tracker instead
|
1913 |
(recommended).
|
|
|
1914 |
Returns:
|
1915 |
`GeneralTracker`: The tracker corresponding to `name` if it exists.
|
|
|
1916 |
Example:
|
|
|
1917 |
```python
|
1918 |
>>> from accelerate import Accelerator
|
|
|
1919 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
1920 |
>>> accelerator.init_trackers("my_project")
|
1921 |
>>> tensorboard_tracker = accelerator.get_tracker("tensorboard")
|
|
|
1928 |
raise ValueError(f"{name} is not an available tracker stored inside the `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.
|
|
|
1935 |
Args:
|
1936 |
values (`dict`):
|
1937 |
Values should be a dictionary-like object containing only types `int`, `float`, or `str`.
|
|
|
1943 |
```python
|
1944 |
{"wandb": {"tags": ["tag_a", "tag_b"]}}
|
1945 |
```
|
|
|
1946 |
Example:
|
|
|
1947 |
```python
|
1948 |
>>> from accelerate import Accelerator
|
|
|
1949 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
1950 |
>>> accelerator.init_trackers("my_project")
|
1951 |
>>> accelerator.log({"loss": 0.5, "accuracy": 0.9})
|
|
|
1953 |
"""
|
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
|
1960 |
called at the end of your script if using experiment tracking.
|
|
|
1961 |
Example:
|
|
|
1962 |
```python
|
1963 |
>>> from accelerate import Accelerator
|
|
|
1964 |
>>> accelerator = Accelerator(log_with="tensorboard")
|
1965 |
>>> accelerator.init_trackers("my_project")
|
1966 |
>>> # Do training
|
|
|
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`.
|
|
|
1975 |
Args:
|
1976 |
obj (`object`): The object to save.
|
1977 |
f (`str` or `os.PathLike`): Where to save the content of `obj`.
|
1978 |
safe_serialization (`bool`, *optional*, defaults to `False`): Whether to save `obj` using `safetensors`
|
|
|
1979 |
Note:
|
1980 |
If `save_on_each_node` was passed in as a `ProjectConfiguration`, will save the object once per node,
|
1981 |
rather than only once on the main node.
|
|
|
1982 |
Example:
|
|
|
1983 |
```python
|
1984 |
>>> from accelerate import Accelerator
|
|
|
1985 |
>>> accelerator = Accelerator()
|
1986 |
>>> arr = [0, 1, 2, 3]
|
1987 |
>>> accelerator.save(arr, "array.pkl")
|
|
|
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,
|
|
|
2002 |
):
|
2003 |
"""
|
2004 |
Save a model so that it can be re-loaded using load_checkpoint_in_model
|
|
|
2005 |
Arguments:
|
2006 |
model: (`torch.nn.Module`):
|
2007 |
Model to be saved. The model can be wrapped or unwraped.
|
|
|
2010 |
max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
|
2011 |
The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
|
2012 |
lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5MB"`).
|
|
|
2013 |
<Tip warning={true}>
|
|
|
2014 |
If a single weight of the model is bigger than `max_shard_size`, it will be in its own checkpoint shard
|
2015 |
which will be bigger than `max_shard_size`.
|
|
|
2016 |
</Tip>
|
|
|
2017 |
safe_serialization (`bool`, *optional*, defaults to `True`):
|
2018 |
Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
|
|
|
2019 |
Example:
|
|
|
2020 |
```python
|
2021 |
>>> from accelerate import Accelerator
|
|
|
2022 |
>>> accelerator = Accelerator()
|
2023 |
>>> model = ...
|
2024 |
>>> accelerator.save_model(model, save_directory)
|
|
|
2027 |
if os.path.isfile(save_directory):
|
2028 |
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
|
2029 |
return
|
|
|
2030 |
os.makedirs(save_directory, exist_ok=True)
|
|
|
2031 |
# get the state_dict of the model
|
2032 |
if any(
|
2033 |
[
|
|
|
2041 |
if any(param.device == torch.device("meta") for param in model.parameters()):
|
2042 |
raise RuntimeError("You can't save the model since some parameters are on the meta device.")
|
2043 |
state_dict = self.get_state_dict(model)
|
|
|
2044 |
if safe_serialization:
|
2045 |
state_dict = clean_state_dict_for_safetensors(state_dict)
|
2046 |
weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
|
|
|
2047 |
# Shard the model if it is too big.
|
2048 |
shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size, weights_name=weights_name)
|
|
|
2049 |
# Clean the folder from a previous save
|
2050 |
for filename in os.listdir(save_directory):
|
2051 |
full_filename = os.path.join(save_directory, filename)
|
2052 |
# If we have a shard file that is not going to be replaced, we delete it, but only from the main process
|
2053 |
# in distributed settings to avoid race conditions.
|
2054 |
weights_no_suffix = weights_name.replace(".bin", "")
|
|
|
2055 |
# make sure that file to be deleted matches format of sharded file, e.g. pytorch_model-00001-of-00005
|
2056 |
filename_no_suffix = filename.replace(".bin", "")
|
2057 |
reg = re.compile(r"(.*?)-\d{5}-of-\d{5}")
|
|
|
2058 |
if (
|
2059 |
filename.startswith(weights_no_suffix)
|
2060 |
and os.path.isfile(full_filename)
|
|
|
2063 |
and PartialState().is_main_process
|
2064 |
):
|
2065 |
os.remove(full_filename)
|
|
|
2066 |
# Save the model
|
2067 |
for shard_file, shard in shards.items():
|
2068 |
self.save(shard, os.path.join(save_directory, shard_file), safe_serialization=safe_serialization)
|
|
|
2069 |
if index is None:
|
2070 |
path_to_weights = os.path.join(save_directory, WEIGHTS_NAME)
|
2071 |
logger.info(f"Model weights saved in {path_to_weights}")
|
|
|
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`].
|
|
|
2087 |
Args:
|
2088 |
hook (`Callable`):
|
2089 |
A function to be called in [`Accelerator.save_state`] before `save_checkpoint`.
|
|
|
2090 |
The hook should have the following signature:
|
|
|
2091 |
`hook(models: list[torch.nn.Module], weights: list[dict[str, torch.Tensor]], input_dir: str) -> None`
|
|
|
2092 |
The `models` argument are the models as saved in the accelerator state under `accelerator._models`, `weigths`
|
2093 |
argument are the state dicts of the `models`, and the `input_dir` argument is the `input_dir` argument passed
|
2094 |
to [`Accelerator.load_state`].
|
|
|
2095 |
<Tip>
|
|
|
2096 |
Should only be used in conjunction with [`Accelerator.register_load_state_pre_hook`]. Can be useful to save
|
2097 |
configurations in addition to model weights. Can also be used to overwrite model saving with a customized
|
2098 |
method. In this case, make sure to remove already loaded weights from the weights list.
|
|
|
2099 |
</Tip>
|
|
|
2100 |
Returns:
|
2101 |
`torch.utils.hooks.RemovableHandle`: a handle that can be used to remove the added hook by calling
|
2102 |
`handle.remove()`
|
|
|
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.
|
|
|
2110 |
If a `ProjectConfiguration` was passed to the `Accelerator` object with `automatic_checkpoint_naming` enabled
|
2111 |
then checkpoints will be saved to `self.project_dir/checkpoints`. If the number of current saves is greater
|
2112 |
than `total_limit` then the oldest save is deleted. Each checkpoint is saved in seperate folders named
|
2113 |
`checkpoint_<iteration>`.
|
|
|
2114 |
Otherwise they are just saved to `output_dir`.
|
|
|
2115 |
<Tip>
|
|
|
2116 |
Should only be used when wanting to save a checkpoint during training and restoring the state in the same
|
2117 |
environment.
|
|
|
2118 |
</Tip>
|
|
|
2119 |
Args:
|
2120 |
output_dir (`str` or `os.PathLike`):
|
2121 |
The name of the folder to save all relevant weights and states.
|
|
|
2124 |
save_model_func_kwargs (`dict`, *optional*):
|
2125 |
Additional keyword arguments for saving model which can be passed to the underlying save function, such
|
2126 |
as optional arguments for DeepSpeed's `save_checkpoint` function.
|
|
|
2127 |
Example:
|
|
|
2128 |
```python
|
2129 |
>>> from accelerate import Accelerator
|
|
|
2130 |
>>> accelerator = Accelerator()
|
2131 |
>>> model, optimizer, lr_scheduler = ...
|
2132 |
>>> model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
|
|
|
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(
|
2150 |
f"Deleting {len(folders) + 1 - self.project_configuration.total_limit} checkpoints to make room for new checkpoint."
|
|
|
2159 |
self.wait_for_everyone()
|
2160 |
os.makedirs(output_dir, exist_ok=True)
|
2161 |
logger.info(f"Saving current state to {output_dir}")
|
|
|
2162 |
if self.distributed_type == DistributedType.TPU:
|
2163 |
# Finish running the previous step before checkpointing
|
2164 |
xm.mark_step()
|
|
|
2165 |
# Save the models taking care of FSDP and DeepSpeed nuances
|
2166 |
weights = []
|
2167 |
for i, model in enumerate(self._models):
|
|
|
2180 |
logger.info(f"Megatron-LM Model , Optimizer and Scheduler saved to output dir {output_dir}")
|
2181 |
else:
|
2182 |
weights.append(self.get_state_dict(model, unwrap=False))
|
|
|
2183 |
# Save the optimizers taking care of FSDP and DeepSpeed nuances
|
2184 |
optimizers = []
|
2185 |
if self.distributed_type == DistributedType.FSDP:
|
|
|
2189 |
logger.info(f"FSDP Optimizer saved to output dir {output_dir}")
|
2190 |
elif self.distributed_type not in [DistributedType.DEEPSPEED, DistributedType.MEGATRON_LM]:
|
2191 |
optimizers = self._optimizers
|
|
|
2192 |
# Save the lr schedulers taking care of DeepSpeed nuances
|
2193 |
schedulers = []
|
2194 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
|
|
2198 |
schedulers.append(scheduler)
|
2199 |
elif self.distributed_type not in [DistributedType.MEGATRON_LM]:
|
2200 |
schedulers = self._schedulers
|
|
|
2201 |
# Save the samplers of the dataloaders
|
2202 |
dataloaders = self._dataloaders
|
|
|
2203 |
# Call model loading hooks that might have been registered with
|
2204 |
# accelerator.register_model_state_hook
|
2205 |
for hook in self._save_model_state_pre_hook.values():
|
2206 |
hook(self._models, weights, output_dir)
|
|
|
2207 |
save_location = save_accelerator_state(
|
2208 |
output_dir,
|
2209 |
weights,
|
|
|
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`].
|
|
|
2225 |
Args:
|
2226 |
hook (`Callable`):
|
2227 |
A function to be called in [`Accelerator.load_state`] before `load_checkpoint`.
|
|
|
2228 |
The hook should have the following signature:
|
|
|
2229 |
`hook(models: list[torch.nn.Module], input_dir: str) -> None`
|
|
|
2230 |
The `models` argument are the models as saved in the accelerator state under `accelerator._models`, and the
|
2231 |
`input_dir` argument is the `input_dir` argument passed to [`Accelerator.load_state`].
|
|
|
2232 |
<Tip>
|
|
|
2233 |
Should only be used in conjunction with [`Accelerator.register_save_state_pre_hook`]. Can be useful to load
|
2234 |
configurations in addition to model weights. Can also be used to overwrite model loading with a customized
|
2235 |
method. In this case, make sure to remove already loaded models from the models list.
|
|
|
2236 |
</Tip>
|
|
|
2237 |
Returns:
|
2238 |
`torch.utils.hooks.RemovableHandle`: a handle that can be used to remove the added hook by calling
|
2239 |
`handle.remove()`
|
|
|
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.
|
|
|
2247 |
<Tip>
|
|
|
2248 |
Should only be used in conjunction with [`Accelerator.save_state`]. If a file is not registered for
|
2249 |
checkpointing, it will not be loaded if stored in the directory.
|
|
|
2250 |
</Tip>
|
|
|
2251 |
Args:
|
2252 |
input_dir (`str` or `os.PathLike`):
|
2253 |
The name of the folder all relevant weights and states were saved in. Can be `None` if
|
|
|
2256 |
Additional keyword arguments for loading model which can be passed to the underlying load function,
|
2257 |
such as optional arguments for DeepSpeed's `load_checkpoint` function or a `map_location` to load the
|
2258 |
model and optimizer on.
|
|
|
2259 |
Example:
|
|
|
2260 |
```python
|
2261 |
>>> from accelerate import Accelerator
|
|
|
2262 |
>>> accelerator = Accelerator()
|
2263 |
>>> model, optimizer, lr_scheduler = ...
|
2264 |
>>> model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
|
|
|
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]
|
2281 |
else:
|
2282 |
raise ValueError("No input_dir provided and automatic checkpoint naming is disabled.")
|
2283 |
logger.info(f"Loading states from {input_dir}")
|
|
|
2284 |
# Load the models taking care of FSDP and DeepSpeed nuances
|
2285 |
models = []
|
2286 |
for i, model in enumerate(self._models):
|
|
|
2299 |
logger.info(f"Megatron-LM Model , Optimizer and Scheduler loaded from input dir {input_dir}")
|
2300 |
else:
|
2301 |
models.append(model)
|
|
|
2302 |
# Load the optimizers taking care of FSDP and DeepSpeed nuances
|
2303 |
optimizers = []
|
2304 |
if self.distributed_type == DistributedType.FSDP:
|
|
|
2308 |
logger.info(f"FSDP Optimizer loaded from input dir {input_dir}")
|
2309 |
elif self.distributed_type not in [DistributedType.DEEPSPEED, DistributedType.MEGATRON_LM]:
|
2310 |
optimizers = self._optimizers
|
|
|
2311 |
# Load the lr schedulers taking care of DeepSpeed nuances
|
2312 |
schedulers = []
|
2313 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
|
|
2317 |
schedulers.append(scheduler)
|
2318 |
elif self.distributed_type not in [DistributedType.MEGATRON_LM]:
|
2319 |
schedulers = self._schedulers
|
|
|
2320 |
dataloaders = self._dataloaders
|
|
|
2321 |
# Call model loading hooks that might have been registered with
|
2322 |
# accelerator.register_model_state_hook
|
2323 |
for hook in self._load_model_state_pre_hook.values():
|
2324 |
hook(models, input_dir)
|
|
|
2325 |
map_location = load_model_func_kwargs.pop("map_location", None)
|
2326 |
if map_location is None:
|
2327 |
if self.num_processes > 1 and self.distributed_type in (
|
|
|
2331 |
map_location = "on_device"
|
2332 |
else:
|
2333 |
map_location = "cpu"
|
|
|
2334 |
load_accelerator_state(
|
2335 |
input_dir,
|
2336 |
models,
|
|
|
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
|
2362 |
method between two trainings with different models/optimizers. Also will reset `Accelerator.step` to 0.
|
|
|
2363 |
Example:
|
|
|
2364 |
```python
|
2365 |
>>> from accelerate import Accelerator
|
|
|
2366 |
>>> accelerator = Accelerator()
|
2367 |
>>> model, optimizer, scheduler = ...
|
2368 |
>>> model, optimizer, scheduler = accelerator.prepare(model, optimizer, scheduler)
|
|
|
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
|
2383 |
garbage collector. You should call this method between two trainings with different models/optimizers.
|
|
|
2384 |
Example:
|
|
|
2385 |
```python
|
2386 |
>>> from accelerate import Accelerator
|
|
|
2387 |
>>> accelerator = Accelerator()
|
2388 |
>>> model, optimizer, scheduler = ...
|
2389 |
>>> model, optimizer, scheduler = accelerator.prepare(model, optimizer, scheduler)
|
|
|
2392 |
```
|
2393 |
"""
|
2394 |
self.free_memory()
|
|
|
2395 |
def _get_named_parameters(self, *args):
|
2396 |
named_parameters = {}
|
2397 |
for obj in args:
|
|
|
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 |
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
|
2421 |
precision.
|
|
|
2422 |
Args:
|
2423 |
model (`torch.nn.Module`):
|
2424 |
A PyTorch model sent through [`Accelerator.prepare`]
|
2425 |
unwrap (`bool`, *optional*, defaults to `True`):
|
2426 |
Whether to return the original underlying state_dict of `model` or to return the wrapped state_dict
|
|
|
2427 |
Returns:
|
2428 |
`dict`: The state dictionary of the model potentially without full precision.
|
|
|
2429 |
Example:
|
|
|
2430 |
```python
|
2431 |
>>> import torch
|
2432 |
>>> from accelerate import Accelerator
|
|
|
2433 |
>>> accelerator = Accelerator()
|
2434 |
>>> net = torch.nn.Linear(2, 2)
|
2435 |
>>> net = accelerator.prepare(net)
|
|
|
2449 |
)
|
2450 |
else:
|
2451 |
from deepspeed.checkpoint.utils import clone_tensors_for_torch_save
|
|
|
2452 |
state_dict = clone_tensors_for_torch_save(self.unwrap_model(model).state_dict())
|
2453 |
elif self.distributed_type == DistributedType.FSDP:
|
2454 |
from torch.distributed.fsdp import FullStateDictConfig, StateDictType
|
2455 |
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
|
|
|
2456 |
full_state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
|
2457 |
with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, full_state_dict_config):
|
2458 |
state_dict = model.state_dict()
|
|
|
2460 |
if unwrap:
|
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`.
|
|
|
2467 |
These should be utilized when the state is being loaded or saved in the same script. It is not designed to be
|
2468 |
used in different scripts.
|
|
|
2469 |
<Tip>
|
|
|
2470 |
Every `object` must have a `load_state_dict` and `state_dict` function to be stored.
|
|
|
2471 |
</Tip>
|
|
|
2472 |
Example:
|
|
|
2473 |
```python
|
2474 |
>>> from accelerate import Accelerator
|
|
|
2475 |
>>> accelerator = Accelerator()
|
2476 |
>>> # Assume `CustomObject` has a `state_dict` and `load_state_dict` function.
|
2477 |
>>> obj = CustomObject()
|
|
|
2489 |
err += f"\n\t- Item at index {index}, `{get_pretty_name(obj)}`"
|
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
|
2496 |
different will happen otherwise.
|
|
|
2497 |
A different `autocast_handler` can be passed in to override the one set in the `Accelerator` object. This is
|
2498 |
useful in blocks under `autocast` where you want to revert to fp32.
|
|
|
2499 |
Example:
|
|
|
2500 |
```python
|
2501 |
>>> from accelerate import Accelerator
|
|
|
2502 |
>>> accelerator = Accelerator(mixed_precision="fp16")
|
2503 |
>>> with accelerator.autocast():
|
2504 |
... train()
|
|
|
2520 |
autocast_context.__enter__()
|
2521 |
yield
|
2522 |
autocast_context.__exit__(*sys.exc_info())
|
|
|
2523 |
@property
|
2524 |
def optimizer_step_was_skipped(self):
|
2525 |
"""
|
|
|
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`.
|
|
|
2536 |
Args:
|
2537 |
dataloader (`torch.utils.data.DataLoader`): The data loader in which to skip batches.
|
2538 |
num_batches (`int`, *optional*, defaults to 0): The number of batches to skip
|
|
|
2539 |
Example:
|
|
|
2540 |
```python
|
2541 |
>>> from accelerate import Accelerator
|
|
|
2542 |
>>> accelerator = Accelerator()
|
2543 |
>>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
|
2544 |
>>> skipped_dataloader = accelerator.skip_first_batches(dataloader, num_batches=2)
|
|
|
2549 |
... loss = loss_func(output, target)
|
2550 |
... accelerator.backward(loss)
|
2551 |
... optimizer.step()
|
|
|
2552 |
>>> # subsequent epochs
|
2553 |
>>> for input, target in dataloader:
|
2554 |
... optimizer.zero_grad()
|
|
|
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`.
|
|
|
2567 |
for m in model.modules():
|
2568 |
if hasattr(m, "hf_device_map") and len(m.hf_device_map) > 1:
|
2569 |
return True
|
|
|
2570 |
return False
|
src/big_modeling.py
CHANGED
@@ -1,86 +1,66 @@
|
|
1 |
logger = logging.getLogger(__name__)
|
2 |
-
|
3 |
-
|
4 |
@contextmanager
|
5 |
def init_empty_weights(include_buffers: bool = None):
|
6 |
"""
|
7 |
A context manager under which models are initialized with all parameters on the meta device, therefore creating an
|
8 |
empty model. Useful when just initializing the model would blow the available RAM.
|
9 |
-
|
10 |
Args:
|
11 |
include_buffers (`bool`, *optional*):
|
12 |
Whether or not to also put all buffers on the meta device while initializing.
|
13 |
-
|
14 |
Example:
|
15 |
-
|
16 |
```python
|
17 |
import torch.nn as nn
|
18 |
from accelerate import init_empty_weights
|
19 |
-
|
20 |
# Initialize a model with 100 billions parameters in no time and without using any RAM.
|
21 |
with init_empty_weights():
|
22 |
tst = nn.Sequential(*[nn.Linear(10000, 10000) for _ in range(1000)])
|
23 |
```
|
24 |
-
|
25 |
<Tip warning={true}>
|
26 |
-
|
27 |
Any model created under this context manager has no weights. As such you can't do something like
|
28 |
`model.to(some_device)` with it. To load weights inside your empty model, see [`load_checkpoint_and_dispatch`].
|
29 |
-
|
30 |
</Tip>
|
31 |
"""
|
32 |
if include_buffers is None:
|
33 |
include_buffers = parse_flag_from_env("ACCELERATE_INIT_INCLUDE_BUFFERS", False)
|
34 |
with init_on_device(torch.device("meta"), include_buffers=include_buffers) as f:
|
35 |
yield f
|
36 |
-
|
37 |
-
|
38 |
@contextmanager
|
39 |
def init_on_device(device: torch.device, include_buffers: bool = None):
|
40 |
"""
|
41 |
A context manager under which models are initialized with all parameters on the specified device.
|
42 |
-
|
43 |
Args:
|
44 |
device (`torch.device`):
|
45 |
Device to initialize all parameters on.
|
46 |
include_buffers (`bool`, *optional*):
|
47 |
Whether or not to also put all buffers on the meta device while initializing.
|
48 |
-
|
49 |
Example:
|
50 |
-
|
51 |
```python
|
52 |
import torch.nn as nn
|
53 |
from accelerate import init_on_device
|
54 |
-
|
55 |
with init_on_device(device=torch.device("cuda")):
|
56 |
tst = nn.Liner(100, 100) # on `cuda` device
|
57 |
```
|
58 |
"""
|
59 |
if include_buffers is None:
|
60 |
include_buffers = parse_flag_from_env("ACCELERATE_INIT_INCLUDE_BUFFERS", False)
|
61 |
-
|
62 |
# TODO(shingjan): remove the torch version check once older versions are deprecated
|
63 |
if is_torch_version(">=", "2.0") and include_buffers:
|
64 |
with device:
|
65 |
yield
|
66 |
return
|
67 |
-
|
68 |
old_register_parameter = nn.Module.register_parameter
|
69 |
if include_buffers:
|
70 |
old_register_buffer = nn.Module.register_buffer
|
71 |
-
|
72 |
def register_empty_parameter(module, name, param):
|
73 |
old_register_parameter(module, name, param)
|
74 |
if param is not None:
|
75 |
param_cls = type(module._parameters[name])
|
76 |
kwargs = module._parameters[name].__dict__
|
77 |
module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
|
78 |
-
|
79 |
def register_empty_buffer(module, name, buffer, persistent=True):
|
80 |
old_register_buffer(module, name, buffer, persistent=persistent)
|
81 |
if buffer is not None:
|
82 |
module._buffers[name] = module._buffers[name].to(device)
|
83 |
-
|
84 |
# Patch tensor creation
|
85 |
if include_buffers:
|
86 |
tensor_constructors_to_patch = {
|
@@ -89,14 +69,11 @@ def init_on_device(device: torch.device, include_buffers: bool = None):
|
|
89 |
}
|
90 |
else:
|
91 |
tensor_constructors_to_patch = {}
|
92 |
-
|
93 |
def patch_tensor_constructor(fn):
|
94 |
def wrapper(*args, **kwargs):
|
95 |
kwargs["device"] = device
|
96 |
return fn(*args, **kwargs)
|
97 |
-
|
98 |
return wrapper
|
99 |
-
|
100 |
try:
|
101 |
nn.Module.register_parameter = register_empty_parameter
|
102 |
if include_buffers:
|
@@ -110,8 +87,6 @@ def init_on_device(device: torch.device, include_buffers: bool = None):
|
|
110 |
nn.Module.register_buffer = old_register_buffer
|
111 |
for torch_function_name, old_torch_function in tensor_constructors_to_patch.items():
|
112 |
setattr(torch, torch_function_name, old_torch_function)
|
113 |
-
|
114 |
-
|
115 |
def cpu_offload(
|
116 |
model: nn.Module,
|
117 |
execution_device: Optional[torch.device] = None,
|
@@ -123,7 +98,6 @@ def cpu_offload(
|
|
123 |
Activates full CPU offload for a model. As a result, all parameters of the model will be offloaded and only one
|
124 |
copy of the state dict of the model will be kept. During the forward pass, parameters will be extracted from that
|
125 |
state dict and put on the execution device passed as they are needed, then offloaded again.
|
126 |
-
|
127 |
Args:
|
128 |
model (`torch.nn.Module`):
|
129 |
The model to offload.
|
@@ -144,7 +118,6 @@ def cpu_offload(
|
|
144 |
execution_device = next(iter(model.parameters())).device
|
145 |
if state_dict is None:
|
146 |
state_dict = {n: p.to("cpu") for n, p in model.state_dict().items()}
|
147 |
-
|
148 |
add_hook_to_module(model, AlignDevicesHook(io_same_device=True), append=True)
|
149 |
attach_align_device_hook(
|
150 |
model,
|
@@ -154,10 +127,7 @@ def cpu_offload(
|
|
154 |
weights_map=state_dict,
|
155 |
preload_module_classes=preload_module_classes,
|
156 |
)
|
157 |
-
|
158 |
return model
|
159 |
-
|
160 |
-
|
161 |
def cpu_offload_with_hook(
|
162 |
model: torch.nn.Module,
|
163 |
execution_device: Optional[Union[int, str, torch.device]] = None,
|
@@ -167,7 +137,6 @@ def cpu_offload_with_hook(
|
|
167 |
Offloads a model on the CPU and puts it back to an execution device when executed. The difference with
|
168 |
[`cpu_offload`] is that the model stays on the execution device after the forward and is only offloaded again when
|
169 |
the `offload` method of the returned `hook` is called. Useful for pipelines running a model in a loop.
|
170 |
-
|
171 |
Args:
|
172 |
model (`torch.nn.Module`):
|
173 |
The model to offload.
|
@@ -177,21 +146,17 @@ def cpu_offload_with_hook(
|
|
177 |
prev_module_hook (`UserCpuOffloadHook`, *optional*):
|
178 |
The hook sent back by this function for a previous model in the pipeline you are running. If passed, its
|
179 |
offload method will be called just before the forward of the model to which this hook is attached.
|
180 |
-
|
181 |
Example:
|
182 |
-
|
183 |
```py
|
184 |
model_1, hook_1 = cpu_offload_with_hook(model_1, cuda_device)
|
185 |
model_2, hook_2 = cpu_offload_with_hook(model_2, cuda_device, prev_module_hook=hook_1)
|
186 |
model_3, hook_3 = cpu_offload_with_hook(model_3, cuda_device, prev_module_hook=hook_2)
|
187 |
-
|
188 |
hid_1 = model_1(input)
|
189 |
for i in range(50):
|
190 |
# model1 is offloaded on the CPU at the first iteration, model 2 stays on the GPU for this whole loop.
|
191 |
hid_2 = model_2(hid_1)
|
192 |
# model2 is offloaded to the CPU just before this forward.
|
193 |
hid_3 = model_3(hid_3)
|
194 |
-
|
195 |
# For model3, you need to manually call the hook offload method.
|
196 |
hook_3.offload()
|
197 |
```
|
@@ -200,8 +165,6 @@ def cpu_offload_with_hook(
|
|
200 |
add_hook_to_module(model, hook, append=True)
|
201 |
user_hook = UserCpuOffloadHook(model, hook)
|
202 |
return model, user_hook
|
203 |
-
|
204 |
-
|
205 |
def disk_offload(
|
206 |
model: nn.Module,
|
207 |
offload_dir: Union[str, os.PathLike],
|
@@ -213,7 +176,6 @@ def disk_offload(
|
|
213 |
Activates full disk offload for a model. As a result, all parameters of the model will be offloaded as
|
214 |
memory-mapped array in a given folder. During the forward pass, parameters will be accessed from that folder and
|
215 |
put on the execution device passed as they are needed, then offloaded again.
|
216 |
-
|
217 |
Args:
|
218 |
model (`torch.nn.Module`): The model to offload.
|
219 |
offload_dir (`str` or `os.PathLike`):
|
@@ -234,7 +196,6 @@ def disk_offload(
|
|
234 |
if execution_device is None:
|
235 |
execution_device = next(iter(model.parameters())).device
|
236 |
weights_map = OffloadedWeightsLoader(save_folder=offload_dir)
|
237 |
-
|
238 |
add_hook_to_module(model, AlignDevicesHook(io_same_device=True), append=True)
|
239 |
attach_align_device_hook(
|
240 |
model,
|
@@ -244,10 +205,7 @@ def disk_offload(
|
|
244 |
weights_map=weights_map,
|
245 |
preload_module_classes=preload_module_classes,
|
246 |
)
|
247 |
-
|
248 |
return model
|
249 |
-
|
250 |
-
|
251 |
def dispatch_model(
|
252 |
model: nn.Module,
|
253 |
device_map: Dict[str, Union[str, int, torch.device]],
|
@@ -263,7 +221,6 @@ def dispatch_model(
|
|
263 |
"""
|
264 |
Dispatches a model according to a given device map. Layers of the model might be spread across GPUs, offloaded on
|
265 |
the CPU or even the disk.
|
266 |
-
|
267 |
Args:
|
268 |
model (`torch.nn.Module`):
|
269 |
The model to dispatch.
|
@@ -295,12 +252,10 @@ def dispatch_model(
|
|
295 |
"""
|
296 |
# Error early if the device map is incomplete.
|
297 |
check_device_map(model, device_map)
|
298 |
-
|
299 |
# for backward compatibility
|
300 |
is_bnb_quantized = (
|
301 |
getattr(model, "is_quantized", False) or getattr(model, "is_loaded_in_8bit", False)
|
302 |
) and getattr(model, "quantization_method", "bitsandbytes") == "bitsandbytes"
|
303 |
-
|
304 |
# We attach hooks if the device_map has at least 2 different devices or if
|
305 |
# force_hooks is set to `True`. Otherwise, the model in already loaded
|
306 |
# in the unique device and the user can decide where to dispatch the model.
|
@@ -311,12 +266,10 @@ def dispatch_model(
|
|
311 |
main_device = "cpu"
|
312 |
else:
|
313 |
main_device = [d for d in device_map.values() if d not in ["cpu", "disk"]][0]
|
314 |
-
|
315 |
if main_device != "cpu":
|
316 |
cpu_modules = [name for name, device in device_map.items() if device == "cpu"]
|
317 |
if state_dict is None and len(cpu_modules) > 0:
|
318 |
state_dict = extract_submodules_state_dict(model.state_dict(), cpu_modules)
|
319 |
-
|
320 |
disk_modules = [name for name, device in device_map.items() if device == "disk"]
|
321 |
if offload_dir is None and offload_index is None and len(disk_modules) > 0:
|
322 |
raise ValueError(
|
@@ -330,7 +283,6 @@ def dispatch_model(
|
|
330 |
):
|
331 |
disk_state_dict = extract_submodules_state_dict(model.state_dict(), disk_modules)
|
332 |
offload_state_dict(offload_dir, disk_state_dict)
|
333 |
-
|
334 |
execution_device = {
|
335 |
name: main_device if device in ["cpu", "disk"] else device for name, device in device_map.items()
|
336 |
}
|
@@ -345,7 +297,6 @@ def dispatch_model(
|
|
345 |
)
|
346 |
else:
|
347 |
weights_map = None
|
348 |
-
|
349 |
tied_params = find_tied_parameters(model)
|
350 |
attach_align_device_hook_on_blocks(
|
351 |
model,
|
@@ -356,7 +307,6 @@ def dispatch_model(
|
|
356 |
skip_keys=skip_keys,
|
357 |
preload_module_classes=preload_module_classes,
|
358 |
)
|
359 |
-
|
360 |
# warn if there is any params on the meta device
|
361 |
offloaded_devices_str = " and ".join(
|
362 |
[device for device in set(device_map.values()) if device in ("cpu", "disk")]
|
@@ -365,10 +315,8 @@ def dispatch_model(
|
|
365 |
logging.warning(
|
366 |
f"Some parameters are on the meta device device because they were offloaded to the {offloaded_devices_str}."
|
367 |
)
|
368 |
-
|
369 |
# Attaching the hook may break tied weights, so we retie them
|
370 |
retie_parameters(model, tied_params)
|
371 |
-
|
372 |
# add warning to cuda and to method
|
373 |
def add_warning(fn, model):
|
374 |
@wraps(fn)
|
@@ -378,15 +326,12 @@ def dispatch_model(
|
|
378 |
if param.device == torch.device("meta"):
|
379 |
raise RuntimeError("You can't move a model that has some modules offloaded to cpu or disk.")
|
380 |
return fn(*args, **kwargs)
|
381 |
-
|
382 |
return wrapper
|
383 |
-
|
384 |
model.to = add_warning(model.to, model)
|
385 |
if is_npu_available():
|
386 |
model.npu = add_warning(model.npu, model)
|
387 |
else:
|
388 |
model.cuda = add_warning(model.cuda, model)
|
389 |
-
|
390 |
else:
|
391 |
device = list(device_map.values())[0]
|
392 |
# `torch.Tensor.to(<int num>)` is not supported by `torch_npu` (see this [issue](https://github.com/Ascend/pytorch/issues/16)).
|
@@ -400,8 +345,6 @@ def dispatch_model(
|
|
400 |
)
|
401 |
model.hf_device_map = device_map
|
402 |
return model
|
403 |
-
|
404 |
-
|
405 |
def load_checkpoint_and_dispatch(
|
406 |
model: nn.Module,
|
407 |
checkpoint: Union[str, os.PathLike],
|
@@ -419,7 +362,6 @@ def load_checkpoint_and_dispatch(
|
|
419 |
"""
|
420 |
Loads a (potentially sharded) checkpoint inside a model, potentially sending weights to a given device as they are
|
421 |
loaded and adds the various hooks that will make this model run properly (even if split across devices).
|
422 |
-
|
423 |
Args:
|
424 |
model (`torch.nn.Module`): The model in which we want to load a checkpoint.
|
425 |
checkpoint (`str` or `os.PathLike`):
|
@@ -430,7 +372,6 @@ def load_checkpoint_and_dispatch(
|
|
430 |
device_map (`Dict[str, Union[int, str, torch.device]]`, *optional*):
|
431 |
A map that specifies where each submodule should go. It doesn't need to be refined to each parameter/buffer
|
432 |
name, once a given module name is inside, every submodule of it will be sent to the same device.
|
433 |
-
|
434 |
To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For more
|
435 |
information about each option see [here](big_modeling#designing-a-device-map).
|
436 |
max_memory (`Dict`, *optional*):
|
@@ -460,23 +401,18 @@ def load_checkpoint_and_dispatch(
|
|
460 |
force_hooks (`bool`, *optional*, defaults to `False`):
|
461 |
Whether or not to force device hooks to be attached to the model even if all layers are dispatched to a
|
462 |
single device.
|
463 |
-
|
464 |
Example:
|
465 |
-
|
466 |
```python
|
467 |
>>> from accelerate import init_empty_weights, load_checkpoint_and_dispatch
|
468 |
>>> from huggingface_hub import hf_hub_download
|
469 |
>>> from transformers import AutoConfig, AutoModelForCausalLM
|
470 |
-
|
471 |
>>> # Download the Weights
|
472 |
>>> checkpoint = "EleutherAI/gpt-j-6B"
|
473 |
>>> weights_location = hf_hub_download(checkpoint, "pytorch_model.bin")
|
474 |
-
|
475 |
>>> # Create a model and initialize it with empty weights
|
476 |
>>> config = AutoConfig.from_pretrained(checkpoint)
|
477 |
>>> with init_empty_weights():
|
478 |
... model = AutoModelForCausalLM.from_config(config)
|
479 |
-
|
480 |
>>> # Load the checkpoint and dispatch it to the right devices
|
481 |
>>> model = load_checkpoint_and_dispatch(
|
482 |
... model, weights_location, device_map="auto", no_split_module_classes=["GPTJBlock"]
|
|
|
1 |
logger = logging.getLogger(__name__)
|
|
|
|
|
2 |
@contextmanager
|
3 |
def init_empty_weights(include_buffers: bool = None):
|
4 |
"""
|
5 |
A context manager under which models are initialized with all parameters on the meta device, therefore creating an
|
6 |
empty model. Useful when just initializing the model would blow the available RAM.
|
|
|
7 |
Args:
|
8 |
include_buffers (`bool`, *optional*):
|
9 |
Whether or not to also put all buffers on the meta device while initializing.
|
|
|
10 |
Example:
|
|
|
11 |
```python
|
12 |
import torch.nn as nn
|
13 |
from accelerate import init_empty_weights
|
|
|
14 |
# Initialize a model with 100 billions parameters in no time and without using any RAM.
|
15 |
with init_empty_weights():
|
16 |
tst = nn.Sequential(*[nn.Linear(10000, 10000) for _ in range(1000)])
|
17 |
```
|
|
|
18 |
<Tip warning={true}>
|
|
|
19 |
Any model created under this context manager has no weights. As such you can't do something like
|
20 |
`model.to(some_device)` with it. To load weights inside your empty model, see [`load_checkpoint_and_dispatch`].
|
|
|
21 |
</Tip>
|
22 |
"""
|
23 |
if include_buffers is None:
|
24 |
include_buffers = parse_flag_from_env("ACCELERATE_INIT_INCLUDE_BUFFERS", False)
|
25 |
with init_on_device(torch.device("meta"), include_buffers=include_buffers) as f:
|
26 |
yield f
|
|
|
|
|
27 |
@contextmanager
|
28 |
def init_on_device(device: torch.device, include_buffers: bool = None):
|
29 |
"""
|
30 |
A context manager under which models are initialized with all parameters on the specified device.
|
|
|
31 |
Args:
|
32 |
device (`torch.device`):
|
33 |
Device to initialize all parameters on.
|
34 |
include_buffers (`bool`, *optional*):
|
35 |
Whether or not to also put all buffers on the meta device while initializing.
|
|
|
36 |
Example:
|
|
|
37 |
```python
|
38 |
import torch.nn as nn
|
39 |
from accelerate import init_on_device
|
|
|
40 |
with init_on_device(device=torch.device("cuda")):
|
41 |
tst = nn.Liner(100, 100) # on `cuda` device
|
42 |
```
|
43 |
"""
|
44 |
if include_buffers is None:
|
45 |
include_buffers = parse_flag_from_env("ACCELERATE_INIT_INCLUDE_BUFFERS", False)
|
|
|
46 |
# TODO(shingjan): remove the torch version check once older versions are deprecated
|
47 |
if is_torch_version(">=", "2.0") and include_buffers:
|
48 |
with device:
|
49 |
yield
|
50 |
return
|
|
|
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:
|
63 |
module._buffers[name] = module._buffers[name].to(device)
|
|
|
64 |
# Patch tensor creation
|
65 |
if include_buffers:
|
66 |
tensor_constructors_to_patch = {
|
|
|
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
|
|
|
77 |
try:
|
78 |
nn.Module.register_parameter = register_empty_parameter
|
79 |
if include_buffers:
|
|
|
87 |
nn.Module.register_buffer = old_register_buffer
|
88 |
for torch_function_name, old_torch_function in tensor_constructors_to_patch.items():
|
89 |
setattr(torch, torch_function_name, old_torch_function)
|
|
|
|
|
90 |
def cpu_offload(
|
91 |
model: nn.Module,
|
92 |
execution_device: Optional[torch.device] = None,
|
|
|
98 |
Activates full CPU offload for a model. As a result, all parameters of the model will be offloaded and only one
|
99 |
copy of the state dict of the model will be kept. During the forward pass, parameters will be extracted from that
|
100 |
state dict and put on the execution device passed as they are needed, then offloaded again.
|
|
|
101 |
Args:
|
102 |
model (`torch.nn.Module`):
|
103 |
The model to offload.
|
|
|
118 |
execution_device = next(iter(model.parameters())).device
|
119 |
if state_dict is None:
|
120 |
state_dict = {n: p.to("cpu") for n, p in model.state_dict().items()}
|
|
|
121 |
add_hook_to_module(model, AlignDevicesHook(io_same_device=True), append=True)
|
122 |
attach_align_device_hook(
|
123 |
model,
|
|
|
127 |
weights_map=state_dict,
|
128 |
preload_module_classes=preload_module_classes,
|
129 |
)
|
|
|
130 |
return model
|
|
|
|
|
131 |
def cpu_offload_with_hook(
|
132 |
model: torch.nn.Module,
|
133 |
execution_device: Optional[Union[int, str, torch.device]] = None,
|
|
|
137 |
Offloads a model on the CPU and puts it back to an execution device when executed. The difference with
|
138 |
[`cpu_offload`] is that the model stays on the execution device after the forward and is only offloaded again when
|
139 |
the `offload` method of the returned `hook` is called. Useful for pipelines running a model in a loop.
|
|
|
140 |
Args:
|
141 |
model (`torch.nn.Module`):
|
142 |
The model to offload.
|
|
|
146 |
prev_module_hook (`UserCpuOffloadHook`, *optional*):
|
147 |
The hook sent back by this function for a previous model in the pipeline you are running. If passed, its
|
148 |
offload method will be called just before the forward of the model to which this hook is attached.
|
|
|
149 |
Example:
|
|
|
150 |
```py
|
151 |
model_1, hook_1 = cpu_offload_with_hook(model_1, cuda_device)
|
152 |
model_2, hook_2 = cpu_offload_with_hook(model_2, cuda_device, prev_module_hook=hook_1)
|
153 |
model_3, hook_3 = cpu_offload_with_hook(model_3, cuda_device, prev_module_hook=hook_2)
|
|
|
154 |
hid_1 = model_1(input)
|
155 |
for i in range(50):
|
156 |
# model1 is offloaded on the CPU at the first iteration, model 2 stays on the GPU for this whole loop.
|
157 |
hid_2 = model_2(hid_1)
|
158 |
# model2 is offloaded to the CPU just before this forward.
|
159 |
hid_3 = model_3(hid_3)
|
|
|
160 |
# For model3, you need to manually call the hook offload method.
|
161 |
hook_3.offload()
|
162 |
```
|
|
|
165 |
add_hook_to_module(model, hook, append=True)
|
166 |
user_hook = UserCpuOffloadHook(model, hook)
|
167 |
return model, user_hook
|
|
|
|
|
168 |
def disk_offload(
|
169 |
model: nn.Module,
|
170 |
offload_dir: Union[str, os.PathLike],
|
|
|
176 |
Activates full disk offload for a model. As a result, all parameters of the model will be offloaded as
|
177 |
memory-mapped array in a given folder. During the forward pass, parameters will be accessed from that folder and
|
178 |
put on the execution device passed as they are needed, then offloaded again.
|
|
|
179 |
Args:
|
180 |
model (`torch.nn.Module`): The model to offload.
|
181 |
offload_dir (`str` or `os.PathLike`):
|
|
|
196 |
if execution_device is None:
|
197 |
execution_device = next(iter(model.parameters())).device
|
198 |
weights_map = OffloadedWeightsLoader(save_folder=offload_dir)
|
|
|
199 |
add_hook_to_module(model, AlignDevicesHook(io_same_device=True), append=True)
|
200 |
attach_align_device_hook(
|
201 |
model,
|
|
|
205 |
weights_map=weights_map,
|
206 |
preload_module_classes=preload_module_classes,
|
207 |
)
|
|
|
208 |
return model
|
|
|
|
|
209 |
def dispatch_model(
|
210 |
model: nn.Module,
|
211 |
device_map: Dict[str, Union[str, int, torch.device]],
|
|
|
221 |
"""
|
222 |
Dispatches a model according to a given device map. Layers of the model might be spread across GPUs, offloaded on
|
223 |
the CPU or even the disk.
|
|
|
224 |
Args:
|
225 |
model (`torch.nn.Module`):
|
226 |
The model to dispatch.
|
|
|
252 |
"""
|
253 |
# Error early if the device map is incomplete.
|
254 |
check_device_map(model, device_map)
|
|
|
255 |
# for backward compatibility
|
256 |
is_bnb_quantized = (
|
257 |
getattr(model, "is_quantized", False) or getattr(model, "is_loaded_in_8bit", False)
|
258 |
) and getattr(model, "quantization_method", "bitsandbytes") == "bitsandbytes"
|
|
|
259 |
# We attach hooks if the device_map has at least 2 different devices or if
|
260 |
# force_hooks is set to `True`. Otherwise, the model in already loaded
|
261 |
# in the unique device and the user can decide where to dispatch the model.
|
|
|
266 |
main_device = "cpu"
|
267 |
else:
|
268 |
main_device = [d for d in device_map.values() if d not in ["cpu", "disk"]][0]
|
|
|
269 |
if main_device != "cpu":
|
270 |
cpu_modules = [name for name, device in device_map.items() if device == "cpu"]
|
271 |
if state_dict is None and len(cpu_modules) > 0:
|
272 |
state_dict = extract_submodules_state_dict(model.state_dict(), cpu_modules)
|
|
|
273 |
disk_modules = [name for name, device in device_map.items() if device == "disk"]
|
274 |
if offload_dir is None and offload_index is None and len(disk_modules) > 0:
|
275 |
raise ValueError(
|
|
|
283 |
):
|
284 |
disk_state_dict = extract_submodules_state_dict(model.state_dict(), disk_modules)
|
285 |
offload_state_dict(offload_dir, disk_state_dict)
|
|
|
286 |
execution_device = {
|
287 |
name: main_device if device in ["cpu", "disk"] else device for name, device in device_map.items()
|
288 |
}
|
|
|
297 |
)
|
298 |
else:
|
299 |
weights_map = None
|
|
|
300 |
tied_params = find_tied_parameters(model)
|
301 |
attach_align_device_hook_on_blocks(
|
302 |
model,
|
|
|
307 |
skip_keys=skip_keys,
|
308 |
preload_module_classes=preload_module_classes,
|
309 |
)
|
|
|
310 |
# warn if there is any params on the meta device
|
311 |
offloaded_devices_str = " and ".join(
|
312 |
[device for device in set(device_map.values()) if device in ("cpu", "disk")]
|
|
|
315 |
logging.warning(
|
316 |
f"Some parameters are on the meta device device because they were offloaded to the {offloaded_devices_str}."
|
317 |
)
|
|
|
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)
|
|
|
326 |
if param.device == torch.device("meta"):
|
327 |
raise RuntimeError("You can't move a model that has some modules offloaded to cpu or disk.")
|
328 |
return fn(*args, **kwargs)
|
|
|
329 |
return wrapper
|
|
|
330 |
model.to = add_warning(model.to, model)
|
331 |
if is_npu_available():
|
332 |
model.npu = add_warning(model.npu, model)
|
333 |
else:
|
334 |
model.cuda = add_warning(model.cuda, model)
|
|
|
335 |
else:
|
336 |
device = list(device_map.values())[0]
|
337 |
# `torch.Tensor.to(<int num>)` is not supported by `torch_npu` (see this [issue](https://github.com/Ascend/pytorch/issues/16)).
|
|
|
345 |
)
|
346 |
model.hf_device_map = device_map
|
347 |
return model
|
|
|
|
|
348 |
def load_checkpoint_and_dispatch(
|
349 |
model: nn.Module,
|
350 |
checkpoint: Union[str, os.PathLike],
|
|
|
362 |
"""
|
363 |
Loads a (potentially sharded) checkpoint inside a model, potentially sending weights to a given device as they are
|
364 |
loaded and adds the various hooks that will make this model run properly (even if split across devices).
|
|
|
365 |
Args:
|
366 |
model (`torch.nn.Module`): The model in which we want to load a checkpoint.
|
367 |
checkpoint (`str` or `os.PathLike`):
|
|
|
372 |
device_map (`Dict[str, Union[int, str, torch.device]]`, *optional*):
|
373 |
A map that specifies where each submodule should go. It doesn't need to be refined to each parameter/buffer
|
374 |
name, once a given module name is inside, every submodule of it will be sent to the same device.
|
|
|
375 |
To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For more
|
376 |
information about each option see [here](big_modeling#designing-a-device-map).
|
377 |
max_memory (`Dict`, *optional*):
|
|
|
401 |
force_hooks (`bool`, *optional*, defaults to `False`):
|
402 |
Whether or not to force device hooks to be attached to the model even if all layers are dispatched to a
|
403 |
single device.
|
|
|
404 |
Example:
|
|
|
405 |
```python
|
406 |
>>> from accelerate import init_empty_weights, load_checkpoint_and_dispatch
|
407 |
>>> from huggingface_hub import hf_hub_download
|
408 |
>>> from transformers import AutoConfig, AutoModelForCausalLM
|
|
|
409 |
>>> # Download the Weights
|
410 |
>>> checkpoint = "EleutherAI/gpt-j-6B"
|
411 |
>>> weights_location = hf_hub_download(checkpoint, "pytorch_model.bin")
|
|
|
412 |
>>> # Create a model and initialize it with empty weights
|
413 |
>>> config = AutoConfig.from_pretrained(checkpoint)
|
414 |
>>> with init_empty_weights():
|
415 |
... model = AutoModelForCausalLM.from_config(config)
|
|
|
416 |
>>> # Load the checkpoint and dispatch it to the right devices
|
417 |
>>> model = load_checkpoint_and_dispatch(
|
418 |
... model, weights_location, device_map="auto", no_split_module_classes=["GPTJBlock"]
|
src/checkpointing.py
CHANGED
@@ -1,6 +1,4 @@
|
|
1 |
logger = get_logger(__name__)
|
2 |
-
|
3 |
-
|
4 |
def save_accelerator_state(
|
5 |
output_dir: str,
|
6 |
model_states: List[dict],
|
@@ -14,14 +12,10 @@ def save_accelerator_state(
|
|
14 |
):
|
15 |
"""
|
16 |
Saves the current states of the models, optimizers, scaler, and RNG generators to a given directory.
|
17 |
-
|
18 |
<Tip>
|
19 |
-
|
20 |
If `safe_serialization` is `True`, models will be saved with `safetensors` while the rest are saved using native
|
21 |
`pickle`.
|
22 |
-
|
23 |
</Tip>
|
24 |
-
|
25 |
Args:
|
26 |
output_dir (`str` or `os.PathLike`):
|
27 |
The name of the folder to save all relevant weights and states.
|
@@ -71,14 +65,11 @@ def save_accelerator_state(
|
|
71 |
output_sampler_file = output_dir.joinpath(sampler_name)
|
72 |
# Only save if we have our custom sampler
|
73 |
from .data_loader import IterableDatasetShard, SeedableRandomSampler
|
74 |
-
|
75 |
if isinstance(dataloader.dataset, IterableDatasetShard):
|
76 |
sampler = dataloader.sampler.sampler
|
77 |
-
|
78 |
if isinstance(sampler, SeedableRandomSampler):
|
79 |
save(sampler, output_sampler_file, save_on_each_node=save_on_each_node, safe_serialization=False)
|
80 |
logger.info(f"Sampler state for dataloader {i} saved in {output_sampler_file}")
|
81 |
-
|
82 |
# GradScaler state
|
83 |
if scaler is not None:
|
84 |
state = scaler.state_dict()
|
@@ -101,8 +92,6 @@ def save_accelerator_state(
|
|
101 |
torch.save(states, output_states_file)
|
102 |
logger.info(f"Random states saved in {output_states_file}")
|
103 |
return output_dir
|
104 |
-
|
105 |
-
|
106 |
def load_accelerator_state(
|
107 |
input_dir,
|
108 |
models,
|
@@ -116,7 +105,6 @@ def load_accelerator_state(
|
|
116 |
):
|
117 |
"""
|
118 |
Loads states of the models, optimizers, scaler, and RNG generators from a given directory.
|
119 |
-
|
120 |
Args:
|
121 |
input_dir (`str` or `os.PathLike`):
|
122 |
The name of the folder to load all relevant weights and states.
|
@@ -143,7 +131,6 @@ def load_accelerator_state(
|
|
143 |
map_location = "cpu"
|
144 |
elif map_location == "on_device":
|
145 |
map_location = PartialState().device
|
146 |
-
|
147 |
input_dir = Path(input_dir)
|
148 |
# Model states
|
149 |
for i, model in enumerate(models):
|
@@ -157,7 +144,6 @@ def load_accelerator_state(
|
|
157 |
state_dict = torch.load(input_model_file, map_location=map_location)
|
158 |
models[i].load_state_dict(state_dict, **load_model_func_kwargs)
|
159 |
logger.info("All model weights loaded successfully")
|
160 |
-
|
161 |
# Optimizer states
|
162 |
for i, opt in enumerate(optimizers):
|
163 |
optimizer_name = f"{OPTIMIZER_NAME}.bin" if i == 0 else f"{OPTIMIZER_NAME}_{i}.bin"
|
@@ -165,33 +151,27 @@ def load_accelerator_state(
|
|
165 |
optimizer_state = torch.load(input_optimizer_file, map_location=map_location)
|
166 |
optimizers[i].load_state_dict(optimizer_state)
|
167 |
logger.info("All optimizer states loaded successfully")
|
168 |
-
|
169 |
# Scheduler states
|
170 |
for i, scheduler in enumerate(schedulers):
|
171 |
scheduler_name = f"{SCHEDULER_NAME}.bin" if i == 0 else f"{SCHEDULER_NAME}_{i}.bin"
|
172 |
input_scheduler_file = input_dir.joinpath(scheduler_name)
|
173 |
scheduler.load_state_dict(torch.load(input_scheduler_file))
|
174 |
logger.info("All scheduler states loaded successfully")
|
175 |
-
|
176 |
for i, dataloader in enumerate(dataloaders):
|
177 |
sampler_name = f"{SAMPLER_NAME}.bin" if i == 0 else f"{SAMPLER_NAME}_{i}.bin"
|
178 |
input_sampler_file = input_dir.joinpath(sampler_name)
|
179 |
# Only load if we have our custom sampler
|
180 |
from .data_loader import IterableDatasetShard, SeedableRandomSampler
|
181 |
-
|
182 |
if isinstance(dataloader.dataset, IterableDatasetShard):
|
183 |
sampler = dataloader.sampler.sampler
|
184 |
-
|
185 |
if isinstance(sampler, SeedableRandomSampler):
|
186 |
dataloader.sampler.sampler = torch.load(input_sampler_file)
|
187 |
logger.info("All dataloader sampler states loaded successfully")
|
188 |
-
|
189 |
# GradScaler state
|
190 |
if scaler is not None:
|
191 |
input_scaler_file = input_dir.joinpath(SCALER_NAME)
|
192 |
scaler.load_state_dict(torch.load(input_scaler_file))
|
193 |
logger.info("GradScaler state loaded successfully")
|
194 |
-
|
195 |
# Random states
|
196 |
try:
|
197 |
states = torch.load(input_dir.joinpath(f"{RNG_STATE_NAME}_{process_index}.pkl"))
|
@@ -207,8 +187,6 @@ def load_accelerator_state(
|
|
207 |
logger.info("All random states loaded successfully")
|
208 |
except Exception:
|
209 |
logger.info("Could not load random states")
|
210 |
-
|
211 |
-
|
212 |
def save_custom_state(obj, path, index: int = 0, save_on_each_node: bool = False):
|
213 |
"""
|
214 |
Saves the state of `obj` to `{path}/custom_checkpoint_{index}.pkl`
|
@@ -217,8 +195,6 @@ def save_custom_state(obj, path, index: int = 0, save_on_each_node: bool = False
|
|
217 |
save_location = Path(path) / f"custom_checkpoint_{index}.pkl"
|
218 |
logger.info(f"Saving the state of {get_pretty_name(obj)} to {save_location}")
|
219 |
save(obj.state_dict(), save_location, save_on_each_node=save_on_each_node)
|
220 |
-
|
221 |
-
|
222 |
def load_custom_state(obj, path, index: int = 0):
|
223 |
"""
|
224 |
Loads the state of `obj` at `{path}/custom_checkpoint_{index}.pkl`
|
|
|
1 |
logger = get_logger(__name__)
|
|
|
|
|
2 |
def save_accelerator_state(
|
3 |
output_dir: str,
|
4 |
model_states: List[dict],
|
|
|
12 |
):
|
13 |
"""
|
14 |
Saves the current states of the models, optimizers, scaler, and RNG generators to a given directory.
|
|
|
15 |
<Tip>
|
|
|
16 |
If `safe_serialization` is `True`, models will be saved with `safetensors` while the rest are saved using native
|
17 |
`pickle`.
|
|
|
18 |
</Tip>
|
|
|
19 |
Args:
|
20 |
output_dir (`str` or `os.PathLike`):
|
21 |
The name of the folder to save all relevant weights and states.
|
|
|
65 |
output_sampler_file = output_dir.joinpath(sampler_name)
|
66 |
# Only save if we have our custom sampler
|
67 |
from .data_loader import IterableDatasetShard, SeedableRandomSampler
|
|
|
68 |
if isinstance(dataloader.dataset, IterableDatasetShard):
|
69 |
sampler = dataloader.sampler.sampler
|
|
|
70 |
if isinstance(sampler, SeedableRandomSampler):
|
71 |
save(sampler, output_sampler_file, save_on_each_node=save_on_each_node, safe_serialization=False)
|
72 |
logger.info(f"Sampler state for dataloader {i} saved in {output_sampler_file}")
|
|
|
73 |
# GradScaler state
|
74 |
if scaler is not None:
|
75 |
state = scaler.state_dict()
|
|
|
92 |
torch.save(states, output_states_file)
|
93 |
logger.info(f"Random states saved in {output_states_file}")
|
94 |
return output_dir
|
|
|
|
|
95 |
def load_accelerator_state(
|
96 |
input_dir,
|
97 |
models,
|
|
|
105 |
):
|
106 |
"""
|
107 |
Loads states of the models, optimizers, scaler, and RNG generators from a given directory.
|
|
|
108 |
Args:
|
109 |
input_dir (`str` or `os.PathLike`):
|
110 |
The name of the folder to load all relevant weights and states.
|
|
|
131 |
map_location = "cpu"
|
132 |
elif map_location == "on_device":
|
133 |
map_location = PartialState().device
|
|
|
134 |
input_dir = Path(input_dir)
|
135 |
# Model states
|
136 |
for i, model in enumerate(models):
|
|
|
144 |
state_dict = torch.load(input_model_file, map_location=map_location)
|
145 |
models[i].load_state_dict(state_dict, **load_model_func_kwargs)
|
146 |
logger.info("All model weights loaded successfully")
|
|
|
147 |
# Optimizer states
|
148 |
for i, opt in enumerate(optimizers):
|
149 |
optimizer_name = f"{OPTIMIZER_NAME}.bin" if i == 0 else f"{OPTIMIZER_NAME}_{i}.bin"
|
|
|
151 |
optimizer_state = torch.load(input_optimizer_file, map_location=map_location)
|
152 |
optimizers[i].load_state_dict(optimizer_state)
|
153 |
logger.info("All optimizer states loaded successfully")
|
|
|
154 |
# Scheduler states
|
155 |
for i, scheduler in enumerate(schedulers):
|
156 |
scheduler_name = f"{SCHEDULER_NAME}.bin" if i == 0 else f"{SCHEDULER_NAME}_{i}.bin"
|
157 |
input_scheduler_file = input_dir.joinpath(scheduler_name)
|
158 |
scheduler.load_state_dict(torch.load(input_scheduler_file))
|
159 |
logger.info("All scheduler states loaded successfully")
|
|
|
160 |
for i, dataloader in enumerate(dataloaders):
|
161 |
sampler_name = f"{SAMPLER_NAME}.bin" if i == 0 else f"{SAMPLER_NAME}_{i}.bin"
|
162 |
input_sampler_file = input_dir.joinpath(sampler_name)
|
163 |
# Only load if we have our custom sampler
|
164 |
from .data_loader import IterableDatasetShard, SeedableRandomSampler
|
|
|
165 |
if isinstance(dataloader.dataset, IterableDatasetShard):
|
166 |
sampler = dataloader.sampler.sampler
|
|
|
167 |
if isinstance(sampler, SeedableRandomSampler):
|
168 |
dataloader.sampler.sampler = torch.load(input_sampler_file)
|
169 |
logger.info("All dataloader sampler states loaded successfully")
|
|
|
170 |
# GradScaler state
|
171 |
if scaler is not None:
|
172 |
input_scaler_file = input_dir.joinpath(SCALER_NAME)
|
173 |
scaler.load_state_dict(torch.load(input_scaler_file))
|
174 |
logger.info("GradScaler state loaded successfully")
|
|
|
175 |
# Random states
|
176 |
try:
|
177 |
states = torch.load(input_dir.joinpath(f"{RNG_STATE_NAME}_{process_index}.pkl"))
|
|
|
187 |
logger.info("All random states loaded successfully")
|
188 |
except Exception:
|
189 |
logger.info("Could not load random states")
|
|
|
|
|
190 |
def save_custom_state(obj, path, index: int = 0, save_on_each_node: bool = False):
|
191 |
"""
|
192 |
Saves the state of `obj` to `{path}/custom_checkpoint_{index}.pkl`
|
|
|
195 |
save_location = Path(path) / f"custom_checkpoint_{index}.pkl"
|
196 |
logger.info(f"Saving the state of {get_pretty_name(obj)} to {save_location}")
|
197 |
save(obj.state_dict(), save_location, save_on_each_node=save_on_each_node)
|
|
|
|
|
198 |
def load_custom_state(obj, path, index: int = 0):
|
199 |
"""
|
200 |
Loads the state of `obj` at `{path}/custom_checkpoint_{index}.pkl`
|
src/commands/accelerate_cli.py
CHANGED
@@ -2,7 +2,6 @@
|
|
2 |
def main():
|
3 |
parser = ArgumentParser("Accelerate CLI tool", usage="accelerate <command> [<args>]", allow_abbrev=False)
|
4 |
subparsers = parser.add_subparsers(help="accelerate command helpers")
|
5 |
-
|
6 |
# Register commands
|
7 |
get_config_parser(subparsers=subparsers)
|
8 |
estimate_command_parser(subparsers=subparsers)
|
@@ -10,17 +9,12 @@ def main():
|
|
10 |
launch_command_parser(subparsers=subparsers)
|
11 |
tpu_command_parser(subparsers=subparsers)
|
12 |
test_command_parser(subparsers=subparsers)
|
13 |
-
|
14 |
# Let's go
|
15 |
args = parser.parse_args()
|
16 |
-
|
17 |
if not hasattr(args, "func"):
|
18 |
parser.print_help()
|
19 |
exit(1)
|
20 |
-
|
21 |
# Run
|
22 |
args.func(args)
|
23 |
-
|
24 |
-
|
25 |
if __name__ == "__main__":
|
26 |
main()
|
|
|
2 |
def main():
|
3 |
parser = ArgumentParser("Accelerate CLI tool", usage="accelerate <command> [<args>]", allow_abbrev=False)
|
4 |
subparsers = parser.add_subparsers(help="accelerate command helpers")
|
|
|
5 |
# Register commands
|
6 |
get_config_parser(subparsers=subparsers)
|
7 |
estimate_command_parser(subparsers=subparsers)
|
|
|
9 |
launch_command_parser(subparsers=subparsers)
|
10 |
tpu_command_parser(subparsers=subparsers)
|
11 |
test_command_parser(subparsers=subparsers)
|
|
|
12 |
# Let's go
|
13 |
args = parser.parse_args()
|
|
|
14 |
if not hasattr(args, "func"):
|
15 |
parser.print_help()
|
16 |
exit(1)
|
|
|
17 |
# Run
|
18 |
args.func(args)
|
|
|
|
|
19 |
if __name__ == "__main__":
|
20 |
main()
|
src/commands/config/cluster.py
CHANGED
@@ -5,7 +5,6 @@ def get_cluster_input():
|
|
5 |
["No distributed training", "multi-CPU", "multi-XPU", "multi-GPU", "multi-NPU", "TPU"],
|
6 |
_convert_distributed_mode,
|
7 |
)
|
8 |
-
|
9 |
machine_rank = 0
|
10 |
num_machines = 1
|
11 |
num_processes = 1
|
@@ -15,7 +14,6 @@ def get_cluster_input():
|
|
15 |
rdzv_backend = "static"
|
16 |
same_network = True
|
17 |
debug = False
|
18 |
-
|
19 |
if distributed_type in [
|
20 |
DistributedType.MULTI_GPU,
|
21 |
DistributedType.MULTI_NPU,
|
@@ -56,7 +54,6 @@ def get_cluster_input():
|
|
56 |
default=False,
|
57 |
error_message="Please enter yes or no.",
|
58 |
)
|
59 |
-
|
60 |
if distributed_type == DistributedType.NO:
|
61 |
use_cpu = _ask_field(
|
62 |
"Do you want to run your training on CPU only (even if a GPU / Apple Silicon / Ascend NPU device is available)? [yes/NO]:",
|
@@ -68,7 +65,6 @@ def get_cluster_input():
|
|
68 |
use_cpu = True
|
69 |
else:
|
70 |
use_cpu = False
|
71 |
-
|
72 |
ipex_config = {}
|
73 |
if use_cpu:
|
74 |
ipex_config["ipex"] = _ask_field(
|
@@ -88,7 +84,6 @@ def get_cluster_input():
|
|
88 |
default=False,
|
89 |
error_message="Please enter yes or no.",
|
90 |
)
|
91 |
-
|
92 |
dynamo_config = {}
|
93 |
use_dynamo = _ask_field(
|
94 |
"Do you wish to optimize your script with torch dynamo?[yes/NO]:",
|
@@ -110,7 +105,6 @@ def get_cluster_input():
|
|
110 |
default=False,
|
111 |
error_message="Please enter yes or no.",
|
112 |
)
|
113 |
-
|
114 |
if use_custom_options:
|
115 |
dynamo_config[prefix + "mode"] = _ask_options(
|
116 |
"Which mode do you want to use?",
|
@@ -130,7 +124,6 @@ def get_cluster_input():
|
|
130 |
default=False,
|
131 |
error_message="Please enter yes or no.",
|
132 |
)
|
133 |
-
|
134 |
use_mps = not use_cpu and is_mps_available()
|
135 |
deepspeed_config = {}
|
136 |
if distributed_type in [DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.NO] and not use_mps:
|
@@ -145,7 +138,6 @@ def get_cluster_input():
|
|
145 |
assert (
|
146 |
is_deepspeed_available()
|
147 |
), "DeepSpeed is not installed => run `pip3 install deepspeed` or build it from source"
|
148 |
-
|
149 |
if distributed_type == DistributedType.DEEPSPEED:
|
150 |
use_deepspeed_config = _ask_field(
|
151 |
"Do you want to specify a json file to a DeepSpeed config? [yes/NO]: ",
|
@@ -166,7 +158,6 @@ def get_cluster_input():
|
|
166 |
int,
|
167 |
default=2,
|
168 |
)
|
169 |
-
|
170 |
deepspeed_devices = ["none", "cpu", "nvme"]
|
171 |
if deepspeed_config["zero_stage"] >= 2:
|
172 |
deepspeed_config["offload_optimizer_device"] = _ask_options(
|
@@ -223,7 +214,6 @@ def get_cluster_input():
|
|
223 |
"When `zero3_init_flag` is set, it requires Transformers to be installed. "
|
224 |
"Please run `pip3 install transformers`."
|
225 |
)
|
226 |
-
|
227 |
if num_machines > 1:
|
228 |
launcher_query = "Which Type of launcher do you want to use?"
|
229 |
deepspeed_config["deepspeed_multinode_launcher"] = _ask_options(
|
@@ -231,7 +221,6 @@ def get_cluster_input():
|
|
231 |
DEEPSPEED_MULTINODE_LAUNCHERS,
|
232 |
lambda x: DEEPSPEED_MULTINODE_LAUNCHERS[int(x)],
|
233 |
)
|
234 |
-
|
235 |
if deepspeed_config["deepspeed_multinode_launcher"] != DEEPSPEED_MULTINODE_LAUNCHERS[1]:
|
236 |
deepspeed_config["deepspeed_hostfile"] = _ask_field(
|
237 |
"DeepSpeed configures multi-node compute resources with hostfile. "
|
@@ -241,7 +230,6 @@ def get_cluster_input():
|
|
241 |
"Please specify the location of hostfile: ",
|
242 |
str,
|
243 |
)
|
244 |
-
|
245 |
is_exclusion_filter = _ask_field(
|
246 |
"Do you want to specify exclusion filter string? [yes/NO]: ",
|
247 |
_convert_yes_no_to_bool,
|
@@ -253,7 +241,6 @@ def get_cluster_input():
|
|
253 |
"DeepSpeed exclusion filter string: ",
|
254 |
str,
|
255 |
)
|
256 |
-
|
257 |
is_inclusion_filter = _ask_field(
|
258 |
"Do you want to specify inclusion filter string? [yes/NO]: ",
|
259 |
_convert_yes_no_to_bool,
|
@@ -265,7 +252,6 @@ def get_cluster_input():
|
|
265 |
"DeepSpeed inclusion filter string: ",
|
266 |
str,
|
267 |
)
|
268 |
-
|
269 |
fsdp_config = {}
|
270 |
if distributed_type in [DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.MULTI_XPU]:
|
271 |
use_fsdp = _ask_field(
|
@@ -354,7 +340,6 @@ def get_cluster_input():
|
|
354 |
default=True,
|
355 |
error_message="Please enter yes or no.",
|
356 |
)
|
357 |
-
|
358 |
megatron_lm_config = {}
|
359 |
if distributed_type in [DistributedType.MULTI_GPU]:
|
360 |
use_megatron_lm = _ask_field(
|
@@ -380,7 +365,6 @@ def get_cluster_input():
|
|
380 |
default=True,
|
381 |
error_message="Please enter yes or no.",
|
382 |
)
|
383 |
-
|
384 |
megatron_lm_config[prefix + "pp_degree"] = _ask_field(
|
385 |
"What is the Pipeline Parallelism degree/size? [1]:",
|
386 |
int,
|
@@ -394,14 +378,12 @@ def get_cluster_input():
|
|
394 |
default=1,
|
395 |
error_message="Please enter an integer.",
|
396 |
)
|
397 |
-
|
398 |
megatron_lm_config[prefix + "recompute_activations"] = _ask_field(
|
399 |
"Do you want to enable selective activation recomputation? [YES/no]: ",
|
400 |
_convert_yes_no_to_bool,
|
401 |
default=True,
|
402 |
error_message="Please enter yes or no.",
|
403 |
)
|
404 |
-
|
405 |
megatron_lm_config[prefix + "use_distributed_optimizer"] = _ask_field(
|
406 |
"Do you want to use distributed optimizer "
|
407 |
"which shards optimizer state and gradients across data parallel ranks? [YES/no]: ",
|
@@ -409,7 +391,6 @@ def get_cluster_input():
|
|
409 |
default=True,
|
410 |
error_message="Please enter yes or no.",
|
411 |
)
|
412 |
-
|
413 |
megatron_lm_config[prefix + "gradient_clipping"] = _ask_field(
|
414 |
"What is the gradient clipping value based on global L2 Norm (0 to disable)? [1.0]: ",
|
415 |
float,
|
@@ -425,7 +406,6 @@ def get_cluster_input():
|
|
425 |
tpu_zone = None
|
426 |
tpu_use_sudo = False
|
427 |
tpu_use_cluster = False
|
428 |
-
|
429 |
if distributed_type in [
|
430 |
DistributedType.MULTI_CPU,
|
431 |
DistributedType.MULTI_XPU,
|
@@ -453,12 +433,10 @@ def get_cluster_input():
|
|
453 |
)
|
454 |
else:
|
455 |
num_processes = 1
|
456 |
-
|
457 |
if (distributed_type == DistributedType.MULTI_GPU) and (num_machines == 1) and (num_processes == 1):
|
458 |
raise ValueError(
|
459 |
f"Specified distributed type {distributed_type} but only using 1 GPU on a single machine. Please select `No distributed training` for the type of machine you are using."
|
460 |
)
|
461 |
-
|
462 |
if (
|
463 |
distributed_type
|
464 |
in [
|
@@ -478,7 +456,6 @@ def get_cluster_input():
|
|
478 |
f"What {machine_type} (by id) should be used for training on this machine as a comma-seperated list? [all]:",
|
479 |
default="all",
|
480 |
)
|
481 |
-
|
482 |
if distributed_type == DistributedType.TPU:
|
483 |
mixed_precision = "no"
|
484 |
main_training_function = _ask_field(
|
@@ -553,7 +530,6 @@ def get_cluster_input():
|
|
553 |
"What environment variables do you wish to set in each pod, seperated by a comma: ",
|
554 |
default="",
|
555 |
).split(",")
|
556 |
-
|
557 |
else:
|
558 |
main_training_function = "main"
|
559 |
if distributed_type == DistributedType.DEEPSPEED and use_deepspeed_config:
|
@@ -564,17 +540,14 @@ def get_cluster_input():
|
|
564 |
["no", "fp16", "bf16", "fp8"],
|
565 |
_convert_mixed_precision,
|
566 |
)
|
567 |
-
|
568 |
if use_dynamo and mixed_precision == "no" and not use_cpu:
|
569 |
print(
|
570 |
"Torch dynamo used without mixed precision requires TF32 to be efficient. Accelerate will enable it by default when launching your scripts."
|
571 |
)
|
572 |
-
|
573 |
if distributed_type == DistributedType.TPU and mixed_precision == "bf16":
|
574 |
tpu_downcast_bf16 = _ask_field(
|
575 |
"Should `torch.float` be cast as `bfloat16` and `torch.double` remain `float32` on TPUs?", default="no"
|
576 |
)
|
577 |
-
|
578 |
return ClusterConfig(
|
579 |
compute_environment=ComputeEnvironment.LOCAL_MACHINE,
|
580 |
distributed_type=distributed_type,
|
|
|
5 |
["No distributed training", "multi-CPU", "multi-XPU", "multi-GPU", "multi-NPU", "TPU"],
|
6 |
_convert_distributed_mode,
|
7 |
)
|
|
|
8 |
machine_rank = 0
|
9 |
num_machines = 1
|
10 |
num_processes = 1
|
|
|
14 |
rdzv_backend = "static"
|
15 |
same_network = True
|
16 |
debug = False
|
|
|
17 |
if distributed_type in [
|
18 |
DistributedType.MULTI_GPU,
|
19 |
DistributedType.MULTI_NPU,
|
|
|
54 |
default=False,
|
55 |
error_message="Please enter yes or no.",
|
56 |
)
|
|
|
57 |
if distributed_type == DistributedType.NO:
|
58 |
use_cpu = _ask_field(
|
59 |
"Do you want to run your training on CPU only (even if a GPU / Apple Silicon / Ascend NPU device is available)? [yes/NO]:",
|
|
|
65 |
use_cpu = True
|
66 |
else:
|
67 |
use_cpu = False
|
|
|
68 |
ipex_config = {}
|
69 |
if use_cpu:
|
70 |
ipex_config["ipex"] = _ask_field(
|
|
|
84 |
default=False,
|
85 |
error_message="Please enter yes or no.",
|
86 |
)
|
|
|
87 |
dynamo_config = {}
|
88 |
use_dynamo = _ask_field(
|
89 |
"Do you wish to optimize your script with torch dynamo?[yes/NO]:",
|
|
|
105 |
default=False,
|
106 |
error_message="Please enter yes or no.",
|
107 |
)
|
|
|
108 |
if use_custom_options:
|
109 |
dynamo_config[prefix + "mode"] = _ask_options(
|
110 |
"Which mode do you want to use?",
|
|
|
124 |
default=False,
|
125 |
error_message="Please enter yes or no.",
|
126 |
)
|
|
|
127 |
use_mps = not use_cpu and is_mps_available()
|
128 |
deepspeed_config = {}
|
129 |
if distributed_type in [DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.NO] and not use_mps:
|
|
|
138 |
assert (
|
139 |
is_deepspeed_available()
|
140 |
), "DeepSpeed is not installed => run `pip3 install deepspeed` or build it from source"
|
|
|
141 |
if distributed_type == DistributedType.DEEPSPEED:
|
142 |
use_deepspeed_config = _ask_field(
|
143 |
"Do you want to specify a json file to a DeepSpeed config? [yes/NO]: ",
|
|
|
158 |
int,
|
159 |
default=2,
|
160 |
)
|
|
|
161 |
deepspeed_devices = ["none", "cpu", "nvme"]
|
162 |
if deepspeed_config["zero_stage"] >= 2:
|
163 |
deepspeed_config["offload_optimizer_device"] = _ask_options(
|
|
|
214 |
"When `zero3_init_flag` is set, it requires Transformers to be installed. "
|
215 |
"Please run `pip3 install transformers`."
|
216 |
)
|
|
|
217 |
if num_machines > 1:
|
218 |
launcher_query = "Which Type of launcher do you want to use?"
|
219 |
deepspeed_config["deepspeed_multinode_launcher"] = _ask_options(
|
|
|
221 |
DEEPSPEED_MULTINODE_LAUNCHERS,
|
222 |
lambda x: DEEPSPEED_MULTINODE_LAUNCHERS[int(x)],
|
223 |
)
|
|
|
224 |
if deepspeed_config["deepspeed_multinode_launcher"] != DEEPSPEED_MULTINODE_LAUNCHERS[1]:
|
225 |
deepspeed_config["deepspeed_hostfile"] = _ask_field(
|
226 |
"DeepSpeed configures multi-node compute resources with hostfile. "
|
|
|
230 |
"Please specify the location of hostfile: ",
|
231 |
str,
|
232 |
)
|
|
|
233 |
is_exclusion_filter = _ask_field(
|
234 |
"Do you want to specify exclusion filter string? [yes/NO]: ",
|
235 |
_convert_yes_no_to_bool,
|
|
|
241 |
"DeepSpeed exclusion filter string: ",
|
242 |
str,
|
243 |
)
|
|
|
244 |
is_inclusion_filter = _ask_field(
|
245 |
"Do you want to specify inclusion filter string? [yes/NO]: ",
|
246 |
_convert_yes_no_to_bool,
|
|
|
252 |
"DeepSpeed inclusion filter string: ",
|
253 |
str,
|
254 |
)
|
|
|
255 |
fsdp_config = {}
|
256 |
if distributed_type in [DistributedType.MULTI_GPU, DistributedType.MULTI_NPU, DistributedType.MULTI_XPU]:
|
257 |
use_fsdp = _ask_field(
|
|
|
340 |
default=True,
|
341 |
error_message="Please enter yes or no.",
|
342 |
)
|
|
|
343 |
megatron_lm_config = {}
|
344 |
if distributed_type in [DistributedType.MULTI_GPU]:
|
345 |
use_megatron_lm = _ask_field(
|
|
|
365 |
default=True,
|
366 |
error_message="Please enter yes or no.",
|
367 |
)
|
|
|
368 |
megatron_lm_config[prefix + "pp_degree"] = _ask_field(
|
369 |
"What is the Pipeline Parallelism degree/size? [1]:",
|
370 |
int,
|
|
|
378 |
default=1,
|
379 |
error_message="Please enter an integer.",
|
380 |
)
|
|
|
381 |
megatron_lm_config[prefix + "recompute_activations"] = _ask_field(
|
382 |
"Do you want to enable selective activation recomputation? [YES/no]: ",
|
383 |
_convert_yes_no_to_bool,
|
384 |
default=True,
|
385 |
error_message="Please enter yes or no.",
|
386 |
)
|
|
|
387 |
megatron_lm_config[prefix + "use_distributed_optimizer"] = _ask_field(
|
388 |
"Do you want to use distributed optimizer "
|
389 |
"which shards optimizer state and gradients across data parallel ranks? [YES/no]: ",
|
|
|
391 |
default=True,
|
392 |
error_message="Please enter yes or no.",
|
393 |
)
|
|
|
394 |
megatron_lm_config[prefix + "gradient_clipping"] = _ask_field(
|
395 |
"What is the gradient clipping value based on global L2 Norm (0 to disable)? [1.0]: ",
|
396 |
float,
|
|
|
406 |
tpu_zone = None
|
407 |
tpu_use_sudo = False
|
408 |
tpu_use_cluster = False
|
|
|
409 |
if distributed_type in [
|
410 |
DistributedType.MULTI_CPU,
|
411 |
DistributedType.MULTI_XPU,
|
|
|
433 |
)
|
434 |
else:
|
435 |
num_processes = 1
|
|
|
436 |
if (distributed_type == DistributedType.MULTI_GPU) and (num_machines == 1) and (num_processes == 1):
|
437 |
raise ValueError(
|
438 |
f"Specified distributed type {distributed_type} but only using 1 GPU on a single machine. Please select `No distributed training` for the type of machine you are using."
|
439 |
)
|
|
|
440 |
if (
|
441 |
distributed_type
|
442 |
in [
|
|
|
456 |
f"What {machine_type} (by id) should be used for training on this machine as a comma-seperated list? [all]:",
|
457 |
default="all",
|
458 |
)
|
|
|
459 |
if distributed_type == DistributedType.TPU:
|
460 |
mixed_precision = "no"
|
461 |
main_training_function = _ask_field(
|
|
|
530 |
"What environment variables do you wish to set in each pod, seperated by a comma: ",
|
531 |
default="",
|
532 |
).split(",")
|
|
|
533 |
else:
|
534 |
main_training_function = "main"
|
535 |
if distributed_type == DistributedType.DEEPSPEED and use_deepspeed_config:
|
|
|
540 |
["no", "fp16", "bf16", "fp8"],
|
541 |
_convert_mixed_precision,
|
542 |
)
|
|
|
543 |
if use_dynamo and mixed_precision == "no" and not use_cpu:
|
544 |
print(
|
545 |
"Torch dynamo used without mixed precision requires TF32 to be efficient. Accelerate will enable it by default when launching your scripts."
|
546 |
)
|
|
|
547 |
if distributed_type == DistributedType.TPU and mixed_precision == "bf16":
|
548 |
tpu_downcast_bf16 = _ask_field(
|
549 |
"Should `torch.float` be cast as `bfloat16` and `torch.double` remain `float32` on TPUs?", default="no"
|
550 |
)
|
|
|
551 |
return ClusterConfig(
|
552 |
compute_environment=ComputeEnvironment.LOCAL_MACHINE,
|
553 |
distributed_type=distributed_type,
|
src/commands/config/config.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
description = "Launches a series of prompts to create and save a `default_config.yaml` configuration file for your training system. Should always be ran first on your machine"
|
3 |
-
|
4 |
-
|
5 |
def get_user_input():
|
6 |
compute_environment = _ask_options(
|
7 |
"In which compute environment are you running?",
|
@@ -13,14 +11,11 @@ def get_user_input():
|
|
13 |
else:
|
14 |
config = get_cluster_input()
|
15 |
return config
|
16 |
-
|
17 |
-
|
18 |
def config_command_parser(subparsers=None):
|
19 |
if subparsers is not None:
|
20 |
parser = subparsers.add_parser("config", description=description)
|
21 |
else:
|
22 |
parser = argparse.ArgumentParser("Accelerate config command", description=description)
|
23 |
-
|
24 |
parser.add_argument(
|
25 |
"--config_file",
|
26 |
default=None,
|
@@ -31,12 +26,9 @@ def config_command_parser(subparsers=None):
|
|
31 |
"with 'huggingface'."
|
32 |
),
|
33 |
)
|
34 |
-
|
35 |
if subparsers is not None:
|
36 |
parser.set_defaults(func=config_command)
|
37 |
return parser
|
38 |
-
|
39 |
-
|
40 |
def config_command(args):
|
41 |
config = get_user_input()
|
42 |
if args.config_file is not None:
|
@@ -45,19 +37,14 @@ def config_command(args):
|
|
45 |
if not os.path.isdir(cache_dir):
|
46 |
os.makedirs(cache_dir)
|
47 |
config_file = default_yaml_config_file
|
48 |
-
|
49 |
if config_file.endswith(".json"):
|
50 |
config.to_json_file(config_file)
|
51 |
else:
|
52 |
config.to_yaml_file(config_file)
|
53 |
print(f"accelerate configuration saved at {config_file}")
|
54 |
-
|
55 |
-
|
56 |
def main():
|
57 |
parser = config_command_parser()
|
58 |
args = parser.parse_args()
|
59 |
config_command(args)
|
60 |
-
|
61 |
-
|
62 |
if __name__ == "__main__":
|
63 |
main()
|
|
|
1 |
#!/usr/bin/env python
|
2 |
description = "Launches a series of prompts to create and save a `default_config.yaml` configuration file for your training system. Should always be ran first on your machine"
|
|
|
|
|
3 |
def get_user_input():
|
4 |
compute_environment = _ask_options(
|
5 |
"In which compute environment are you running?",
|
|
|
11 |
else:
|
12 |
config = get_cluster_input()
|
13 |
return config
|
|
|
|
|
14 |
def config_command_parser(subparsers=None):
|
15 |
if subparsers is not None:
|
16 |
parser = subparsers.add_parser("config", description=description)
|
17 |
else:
|
18 |
parser = argparse.ArgumentParser("Accelerate config command", description=description)
|
|
|
19 |
parser.add_argument(
|
20 |
"--config_file",
|
21 |
default=None,
|
|
|
26 |
"with 'huggingface'."
|
27 |
),
|
28 |
)
|
|
|
29 |
if subparsers is not None:
|
30 |
parser.set_defaults(func=config_command)
|
31 |
return parser
|
|
|
|
|
32 |
def config_command(args):
|
33 |
config = get_user_input()
|
34 |
if args.config_file is not None:
|
|
|
37 |
if not os.path.isdir(cache_dir):
|
38 |
os.makedirs(cache_dir)
|
39 |
config_file = default_yaml_config_file
|
|
|
40 |
if config_file.endswith(".json"):
|
41 |
config.to_json_file(config_file)
|
42 |
else:
|
43 |
config.to_yaml_file(config_file)
|
44 |
print(f"accelerate configuration saved at {config_file}")
|
|
|
|
|
45 |
def main():
|
46 |
parser = config_command_parser()
|
47 |
args = parser.parse_args()
|
48 |
config_command(args)
|
|
|
|
|
49 |
if __name__ == "__main__":
|
50 |
main()
|
src/commands/config/config_args.py
CHANGED
@@ -5,14 +5,11 @@ hf_cache_home = os.path.expanduser(
|
|
5 |
cache_dir = os.path.join(hf_cache_home, "accelerate")
|
6 |
default_json_config_file = os.path.join(cache_dir, "default_config.yaml")
|
7 |
default_yaml_config_file = os.path.join(cache_dir, "default_config.yaml")
|
8 |
-
|
9 |
# For backward compatibility: the default config is the json one if it's the only existing file.
|
10 |
if os.path.isfile(default_yaml_config_file) or not os.path.isfile(default_json_config_file):
|
11 |
default_config_file = default_yaml_config_file
|
12 |
else:
|
13 |
default_config_file = default_json_config_file
|
14 |
-
|
15 |
-
|
16 |
def load_config_from_file(config_file):
|
17 |
if config_file is not None:
|
18 |
if not os.path.isfile(config_file):
|
@@ -43,8 +40,6 @@ def load_config_from_file(config_file):
|
|
43 |
else:
|
44 |
config_class = SageMakerConfig
|
45 |
return config_class.from_yaml_file(yaml_file=config_file)
|
46 |
-
|
47 |
-
|
48 |
@dataclass
|
49 |
class BaseConfig:
|
50 |
compute_environment: ComputeEnvironment
|
@@ -52,7 +47,6 @@ class BaseConfig:
|
|
52 |
mixed_precision: str
|
53 |
use_cpu: bool
|
54 |
debug: bool
|
55 |
-
|
56 |
def to_dict(self):
|
57 |
result = self.__dict__
|
58 |
# For serialization, it's best to convert Enums to strings (or their underlying value type).
|
@@ -63,7 +57,6 @@ class BaseConfig:
|
|
63 |
result[key] = None
|
64 |
result = {k: v for k, v in result.items() if v is not None}
|
65 |
return result
|
66 |
-
|
67 |
@classmethod
|
68 |
def from_json_file(cls, json_file=None):
|
69 |
json_file = default_json_config_file if json_file is None else json_file
|
@@ -88,14 +81,11 @@ class BaseConfig:
|
|
88 |
f"The config file at {json_file} had unknown keys ({extra_keys}), please try upgrading your `accelerate`"
|
89 |
" version or fix (and potentially remove) these keys from your config file."
|
90 |
)
|
91 |
-
|
92 |
return cls(**config_dict)
|
93 |
-
|
94 |
def to_json_file(self, json_file):
|
95 |
with open(json_file, "w", encoding="utf-8") as f:
|
96 |
content = json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
|
97 |
f.write(content)
|
98 |
-
|
99 |
@classmethod
|
100 |
def from_yaml_file(cls, yaml_file=None):
|
101 |
yaml_file = default_yaml_config_file if yaml_file is None else yaml_file
|
@@ -123,11 +113,9 @@ class BaseConfig:
|
|
123 |
" version or fix (and potentially remove) these keys from your config file."
|
124 |
)
|
125 |
return cls(**config_dict)
|
126 |
-
|
127 |
def to_yaml_file(self, yaml_file):
|
128 |
with open(yaml_file, "w", encoding="utf-8") as f:
|
129 |
yaml.safe_dump(self.to_dict(), f)
|
130 |
-
|
131 |
def __post_init__(self):
|
132 |
if isinstance(self.compute_environment, str):
|
133 |
self.compute_environment = ComputeEnvironment(self.compute_environment)
|
@@ -138,8 +126,6 @@ class BaseConfig:
|
|
138 |
self.distributed_type = DistributedType(self.distributed_type)
|
139 |
if self.dynamo_config is None:
|
140 |
self.dynamo_config = {}
|
141 |
-
|
142 |
-
|
143 |
@dataclass
|
144 |
class ClusterConfig(BaseConfig):
|
145 |
num_processes: int
|
@@ -151,7 +137,6 @@ class ClusterConfig(BaseConfig):
|
|
151 |
rdzv_backend: Optional[str] = "static"
|
152 |
same_network: Optional[bool] = False
|
153 |
main_training_function: str = "main"
|
154 |
-
|
155 |
# args for deepspeed_plugin
|
156 |
deepspeed_config: dict = None
|
157 |
# args for fsdp
|
@@ -162,7 +147,6 @@ class ClusterConfig(BaseConfig):
|
|
162 |
ipex_config: dict = None
|
163 |
# args for TPU
|
164 |
downcast_bf16: bool = False
|
165 |
-
|
166 |
# args for TPU pods
|
167 |
tpu_name: str = None
|
168 |
tpu_zone: str = None
|
@@ -172,10 +156,8 @@ class ClusterConfig(BaseConfig):
|
|
172 |
commands: List[str] = None
|
173 |
tpu_vm: List[str] = None
|
174 |
tpu_env: List[str] = None
|
175 |
-
|
176 |
# args for dynamo
|
177 |
dynamo_config: dict = None
|
178 |
-
|
179 |
def __post_init__(self):
|
180 |
if self.deepspeed_config is None:
|
181 |
self.deepspeed_config = {}
|
@@ -186,8 +168,6 @@ class ClusterConfig(BaseConfig):
|
|
186 |
if self.ipex_config is None:
|
187 |
self.ipex_config = {}
|
188 |
return super().__post_init__()
|
189 |
-
|
190 |
-
|
191 |
@dataclass
|
192 |
class SageMakerConfig(BaseConfig):
|
193 |
ec2_instance_type: str
|
|
|
5 |
cache_dir = os.path.join(hf_cache_home, "accelerate")
|
6 |
default_json_config_file = os.path.join(cache_dir, "default_config.yaml")
|
7 |
default_yaml_config_file = os.path.join(cache_dir, "default_config.yaml")
|
|
|
8 |
# For backward compatibility: the default config is the json one if it's the only existing file.
|
9 |
if os.path.isfile(default_yaml_config_file) or not os.path.isfile(default_json_config_file):
|
10 |
default_config_file = default_yaml_config_file
|
11 |
else:
|
12 |
default_config_file = default_json_config_file
|
|
|
|
|
13 |
def load_config_from_file(config_file):
|
14 |
if config_file is not None:
|
15 |
if not os.path.isfile(config_file):
|
|
|
40 |
else:
|
41 |
config_class = SageMakerConfig
|
42 |
return config_class.from_yaml_file(yaml_file=config_file)
|
|
|
|
|
43 |
@dataclass
|
44 |
class BaseConfig:
|
45 |
compute_environment: ComputeEnvironment
|
|
|
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).
|
|
|
57 |
result[key] = None
|
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
|
|
|
81 |
f"The config file at {json_file} had unknown keys ({extra_keys}), please try upgrading your `accelerate`"
|
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
|
|
|
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)
|
|
|
126 |
self.distributed_type = DistributedType(self.distributed_type)
|
127 |
if self.dynamo_config is None:
|
128 |
self.dynamo_config = {}
|
|
|
|
|
129 |
@dataclass
|
130 |
class ClusterConfig(BaseConfig):
|
131 |
num_processes: int
|
|
|
137 |
rdzv_backend: Optional[str] = "static"
|
138 |
same_network: Optional[bool] = False
|
139 |
main_training_function: str = "main"
|
|
|
140 |
# args for deepspeed_plugin
|
141 |
deepspeed_config: dict = None
|
142 |
# args for fsdp
|
|
|
147 |
ipex_config: dict = None
|
148 |
# args for TPU
|
149 |
downcast_bf16: bool = False
|
|
|
150 |
# args for TPU pods
|
151 |
tpu_name: str = None
|
152 |
tpu_zone: str = None
|
|
|
156 |
commands: List[str] = None
|
157 |
tpu_vm: List[str] = None
|
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 = {}
|
|
|
168 |
if self.ipex_config is None:
|
169 |
self.ipex_config = {}
|
170 |
return super().__post_init__()
|
|
|
|
|
171 |
@dataclass
|
172 |
class SageMakerConfig(BaseConfig):
|
173 |
ec2_instance_type: str
|
src/commands/config/config_utils.py
CHANGED
@@ -13,8 +13,6 @@ DYNAMO_BACKENDS = [
|
|
13 |
"IPEX",
|
14 |
"TVM",
|
15 |
]
|
16 |
-
|
17 |
-
|
18 |
def _ask_field(input_text, convert_value=None, default=None, error_message=None):
|
19 |
ask_again = True
|
20 |
while ask_again:
|
@@ -26,43 +24,27 @@ def _ask_field(input_text, convert_value=None, default=None, error_message=None)
|
|
26 |
except Exception:
|
27 |
if error_message is not None:
|
28 |
print(error_message)
|
29 |
-
|
30 |
-
|
31 |
def _ask_options(input_text, options=[], convert_value=None, default=0):
|
32 |
menu = BulletMenu(input_text, options)
|
33 |
result = menu.run(default_choice=default)
|
34 |
return convert_value(result) if convert_value is not None else result
|
35 |
-
|
36 |
-
|
37 |
def _convert_compute_environment(value):
|
38 |
value = int(value)
|
39 |
return ComputeEnvironment(["LOCAL_MACHINE", "AMAZON_SAGEMAKER"][value])
|
40 |
-
|
41 |
-
|
42 |
def _convert_distributed_mode(value):
|
43 |
value = int(value)
|
44 |
return DistributedType(["NO", "MULTI_CPU", "MULTI_XPU", "MULTI_GPU", "MULTI_NPU", "TPU"][value])
|
45 |
-
|
46 |
-
|
47 |
def _convert_dynamo_backend(value):
|
48 |
value = int(value)
|
49 |
return DynamoBackend(DYNAMO_BACKENDS[value]).value
|
50 |
-
|
51 |
-
|
52 |
def _convert_mixed_precision(value):
|
53 |
value = int(value)
|
54 |
return PrecisionType(["no", "fp16", "bf16", "fp8"][value])
|
55 |
-
|
56 |
-
|
57 |
def _convert_sagemaker_distributed_mode(value):
|
58 |
value = int(value)
|
59 |
return SageMakerDistributedType(["NO", "DATA_PARALLEL", "MODEL_PARALLEL"][value])
|
60 |
-
|
61 |
-
|
62 |
def _convert_yes_no_to_bool(value):
|
63 |
return {"yes": True, "no": False}[value.lower()]
|
64 |
-
|
65 |
-
|
66 |
class SubcommandHelpFormatter(argparse.RawDescriptionHelpFormatter):
|
67 |
"""
|
68 |
A custom formatter that will remove the usage line from the help message for subcommands.
|
|
|
13 |
"IPEX",
|
14 |
"TVM",
|
15 |
]
|
|
|
|
|
16 |
def _ask_field(input_text, convert_value=None, default=None, error_message=None):
|
17 |
ask_again = True
|
18 |
while ask_again:
|
|
|
24 |
except Exception:
|
25 |
if error_message is not None:
|
26 |
print(error_message)
|
|
|
|
|
27 |
def _ask_options(input_text, options=[], convert_value=None, default=0):
|
28 |
menu = BulletMenu(input_text, options)
|
29 |
result = menu.run(default_choice=default)
|
30 |
return convert_value(result) if convert_value is not None else result
|
|
|
|
|
31 |
def _convert_compute_environment(value):
|
32 |
value = int(value)
|
33 |
return ComputeEnvironment(["LOCAL_MACHINE", "AMAZON_SAGEMAKER"][value])
|
|
|
|
|
34 |
def _convert_distributed_mode(value):
|
35 |
value = int(value)
|
36 |
return DistributedType(["NO", "MULTI_CPU", "MULTI_XPU", "MULTI_GPU", "MULTI_NPU", "TPU"][value])
|
|
|
|
|
37 |
def _convert_dynamo_backend(value):
|
38 |
value = int(value)
|
39 |
return DynamoBackend(DYNAMO_BACKENDS[value]).value
|
|
|
|
|
40 |
def _convert_mixed_precision(value):
|
41 |
value = int(value)
|
42 |
return PrecisionType(["no", "fp16", "bf16", "fp8"][value])
|
|
|
|
|
43 |
def _convert_sagemaker_distributed_mode(value):
|
44 |
value = int(value)
|
45 |
return SageMakerDistributedType(["NO", "DATA_PARALLEL", "MODEL_PARALLEL"][value])
|
|
|
|
|
46 |
def _convert_yes_no_to_bool(value):
|
47 |
return {"yes": True, "no": False}[value.lower()]
|
|
|
|
|
48 |
class SubcommandHelpFormatter(argparse.RawDescriptionHelpFormatter):
|
49 |
"""
|
50 |
A custom formatter that will remove the usage line from the help message for subcommands.
|
src/commands/config/default.py
CHANGED
@@ -1,12 +1,9 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
description = "Create a default config file for Accelerate with only a few flags set."
|
3 |
-
|
4 |
-
|
5 |
def write_basic_config(mixed_precision="no", save_location: str = default_json_config_file, use_xpu: bool = False):
|
6 |
"""
|
7 |
Creates and saves a basic cluster config to be used on a local machine with potentially multiple GPUs. Will also
|
8 |
set CPU if it is a CPU-only machine.
|
9 |
-
|
10 |
Args:
|
11 |
mixed_precision (`str`, *optional*, defaults to "no"):
|
12 |
Mixed Precision to use. Should be one of "no", "fp16", or "bf16"
|
@@ -66,8 +63,6 @@ def write_basic_config(mixed_precision="no", save_location: str = default_json_c
|
|
66 |
config = ClusterConfig(**config)
|
67 |
config.to_json_file(path)
|
68 |
return path
|
69 |
-
|
70 |
-
|
71 |
def default_command_parser(parser, parents):
|
72 |
parser = parser.add_parser("default", parents=parents, help=description, formatter_class=SubcommandHelpFormatter)
|
73 |
parser.add_argument(
|
@@ -81,7 +76,6 @@ def default_command_parser(parser, parents):
|
|
81 |
),
|
82 |
dest="save_location",
|
83 |
)
|
84 |
-
|
85 |
parser.add_argument(
|
86 |
"--mixed_precision",
|
87 |
choices=["no", "fp16", "bf16"],
|
@@ -93,8 +87,6 @@ def default_command_parser(parser, parents):
|
|
93 |
)
|
94 |
parser.set_defaults(func=default_config_command)
|
95 |
return parser
|
96 |
-
|
97 |
-
|
98 |
def default_config_command(args):
|
99 |
config_file = write_basic_config(args.mixed_precision, args.save_location)
|
100 |
if config_file:
|
|
|
1 |
#!/usr/bin/env python
|
2 |
description = "Create a default config file for Accelerate with only a few flags set."
|
|
|
|
|
3 |
def write_basic_config(mixed_precision="no", save_location: str = default_json_config_file, use_xpu: bool = False):
|
4 |
"""
|
5 |
Creates and saves a basic cluster config to be used on a local machine with potentially multiple GPUs. Will also
|
6 |
set CPU if it is a CPU-only machine.
|
|
|
7 |
Args:
|
8 |
mixed_precision (`str`, *optional*, defaults to "no"):
|
9 |
Mixed Precision to use. Should be one of "no", "fp16", or "bf16"
|
|
|
63 |
config = ClusterConfig(**config)
|
64 |
config.to_json_file(path)
|
65 |
return path
|
|
|
|
|
66 |
def default_command_parser(parser, parents):
|
67 |
parser = parser.add_parser("default", parents=parents, help=description, formatter_class=SubcommandHelpFormatter)
|
68 |
parser.add_argument(
|
|
|
76 |
),
|
77 |
dest="save_location",
|
78 |
)
|
|
|
79 |
parser.add_argument(
|
80 |
"--mixed_precision",
|
81 |
choices=["no", "fp16", "bf16"],
|
|
|
87 |
)
|
88 |
parser.set_defaults(func=default_config_command)
|
89 |
return parser
|
|
|
|
|
90 |
def default_config_command(args):
|
91 |
config_file = write_basic_config(args.mixed_precision, args.save_location)
|
92 |
if config_file:
|
src/commands/config/sagemaker.py
CHANGED
@@ -1,7 +1,6 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
def _create_iam_role_for_sagemaker(role_name):
|
3 |
iam_client = boto3.client("iam")
|
4 |
-
|
5 |
sagemaker_trust_policy = {
|
6 |
"Version": "2012-10-17",
|
7 |
"Statement": [
|
@@ -51,13 +50,9 @@ def _create_iam_role_for_sagemaker(role_name):
|
|
51 |
)
|
52 |
except iam_client.exceptions.EntityAlreadyExistsException:
|
53 |
print(f"role {role_name} already exists. Using existing one")
|
54 |
-
|
55 |
-
|
56 |
def _get_iam_role_arn(role_name):
|
57 |
iam_client = boto3.client("iam")
|
58 |
return iam_client.get_role(RoleName=role_name)["Role"]["Arn"]
|
59 |
-
|
60 |
-
|
61 |
def get_sagemaker_input():
|
62 |
credentials_configuration = _ask_options(
|
63 |
"How do you want to authorize?",
|
@@ -75,13 +70,10 @@ def get_sagemaker_input():
|
|
75 |
)
|
76 |
aws_access_key_id = _ask_field("AWS Access Key ID: ")
|
77 |
os.environ["AWS_ACCESS_KEY_ID"] = aws_access_key_id
|
78 |
-
|
79 |
aws_secret_access_key = _ask_field("AWS Secret Access Key: ")
|
80 |
os.environ["AWS_SECRET_ACCESS_KEY"] = aws_secret_access_key
|
81 |
-
|
82 |
aws_region = _ask_field("Enter your AWS Region: [us-east-1]", default="us-east-1")
|
83 |
os.environ["AWS_DEFAULT_REGION"] = aws_region
|
84 |
-
|
85 |
role_management = _ask_options(
|
86 |
"Do you already have an IAM Role for executing Amazon SageMaker Training Jobs?",
|
87 |
["Provide IAM Role name", "Create new IAM role using credentials"],
|
@@ -93,7 +85,6 @@ def get_sagemaker_input():
|
|
93 |
iam_role_name = "accelerate_sagemaker_execution_role"
|
94 |
print(f'Accelerate will create an iam role "{iam_role_name}" using the provided credentials')
|
95 |
_create_iam_role_for_sagemaker(iam_role_name)
|
96 |
-
|
97 |
is_custom_docker_image = _ask_field(
|
98 |
"Do you want to use custom Docker image? [yes/NO]: ",
|
99 |
_convert_yes_no_to_bool,
|
@@ -103,7 +94,6 @@ def get_sagemaker_input():
|
|
103 |
docker_image = None
|
104 |
if is_custom_docker_image:
|
105 |
docker_image = _ask_field("Enter your Docker image: ", lambda x: str(x).lower())
|
106 |
-
|
107 |
is_sagemaker_inputs_enabled = _ask_field(
|
108 |
"Do you want to provide SageMaker input channels with data locations? [yes/NO]: ",
|
109 |
_convert_yes_no_to_bool,
|
@@ -116,7 +106,6 @@ def get_sagemaker_input():
|
|
116 |
"Enter the path to the SageMaker inputs TSV file with columns (channel_name, data_location): ",
|
117 |
lambda x: str(x).lower(),
|
118 |
)
|
119 |
-
|
120 |
is_sagemaker_metrics_enabled = _ask_field(
|
121 |
"Do you want to enable SageMaker metrics? [yes/NO]: ",
|
122 |
_convert_yes_no_to_bool,
|
@@ -129,7 +118,6 @@ def get_sagemaker_input():
|
|
129 |
"Enter the path to the SageMaker metrics TSV file with columns (metric_name, metric_regex): ",
|
130 |
lambda x: str(x).lower(),
|
131 |
)
|
132 |
-
|
133 |
distributed_type = _ask_options(
|
134 |
"What is the distributed mode?",
|
135 |
["No distributed training", "Data parallelism"],
|
@@ -156,7 +144,6 @@ def get_sagemaker_input():
|
|
156 |
default=False,
|
157 |
error_message="Please enter yes or no.",
|
158 |
)
|
159 |
-
|
160 |
if use_custom_options:
|
161 |
dynamo_config[prefix + "mode"] = _ask_options(
|
162 |
"Which mode do you want to use?",
|
@@ -184,7 +171,6 @@ def get_sagemaker_input():
|
|
184 |
else:
|
185 |
ec2_instance_query += "? [ml.p3.2xlarge]:"
|
186 |
ec2_instance_type = _ask_field(ec2_instance_query, lambda x: str(x).lower(), default="ml.p3.2xlarge")
|
187 |
-
|
188 |
debug = False
|
189 |
if distributed_type != SageMakerDistributedType.NO:
|
190 |
debug = _ask_field(
|
@@ -193,7 +179,6 @@ def get_sagemaker_input():
|
|
193 |
default=False,
|
194 |
error_message="Please enter yes or no.",
|
195 |
)
|
196 |
-
|
197 |
num_machines = 1
|
198 |
if distributed_type in (SageMakerDistributedType.DATA_PARALLEL, SageMakerDistributedType.MODEL_PARALLEL):
|
199 |
num_machines = _ask_field(
|
@@ -201,18 +186,15 @@ def get_sagemaker_input():
|
|
201 |
int,
|
202 |
default=1,
|
203 |
)
|
204 |
-
|
205 |
mixed_precision = _ask_options(
|
206 |
"Do you wish to use FP16 or BF16 (mixed precision)?",
|
207 |
["no", "fp16", "bf16", "fp8"],
|
208 |
_convert_mixed_precision,
|
209 |
)
|
210 |
-
|
211 |
if use_dynamo and mixed_precision == "no":
|
212 |
print(
|
213 |
"Torch dynamo used without mixed precision requires TF32 to be efficient. Accelerate will enable it by default when launching your scripts."
|
214 |
)
|
215 |
-
|
216 |
return SageMakerConfig(
|
217 |
image_uri=docker_image,
|
218 |
compute_environment=ComputeEnvironment.AMAZON_SAGEMAKER,
|
|
|
1 |
#!/usr/bin/env python
|
2 |
def _create_iam_role_for_sagemaker(role_name):
|
3 |
iam_client = boto3.client("iam")
|
|
|
4 |
sagemaker_trust_policy = {
|
5 |
"Version": "2012-10-17",
|
6 |
"Statement": [
|
|
|
50 |
)
|
51 |
except iam_client.exceptions.EntityAlreadyExistsException:
|
52 |
print(f"role {role_name} already exists. Using existing one")
|
|
|
|
|
53 |
def _get_iam_role_arn(role_name):
|
54 |
iam_client = boto3.client("iam")
|
55 |
return iam_client.get_role(RoleName=role_name)["Role"]["Arn"]
|
|
|
|
|
56 |
def get_sagemaker_input():
|
57 |
credentials_configuration = _ask_options(
|
58 |
"How do you want to authorize?",
|
|
|
70 |
)
|
71 |
aws_access_key_id = _ask_field("AWS Access Key ID: ")
|
72 |
os.environ["AWS_ACCESS_KEY_ID"] = aws_access_key_id
|
|
|
73 |
aws_secret_access_key = _ask_field("AWS Secret Access Key: ")
|
74 |
os.environ["AWS_SECRET_ACCESS_KEY"] = aws_secret_access_key
|
|
|
75 |
aws_region = _ask_field("Enter your AWS Region: [us-east-1]", default="us-east-1")
|
76 |
os.environ["AWS_DEFAULT_REGION"] = aws_region
|
|
|
77 |
role_management = _ask_options(
|
78 |
"Do you already have an IAM Role for executing Amazon SageMaker Training Jobs?",
|
79 |
["Provide IAM Role name", "Create new IAM role using credentials"],
|
|
|
85 |
iam_role_name = "accelerate_sagemaker_execution_role"
|
86 |
print(f'Accelerate will create an iam role "{iam_role_name}" using the provided credentials')
|
87 |
_create_iam_role_for_sagemaker(iam_role_name)
|
|
|
88 |
is_custom_docker_image = _ask_field(
|
89 |
"Do you want to use custom Docker image? [yes/NO]: ",
|
90 |
_convert_yes_no_to_bool,
|
|
|
94 |
docker_image = None
|
95 |
if is_custom_docker_image:
|
96 |
docker_image = _ask_field("Enter your Docker image: ", lambda x: str(x).lower())
|
|
|
97 |
is_sagemaker_inputs_enabled = _ask_field(
|
98 |
"Do you want to provide SageMaker input channels with data locations? [yes/NO]: ",
|
99 |
_convert_yes_no_to_bool,
|
|
|
106 |
"Enter the path to the SageMaker inputs TSV file with columns (channel_name, data_location): ",
|
107 |
lambda x: str(x).lower(),
|
108 |
)
|
|
|
109 |
is_sagemaker_metrics_enabled = _ask_field(
|
110 |
"Do you want to enable SageMaker metrics? [yes/NO]: ",
|
111 |
_convert_yes_no_to_bool,
|
|
|
118 |
"Enter the path to the SageMaker metrics TSV file with columns (metric_name, metric_regex): ",
|
119 |
lambda x: str(x).lower(),
|
120 |
)
|
|
|
121 |
distributed_type = _ask_options(
|
122 |
"What is the distributed mode?",
|
123 |
["No distributed training", "Data parallelism"],
|
|
|
144 |
default=False,
|
145 |
error_message="Please enter yes or no.",
|
146 |
)
|
|
|
147 |
if use_custom_options:
|
148 |
dynamo_config[prefix + "mode"] = _ask_options(
|
149 |
"Which mode do you want to use?",
|
|
|
171 |
else:
|
172 |
ec2_instance_query += "? [ml.p3.2xlarge]:"
|
173 |
ec2_instance_type = _ask_field(ec2_instance_query, lambda x: str(x).lower(), default="ml.p3.2xlarge")
|
|
|
174 |
debug = False
|
175 |
if distributed_type != SageMakerDistributedType.NO:
|
176 |
debug = _ask_field(
|
|
|
179 |
default=False,
|
180 |
error_message="Please enter yes or no.",
|
181 |
)
|
|
|
182 |
num_machines = 1
|
183 |
if distributed_type in (SageMakerDistributedType.DATA_PARALLEL, SageMakerDistributedType.MODEL_PARALLEL):
|
184 |
num_machines = _ask_field(
|
|
|
186 |
int,
|
187 |
default=1,
|
188 |
)
|
|
|
189 |
mixed_precision = _ask_options(
|
190 |
"Do you wish to use FP16 or BF16 (mixed precision)?",
|
191 |
["no", "fp16", "bf16", "fp8"],
|
192 |
_convert_mixed_precision,
|
193 |
)
|
|
|
194 |
if use_dynamo and mixed_precision == "no":
|
195 |
print(
|
196 |
"Torch dynamo used without mixed precision requires TF32 to be efficient. Accelerate will enable it by default when launching your scripts."
|
197 |
)
|
|
|
198 |
return SageMakerConfig(
|
199 |
image_uri=docker_image,
|
200 |
compute_environment=ComputeEnvironment.AMAZON_SAGEMAKER,
|
src/commands/config/update.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
description = "Update an existing config file with the latest defaults while maintaining the old configuration."
|
3 |
-
|
4 |
-
|
5 |
def update_config(args):
|
6 |
"""
|
7 |
Update an existing config file with the latest defaults while maintaining the old configuration.
|
@@ -12,14 +10,11 @@ def update_config(args):
|
|
12 |
elif not Path(config_file).exists():
|
13 |
raise ValueError(f"The passed config file located at {config_file} doesn't exist.")
|
14 |
config = load_config_from_file(config_file)
|
15 |
-
|
16 |
if config_file.endswith(".json"):
|
17 |
config.to_json_file(config_file)
|
18 |
else:
|
19 |
config.to_yaml_file(config_file)
|
20 |
return config_file
|
21 |
-
|
22 |
-
|
23 |
def update_command_parser(parser, parents):
|
24 |
parser = parser.add_parser("update", parents=parents, help=description, formatter_class=SubcommandHelpFormatter)
|
25 |
parser.add_argument(
|
@@ -32,11 +27,8 @@ def update_command_parser(parser, parents):
|
|
32 |
"with 'huggingface'."
|
33 |
),
|
34 |
)
|
35 |
-
|
36 |
parser.set_defaults(func=update_config_command)
|
37 |
return parser
|
38 |
-
|
39 |
-
|
40 |
def update_config_command(args):
|
41 |
config_file = update_config(args)
|
42 |
print(f"Sucessfully updated the configuration file at {config_file}.")
|
|
|
1 |
#!/usr/bin/env python
|
2 |
description = "Update an existing config file with the latest defaults while maintaining the old configuration."
|
|
|
|
|
3 |
def update_config(args):
|
4 |
"""
|
5 |
Update an existing config file with the latest defaults while maintaining the old configuration.
|
|
|
10 |
elif not Path(config_file).exists():
|
11 |
raise ValueError(f"The passed config file located at {config_file} doesn't exist.")
|
12 |
config = load_config_from_file(config_file)
|
|
|
13 |
if config_file.endswith(".json"):
|
14 |
config.to_json_file(config_file)
|
15 |
else:
|
16 |
config.to_yaml_file(config_file)
|
17 |
return config_file
|
|
|
|
|
18 |
def update_command_parser(parser, parents):
|
19 |
parser = parser.add_parser("update", parents=parents, help=description, formatter_class=SubcommandHelpFormatter)
|
20 |
parser.add_argument(
|
|
|
27 |
"with 'huggingface'."
|
28 |
),
|
29 |
)
|
|
|
30 |
parser.set_defaults(func=update_config_command)
|
31 |
return parser
|
|
|
|
|
32 |
def update_config_command(args):
|
33 |
config_file = update_config(args)
|
34 |
print(f"Sucessfully updated the configuration file at {config_file}.")
|
src/commands/env.py
CHANGED
@@ -4,27 +4,21 @@ def env_command_parser(subparsers=None):
|
|
4 |
parser = subparsers.add_parser("env")
|
5 |
else:
|
6 |
parser = argparse.ArgumentParser("Accelerate env command")
|
7 |
-
|
8 |
parser.add_argument(
|
9 |
"--config_file", default=None, help="The config file to use for the default values in the launching script."
|
10 |
)
|
11 |
-
|
12 |
if subparsers is not None:
|
13 |
parser.set_defaults(func=env_command)
|
14 |
return parser
|
15 |
-
|
16 |
-
|
17 |
def env_command(args):
|
18 |
pt_version = torch.__version__
|
19 |
pt_cuda_available = torch.cuda.is_available()
|
20 |
pt_xpu_available = is_xpu_available()
|
21 |
pt_npu_available = is_npu_available()
|
22 |
-
|
23 |
accelerate_config = "Not found"
|
24 |
# Get the default from the config file.
|
25 |
if args.config_file is not None or os.path.isfile(default_config_file):
|
26 |
accelerate_config = load_config_from_file(args.config_file).to_dict()
|
27 |
-
|
28 |
info = {
|
29 |
"`Accelerate` version": version,
|
30 |
"Platform": platform.platform(),
|
@@ -37,10 +31,8 @@ def env_command(args):
|
|
37 |
}
|
38 |
if pt_cuda_available:
|
39 |
info["GPU type"] = torch.cuda.get_device_name()
|
40 |
-
|
41 |
print("\nCopy-and-paste the text below in your GitHub issue\n")
|
42 |
print("\n".join([f"- {prop}: {val}" for prop, val in info.items()]))
|
43 |
-
|
44 |
print("- `Accelerate` default config:" if args.config_file is None else "- `Accelerate` config passed:")
|
45 |
accelerate_config_str = (
|
46 |
"\n".join([f"\t- {prop}: {val}" for prop, val in accelerate_config.items()])
|
@@ -48,18 +40,12 @@ def env_command(args):
|
|
48 |
else f"\t{accelerate_config}"
|
49 |
)
|
50 |
print(accelerate_config_str)
|
51 |
-
|
52 |
info["`Accelerate` configs"] = accelerate_config
|
53 |
-
|
54 |
return info
|
55 |
-
|
56 |
-
|
57 |
def main() -> int:
|
58 |
parser = env_command_parser()
|
59 |
args = parser.parse_args()
|
60 |
env_command(args)
|
61 |
return 0
|
62 |
-
|
63 |
-
|
64 |
if __name__ == "__main__":
|
65 |
raise SystemExit(main())
|
|
|
4 |
parser = subparsers.add_parser("env")
|
5 |
else:
|
6 |
parser = argparse.ArgumentParser("Accelerate env command")
|
|
|
7 |
parser.add_argument(
|
8 |
"--config_file", default=None, help="The config file to use for the default values in the launching script."
|
9 |
)
|
|
|
10 |
if subparsers is not None:
|
11 |
parser.set_defaults(func=env_command)
|
12 |
return parser
|
|
|
|
|
13 |
def env_command(args):
|
14 |
pt_version = torch.__version__
|
15 |
pt_cuda_available = torch.cuda.is_available()
|
16 |
pt_xpu_available = is_xpu_available()
|
17 |
pt_npu_available = is_npu_available()
|
|
|
18 |
accelerate_config = "Not found"
|
19 |
# Get the default from the config file.
|
20 |
if args.config_file is not None or os.path.isfile(default_config_file):
|
21 |
accelerate_config = load_config_from_file(args.config_file).to_dict()
|
|
|
22 |
info = {
|
23 |
"`Accelerate` version": version,
|
24 |
"Platform": platform.platform(),
|
|
|
31 |
}
|
32 |
if pt_cuda_available:
|
33 |
info["GPU type"] = torch.cuda.get_device_name()
|
|
|
34 |
print("\nCopy-and-paste the text below in your GitHub issue\n")
|
35 |
print("\n".join([f"- {prop}: {val}" for prop, val in info.items()]))
|
|
|
36 |
print("- `Accelerate` default config:" if args.config_file is None else "- `Accelerate` config passed:")
|
37 |
accelerate_config_str = (
|
38 |
"\n".join([f"\t- {prop}: {val}" for prop, val in accelerate_config.items()])
|
|
|
40 |
else f"\t{accelerate_config}"
|
41 |
)
|
42 |
print(accelerate_config_str)
|
|
|
43 |
info["`Accelerate` configs"] = accelerate_config
|
|
|
44 |
return info
|
|
|
|
|
45 |
def main() -> int:
|
46 |
parser = env_command_parser()
|
47 |
args = parser.parse_args()
|
48 |
env_command(args)
|
49 |
return 0
|
|
|
|
|
50 |
if __name__ == "__main__":
|
51 |
raise SystemExit(main())
|
src/commands/estimate.py
CHANGED
@@ -7,8 +7,6 @@ def verify_on_hub(repo: str, token: str = None):
|
|
7 |
return "gated"
|
8 |
except RepositoryNotFoundError:
|
9 |
return "repo"
|
10 |
-
|
11 |
-
|
12 |
def check_has_model(error):
|
13 |
"""
|
14 |
Checks what library spawned `error` when a model is not found
|
@@ -23,12 +21,9 @@ def check_has_model(error):
|
|
23 |
return "transformers"
|
24 |
else:
|
25 |
return "unknown"
|
26 |
-
|
27 |
-
|
28 |
def create_empty_model(model_name: str, library_name: str, trust_remote_code: bool = False, access_token: str = None):
|
29 |
"""
|
30 |
Creates an empty model from its parent library on the `Hub` to calculate the overall memory consumption.
|
31 |
-
|
32 |
Args:
|
33 |
model_name (`str`):
|
34 |
The model name on the Hub
|
@@ -41,10 +36,8 @@ def create_empty_model(model_name: str, library_name: str, trust_remote_code: bo
|
|
41 |
execute code present on the Hub on your local machine.
|
42 |
access_token (`str`, `optional`, defaults to `None`):
|
43 |
The access token to use to access private or gated models on the Hub. (for use on the Gradio app)
|
44 |
-
|
45 |
Returns:
|
46 |
`torch.nn.Module`: The torch model that has been initialized on the `meta` device.
|
47 |
-
|
48 |
"""
|
49 |
model_info = verify_on_hub(model_name, access_token)
|
50 |
# Simplified errors
|
@@ -69,10 +62,8 @@ def create_empty_model(model_name: str, library_name: str, trust_remote_code: bo
|
|
69 |
f"To check `{model_name}`, `transformers` must be installed. Please install it via `pip install transformers`"
|
70 |
)
|
71 |
print(f"Loading pretrained config for `{model_name}` from `transformers`...")
|
72 |
-
|
73 |
auto_map = model_info.config.get("auto_map", False)
|
74 |
config = AutoConfig.from_pretrained(model_name, trust_remote_code=trust_remote_code)
|
75 |
-
|
76 |
with init_empty_weights():
|
77 |
# remote code could specify a specific `AutoModel` class in the `auto_map`
|
78 |
constructor = AutoModel
|
@@ -98,8 +89,6 @@ def create_empty_model(model_name: str, library_name: str, trust_remote_code: bo
|
|
98 |
f"Library `{library_name}` is not supported yet, please open an issue on GitHub for us to add support."
|
99 |
)
|
100 |
return model
|
101 |
-
|
102 |
-
|
103 |
def create_ascii_table(headers: list, rows: list, title: str):
|
104 |
"Creates a pretty table from a list of rows, minimal version of `tabulate`."
|
105 |
sep_char, in_between = "│", "─"
|
@@ -108,20 +97,15 @@ def create_ascii_table(headers: list, rows: list, title: str):
|
|
108 |
column_values = [row[i] for row in rows] + [headers[i]]
|
109 |
max_column_width = max(len(value) for value in column_values)
|
110 |
column_widths.append(max_column_width)
|
111 |
-
|
112 |
formats = [f"%{column_widths[i]}s" for i in range(len(rows[0]))]
|
113 |
-
|
114 |
pattern = f"{sep_char}{sep_char.join(formats)}{sep_char}"
|
115 |
diff = 0
|
116 |
-
|
117 |
def make_row(left_char, middle_char, right_char):
|
118 |
return f"{left_char}{middle_char.join([in_between * n for n in column_widths])}{in_between * diff}{right_char}"
|
119 |
-
|
120 |
separator = make_row("├", "┼", "┤")
|
121 |
if len(title) > sum(column_widths):
|
122 |
diff = abs(len(title) - len(separator))
|
123 |
column_widths[-1] += diff
|
124 |
-
|
125 |
# Update with diff
|
126 |
separator = make_row("├", "┼", "┤")
|
127 |
initial_rows = [
|
@@ -137,16 +121,12 @@ def create_ascii_table(headers: list, rows: list, title: str):
|
|
137 |
centered_line = [t.center(column_widths[i]) for i, t in enumerate(line)]
|
138 |
table += f"{pattern % tuple(centered_line)}\n"
|
139 |
table += f'└{"┴".join([in_between * n for n in column_widths])}┘'
|
140 |
-
|
141 |
return table
|
142 |
-
|
143 |
-
|
144 |
def estimate_command_parser(subparsers=None):
|
145 |
if subparsers is not None:
|
146 |
parser = subparsers.add_parser("estimate-memory")
|
147 |
else:
|
148 |
parser = argparse.ArgumentParser(description="Model size estimator for fitting a model onto CUDA memory.")
|
149 |
-
|
150 |
parser.add_argument("model_name", type=str, help="The model name on the Hugging Face Hub.")
|
151 |
parser.add_argument(
|
152 |
"--library_name",
|
@@ -169,12 +149,9 @@ def estimate_command_parser(subparsers=None):
|
|
169 |
should only be used for repositories you trust and in which you have read the code, as it will execute
|
170 |
code present on the Hub on your local machine.""",
|
171 |
)
|
172 |
-
|
173 |
if subparsers is not None:
|
174 |
parser.set_defaults(func=estimate_command)
|
175 |
return parser
|
176 |
-
|
177 |
-
|
178 |
def gather_data(args):
|
179 |
"Creates an empty model and gathers the data for the sizes"
|
180 |
try:
|
@@ -188,11 +165,8 @@ def gather_data(args):
|
|
188 |
f"Tried to load `{args.model_name}` with `{library}` but a possible model to load was not found inside the repo."
|
189 |
)
|
190 |
raise e
|
191 |
-
|
192 |
total_size, largest_layer = calculate_maximum_sizes(model)
|
193 |
-
|
194 |
data = []
|
195 |
-
|
196 |
for dtype in args.dtypes:
|
197 |
dtype_total_size = total_size
|
198 |
dtype_largest_layer = largest_layer[0]
|
@@ -208,27 +182,19 @@ def gather_data(args):
|
|
208 |
dtype_training_size = dtype_total_size * 4
|
209 |
data.append([dtype, dtype_largest_layer, dtype_total_size, dtype_training_size])
|
210 |
return data
|
211 |
-
|
212 |
-
|
213 |
def estimate_command(args):
|
214 |
data = gather_data(args)
|
215 |
for row in data:
|
216 |
for i, item in enumerate(row):
|
217 |
if isinstance(item, (int, float)):
|
218 |
row[i] = convert_bytes(item)
|
219 |
-
|
220 |
headers = ["dtype", "Largest Layer", "Total Size", "Training using Adam"]
|
221 |
-
|
222 |
title = f"Memory Usage for loading `{args.model_name}`"
|
223 |
table = create_ascii_table(headers, data, title)
|
224 |
print(table)
|
225 |
-
|
226 |
-
|
227 |
def main():
|
228 |
parser = estimate_command_parser()
|
229 |
args = parser.parse_args()
|
230 |
estimate_command(args)
|
231 |
-
|
232 |
-
|
233 |
if __name__ == "__main__":
|
234 |
main()
|
|
|
7 |
return "gated"
|
8 |
except RepositoryNotFoundError:
|
9 |
return "repo"
|
|
|
|
|
10 |
def check_has_model(error):
|
11 |
"""
|
12 |
Checks what library spawned `error` when a model is not found
|
|
|
21 |
return "transformers"
|
22 |
else:
|
23 |
return "unknown"
|
|
|
|
|
24 |
def create_empty_model(model_name: str, library_name: str, trust_remote_code: bool = False, access_token: str = None):
|
25 |
"""
|
26 |
Creates an empty model from its parent library on the `Hub` to calculate the overall memory consumption.
|
|
|
27 |
Args:
|
28 |
model_name (`str`):
|
29 |
The model name on the Hub
|
|
|
36 |
execute code present on the Hub on your local machine.
|
37 |
access_token (`str`, `optional`, defaults to `None`):
|
38 |
The access token to use to access private or gated models on the Hub. (for use on the Gradio app)
|
|
|
39 |
Returns:
|
40 |
`torch.nn.Module`: The torch model that has been initialized on the `meta` device.
|
|
|
41 |
"""
|
42 |
model_info = verify_on_hub(model_name, access_token)
|
43 |
# Simplified errors
|
|
|
62 |
f"To check `{model_name}`, `transformers` must be installed. Please install it via `pip install transformers`"
|
63 |
)
|
64 |
print(f"Loading pretrained config for `{model_name}` from `transformers`...")
|
|
|
65 |
auto_map = model_info.config.get("auto_map", False)
|
66 |
config = AutoConfig.from_pretrained(model_name, trust_remote_code=trust_remote_code)
|
|
|
67 |
with init_empty_weights():
|
68 |
# remote code could specify a specific `AutoModel` class in the `auto_map`
|
69 |
constructor = AutoModel
|
|
|
89 |
f"Library `{library_name}` is not supported yet, please open an issue on GitHub for us to add support."
|
90 |
)
|
91 |
return model
|
|
|
|
|
92 |
def create_ascii_table(headers: list, rows: list, title: str):
|
93 |
"Creates a pretty table from a list of rows, minimal version of `tabulate`."
|
94 |
sep_char, in_between = "│", "─"
|
|
|
97 |
column_values = [row[i] for row in rows] + [headers[i]]
|
98 |
max_column_width = max(len(value) for value in column_values)
|
99 |
column_widths.append(max_column_width)
|
|
|
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("├", "┼", "┤")
|
106 |
if len(title) > sum(column_widths):
|
107 |
diff = abs(len(title) - len(separator))
|
108 |
column_widths[-1] += diff
|
|
|
109 |
# Update with diff
|
110 |
separator = make_row("├", "┼", "┤")
|
111 |
initial_rows = [
|
|
|
121 |
centered_line = [t.center(column_widths[i]) for i, t in enumerate(line)]
|
122 |
table += f"{pattern % tuple(centered_line)}\n"
|
123 |
table += f'└{"┴".join([in_between * n for n in column_widths])}┘'
|
|
|
124 |
return table
|
|
|
|
|
125 |
def estimate_command_parser(subparsers=None):
|
126 |
if subparsers is not None:
|
127 |
parser = subparsers.add_parser("estimate-memory")
|
128 |
else:
|
129 |
parser = argparse.ArgumentParser(description="Model size estimator for fitting a model onto CUDA memory.")
|
|
|
130 |
parser.add_argument("model_name", type=str, help="The model name on the Hugging Face Hub.")
|
131 |
parser.add_argument(
|
132 |
"--library_name",
|
|
|
149 |
should only be used for repositories you trust and in which you have read the code, as it will execute
|
150 |
code present on the Hub on your local machine.""",
|
151 |
)
|
|
|
152 |
if subparsers is not None:
|
153 |
parser.set_defaults(func=estimate_command)
|
154 |
return parser
|
|
|
|
|
155 |
def gather_data(args):
|
156 |
"Creates an empty model and gathers the data for the sizes"
|
157 |
try:
|
|
|
165 |
f"Tried to load `{args.model_name}` with `{library}` but a possible model to load was not found inside the repo."
|
166 |
)
|
167 |
raise e
|
|
|
168 |
total_size, largest_layer = calculate_maximum_sizes(model)
|
|
|
169 |
data = []
|
|
|
170 |
for dtype in args.dtypes:
|
171 |
dtype_total_size = total_size
|
172 |
dtype_largest_layer = largest_layer[0]
|
|
|
182 |
dtype_training_size = dtype_total_size * 4
|
183 |
data.append([dtype, dtype_largest_layer, dtype_total_size, dtype_training_size])
|
184 |
return data
|
|
|
|
|
185 |
def estimate_command(args):
|
186 |
data = gather_data(args)
|
187 |
for row in data:
|
188 |
for i, item in enumerate(row):
|
189 |
if isinstance(item, (int, float)):
|
190 |
row[i] = convert_bytes(item)
|
|
|
191 |
headers = ["dtype", "Largest Layer", "Total Size", "Training using Adam"]
|
|
|
192 |
title = f"Memory Usage for loading `{args.model_name}`"
|
193 |
table = create_ascii_table(headers, data, title)
|
194 |
print(table)
|
|
|
|
|
195 |
def main():
|
196 |
parser = estimate_command_parser()
|
197 |
args = parser.parse_args()
|
198 |
estimate_command(args)
|
|
|
|
|
199 |
if __name__ == "__main__":
|
200 |
main()
|
src/commands/launch.py
CHANGED
@@ -1,6 +1,5 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
logger = logging.getLogger(__name__)
|
3 |
-
|
4 |
options_to_group = {
|
5 |
"--multi-gpu": "Distributed GPUs",
|
6 |
"--tpu": "TPU",
|
@@ -8,14 +7,10 @@ options_to_group = {
|
|
8 |
"--use_fsdp": "FSDP Arguments",
|
9 |
"--use_megatron_lm": "Megatron-LM Arguments",
|
10 |
}
|
11 |
-
|
12 |
-
|
13 |
def clean_option(option):
|
14 |
"Finds all cases of - after the first two characters and changes them to _"
|
15 |
if option.startswith("--"):
|
16 |
return option[:3] + option[3:].replace("-", "_")
|
17 |
-
|
18 |
-
|
19 |
class _CustomHelpAction(argparse._HelpAction):
|
20 |
"""
|
21 |
This is a custom help action that will hide all arguments that are not used in the command line when the help is
|
@@ -59,19 +54,14 @@ class _CustomHelpAction(argparse._HelpAction):
|
|
59 |
# If all arguments in the group are hidden, hide the group
|
60 |
if all([arg.help == argparse.SUPPRESS for arg in group._group_actions]):
|
61 |
parser._action_groups.remove(group)
|
62 |
-
|
63 |
super().__call__(parser, namespace, values, option_string)
|
64 |
-
|
65 |
-
|
66 |
def launch_command_parser(subparsers=None):
|
67 |
if subparsers is not None:
|
68 |
parser = subparsers.add_parser("launch", add_help=False, allow_abbrev=False)
|
69 |
else:
|
70 |
parser = argparse.ArgumentParser("Accelerate launch command", add_help=False, allow_abbrev=False)
|
71 |
-
|
72 |
parser.register("action", "help", _CustomHelpAction)
|
73 |
parser.add_argument("-h", "--help", action="help", help="Show this help message and exit.")
|
74 |
-
|
75 |
parser.add_argument(
|
76 |
"--config_file", default=None, help="The config file to use for the default values in the launching script."
|
77 |
)
|
@@ -103,7 +93,6 @@ def launch_command_parser(subparsers=None):
|
|
103 |
action="store_true",
|
104 |
help="Whether or not this should launch a Intel PyTorch Extension (IPEX) training.",
|
105 |
)
|
106 |
-
|
107 |
# Resource selection arguments
|
108 |
resource_args = parser.add_argument_group(
|
109 |
"Resource Selection Arguments", "Arguments for fine-tuning how available hardware should be used."
|
@@ -128,7 +117,6 @@ def launch_command_parser(subparsers=None):
|
|
128 |
default=None,
|
129 |
help="The number of CPU threads per process. Can be tuned for optimal performance.",
|
130 |
)
|
131 |
-
|
132 |
# Dynamo arguments
|
133 |
resource_args.add_argument(
|
134 |
"--dynamo_backend",
|
@@ -156,7 +144,6 @@ def launch_command_parser(subparsers=None):
|
|
156 |
action="store_true",
|
157 |
help="Whether to enable dynamic shape tracing.",
|
158 |
)
|
159 |
-
|
160 |
# Training Paradigm arguments
|
161 |
paradigm_args = parser.add_argument_group(
|
162 |
"Training Paradigm Arguments", "Arguments for selecting which training paradigm to be used."
|
@@ -185,7 +172,6 @@ def launch_command_parser(subparsers=None):
|
|
185 |
action="store_true",
|
186 |
help="Whether to use IPEX plugin to speed up training on XPU specifically.",
|
187 |
)
|
188 |
-
|
189 |
# distributed GPU training arguments
|
190 |
distributed_args = parser.add_argument_group("Distributed GPUs", "Arguments related to distributed GPU training.")
|
191 |
distributed_args.add_argument(
|
@@ -260,7 +246,6 @@ def launch_command_parser(subparsers=None):
|
|
260 |
action="store_true",
|
261 |
help="Skip prepending the training script with 'python' - just execute it directly. Useful when the script is not a Python script.",
|
262 |
)
|
263 |
-
|
264 |
# TPU arguments
|
265 |
tpu_args = parser.add_argument_group("TPU", "Arguments related to TPU.")
|
266 |
tpu_args.add_argument(
|
@@ -306,7 +291,6 @@ def launch_command_parser(subparsers=None):
|
|
306 |
action="store_true",
|
307 |
help="Whether when using bf16 precision on TPUs if both float and double tensors are cast to bfloat16 or if double tensors remain as float32.",
|
308 |
)
|
309 |
-
|
310 |
# DeepSpeed arguments
|
311 |
deepspeed_args = parser.add_argument_group("DeepSpeed Arguments", "Arguments related to DeepSpeed.")
|
312 |
deepspeed_args.add_argument(
|
@@ -402,7 +386,6 @@ def launch_command_parser(subparsers=None):
|
|
402 |
type=str,
|
403 |
help="DeepSpeed multi-node launcher to use. If unspecified, will default to `pdsh`.",
|
404 |
)
|
405 |
-
|
406 |
# fsdp arguments
|
407 |
fsdp_args = parser.add_argument_group("FSDP Arguments", "Arguments related to Fully Shared Data Parallelism.")
|
408 |
fsdp_args.add_argument(
|
@@ -483,7 +466,6 @@ def launch_command_parser(subparsers=None):
|
|
483 |
help="If True, each individually wrapped FSDP unit will broadcast module parameters from rank 0."
|
484 |
" (useful only when `use_fsdp` flag is passed).",
|
485 |
)
|
486 |
-
|
487 |
# megatron_lm args
|
488 |
megatron_lm_args = parser.add_argument_group("Megatron-LM Arguments", "Arguments related to Megatron-LM.")
|
489 |
megatron_lm_args.add_argument(
|
@@ -533,7 +515,6 @@ def launch_command_parser(subparsers=None):
|
|
533 |
help="Megatron-LM's gradient clipping value based on global L2 Norm (0 to disable). "
|
534 |
"(useful only when `use_megatron_lm` flag is passed).",
|
535 |
)
|
536 |
-
|
537 |
# AWS arguments
|
538 |
aws_args = parser.add_argument_group("AWS Arguments", "Arguments related to AWS.")
|
539 |
aws_args.add_argument(
|
@@ -561,18 +542,13 @@ def launch_command_parser(subparsers=None):
|
|
561 |
"script."
|
562 |
),
|
563 |
)
|
564 |
-
|
565 |
# Other arguments of the training scripts
|
566 |
parser.add_argument("training_script_args", nargs=argparse.REMAINDER, help="Arguments of the training script.")
|
567 |
-
|
568 |
if subparsers is not None:
|
569 |
parser.set_defaults(func=launch_command)
|
570 |
return parser
|
571 |
-
|
572 |
-
|
573 |
def simple_launcher(args):
|
574 |
cmd, current_env = prepare_simple_launcher_cmd_env(args)
|
575 |
-
|
576 |
process = subprocess.Popen(cmd, env=current_env)
|
577 |
process.wait()
|
578 |
if process.returncode != 0:
|
@@ -580,11 +556,8 @@ def simple_launcher(args):
|
|
580 |
raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
|
581 |
else:
|
582 |
sys.exit(1)
|
583 |
-
|
584 |
-
|
585 |
def multi_gpu_launcher(args):
|
586 |
import torch.distributed.run as distrib_run
|
587 |
-
|
588 |
current_env = prepare_multi_gpu_env(args)
|
589 |
if not check_cuda_p2p_ib_support():
|
590 |
message = "Using RTX 3090 or 4000 series which doesn't support faster communication speedups. Ensuring P2P and IB communications are disabled."
|
@@ -597,7 +570,6 @@ def multi_gpu_launcher(args):
|
|
597 |
warn = True
|
598 |
if warn:
|
599 |
logger.warning(message)
|
600 |
-
|
601 |
debug = getattr(args, "debug", False)
|
602 |
args = _filter_args(
|
603 |
args,
|
@@ -614,14 +586,10 @@ def multi_gpu_launcher(args):
|
|
614 |
console.print_exception(suppress=[__file__], show_locals=False)
|
615 |
else:
|
616 |
raise
|
617 |
-
|
618 |
-
|
619 |
def deepspeed_launcher(args):
|
620 |
import torch.distributed.run as distrib_run
|
621 |
-
|
622 |
if not is_deepspeed_available():
|
623 |
raise ImportError("DeepSpeed is not installed => run `pip3 install deepspeed` or build it from source.")
|
624 |
-
|
625 |
cmd, current_env = prepare_deepspeed_cmd_env(args)
|
626 |
if not check_cuda_p2p_ib_support():
|
627 |
message = "Using RTX 3090 or 4000 series which doesn't support faster communication speedups. Ensuring P2P and IB communications are disabled."
|
@@ -634,14 +602,12 @@ def deepspeed_launcher(args):
|
|
634 |
warn = True
|
635 |
if warn:
|
636 |
logger.warning(message)
|
637 |
-
|
638 |
if args.num_machines > 1 and args.deepspeed_multinode_launcher != DEEPSPEED_MULTINODE_LAUNCHERS[1]:
|
639 |
with open(".deepspeed_env", "a") as f:
|
640 |
for key, value in current_env.items():
|
641 |
if ";" in value or " " in value:
|
642 |
continue
|
643 |
f.write(f"{key}={value}\n")
|
644 |
-
|
645 |
process = subprocess.Popen(cmd, env=current_env)
|
646 |
process.wait()
|
647 |
if process.returncode != 0:
|
@@ -666,16 +632,11 @@ def deepspeed_launcher(args):
|
|
666 |
console.print_exception(suppress=[__file__], show_locals=False)
|
667 |
else:
|
668 |
raise
|
669 |
-
|
670 |
-
|
671 |
def tpu_launcher(args):
|
672 |
import torch_xla.distributed.xla_multiprocessing as xmp
|
673 |
-
|
674 |
if args.no_python:
|
675 |
raise ValueError("--no_python cannot be used with TPU launcher")
|
676 |
-
|
677 |
args, current_env = prepare_tpu(args, {})
|
678 |
-
|
679 |
if args.module:
|
680 |
mod_name = args.training_script
|
681 |
else:
|
@@ -683,40 +644,31 @@ def tpu_launcher(args):
|
|
683 |
script_path = Path(args.training_script)
|
684 |
sys.path.append(str(script_path.parent.resolve()))
|
685 |
mod_name = script_path.stem
|
686 |
-
|
687 |
mod = importlib.import_module(mod_name)
|
688 |
if not hasattr(mod, args.main_training_function):
|
689 |
raise ValueError(
|
690 |
f"Your training script should have a function named {args.main_training_function}, or you should pass a "
|
691 |
"different value to `--main_training_function`."
|
692 |
)
|
693 |
-
|
694 |
# Patch sys.argv
|
695 |
sys.argv = [mod.__file__] + args.training_script_args
|
696 |
-
|
697 |
main_function = getattr(mod, args.main_training_function)
|
698 |
with patch_environment(**current_env):
|
699 |
xmp.spawn(PrepareForLaunch(main_function), args=(), nprocs=args.num_processes)
|
700 |
-
|
701 |
-
|
702 |
def tpu_pod_launcher(args):
|
703 |
from torch_xla.distributed import xla_dist
|
704 |
-
|
705 |
current_env = {}
|
706 |
args, current_env = prepare_tpu(args, current_env, True)
|
707 |
debug = getattr(args, "debug", False)
|
708 |
-
|
709 |
training_script = args.training_script
|
710 |
training_script_args = args.training_script_args
|
711 |
new_args = _filter_args(
|
712 |
args, xla_dist.get_args_parser(), ["--tpu", args.tpu_name, "--positional", "", "--restart-tpuvm-pod-server"]
|
713 |
)
|
714 |
-
|
715 |
if args.tpu_use_sudo:
|
716 |
new_cmd = ["sudo"]
|
717 |
else:
|
718 |
new_cmd = []
|
719 |
-
|
720 |
new_cmd += [
|
721 |
"accelerate-launch",
|
722 |
"--tpu",
|
@@ -733,7 +685,6 @@ def tpu_pod_launcher(args):
|
|
733 |
str(args.main_training_function),
|
734 |
training_script,
|
735 |
] + training_script_args
|
736 |
-
|
737 |
new_args.positional = new_cmd
|
738 |
bad_flags = ""
|
739 |
for arg in vars(new_args):
|
@@ -756,8 +707,6 @@ def tpu_pod_launcher(args):
|
|
756 |
console.print_exception(suppress=[__file__], show_locals=False)
|
757 |
else:
|
758 |
raise
|
759 |
-
|
760 |
-
|
761 |
def sagemaker_launcher(sagemaker_config: SageMakerConfig, args):
|
762 |
if not is_sagemaker_available():
|
763 |
raise ImportError(
|
@@ -767,17 +716,11 @@ def sagemaker_launcher(sagemaker_config: SageMakerConfig, args):
|
|
767 |
raise ValueError(
|
768 |
"SageMaker requires a python training script file and cannot be used with --module or --no_python"
|
769 |
)
|
770 |
-
|
771 |
from sagemaker.huggingface import HuggingFace
|
772 |
-
|
773 |
args, sagemaker_inputs = prepare_sagemager_args_inputs(sagemaker_config, args)
|
774 |
-
|
775 |
huggingface_estimator = HuggingFace(**args)
|
776 |
-
|
777 |
huggingface_estimator.fit(inputs=sagemaker_inputs)
|
778 |
print(f"You can find your model data at: {huggingface_estimator.model_data}")
|
779 |
-
|
780 |
-
|
781 |
def _validate_launch_command(args):
|
782 |
# Sanity checks
|
783 |
if sum([args.multi_gpu, args.cpu, args.tpu, args.use_deepspeed, args.use_fsdp]) > 1:
|
@@ -786,7 +729,6 @@ def _validate_launch_command(args):
|
|
786 |
)
|
787 |
if args.multi_gpu and (args.num_processes is not None) and (args.num_processes < 2):
|
788 |
raise ValueError("You need to use at least 2 processes to use `--multi_gpu`.")
|
789 |
-
|
790 |
defaults = None
|
791 |
warned = []
|
792 |
mp_from_config_flag = False
|
@@ -817,10 +759,8 @@ def _validate_launch_command(args):
|
|
817 |
args.gpu_ids = defaults.gpu_ids
|
818 |
else:
|
819 |
args.gpu_ids = "all"
|
820 |
-
|
821 |
if args.multi_gpu and args.num_machines is None:
|
822 |
args.num_machines = defaults.num_machines
|
823 |
-
|
824 |
if len(args.gpu_ids.split(",")) < 2 and (args.gpu_ids != "all") and args.multi_gpu and args.num_machines <= 1:
|
825 |
raise ValueError(
|
826 |
"Less than two GPU ids were configured and tried to run on on multiple GPUs. "
|
@@ -844,7 +784,6 @@ def _validate_launch_command(args):
|
|
844 |
for k in defaults.ipex_config:
|
845 |
setattr(args, k, defaults.ipex_config[k])
|
846 |
continue
|
847 |
-
|
848 |
# Those args are handled separately
|
849 |
if (
|
850 |
name not in ["compute_environment", "mixed_precision", "distributed_type"]
|
@@ -853,7 +792,6 @@ def _validate_launch_command(args):
|
|
853 |
setattr(args, name, attr)
|
854 |
if not args.debug:
|
855 |
args.debug = defaults.debug
|
856 |
-
|
857 |
if not args.mixed_precision:
|
858 |
if defaults.mixed_precision is None:
|
859 |
args.mixed_precision = "no"
|
@@ -869,7 +807,6 @@ def _validate_launch_command(args):
|
|
869 |
native_amp = is_bf16_available(True)
|
870 |
if args.mixed_precision == "bf16" and not native_amp and not (args.tpu and is_tpu_available()):
|
871 |
raise ValueError(err.format(mode="bf16", requirement="PyTorch >= 1.10 and a supported device."))
|
872 |
-
|
873 |
# Silently set the default here
|
874 |
if args.dynamo_backend is None:
|
875 |
args.dynamo_backend = "no"
|
@@ -907,7 +844,6 @@ def _validate_launch_command(args):
|
|
907 |
args.dynamo_backend = "no"
|
908 |
if args.debug:
|
909 |
logger.debug("Running script in debug mode, expect distributed operations to be slightly slower.")
|
910 |
-
|
911 |
is_aws_env_disabled = defaults is None or (
|
912 |
defaults is not None and defaults.compute_environment != ComputeEnvironment.AMAZON_SAGEMAKER
|
913 |
)
|
@@ -923,7 +859,6 @@ def _validate_launch_command(args):
|
|
923 |
warned.append(
|
924 |
f"\t`--num_cpu_threads_per_process` was set to `{args.num_cpu_threads_per_process}` to improve out-of-box performance when training on CPUs"
|
925 |
)
|
926 |
-
|
927 |
if any(warned):
|
928 |
message = "The following values were not passed to `accelerate launch` and had defaults used instead:\n"
|
929 |
message += "\n".join(warned)
|
@@ -932,8 +867,6 @@ def _validate_launch_command(args):
|
|
932 |
)
|
933 |
logger.warning(message)
|
934 |
return args, defaults, mp_from_config_flag
|
935 |
-
|
936 |
-
|
937 |
def launch_command(args):
|
938 |
args, defaults, mp_from_config_flag = _validate_launch_command(args)
|
939 |
# Use the proper launcher
|
@@ -958,13 +891,9 @@ def launch_command(args):
|
|
958 |
sagemaker_launcher(defaults, args)
|
959 |
else:
|
960 |
simple_launcher(args)
|
961 |
-
|
962 |
-
|
963 |
def main():
|
964 |
parser = launch_command_parser()
|
965 |
args = parser.parse_args()
|
966 |
launch_command(args)
|
967 |
-
|
968 |
-
|
969 |
if __name__ == "__main__":
|
970 |
main()
|
|
|
1 |
#!/usr/bin/env python
|
2 |
logger = logging.getLogger(__name__)
|
|
|
3 |
options_to_group = {
|
4 |
"--multi-gpu": "Distributed GPUs",
|
5 |
"--tpu": "TPU",
|
|
|
7 |
"--use_fsdp": "FSDP Arguments",
|
8 |
"--use_megatron_lm": "Megatron-LM Arguments",
|
9 |
}
|
|
|
|
|
10 |
def clean_option(option):
|
11 |
"Finds all cases of - after the first two characters and changes them to _"
|
12 |
if option.startswith("--"):
|
13 |
return option[:3] + option[3:].replace("-", "_")
|
|
|
|
|
14 |
class _CustomHelpAction(argparse._HelpAction):
|
15 |
"""
|
16 |
This is a custom help action that will hide all arguments that are not used in the command line when the help is
|
|
|
54 |
# If all arguments in the group are hidden, hide the group
|
55 |
if all([arg.help == argparse.SUPPRESS for arg in group._group_actions]):
|
56 |
parser._action_groups.remove(group)
|
|
|
57 |
super().__call__(parser, namespace, values, option_string)
|
|
|
|
|
58 |
def launch_command_parser(subparsers=None):
|
59 |
if subparsers is not None:
|
60 |
parser = subparsers.add_parser("launch", add_help=False, allow_abbrev=False)
|
61 |
else:
|
62 |
parser = argparse.ArgumentParser("Accelerate launch command", add_help=False, allow_abbrev=False)
|
|
|
63 |
parser.register("action", "help", _CustomHelpAction)
|
64 |
parser.add_argument("-h", "--help", action="help", help="Show this help message and exit.")
|
|
|
65 |
parser.add_argument(
|
66 |
"--config_file", default=None, help="The config file to use for the default values in the launching script."
|
67 |
)
|
|
|
93 |
action="store_true",
|
94 |
help="Whether or not this should launch a Intel PyTorch Extension (IPEX) training.",
|
95 |
)
|
|
|
96 |
# Resource selection arguments
|
97 |
resource_args = parser.add_argument_group(
|
98 |
"Resource Selection Arguments", "Arguments for fine-tuning how available hardware should be used."
|
|
|
117 |
default=None,
|
118 |
help="The number of CPU threads per process. Can be tuned for optimal performance.",
|
119 |
)
|
|
|
120 |
# Dynamo arguments
|
121 |
resource_args.add_argument(
|
122 |
"--dynamo_backend",
|
|
|
144 |
action="store_true",
|
145 |
help="Whether to enable dynamic shape tracing.",
|
146 |
)
|
|
|
147 |
# Training Paradigm arguments
|
148 |
paradigm_args = parser.add_argument_group(
|
149 |
"Training Paradigm Arguments", "Arguments for selecting which training paradigm to be used."
|
|
|
172 |
action="store_true",
|
173 |
help="Whether to use IPEX plugin to speed up training on XPU specifically.",
|
174 |
)
|
|
|
175 |
# distributed GPU training arguments
|
176 |
distributed_args = parser.add_argument_group("Distributed GPUs", "Arguments related to distributed GPU training.")
|
177 |
distributed_args.add_argument(
|
|
|
246 |
action="store_true",
|
247 |
help="Skip prepending the training script with 'python' - just execute it directly. Useful when the script is not a Python script.",
|
248 |
)
|
|
|
249 |
# TPU arguments
|
250 |
tpu_args = parser.add_argument_group("TPU", "Arguments related to TPU.")
|
251 |
tpu_args.add_argument(
|
|
|
291 |
action="store_true",
|
292 |
help="Whether when using bf16 precision on TPUs if both float and double tensors are cast to bfloat16 or if double tensors remain as float32.",
|
293 |
)
|
|
|
294 |
# DeepSpeed arguments
|
295 |
deepspeed_args = parser.add_argument_group("DeepSpeed Arguments", "Arguments related to DeepSpeed.")
|
296 |
deepspeed_args.add_argument(
|
|
|
386 |
type=str,
|
387 |
help="DeepSpeed multi-node launcher to use. If unspecified, will default to `pdsh`.",
|
388 |
)
|
|
|
389 |
# fsdp arguments
|
390 |
fsdp_args = parser.add_argument_group("FSDP Arguments", "Arguments related to Fully Shared Data Parallelism.")
|
391 |
fsdp_args.add_argument(
|
|
|
466 |
help="If True, each individually wrapped FSDP unit will broadcast module parameters from rank 0."
|
467 |
" (useful only when `use_fsdp` flag is passed).",
|
468 |
)
|
|
|
469 |
# megatron_lm args
|
470 |
megatron_lm_args = parser.add_argument_group("Megatron-LM Arguments", "Arguments related to Megatron-LM.")
|
471 |
megatron_lm_args.add_argument(
|
|
|
515 |
help="Megatron-LM's gradient clipping value based on global L2 Norm (0 to disable). "
|
516 |
"(useful only when `use_megatron_lm` flag is passed).",
|
517 |
)
|
|
|
518 |
# AWS arguments
|
519 |
aws_args = parser.add_argument_group("AWS Arguments", "Arguments related to AWS.")
|
520 |
aws_args.add_argument(
|
|
|
542 |
"script."
|
543 |
),
|
544 |
)
|
|
|
545 |
# Other arguments of the training scripts
|
546 |
parser.add_argument("training_script_args", nargs=argparse.REMAINDER, help="Arguments of the training script.")
|
|
|
547 |
if subparsers is not None:
|
548 |
parser.set_defaults(func=launch_command)
|
549 |
return parser
|
|
|
|
|
550 |
def simple_launcher(args):
|
551 |
cmd, current_env = prepare_simple_launcher_cmd_env(args)
|
|
|
552 |
process = subprocess.Popen(cmd, env=current_env)
|
553 |
process.wait()
|
554 |
if process.returncode != 0:
|
|
|
556 |
raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
|
557 |
else:
|
558 |
sys.exit(1)
|
|
|
|
|
559 |
def multi_gpu_launcher(args):
|
560 |
import torch.distributed.run as distrib_run
|
|
|
561 |
current_env = prepare_multi_gpu_env(args)
|
562 |
if not check_cuda_p2p_ib_support():
|
563 |
message = "Using RTX 3090 or 4000 series which doesn't support faster communication speedups. Ensuring P2P and IB communications are disabled."
|
|
|
570 |
warn = True
|
571 |
if warn:
|
572 |
logger.warning(message)
|
|
|
573 |
debug = getattr(args, "debug", False)
|
574 |
args = _filter_args(
|
575 |
args,
|
|
|
586 |
console.print_exception(suppress=[__file__], show_locals=False)
|
587 |
else:
|
588 |
raise
|
|
|
|
|
589 |
def deepspeed_launcher(args):
|
590 |
import torch.distributed.run as distrib_run
|
|
|
591 |
if not is_deepspeed_available():
|
592 |
raise ImportError("DeepSpeed is not installed => run `pip3 install deepspeed` or build it from source.")
|
|
|
593 |
cmd, current_env = prepare_deepspeed_cmd_env(args)
|
594 |
if not check_cuda_p2p_ib_support():
|
595 |
message = "Using RTX 3090 or 4000 series which doesn't support faster communication speedups. Ensuring P2P and IB communications are disabled."
|
|
|
602 |
warn = True
|
603 |
if warn:
|
604 |
logger.warning(message)
|
|
|
605 |
if args.num_machines > 1 and args.deepspeed_multinode_launcher != DEEPSPEED_MULTINODE_LAUNCHERS[1]:
|
606 |
with open(".deepspeed_env", "a") as f:
|
607 |
for key, value in current_env.items():
|
608 |
if ";" in value or " " in value:
|
609 |
continue
|
610 |
f.write(f"{key}={value}\n")
|
|
|
611 |
process = subprocess.Popen(cmd, env=current_env)
|
612 |
process.wait()
|
613 |
if process.returncode != 0:
|
|
|
632 |
console.print_exception(suppress=[__file__], show_locals=False)
|
633 |
else:
|
634 |
raise
|
|
|
|
|
635 |
def tpu_launcher(args):
|
636 |
import torch_xla.distributed.xla_multiprocessing as xmp
|
|
|
637 |
if args.no_python:
|
638 |
raise ValueError("--no_python cannot be used with TPU launcher")
|
|
|
639 |
args, current_env = prepare_tpu(args, {})
|
|
|
640 |
if args.module:
|
641 |
mod_name = args.training_script
|
642 |
else:
|
|
|
644 |
script_path = Path(args.training_script)
|
645 |
sys.path.append(str(script_path.parent.resolve()))
|
646 |
mod_name = script_path.stem
|
|
|
647 |
mod = importlib.import_module(mod_name)
|
648 |
if not hasattr(mod, args.main_training_function):
|
649 |
raise ValueError(
|
650 |
f"Your training script should have a function named {args.main_training_function}, or you should pass a "
|
651 |
"different value to `--main_training_function`."
|
652 |
)
|
|
|
653 |
# Patch sys.argv
|
654 |
sys.argv = [mod.__file__] + args.training_script_args
|
|
|
655 |
main_function = getattr(mod, args.main_training_function)
|
656 |
with patch_environment(**current_env):
|
657 |
xmp.spawn(PrepareForLaunch(main_function), args=(), nprocs=args.num_processes)
|
|
|
|
|
658 |
def tpu_pod_launcher(args):
|
659 |
from torch_xla.distributed import xla_dist
|
|
|
660 |
current_env = {}
|
661 |
args, current_env = prepare_tpu(args, current_env, True)
|
662 |
debug = getattr(args, "debug", False)
|
|
|
663 |
training_script = args.training_script
|
664 |
training_script_args = args.training_script_args
|
665 |
new_args = _filter_args(
|
666 |
args, xla_dist.get_args_parser(), ["--tpu", args.tpu_name, "--positional", "", "--restart-tpuvm-pod-server"]
|
667 |
)
|
|
|
668 |
if args.tpu_use_sudo:
|
669 |
new_cmd = ["sudo"]
|
670 |
else:
|
671 |
new_cmd = []
|
|
|
672 |
new_cmd += [
|
673 |
"accelerate-launch",
|
674 |
"--tpu",
|
|
|
685 |
str(args.main_training_function),
|
686 |
training_script,
|
687 |
] + training_script_args
|
|
|
688 |
new_args.positional = new_cmd
|
689 |
bad_flags = ""
|
690 |
for arg in vars(new_args):
|
|
|
707 |
console.print_exception(suppress=[__file__], show_locals=False)
|
708 |
else:
|
709 |
raise
|
|
|
|
|
710 |
def sagemaker_launcher(sagemaker_config: SageMakerConfig, args):
|
711 |
if not is_sagemaker_available():
|
712 |
raise ImportError(
|
|
|
716 |
raise ValueError(
|
717 |
"SageMaker requires a python training script file and cannot be used with --module or --no_python"
|
718 |
)
|
|
|
719 |
from sagemaker.huggingface import HuggingFace
|
|
|
720 |
args, sagemaker_inputs = prepare_sagemager_args_inputs(sagemaker_config, args)
|
|
|
721 |
huggingface_estimator = HuggingFace(**args)
|
|
|
722 |
huggingface_estimator.fit(inputs=sagemaker_inputs)
|
723 |
print(f"You can find your model data at: {huggingface_estimator.model_data}")
|
|
|
|
|
724 |
def _validate_launch_command(args):
|
725 |
# Sanity checks
|
726 |
if sum([args.multi_gpu, args.cpu, args.tpu, args.use_deepspeed, args.use_fsdp]) > 1:
|
|
|
729 |
)
|
730 |
if args.multi_gpu and (args.num_processes is not None) and (args.num_processes < 2):
|
731 |
raise ValueError("You need to use at least 2 processes to use `--multi_gpu`.")
|
|
|
732 |
defaults = None
|
733 |
warned = []
|
734 |
mp_from_config_flag = False
|
|
|
759 |
args.gpu_ids = defaults.gpu_ids
|
760 |
else:
|
761 |
args.gpu_ids = "all"
|
|
|
762 |
if args.multi_gpu and args.num_machines is None:
|
763 |
args.num_machines = defaults.num_machines
|
|
|
764 |
if len(args.gpu_ids.split(",")) < 2 and (args.gpu_ids != "all") and args.multi_gpu and args.num_machines <= 1:
|
765 |
raise ValueError(
|
766 |
"Less than two GPU ids were configured and tried to run on on multiple GPUs. "
|
|
|
784 |
for k in defaults.ipex_config:
|
785 |
setattr(args, k, defaults.ipex_config[k])
|
786 |
continue
|
|
|
787 |
# Those args are handled separately
|
788 |
if (
|
789 |
name not in ["compute_environment", "mixed_precision", "distributed_type"]
|
|
|
792 |
setattr(args, name, attr)
|
793 |
if not args.debug:
|
794 |
args.debug = defaults.debug
|
|
|
795 |
if not args.mixed_precision:
|
796 |
if defaults.mixed_precision is None:
|
797 |
args.mixed_precision = "no"
|
|
|
807 |
native_amp = is_bf16_available(True)
|
808 |
if args.mixed_precision == "bf16" and not native_amp and not (args.tpu and is_tpu_available()):
|
809 |
raise ValueError(err.format(mode="bf16", requirement="PyTorch >= 1.10 and a supported device."))
|
|
|
810 |
# Silently set the default here
|
811 |
if args.dynamo_backend is None:
|
812 |
args.dynamo_backend = "no"
|
|
|
844 |
args.dynamo_backend = "no"
|
845 |
if args.debug:
|
846 |
logger.debug("Running script in debug mode, expect distributed operations to be slightly slower.")
|
|
|
847 |
is_aws_env_disabled = defaults is None or (
|
848 |
defaults is not None and defaults.compute_environment != ComputeEnvironment.AMAZON_SAGEMAKER
|
849 |
)
|
|
|
859 |
warned.append(
|
860 |
f"\t`--num_cpu_threads_per_process` was set to `{args.num_cpu_threads_per_process}` to improve out-of-box performance when training on CPUs"
|
861 |
)
|
|
|
862 |
if any(warned):
|
863 |
message = "The following values were not passed to `accelerate launch` and had defaults used instead:\n"
|
864 |
message += "\n".join(warned)
|
|
|
867 |
)
|
868 |
logger.warning(message)
|
869 |
return args, defaults, mp_from_config_flag
|
|
|
|
|
870 |
def launch_command(args):
|
871 |
args, defaults, mp_from_config_flag = _validate_launch_command(args)
|
872 |
# Use the proper launcher
|
|
|
891 |
sagemaker_launcher(defaults, args)
|
892 |
else:
|
893 |
simple_launcher(args)
|
|
|
|
|
894 |
def main():
|
895 |
parser = launch_command_parser()
|
896 |
args = parser.parse_args()
|
897 |
launch_command(args)
|
|
|
|
|
898 |
if __name__ == "__main__":
|
899 |
main()
|
src/commands/test.py
CHANGED
@@ -4,7 +4,6 @@ def test_command_parser(subparsers=None):
|
|
4 |
parser = subparsers.add_parser("test")
|
5 |
else:
|
6 |
parser = argparse.ArgumentParser("Accelerate test command")
|
7 |
-
|
8 |
parser.add_argument(
|
9 |
"--config_file",
|
10 |
default=None,
|
@@ -15,31 +14,22 @@ def test_command_parser(subparsers=None):
|
|
15 |
"with 'huggingface'."
|
16 |
),
|
17 |
)
|
18 |
-
|
19 |
if subparsers is not None:
|
20 |
parser.set_defaults(func=test_command)
|
21 |
return parser
|
22 |
-
|
23 |
-
|
24 |
def test_command(args):
|
25 |
script_name = os.path.sep.join(__file__.split(os.path.sep)[:-2] + ["test_utils", "scripts", "test_script.py"])
|
26 |
-
|
27 |
if args.config_file is None:
|
28 |
test_args = script_name
|
29 |
else:
|
30 |
test_args = f"--config_file={args.config_file} {script_name}"
|
31 |
-
|
32 |
cmd = ["accelerate-launch"] + test_args.split()
|
33 |
result = execute_subprocess_async(cmd, env=os.environ.copy())
|
34 |
if result.returncode == 0:
|
35 |
print("Test is a success! You are ready for your distributed training!")
|
36 |
-
|
37 |
-
|
38 |
def main():
|
39 |
parser = test_command_parser()
|
40 |
args = parser.parse_args()
|
41 |
test_command(args)
|
42 |
-
|
43 |
-
|
44 |
if __name__ == "__main__":
|
45 |
main()
|
|
|
4 |
parser = subparsers.add_parser("test")
|
5 |
else:
|
6 |
parser = argparse.ArgumentParser("Accelerate test command")
|
|
|
7 |
parser.add_argument(
|
8 |
"--config_file",
|
9 |
default=None,
|
|
|
14 |
"with 'huggingface'."
|
15 |
),
|
16 |
)
|
|
|
17 |
if subparsers is not None:
|
18 |
parser.set_defaults(func=test_command)
|
19 |
return parser
|
|
|
|
|
20 |
def test_command(args):
|
21 |
script_name = os.path.sep.join(__file__.split(os.path.sep)[:-2] + ["test_utils", "scripts", "test_script.py"])
|
|
|
22 |
if args.config_file is None:
|
23 |
test_args = script_name
|
24 |
else:
|
25 |
test_args = f"--config_file={args.config_file} {script_name}"
|
|
|
26 |
cmd = ["accelerate-launch"] + test_args.split()
|
27 |
result = execute_subprocess_async(cmd, env=os.environ.copy())
|
28 |
if result.returncode == 0:
|
29 |
print("Test is a success! You are ready for your distributed training!")
|
|
|
|
|
30 |
def main():
|
31 |
parser = test_command_parser()
|
32 |
args = parser.parse_args()
|
33 |
test_command(args)
|
|
|
|
|
34 |
if __name__ == "__main__":
|
35 |
main()
|
src/commands/tpu.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
_description = "Run commands across TPU VMs for initial setup before running `accelerate launch`."
|
3 |
-
|
4 |
-
|
5 |
def tpu_command_parser(subparsers=None):
|
6 |
if subparsers is not None:
|
7 |
parser = subparsers.add_parser("tpu-config", description=_description)
|
@@ -57,15 +55,11 @@ def tpu_command_parser(subparsers=None):
|
|
57 |
pod_args.add_argument(
|
58 |
"--debug", action="store_true", help="If set, will print the command that would be run instead of running it."
|
59 |
)
|
60 |
-
|
61 |
if subparsers is not None:
|
62 |
parser.set_defaults(func=tpu_command_launcher)
|
63 |
return parser
|
64 |
-
|
65 |
-
|
66 |
def tpu_command_launcher(args):
|
67 |
defaults = None
|
68 |
-
|
69 |
# Get the default from the config file if it exists.
|
70 |
if args.config_file is not None or os.path.isfile(default_config_file):
|
71 |
defaults = load_config_from_file(args.config_file)
|
@@ -83,14 +77,11 @@ def tpu_command_launcher(args):
|
|
83 |
args.accelerate_version = "accelerate -U"
|
84 |
elif isinstance(parse(args.accelerate_version), Version):
|
85 |
args.accelerate_version = f"accelerate=={args.accelerate_version}"
|
86 |
-
|
87 |
if not args.command_file and not args.command:
|
88 |
raise ValueError("You must specify either a command file or a command to run on the pod.")
|
89 |
-
|
90 |
if args.command_file:
|
91 |
with open(args.command_file, "r") as f:
|
92 |
args.command = [f.read().splitlines()]
|
93 |
-
|
94 |
# To turn list of lists into list of strings
|
95 |
if isinstance(args.command[0], list):
|
96 |
args.command = [line for cmd in args.command for line in cmd]
|
@@ -100,7 +91,6 @@ def tpu_command_launcher(args):
|
|
100 |
new_cmd += [f"pip install {args.accelerate_version}"]
|
101 |
new_cmd += args.command
|
102 |
args.command = "; ".join(new_cmd)
|
103 |
-
|
104 |
# Then send it to gcloud
|
105 |
# Eventually try to use google-api-core to do this instead of subprocess
|
106 |
cmd = ["gcloud"]
|
@@ -124,10 +114,7 @@ def tpu_command_launcher(args):
|
|
124 |
return
|
125 |
subprocess.run(cmd)
|
126 |
print("Successfully setup pod.")
|
127 |
-
|
128 |
-
|
129 |
def main():
|
130 |
parser = tpu_command_parser()
|
131 |
args = parser.parse_args()
|
132 |
-
|
133 |
tpu_command_launcher(args)
|
|
|
1 |
#!/usr/bin/env python
|
2 |
_description = "Run commands across TPU VMs for initial setup before running `accelerate launch`."
|
|
|
|
|
3 |
def tpu_command_parser(subparsers=None):
|
4 |
if subparsers is not None:
|
5 |
parser = subparsers.add_parser("tpu-config", description=_description)
|
|
|
55 |
pod_args.add_argument(
|
56 |
"--debug", action="store_true", help="If set, will print the command that would be run instead of running it."
|
57 |
)
|
|
|
58 |
if subparsers is not None:
|
59 |
parser.set_defaults(func=tpu_command_launcher)
|
60 |
return parser
|
|
|
|
|
61 |
def tpu_command_launcher(args):
|
62 |
defaults = None
|
|
|
63 |
# Get the default from the config file if it exists.
|
64 |
if args.config_file is not None or os.path.isfile(default_config_file):
|
65 |
defaults = load_config_from_file(args.config_file)
|
|
|
77 |
args.accelerate_version = "accelerate -U"
|
78 |
elif isinstance(parse(args.accelerate_version), Version):
|
79 |
args.accelerate_version = f"accelerate=={args.accelerate_version}"
|
|
|
80 |
if not args.command_file and not args.command:
|
81 |
raise ValueError("You must specify either a command file or a command to run on the pod.")
|
|
|
82 |
if args.command_file:
|
83 |
with open(args.command_file, "r") as f:
|
84 |
args.command = [f.read().splitlines()]
|
|
|
85 |
# To turn list of lists into list of strings
|
86 |
if isinstance(args.command[0], list):
|
87 |
args.command = [line for cmd in args.command for line in cmd]
|
|
|
91 |
new_cmd += [f"pip install {args.accelerate_version}"]
|
92 |
new_cmd += args.command
|
93 |
args.command = "; ".join(new_cmd)
|
|
|
94 |
# Then send it to gcloud
|
95 |
# Eventually try to use google-api-core to do this instead of subprocess
|
96 |
cmd = ["gcloud"]
|
|
|
114 |
return
|
115 |
subprocess.run(cmd)
|
116 |
print("Successfully setup pod.")
|
|
|
|
|
117 |
def main():
|
118 |
parser = tpu_command_parser()
|
119 |
args = parser.parse_args()
|
|
|
120 |
tpu_command_launcher(args)
|
src/data_loader.py
CHANGED
@@ -1,5 +1,4 @@
|
|
1 |
logger = get_logger(__name__)
|
2 |
-
|
3 |
# kwargs of the DataLoader in min version 1.4.0.
|
4 |
_PYTORCH_DATALOADER_KWARGS = {
|
5 |
"batch_size": 1,
|
@@ -17,22 +16,16 @@ _PYTORCH_DATALOADER_KWARGS = {
|
|
17 |
"prefetch_factor": 2,
|
18 |
"persistent_workers": False,
|
19 |
}
|
20 |
-
|
21 |
# kwargs added after by version
|
22 |
_PYTORCH_DATALOADER_ADDITIONAL_KWARGS = {}
|
23 |
-
|
24 |
for v, additional_kwargs in _PYTORCH_DATALOADER_ADDITIONAL_KWARGS.items():
|
25 |
if is_torch_version(">=", v):
|
26 |
_PYTORCH_DATALOADER_KWARGS.update(additional_kwargs)
|
27 |
-
|
28 |
-
|
29 |
class SeedableRandomSampler(RandomSampler):
|
30 |
"""
|
31 |
Same as a random sampler, except that in `__iter__` a seed can be used.
|
32 |
-
|
33 |
Needed specifically in distributed cases, when the random generator for each GPU needs to start from the same seed
|
34 |
and be fully reproducable on multiple iterations.
|
35 |
-
|
36 |
If a custom `generator` is passed, it will rely on its initial seed as well as the current iteration it is on
|
37 |
(stored in `self.epoch`).
|
38 |
"""
|
@@ -40,7 +33,6 @@ class SeedableRandomSampler(RandomSampler):
|
|
40 |
super().__init__(*args, **kwargs)
|
41 |
self.epoch = 0
|
42 |
self.seed = torch.random.initial_seed()
|
43 |
-
|
44 |
def __iter__(self):
|
45 |
if self.generator is None:
|
46 |
self.generator = torch.Generator()
|
@@ -51,19 +43,15 @@ class SeedableRandomSampler(RandomSampler):
|
|
51 |
self.generator.manual_seed(seed)
|
52 |
yield from super().__iter__()
|
53 |
self.set_epoch(self.epoch + 1)
|
54 |
-
|
55 |
def set_epoch(self, epoch: int):
|
56 |
"Sets the current iteration of the sampler."
|
57 |
self.epoch = epoch
|
58 |
-
|
59 |
-
|
60 |
class BatchSamplerShard(BatchSampler):
|
61 |
"""
|
62 |
Wraps a PyTorch `BatchSampler` to generate batches for one of the processes only. Instances of this class will
|
63 |
always yield a number of batches that is a round multiple of `num_processes` and that all have the same size.
|
64 |
Depending on the value of the `drop_last` attribute of the batch sampler passed, it will either stop the iteration
|
65 |
at the first batch that would be too small / not present on all processes or loop with indices from the beginning.
|
66 |
-
|
67 |
Args:
|
68 |
batch_sampler (`torch.utils.data.sampler.BatchSampler`):
|
69 |
The batch sampler to split in several shards.
|
@@ -74,9 +62,7 @@ class BatchSamplerShard(BatchSampler):
|
|
74 |
split_batches (`bool`, *optional*, defaults to `False`):
|
75 |
Whether the shards should be created by splitting a batch to give a piece of it on each process, or by
|
76 |
yielding different full batches on each process.
|
77 |
-
|
78 |
On two processes with a sampler of `[[0, 1, 2, 3], [4, 5, 6, 7]]`, this will result in:
|
79 |
-
|
80 |
- the sampler on process 0 to yield `[0, 1, 2, 3]` and the sampler on process 1 to yield `[4, 5, 6, 7]` if
|
81 |
this argument is set to `False`.
|
82 |
- the sampler on process 0 to yield `[0, 1]` then `[4, 5]` and the sampler on process 1 to yield `[2, 3]`
|
@@ -84,14 +70,10 @@ class BatchSamplerShard(BatchSampler):
|
|
84 |
even_batches (`bool`, *optional*, defaults to `True`):
|
85 |
Whether or not to loop back at the beginning of the sampler when the number of samples is not a round
|
86 |
multiple of (original batch size / number of processes).
|
87 |
-
|
88 |
<Tip warning={true}>
|
89 |
-
|
90 |
`BatchSampler`s with varying batch sizes are not enabled by default. To enable this behaviour, set `even_batches`
|
91 |
equal to `False`
|
92 |
-
|
93 |
</Tip>"""
|
94 |
-
|
95 |
def __init__(
|
96 |
self,
|
97 |
batch_sampler: BatchSampler,
|
@@ -117,11 +99,9 @@ class BatchSamplerShard(BatchSampler):
|
|
117 |
"You need to use `even_batches=False` when the batch sampler has no batch size. If you "
|
118 |
"are not calling this method directly, set `accelerator.even_batches=False` instead."
|
119 |
)
|
120 |
-
|
121 |
@property
|
122 |
def total_length(self):
|
123 |
return len(self.batch_sampler)
|
124 |
-
|
125 |
def __len__(self):
|
126 |
if self.split_batches:
|
127 |
# Split batches does not change the length of the batch sampler
|
@@ -139,10 +119,8 @@ class BatchSamplerShard(BatchSampler):
|
|
139 |
else:
|
140 |
# Otherwise it depends on the process index.
|
141 |
return length + 1 if self.process_index < len(self.batch_sampler) % self.num_processes else length
|
142 |
-
|
143 |
def __iter__(self):
|
144 |
return self._iter_with_split() if self.split_batches else self._iter_with_no_split()
|
145 |
-
|
146 |
def _iter_with_split(self):
|
147 |
initial_data = []
|
148 |
batch_length = self.batch_sampler.batch_size // self.num_processes
|
@@ -152,7 +130,6 @@ class BatchSamplerShard(BatchSampler):
|
|
152 |
if len(batch) == self.batch_size:
|
153 |
# If the batch is full, we yield the part of it this process is responsible of.
|
154 |
yield batch[batch_length * self.process_index : batch_length * (self.process_index + 1)]
|
155 |
-
|
156 |
# If drop_last is True of the last batch was full, iteration is over, otherwise...
|
157 |
if not self.drop_last and len(initial_data) > 0 and len(batch) < self.batch_size:
|
158 |
if not self.even_batches:
|
@@ -164,7 +141,6 @@ class BatchSamplerShard(BatchSampler):
|
|
164 |
initial_data += initial_data
|
165 |
batch = batch + initial_data
|
166 |
yield batch[batch_length * self.process_index : batch_length * (self.process_index + 1)]
|
167 |
-
|
168 |
def _iter_with_no_split(self):
|
169 |
initial_data = []
|
170 |
batch_to_yield = []
|
@@ -181,7 +157,6 @@ class BatchSamplerShard(BatchSampler):
|
|
181 |
):
|
182 |
yield batch_to_yield
|
183 |
batch_to_yield = []
|
184 |
-
|
185 |
# If drop_last is True, iteration is over, otherwise...
|
186 |
if not self.drop_last and len(initial_data) > 0:
|
187 |
if not self.even_batches:
|
@@ -191,16 +166,13 @@ class BatchSamplerShard(BatchSampler):
|
|
191 |
# ... we yield the complete batch we had saved before if it has the proper length
|
192 |
if len(batch_to_yield) == self.batch_size:
|
193 |
yield batch_to_yield
|
194 |
-
|
195 |
# For degenerate cases where the dataset has less than num_process * batch_size samples
|
196 |
while len(initial_data) < self.num_processes * self.batch_size:
|
197 |
initial_data += initial_data
|
198 |
-
|
199 |
# If the last batch seen was of the proper size, it has been yielded by its process so we move to the next
|
200 |
if len(batch) == self.batch_size:
|
201 |
batch = []
|
202 |
idx += 1
|
203 |
-
|
204 |
# Make sure we yield a multiple of self.num_processes batches
|
205 |
cycle_index = 0
|
206 |
while idx % self.num_processes != 0 or len(batch) > 0:
|
@@ -211,8 +183,6 @@ class BatchSamplerShard(BatchSampler):
|
|
211 |
cycle_index = end_index
|
212 |
batch = []
|
213 |
idx += 1
|
214 |
-
|
215 |
-
|
216 |
class IterableDatasetShard(IterableDataset):
|
217 |
"""
|
218 |
Wraps a PyTorch `IterableDataset` to generate samples for one of the processes only. Instances of this class will
|
@@ -220,7 +190,6 @@ class IterableDatasetShard(IterableDataset):
|
|
220 |
`split_batches`, this is either `batch_size` or `batch_size x num_processes`). Depending on the value of the
|
221 |
`drop_last` attribute of the batch sampler passed, it will either stop the iteration at the first batch that would
|
222 |
be too small or loop with indices from the beginning.
|
223 |
-
|
224 |
Args:
|
225 |
dataset (`torch.utils.data.dataset.IterableDataset`):
|
226 |
The batch sampler to split in several shards.
|
@@ -237,9 +206,7 @@ class IterableDatasetShard(IterableDataset):
|
|
237 |
split_batches (`bool`, *optional*, defaults to `False`):
|
238 |
Whether the shards should be created by splitting a batch to give a piece of it on each process, or by
|
239 |
yielding different full batches on each process.
|
240 |
-
|
241 |
On two processes with an iterable dataset yielding of `[0, 1, 2, 3, 4, 5, 6, 7]`, this will result in:
|
242 |
-
|
243 |
- the shard on process 0 to yield `[0, 1, 2, 3]` and the shard on process 1 to yield `[4, 5, 6, 7]` if this
|
244 |
argument is set to `False`.
|
245 |
- the shard on process 0 to yield `[0, 1, 4, 5]` and the sampler on process 1 to yield `[2, 3, 6, 7]` if
|
@@ -265,19 +232,16 @@ class IterableDatasetShard(IterableDataset):
|
|
265 |
self.num_processes = num_processes
|
266 |
self.process_index = process_index
|
267 |
self.split_batches = split_batches
|
268 |
-
|
269 |
def set_epoch(self, epoch):
|
270 |
self.epoch = epoch
|
271 |
if hasattr(self.dataset, "set_epoch"):
|
272 |
self.dataset.set_epoch(epoch)
|
273 |
-
|
274 |
def __len__(self):
|
275 |
# We will just raise the downstream error if the underlying dataset is not sized
|
276 |
if self.drop_last:
|
277 |
return (len(self.dataset) // (self.batch_size * self.num_processes)) * self.batch_size
|
278 |
else:
|
279 |
return math.ceil(len(self.dataset) / (self.batch_size * self.num_processes)) * self.batch_size
|
280 |
-
|
281 |
def __iter__(self):
|
282 |
if (
|
283 |
not hasattr(self.dataset, "set_epoch")
|
@@ -288,7 +252,6 @@ class IterableDatasetShard(IterableDataset):
|
|
288 |
real_batch_size = self.batch_size if self.split_batches else (self.batch_size * self.num_processes)
|
289 |
process_batch_size = (self.batch_size // self.num_processes) if self.split_batches else self.batch_size
|
290 |
process_slice = range(self.process_index * process_batch_size, (self.process_index + 1) * process_batch_size)
|
291 |
-
|
292 |
first_batch = None
|
293 |
current_batch = []
|
294 |
for element in self.dataset:
|
@@ -300,7 +263,6 @@ class IterableDatasetShard(IterableDataset):
|
|
300 |
if first_batch is None:
|
301 |
first_batch = current_batch.copy()
|
302 |
current_batch = []
|
303 |
-
|
304 |
# Finished if drop_last is True, otherwise complete the last batch with elements from the beginning.
|
305 |
if not self.drop_last and len(current_batch) > 0:
|
306 |
if first_batch is None:
|
@@ -309,29 +271,22 @@ class IterableDatasetShard(IterableDataset):
|
|
309 |
current_batch += first_batch
|
310 |
for i in process_slice:
|
311 |
yield current_batch[i]
|
312 |
-
|
313 |
-
|
314 |
class DataLoaderStateMixin:
|
315 |
"""
|
316 |
Mixin class that adds a state to a `DataLoader` to keep track of the status inside the dataloader such as at the
|
317 |
end of the iteration, the number of items in the dataset in the last batch relative to the batch size, and other
|
318 |
useful information that might be needed.
|
319 |
-
|
320 |
**Available attributes:**
|
321 |
-
|
322 |
- **end_of_dataloader** (`bool`) -- Whether at the last iteration or batch
|
323 |
- **remainder** (`int`) -- The number of items that are remaining in the last batch, relative to the total
|
324 |
batch size
|
325 |
-
|
326 |
"""
|
327 |
def __init_subclass__(cls, **kwargs):
|
328 |
cls.end_of_dataloader = False
|
329 |
cls.remainder = -1
|
330 |
-
|
331 |
def reset(self):
|
332 |
self.end_of_dataloader = False
|
333 |
self.remainder = -1
|
334 |
-
|
335 |
def begin(self):
|
336 |
"Prepares the gradient state for the current dataloader"
|
337 |
self.reset()
|
@@ -340,16 +295,12 @@ class DataLoaderStateMixin:
|
|
340 |
length = getattr(self.dataset, "total_dataset_length", len(self.dataset))
|
341 |
self.remainder = length % self.total_batch_size
|
342 |
self.gradient_state._add_dataloader(self)
|
343 |
-
|
344 |
def end(self):
|
345 |
"Cleans up the gradient state after exiting the dataloader"
|
346 |
self.gradient_state._remove_dataloader(self)
|
347 |
-
|
348 |
-
|
349 |
class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
350 |
"""
|
351 |
Subclass of a PyTorch `DataLoader` that will deal with device placement and current distributed setup.
|
352 |
-
|
353 |
Args:
|
354 |
dataset (`torch.utils.data.dataset.Dataset`):
|
355 |
The dataset to use to build this datalaoder.
|
@@ -358,7 +309,6 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
358 |
rng_types (list of `str` or [`~utils.RNGType`]):
|
359 |
The list of random number generators to synchronize at the beginning of each iteration. Should be one or
|
360 |
several of:
|
361 |
-
|
362 |
- `"torch"`: the base torch random number generator
|
363 |
- `"cuda"`: the CUDA random number generator (GPU only)
|
364 |
- `"xla"`: the XLA random number generator (TPU only)
|
@@ -369,13 +319,10 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
369 |
The number of batches to skip at the beginning.
|
370 |
kwargs:
|
371 |
All other keyword arguments to pass to the regular `DataLoader` initialization.
|
372 |
-
|
373 |
**Available attributes:**
|
374 |
-
|
375 |
- **total_batch_size** (`int`) -- Total batch size of the dataloader across all processes.
|
376 |
Equal to the original batch size when `split_batches=True`; otherwise the original batch size * the total
|
377 |
number of processes
|
378 |
-
|
379 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
380 |
"""
|
381 |
def __init__(
|
@@ -396,12 +343,10 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
396 |
self.gradient_state = GradientState()
|
397 |
self._drop_last = _drop_last
|
398 |
self.iteration = 0
|
399 |
-
|
400 |
def __iter__(self):
|
401 |
if self.rng_types is not None:
|
402 |
synchronize_rng_states(self.rng_types, self.synchronized_generator)
|
403 |
self.begin()
|
404 |
-
|
405 |
self.set_epoch(self.iteration)
|
406 |
dataloader_iter = super().__iter__()
|
407 |
# We iterate one batch ahead to check when we are at the end
|
@@ -409,7 +354,6 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
409 |
current_batch = next(dataloader_iter)
|
410 |
except StopIteration:
|
411 |
yield
|
412 |
-
|
413 |
batch_index = 0
|
414 |
while True:
|
415 |
try:
|
@@ -426,10 +370,8 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
426 |
if batch_index >= self.skip_batches:
|
427 |
yield current_batch
|
428 |
break
|
429 |
-
|
430 |
self.iteration += 1
|
431 |
self.end()
|
432 |
-
|
433 |
def set_epoch(self, epoch: int):
|
434 |
# In case it is manually passed in, the user can set it to what they like
|
435 |
if self.iteration != epoch:
|
@@ -440,7 +382,6 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
440 |
# or in general HF datasets `Datasets`
|
441 |
elif hasattr(self.dataset, "set_epoch"):
|
442 |
self.dataset.set_epoch(epoch)
|
443 |
-
|
444 |
@property
|
445 |
def total_batch_size(self):
|
446 |
batch_sampler = self.sampler if isinstance(self.sampler, BatchSampler) else self.batch_sampler
|
@@ -449,63 +390,47 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
449 |
if getattr(batch_sampler, "split_batches", False)
|
450 |
else (batch_sampler.batch_size * getattr(batch_sampler, "num_processes", 1))
|
451 |
)
|
452 |
-
|
453 |
@property
|
454 |
def total_dataset_length(self):
|
455 |
if hasattr(self.dataset, "total_length"):
|
456 |
return self.dataset.total_length
|
457 |
else:
|
458 |
return len(self.dataset)
|
459 |
-
|
460 |
-
|
461 |
if is_tpu_available(check_device=False):
|
462 |
import torch_xla.distributed.parallel_loader as xpl
|
463 |
-
|
464 |
class MpDeviceLoaderWrapper(xpl.MpDeviceLoader):
|
465 |
"""
|
466 |
Wrapper for the xpl.MpDeviceLoader class that knows the total batch size.
|
467 |
-
|
468 |
XLA preloading threads will all call DataLoaderShard's __iter__(). Remove rng_types from DataLoaderShard to
|
469 |
prevent it from using the XLA device in the preloading threads, and synchronize the RNG once from the main
|
470 |
thread only.
|
471 |
-
|
472 |
**Available attributes:**
|
473 |
-
|
474 |
- **total_batch_size** (`int`) -- Total batch size of the dataloader across all processes.
|
475 |
Equal to the original batch size when `split_batches=True`; otherwise the original batch size * the total
|
476 |
number of processes
|
477 |
-
|
478 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
479 |
"""
|
480 |
def __init__(self, dataloader: DataLoaderShard, device: torch.device):
|
481 |
super().__init__(dataloader, device)
|
482 |
self._rng_types = self._loader.rng_types
|
483 |
self._loader.rng_types = None
|
484 |
-
|
485 |
def __iter__(self):
|
486 |
if self._rng_types is not None:
|
487 |
synchronize_rng_states(self._rng_types, self._loader.synchronized_generator)
|
488 |
-
|
489 |
return super().__iter__()
|
490 |
-
|
491 |
@property
|
492 |
def total_batch_size(self):
|
493 |
return self._loader.total_batch_size
|
494 |
-
|
495 |
@property
|
496 |
def total_dataset_length(self):
|
497 |
return self._loader.total_dataset_length
|
498 |
-
|
499 |
@property
|
500 |
def batch_sampler(self):
|
501 |
return self._loader.batch_sampler
|
502 |
-
|
503 |
-
|
504 |
class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
505 |
"""
|
506 |
Subclass of a PyTorch `DataLoader` that will iterate and preprocess on process 0 only, then dispatch on each
|
507 |
process their part of the batch.
|
508 |
-
|
509 |
Args:
|
510 |
split_batches (`bool`, *optional*, defaults to `False`):
|
511 |
Whether the resulting `DataLoader` should split the batches of the original data loader across devices or
|
@@ -516,13 +441,10 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
516 |
size of the `dataloader` is a round multiple of `batch_size`.
|
517 |
skip_batches (`int`, *optional*, defaults to 0):
|
518 |
The number of batches to skip at the beginning of an iteration.
|
519 |
-
|
520 |
**Available attributes:**
|
521 |
-
|
522 |
- **total_batch_size** (`int`) -- Total batch size of the dataloader across all processes.
|
523 |
Equal to the original batch size when `split_batches=True`; otherwise the original batch size * the total
|
524 |
number of processes
|
525 |
-
|
526 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
527 |
"""
|
528 |
def __init__(
|
@@ -531,7 +453,6 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
531 |
shuffle = False
|
532 |
if is_torch_version(">=", "1.11.0"):
|
533 |
from torch.utils.data.datapipes.iter.combinatorics import ShufflerIterDataPipe
|
534 |
-
|
535 |
# We need to save the shuffling state of the DataPipe
|
536 |
if isinstance(dataset, ShufflerIterDataPipe):
|
537 |
shuffle = dataset._shuffle_enabled
|
@@ -539,15 +460,12 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
539 |
self.split_batches = split_batches
|
540 |
if shuffle:
|
541 |
torch.utils.data.graph_settings.apply_shuffle_settings(dataset, shuffle=shuffle)
|
542 |
-
|
543 |
self.gradient_state = GradientState()
|
544 |
self.state = AcceleratorState()
|
545 |
self._drop_last = _drop_last
|
546 |
self.skip_batches = skip_batches
|
547 |
-
|
548 |
self.slice_fn = slice_tensors if slice_fn is None else slice_fn
|
549 |
self.iteration = 0
|
550 |
-
|
551 |
def _fetch_batches(self, iterator):
|
552 |
batches, batch = None, None
|
553 |
# On process 0, we gather the batch to dispatch.
|
@@ -584,7 +502,6 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
584 |
batch_info = [None, True]
|
585 |
broadcast_object_list(batch_info)
|
586 |
return batch, batch_info
|
587 |
-
|
588 |
def __iter__(self):
|
589 |
self.begin()
|
590 |
self.set_epoch(self.iteration)
|
@@ -603,14 +520,12 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
603 |
batch_index = 0
|
604 |
while not stop_iteration:
|
605 |
batch, batch_info = next_batch, next_batch_info
|
606 |
-
|
607 |
if self.state.process_index != 0:
|
608 |
# Initialize tensors on other processes than process 0.
|
609 |
batch = initialize_tensors(batch_info[0])
|
610 |
batch = send_to_device(batch, self.state.device)
|
611 |
# Broadcast the batch before splitting it.
|
612 |
batch = broadcast(batch, from_process=0)
|
613 |
-
|
614 |
if not self._drop_last and first_batch is None:
|
615 |
# We keep at least num processes elements of the first batch to be able to complete the last batch
|
616 |
first_batch = self.slice_fn(
|
@@ -619,15 +534,12 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
619 |
process_index=self.state.process_index,
|
620 |
num_processes=self.state.num_processes,
|
621 |
)
|
622 |
-
|
623 |
if batch is None:
|
624 |
raise ValueError(
|
625 |
f"Batch does not contain any data (`{batch}`). At the end of all iterable data available before expected stop iteration."
|
626 |
)
|
627 |
-
|
628 |
observed_batch_size = find_batch_size(batch)
|
629 |
batch_size = observed_batch_size // self.state.num_processes
|
630 |
-
|
631 |
stop_iteration = self._stop_iteration
|
632 |
if not stop_iteration:
|
633 |
# We may still be at the end of the dataloader without knowing it yet: if there is nothing left in
|
@@ -636,13 +548,11 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
636 |
# next_batch_info[0] is None when there are no more batches, otherwise we still need to process them.
|
637 |
if self._stop_iteration and next_batch_info[0] is None:
|
638 |
stop_iteration = True
|
639 |
-
|
640 |
if not self._drop_last and stop_iteration and observed_batch_size % self.state.num_processes != 0:
|
641 |
# If the last batch is not complete, let's add the first batch to it.
|
642 |
batch = concatenate([batch, first_batch], dim=0)
|
643 |
# Batch size computation above is wrong, it's off by 1 so we fix it.
|
644 |
batch_size += 1
|
645 |
-
|
646 |
data_slice = slice(self.state.process_index * batch_size, (self.state.process_index + 1) * batch_size)
|
647 |
batch = self.slice_fn(
|
648 |
batch,
|
@@ -650,7 +560,6 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
650 |
process_index=self.state.process_index,
|
651 |
num_processes=self.state.num_processes,
|
652 |
)
|
653 |
-
|
654 |
if stop_iteration:
|
655 |
self.end_of_dataloader = True
|
656 |
self.remainder = observed_batch_size
|
@@ -659,7 +568,6 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
659 |
batch_index += 1
|
660 |
self.iteration += 1
|
661 |
self.end()
|
662 |
-
|
663 |
def set_epoch(self, epoch: int):
|
664 |
# In case it is manually passed in, the user can set it to what they like
|
665 |
if self.iteration != epoch:
|
@@ -668,7 +576,6 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
668 |
self.batch_sampler.sampler.set_epoch(epoch)
|
669 |
elif hasattr(self.dataset, "set_epoch"):
|
670 |
self.dataset.set_epoch(epoch)
|
671 |
-
|
672 |
def __len__(self):
|
673 |
whole_length = super().__len__()
|
674 |
if self.split_batches:
|
@@ -677,18 +584,14 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
677 |
return whole_length // self.state.num_processes
|
678 |
else:
|
679 |
return math.ceil(whole_length / self.state.num_processes)
|
680 |
-
|
681 |
@property
|
682 |
def total_batch_size(self):
|
683 |
return (
|
684 |
self.dataset.batch_size if self.split_batches else (self.dataset.batch_size * self.dataset.num_processes)
|
685 |
)
|
686 |
-
|
687 |
@property
|
688 |
def total_dataset_length(self):
|
689 |
return len(self.dataset)
|
690 |
-
|
691 |
-
|
692 |
def prepare_data_loader(
|
693 |
dataloader: DataLoader,
|
694 |
device: Optional[torch.device] = None,
|
@@ -703,10 +606,8 @@ def prepare_data_loader(
|
|
703 |
) -> DataLoader:
|
704 |
"""
|
705 |
Wraps a PyTorch `DataLoader` to generate batches for one of the processes only.
|
706 |
-
|
707 |
Depending on the value of the `drop_last` attribute of the `dataloader` passed, it will either stop the iteration
|
708 |
at the first batch that would be too small / not present on all processes or loop with indices from the beginning.
|
709 |
-
|
710 |
Args:
|
711 |
dataloader (`torch.utils.data.dataloader.DataLoader`):
|
712 |
The data loader to split across several devices.
|
@@ -721,11 +622,9 @@ def prepare_data_loader(
|
|
721 |
Whether the resulting `DataLoader` should split the batches of the original data loader across devices or
|
722 |
yield full batches (in which case it will yield batches starting at the `process_index`-th and advancing of
|
723 |
`num_processes` batches at each iteration).
|
724 |
-
|
725 |
Another way to see this is that the observed batch size will be the same as the initial `dataloader` if
|
726 |
this option is set to `True`, the batch size of the initial `dataloader` multiplied by `num_processes`
|
727 |
otherwise.
|
728 |
-
|
729 |
Setting this option to `True` requires that the batch size of the `dataloader` is a round multiple of
|
730 |
`batch_size`.
|
731 |
put_on_device (`bool`, *optional*, defaults to `False`):
|
@@ -734,13 +633,11 @@ def prepare_data_loader(
|
|
734 |
rng_types (list of `str` or [`~utils.RNGType`]):
|
735 |
The list of random number generators to synchronize at the beginning of each iteration. Should be one or
|
736 |
several of:
|
737 |
-
|
738 |
- `"torch"`: the base torch random number generator
|
739 |
- `"cuda"`: the CUDA random number generator (GPU only)
|
740 |
- `"xla"`: the XLA random number generator (TPU only)
|
741 |
- `"generator"`: the `torch.Generator` of the sampler (or batch sampler if there is no sampler in your
|
742 |
dataloader) or of the iterable dataset (if it exists) if the underlying dataset is of that type.
|
743 |
-
|
744 |
dispatch_batches (`bool`, *optional*):
|
745 |
If set to `True`, the datalaoder prepared is only iterated through on the main process and then the batches
|
746 |
are split and broadcast to each process. Will default to `True` when the underlying dataset is an
|
@@ -753,15 +650,11 @@ def prepare_data_loader(
|
|
753 |
If passed, this function will be used to slice tensors across `num_processes`. Will default to
|
754 |
[`~utils.slice_tensors`]. This argument is used only when `dispatch_batches` is set to `True` and will be
|
755 |
ignored otherwise.
|
756 |
-
|
757 |
Returns:
|
758 |
`torch.utils.data.dataloader.DataLoader`: A new data loader that will yield the portion of the batches
|
759 |
-
|
760 |
<Tip warning={true}>
|
761 |
-
|
762 |
`BatchSampler`s with varying batch sizes are not enabled by default. To enable this behaviour, set `even_batches`
|
763 |
equal to `False`
|
764 |
-
|
765 |
</Tip>
|
766 |
"""
|
767 |
if dispatch_batches is None:
|
@@ -769,7 +662,6 @@ def prepare_data_loader(
|
|
769 |
dispatch_batches = False
|
770 |
else:
|
771 |
dispatch_batches = isinstance(dataloader.dataset, IterableDataset)
|
772 |
-
|
773 |
if dispatch_batches and not put_on_device:
|
774 |
raise ValueError("Using `dispatch_batches=True` requires `put_on_device=True`.")
|
775 |
# Grab defaults from AcceleratorState
|
@@ -778,14 +670,12 @@ def prepare_data_loader(
|
|
778 |
num_processes = state.num_processes
|
779 |
if process_index is None:
|
780 |
process_index = state.process_index
|
781 |
-
|
782 |
# Sanity check
|
783 |
if split_batches and dataloader.batch_size > 1 and dataloader.batch_size % num_processes != 0:
|
784 |
raise ValueError(
|
785 |
f"To use a `DataLoader` in `split_batches` mode, the batch size ({dataloader.batch_size}) "
|
786 |
f"needs to be a round multiple of the number of processes ({num_processes})."
|
787 |
)
|
788 |
-
|
789 |
new_dataset = dataloader.dataset
|
790 |
# Iterable dataset doesn't like batch_sampler, but data_loader creates a default one for it
|
791 |
new_batch_sampler = dataloader.batch_sampler if not isinstance(new_dataset, IterableDataset) else None
|
@@ -807,7 +697,6 @@ def prepare_data_loader(
|
|
807 |
num_samples=sampler._num_samples,
|
808 |
generator=getattr(sampler, "generator", torch.Generator()),
|
809 |
)
|
810 |
-
|
811 |
# No change if no multiprocess
|
812 |
if (num_processes != 1 or state.distributed_type == DistributedType.MEGATRON_LM) and not dispatch_batches:
|
813 |
if isinstance(new_dataset, IterableDataset):
|
@@ -830,7 +719,6 @@ def prepare_data_loader(
|
|
830 |
split_batches=split_batches,
|
831 |
even_batches=even_batches,
|
832 |
)
|
833 |
-
|
834 |
# We ignore all of those since they are all dealt with by our new_batch_sampler
|
835 |
ignore_kwargs = [
|
836 |
"batch_size",
|
@@ -839,16 +727,13 @@ def prepare_data_loader(
|
|
839 |
"batch_sampler",
|
840 |
"drop_last",
|
841 |
]
|
842 |
-
|
843 |
if rng_types is not None and synchronized_generator is None and "generator" in rng_types:
|
844 |
rng_types.remove("generator")
|
845 |
-
|
846 |
kwargs = {
|
847 |
k: getattr(dataloader, k, _PYTORCH_DATALOADER_KWARGS[k])
|
848 |
for k in _PYTORCH_DATALOADER_KWARGS
|
849 |
if k not in ignore_kwargs
|
850 |
}
|
851 |
-
|
852 |
# Need to provide batch_size as batch_sampler is None for Iterable dataset
|
853 |
if new_batch_sampler is None:
|
854 |
kwargs["drop_last"] = dataloader.drop_last
|
@@ -891,12 +776,9 @@ def prepare_data_loader(
|
|
891 |
_drop_last=dataloader.drop_last,
|
892 |
**kwargs,
|
893 |
)
|
894 |
-
|
895 |
if state.distributed_type == DistributedType.TPU:
|
896 |
return MpDeviceLoaderWrapper(dataloader, device)
|
897 |
return dataloader
|
898 |
-
|
899 |
-
|
900 |
class SkipBatchSampler(BatchSampler):
|
901 |
"""
|
902 |
A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
|
@@ -904,24 +786,18 @@ class SkipBatchSampler(BatchSampler):
|
|
904 |
def __init__(self, batch_sampler, skip_batches=0):
|
905 |
self.batch_sampler = batch_sampler
|
906 |
self.skip_batches = skip_batches
|
907 |
-
|
908 |
def __iter__(self):
|
909 |
for index, samples in enumerate(self.batch_sampler):
|
910 |
if index >= self.skip_batches:
|
911 |
yield samples
|
912 |
-
|
913 |
@property
|
914 |
def total_length(self):
|
915 |
return len(self.batch_sampler)
|
916 |
-
|
917 |
def __len__(self):
|
918 |
return len(self.batch_sampler) - self.skip_batches
|
919 |
-
|
920 |
-
|
921 |
class SkipDataLoader(DataLoader):
|
922 |
"""
|
923 |
Subclass of a PyTorch `DataLoader` that will skip the first batches.
|
924 |
-
|
925 |
Args:
|
926 |
dataset (`torch.utils.data.dataset.Dataset`):
|
927 |
The dataset to use to build this datalaoder.
|
@@ -933,13 +809,10 @@ class SkipDataLoader(DataLoader):
|
|
933 |
def __init__(self, dataset, skip_batches=0, **kwargs):
|
934 |
super().__init__(dataset, **kwargs)
|
935 |
self.skip_batches = skip_batches
|
936 |
-
|
937 |
def __iter__(self):
|
938 |
for index, batch in enumerate(super().__iter__()):
|
939 |
if index >= self.skip_batches:
|
940 |
yield batch
|
941 |
-
|
942 |
-
|
943 |
def skip_first_batches(dataloader, num_batches=0):
|
944 |
"""
|
945 |
Creates a `torch.utils.data.DataLoader` that will efficiently skip the first `num_batches`.
|
@@ -952,7 +825,6 @@ def skip_first_batches(dataloader, num_batches=0):
|
|
952 |
sampler_is_batch_sampler = isinstance(dataloader.sampler, BatchSampler)
|
953 |
batch_sampler = dataloader.sampler if sampler_is_batch_sampler else dataloader.batch_sampler
|
954 |
new_batch_sampler = SkipBatchSampler(batch_sampler, skip_batches=num_batches)
|
955 |
-
|
956 |
# We ignore all of those since they are all dealt with by our new_batch_sampler
|
957 |
ignore_kwargs = [
|
958 |
"batch_size",
|
@@ -961,18 +833,15 @@ def skip_first_batches(dataloader, num_batches=0):
|
|
961 |
"batch_sampler",
|
962 |
"drop_last",
|
963 |
]
|
964 |
-
|
965 |
kwargs = {
|
966 |
k: getattr(dataloader, k, _PYTORCH_DATALOADER_KWARGS[k])
|
967 |
for k in _PYTORCH_DATALOADER_KWARGS
|
968 |
if k not in ignore_kwargs
|
969 |
}
|
970 |
-
|
971 |
# Need to provide batch_size as batch_sampler is None for Iterable dataset
|
972 |
if new_batch_sampler is None:
|
973 |
kwargs["drop_last"] = dataloader.drop_last
|
974 |
kwargs["batch_size"] = dataloader.batch_size
|
975 |
-
|
976 |
if isinstance(dataloader, DataLoaderDispatcher):
|
977 |
if new_batch_sampler is None:
|
978 |
# Need to manually skip batches in the dataloader
|
@@ -1006,5 +875,4 @@ def skip_first_batches(dataloader, num_batches=0):
|
|
1006 |
dataloader = SkipDataLoader(dataset, skip_batches=num_batches, **kwargs)
|
1007 |
else:
|
1008 |
dataloader = DataLoader(dataset, batch_sampler=new_batch_sampler, **kwargs)
|
1009 |
-
|
1010 |
return dataloader
|
|
|
1 |
logger = get_logger(__name__)
|
|
|
2 |
# kwargs of the DataLoader in min version 1.4.0.
|
3 |
_PYTORCH_DATALOADER_KWARGS = {
|
4 |
"batch_size": 1,
|
|
|
16 |
"prefetch_factor": 2,
|
17 |
"persistent_workers": False,
|
18 |
}
|
|
|
19 |
# kwargs added after by version
|
20 |
_PYTORCH_DATALOADER_ADDITIONAL_KWARGS = {}
|
|
|
21 |
for v, additional_kwargs in _PYTORCH_DATALOADER_ADDITIONAL_KWARGS.items():
|
22 |
if is_torch_version(">=", v):
|
23 |
_PYTORCH_DATALOADER_KWARGS.update(additional_kwargs)
|
|
|
|
|
24 |
class SeedableRandomSampler(RandomSampler):
|
25 |
"""
|
26 |
Same as a random sampler, except that in `__iter__` a seed can be used.
|
|
|
27 |
Needed specifically in distributed cases, when the random generator for each GPU needs to start from the same seed
|
28 |
and be fully reproducable on multiple iterations.
|
|
|
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 |
"""
|
|
|
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 |
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
|
|
|
|
|
49 |
class BatchSamplerShard(BatchSampler):
|
50 |
"""
|
51 |
Wraps a PyTorch `BatchSampler` to generate batches for one of the processes only. Instances of this class will
|
52 |
always yield a number of batches that is a round multiple of `num_processes` and that all have the same size.
|
53 |
Depending on the value of the `drop_last` attribute of the batch sampler passed, it will either stop the iteration
|
54 |
at the first batch that would be too small / not present on all processes or loop with indices from the beginning.
|
|
|
55 |
Args:
|
56 |
batch_sampler (`torch.utils.data.sampler.BatchSampler`):
|
57 |
The batch sampler to split in several shards.
|
|
|
62 |
split_batches (`bool`, *optional*, defaults to `False`):
|
63 |
Whether the shards should be created by splitting a batch to give a piece of it on each process, or by
|
64 |
yielding different full batches on each process.
|
|
|
65 |
On two processes with a sampler of `[[0, 1, 2, 3], [4, 5, 6, 7]]`, this will result in:
|
|
|
66 |
- the sampler on process 0 to yield `[0, 1, 2, 3]` and the sampler on process 1 to yield `[4, 5, 6, 7]` if
|
67 |
this argument is set to `False`.
|
68 |
- the sampler on process 0 to yield `[0, 1]` then `[4, 5]` and the sampler on process 1 to yield `[2, 3]`
|
|
|
70 |
even_batches (`bool`, *optional*, defaults to `True`):
|
71 |
Whether or not to loop back at the beginning of the sampler when the number of samples is not a round
|
72 |
multiple of (original batch size / number of processes).
|
|
|
73 |
<Tip warning={true}>
|
|
|
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,
|
|
|
99 |
"You need to use `even_batches=False` when the batch sampler has no batch size. If you "
|
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 |
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
|
|
|
130 |
if len(batch) == self.batch_size:
|
131 |
# If the batch is full, we yield the part of it this process is responsible of.
|
132 |
yield batch[batch_length * self.process_index : batch_length * (self.process_index + 1)]
|
|
|
133 |
# If drop_last is True of the last batch was full, iteration is over, otherwise...
|
134 |
if not self.drop_last and len(initial_data) > 0 and len(batch) < self.batch_size:
|
135 |
if not self.even_batches:
|
|
|
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 = []
|
|
|
157 |
):
|
158 |
yield batch_to_yield
|
159 |
batch_to_yield = []
|
|
|
160 |
# If drop_last is True, iteration is over, otherwise...
|
161 |
if not self.drop_last and len(initial_data) > 0:
|
162 |
if not self.even_batches:
|
|
|
166 |
# ... we yield the complete batch we had saved before if it has the proper length
|
167 |
if len(batch_to_yield) == self.batch_size:
|
168 |
yield batch_to_yield
|
|
|
169 |
# For degenerate cases where the dataset has less than num_process * batch_size samples
|
170 |
while len(initial_data) < self.num_processes * self.batch_size:
|
171 |
initial_data += initial_data
|
|
|
172 |
# If the last batch seen was of the proper size, it has been yielded by its process so we move to the next
|
173 |
if len(batch) == self.batch_size:
|
174 |
batch = []
|
175 |
idx += 1
|
|
|
176 |
# Make sure we yield a multiple of self.num_processes batches
|
177 |
cycle_index = 0
|
178 |
while idx % self.num_processes != 0 or len(batch) > 0:
|
|
|
183 |
cycle_index = end_index
|
184 |
batch = []
|
185 |
idx += 1
|
|
|
|
|
186 |
class IterableDatasetShard(IterableDataset):
|
187 |
"""
|
188 |
Wraps a PyTorch `IterableDataset` to generate samples for one of the processes only. Instances of this class will
|
|
|
190 |
`split_batches`, this is either `batch_size` or `batch_size x num_processes`). Depending on the value of the
|
191 |
`drop_last` attribute of the batch sampler passed, it will either stop the iteration at the first batch that would
|
192 |
be too small or loop with indices from the beginning.
|
|
|
193 |
Args:
|
194 |
dataset (`torch.utils.data.dataset.IterableDataset`):
|
195 |
The batch sampler to split in several shards.
|
|
|
206 |
split_batches (`bool`, *optional*, defaults to `False`):
|
207 |
Whether the shards should be created by splitting a batch to give a piece of it on each process, or by
|
208 |
yielding different full batches on each process.
|
|
|
209 |
On two processes with an iterable dataset yielding of `[0, 1, 2, 3, 4, 5, 6, 7]`, this will result in:
|
|
|
210 |
- the shard on process 0 to yield `[0, 1, 2, 3]` and the shard on process 1 to yield `[4, 5, 6, 7]` if this
|
211 |
argument is set to `False`.
|
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
|
|
|
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")
|
|
|
252 |
real_batch_size = self.batch_size if self.split_batches else (self.batch_size * self.num_processes)
|
253 |
process_batch_size = (self.batch_size // self.num_processes) if self.split_batches else self.batch_size
|
254 |
process_slice = range(self.process_index * process_batch_size, (self.process_index + 1) * process_batch_size)
|
|
|
255 |
first_batch = None
|
256 |
current_batch = []
|
257 |
for element in self.dataset:
|
|
|
263 |
if first_batch is None:
|
264 |
first_batch = current_batch.copy()
|
265 |
current_batch = []
|
|
|
266 |
# Finished if drop_last is True, otherwise complete the last batch with elements from the beginning.
|
267 |
if not self.drop_last and len(current_batch) > 0:
|
268 |
if first_batch is None:
|
|
|
271 |
current_batch += first_batch
|
272 |
for i in process_slice:
|
273 |
yield current_batch[i]
|
|
|
|
|
274 |
class DataLoaderStateMixin:
|
275 |
"""
|
276 |
Mixin class that adds a state to a `DataLoader` to keep track of the status inside the dataloader such as at the
|
277 |
end of the iteration, the number of items in the dataset in the last batch relative to the batch size, and other
|
278 |
useful information that might be needed.
|
|
|
279 |
**Available attributes:**
|
|
|
280 |
- **end_of_dataloader** (`bool`) -- Whether at the last iteration or batch
|
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 |
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)
|
|
|
|
|
301 |
class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
302 |
"""
|
303 |
Subclass of a PyTorch `DataLoader` that will deal with device placement and current distributed setup.
|
|
|
304 |
Args:
|
305 |
dataset (`torch.utils.data.dataset.Dataset`):
|
306 |
The dataset to use to build this datalaoder.
|
|
|
309 |
rng_types (list of `str` or [`~utils.RNGType`]):
|
310 |
The list of random number generators to synchronize at the beginning of each iteration. Should be one or
|
311 |
several of:
|
|
|
312 |
- `"torch"`: the base torch random number generator
|
313 |
- `"cuda"`: the CUDA random number generator (GPU only)
|
314 |
- `"xla"`: the XLA random number generator (TPU only)
|
|
|
319 |
The number of batches to skip at the beginning.
|
320 |
kwargs:
|
321 |
All other keyword arguments to pass to the regular `DataLoader` initialization.
|
|
|
322 |
**Available attributes:**
|
|
|
323 |
- **total_batch_size** (`int`) -- Total batch size of the dataloader across all processes.
|
324 |
Equal to the original batch size when `split_batches=True`; otherwise the original batch size * the total
|
325 |
number of processes
|
|
|
326 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
327 |
"""
|
328 |
def __init__(
|
|
|
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)
|
349 |
self.begin()
|
|
|
350 |
self.set_epoch(self.iteration)
|
351 |
dataloader_iter = super().__iter__()
|
352 |
# We iterate one batch ahead to check when we are at the end
|
|
|
354 |
current_batch = next(dataloader_iter)
|
355 |
except StopIteration:
|
356 |
yield
|
|
|
357 |
batch_index = 0
|
358 |
while True:
|
359 |
try:
|
|
|
370 |
if batch_index >= self.skip_batches:
|
371 |
yield current_batch
|
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:
|
|
|
382 |
# or in general HF datasets `Datasets`
|
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
|
|
|
390 |
if getattr(batch_sampler, "split_batches", False)
|
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
|
397 |
else:
|
398 |
return len(self.dataset)
|
|
|
|
|
399 |
if is_tpu_available(check_device=False):
|
400 |
import torch_xla.distributed.parallel_loader as xpl
|
|
|
401 |
class MpDeviceLoaderWrapper(xpl.MpDeviceLoader):
|
402 |
"""
|
403 |
Wrapper for the xpl.MpDeviceLoader class that knows the total batch size.
|
|
|
404 |
XLA preloading threads will all call DataLoaderShard's __iter__(). Remove rng_types from DataLoaderShard to
|
405 |
prevent it from using the XLA device in the preloading threads, and synchronize the RNG once from the main
|
406 |
thread only.
|
|
|
407 |
**Available attributes:**
|
|
|
408 |
- **total_batch_size** (`int`) -- Total batch size of the dataloader across all processes.
|
409 |
Equal to the original batch size when `split_batches=True`; otherwise the original batch size * the total
|
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 |
"""
|
432 |
Subclass of a PyTorch `DataLoader` that will iterate and preprocess on process 0 only, then dispatch on each
|
433 |
process their part of the batch.
|
|
|
434 |
Args:
|
435 |
split_batches (`bool`, *optional*, defaults to `False`):
|
436 |
Whether the resulting `DataLoader` should split the batches of the original data loader across devices or
|
|
|
441 |
size of the `dataloader` is a round multiple of `batch_size`.
|
442 |
skip_batches (`int`, *optional*, defaults to 0):
|
443 |
The number of batches to skip at the beginning of an iteration.
|
|
|
444 |
**Available attributes:**
|
|
|
445 |
- **total_batch_size** (`int`) -- Total batch size of the dataloader across all processes.
|
446 |
Equal to the original batch size when `split_batches=True`; otherwise the original batch size * the total
|
447 |
number of processes
|
|
|
448 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
449 |
"""
|
450 |
def __init__(
|
|
|
453 |
shuffle = False
|
454 |
if is_torch_version(">=", "1.11.0"):
|
455 |
from torch.utils.data.datapipes.iter.combinatorics import ShufflerIterDataPipe
|
|
|
456 |
# We need to save the shuffling state of the DataPipe
|
457 |
if isinstance(dataset, ShufflerIterDataPipe):
|
458 |
shuffle = dataset._shuffle_enabled
|
|
|
460 |
self.split_batches = split_batches
|
461 |
if shuffle:
|
462 |
torch.utils.data.graph_settings.apply_shuffle_settings(dataset, shuffle=shuffle)
|
|
|
463 |
self.gradient_state = GradientState()
|
464 |
self.state = AcceleratorState()
|
465 |
self._drop_last = _drop_last
|
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 |
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)
|
|
|
520 |
batch_index = 0
|
521 |
while not stop_iteration:
|
522 |
batch, batch_info = next_batch, next_batch_info
|
|
|
523 |
if self.state.process_index != 0:
|
524 |
# Initialize tensors on other processes than process 0.
|
525 |
batch = initialize_tensors(batch_info[0])
|
526 |
batch = send_to_device(batch, self.state.device)
|
527 |
# Broadcast the batch before splitting it.
|
528 |
batch = broadcast(batch, from_process=0)
|
|
|
529 |
if not self._drop_last and first_batch is None:
|
530 |
# We keep at least num processes elements of the first batch to be able to complete the last batch
|
531 |
first_batch = self.slice_fn(
|
|
|
534 |
process_index=self.state.process_index,
|
535 |
num_processes=self.state.num_processes,
|
536 |
)
|
|
|
537 |
if batch is None:
|
538 |
raise ValueError(
|
539 |
f"Batch does not contain any data (`{batch}`). At the end of all iterable data available before expected stop iteration."
|
540 |
)
|
|
|
541 |
observed_batch_size = find_batch_size(batch)
|
542 |
batch_size = observed_batch_size // self.state.num_processes
|
|
|
543 |
stop_iteration = self._stop_iteration
|
544 |
if not stop_iteration:
|
545 |
# We may still be at the end of the dataloader without knowing it yet: if there is nothing left in
|
|
|
548 |
# next_batch_info[0] is None when there are no more batches, otherwise we still need to process them.
|
549 |
if self._stop_iteration and next_batch_info[0] is None:
|
550 |
stop_iteration = True
|
|
|
551 |
if not self._drop_last and stop_iteration and observed_batch_size % self.state.num_processes != 0:
|
552 |
# If the last batch is not complete, let's add the first batch to it.
|
553 |
batch = concatenate([batch, first_batch], dim=0)
|
554 |
# Batch size computation above is wrong, it's off by 1 so we fix it.
|
555 |
batch_size += 1
|
|
|
556 |
data_slice = slice(self.state.process_index * batch_size, (self.state.process_index + 1) * batch_size)
|
557 |
batch = self.slice_fn(
|
558 |
batch,
|
|
|
560 |
process_index=self.state.process_index,
|
561 |
num_processes=self.state.num_processes,
|
562 |
)
|
|
|
563 |
if stop_iteration:
|
564 |
self.end_of_dataloader = True
|
565 |
self.remainder = observed_batch_size
|
|
|
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 |
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:
|
|
|
584 |
return whole_length // self.state.num_processes
|
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(
|
596 |
dataloader: DataLoader,
|
597 |
device: Optional[torch.device] = None,
|
|
|
606 |
) -> DataLoader:
|
607 |
"""
|
608 |
Wraps a PyTorch `DataLoader` to generate batches for one of the processes only.
|
|
|
609 |
Depending on the value of the `drop_last` attribute of the `dataloader` passed, it will either stop the iteration
|
610 |
at the first batch that would be too small / not present on all processes or loop with indices from the beginning.
|
|
|
611 |
Args:
|
612 |
dataloader (`torch.utils.data.dataloader.DataLoader`):
|
613 |
The data loader to split across several devices.
|
|
|
622 |
Whether the resulting `DataLoader` should split the batches of the original data loader across devices or
|
623 |
yield full batches (in which case it will yield batches starting at the `process_index`-th and advancing of
|
624 |
`num_processes` batches at each iteration).
|
|
|
625 |
Another way to see this is that the observed batch size will be the same as the initial `dataloader` if
|
626 |
this option is set to `True`, the batch size of the initial `dataloader` multiplied by `num_processes`
|
627 |
otherwise.
|
|
|
628 |
Setting this option to `True` requires that the batch size of the `dataloader` is a round multiple of
|
629 |
`batch_size`.
|
630 |
put_on_device (`bool`, *optional*, defaults to `False`):
|
|
|
633 |
rng_types (list of `str` or [`~utils.RNGType`]):
|
634 |
The list of random number generators to synchronize at the beginning of each iteration. Should be one or
|
635 |
several of:
|
|
|
636 |
- `"torch"`: the base torch random number generator
|
637 |
- `"cuda"`: the CUDA random number generator (GPU only)
|
638 |
- `"xla"`: the XLA random number generator (TPU only)
|
639 |
- `"generator"`: the `torch.Generator` of the sampler (or batch sampler if there is no sampler in your
|
640 |
dataloader) or of the iterable dataset (if it exists) if the underlying dataset is of that type.
|
|
|
641 |
dispatch_batches (`bool`, *optional*):
|
642 |
If set to `True`, the datalaoder prepared is only iterated through on the main process and then the batches
|
643 |
are split and broadcast to each process. Will default to `True` when the underlying dataset is an
|
|
|
650 |
If passed, this function will be used to slice tensors across `num_processes`. Will default to
|
651 |
[`~utils.slice_tensors`]. This argument is used only when `dispatch_batches` is set to `True` and will be
|
652 |
ignored otherwise.
|
|
|
653 |
Returns:
|
654 |
`torch.utils.data.dataloader.DataLoader`: A new data loader that will yield the portion of the batches
|
|
|
655 |
<Tip warning={true}>
|
|
|
656 |
`BatchSampler`s with varying batch sizes are not enabled by default. To enable this behaviour, set `even_batches`
|
657 |
equal to `False`
|
|
|
658 |
</Tip>
|
659 |
"""
|
660 |
if dispatch_batches is None:
|
|
|
662 |
dispatch_batches = False
|
663 |
else:
|
664 |
dispatch_batches = isinstance(dataloader.dataset, IterableDataset)
|
|
|
665 |
if dispatch_batches and not put_on_device:
|
666 |
raise ValueError("Using `dispatch_batches=True` requires `put_on_device=True`.")
|
667 |
# Grab defaults from AcceleratorState
|
|
|
670 |
num_processes = state.num_processes
|
671 |
if process_index is None:
|
672 |
process_index = state.process_index
|
|
|
673 |
# Sanity check
|
674 |
if split_batches and dataloader.batch_size > 1 and dataloader.batch_size % num_processes != 0:
|
675 |
raise ValueError(
|
676 |
f"To use a `DataLoader` in `split_batches` mode, the batch size ({dataloader.batch_size}) "
|
677 |
f"needs to be a round multiple of the number of processes ({num_processes})."
|
678 |
)
|
|
|
679 |
new_dataset = dataloader.dataset
|
680 |
# Iterable dataset doesn't like batch_sampler, but data_loader creates a default one for it
|
681 |
new_batch_sampler = dataloader.batch_sampler if not isinstance(new_dataset, IterableDataset) else None
|
|
|
697 |
num_samples=sampler._num_samples,
|
698 |
generator=getattr(sampler, "generator", torch.Generator()),
|
699 |
)
|
|
|
700 |
# No change if no multiprocess
|
701 |
if (num_processes != 1 or state.distributed_type == DistributedType.MEGATRON_LM) and not dispatch_batches:
|
702 |
if isinstance(new_dataset, IterableDataset):
|
|
|
719 |
split_batches=split_batches,
|
720 |
even_batches=even_batches,
|
721 |
)
|
|
|
722 |
# We ignore all of those since they are all dealt with by our new_batch_sampler
|
723 |
ignore_kwargs = [
|
724 |
"batch_size",
|
|
|
727 |
"batch_sampler",
|
728 |
"drop_last",
|
729 |
]
|
|
|
730 |
if rng_types is not None and synchronized_generator is None and "generator" in rng_types:
|
731 |
rng_types.remove("generator")
|
|
|
732 |
kwargs = {
|
733 |
k: getattr(dataloader, k, _PYTORCH_DATALOADER_KWARGS[k])
|
734 |
for k in _PYTORCH_DATALOADER_KWARGS
|
735 |
if k not in ignore_kwargs
|
736 |
}
|
|
|
737 |
# Need to provide batch_size as batch_sampler is None for Iterable dataset
|
738 |
if new_batch_sampler is None:
|
739 |
kwargs["drop_last"] = dataloader.drop_last
|
|
|
776 |
_drop_last=dataloader.drop_last,
|
777 |
**kwargs,
|
778 |
)
|
|
|
779 |
if state.distributed_type == DistributedType.TPU:
|
780 |
return MpDeviceLoaderWrapper(dataloader, device)
|
781 |
return dataloader
|
|
|
|
|
782 |
class SkipBatchSampler(BatchSampler):
|
783 |
"""
|
784 |
A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
|
|
|
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):
|
799 |
"""
|
800 |
Subclass of a PyTorch `DataLoader` that will skip the first batches.
|
|
|
801 |
Args:
|
802 |
dataset (`torch.utils.data.dataset.Dataset`):
|
803 |
The dataset to use to build this datalaoder.
|
|
|
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:
|
815 |
yield batch
|
|
|
|
|
816 |
def skip_first_batches(dataloader, num_batches=0):
|
817 |
"""
|
818 |
Creates a `torch.utils.data.DataLoader` that will efficiently skip the first `num_batches`.
|
|
|
825 |
sampler_is_batch_sampler = isinstance(dataloader.sampler, BatchSampler)
|
826 |
batch_sampler = dataloader.sampler if sampler_is_batch_sampler else dataloader.batch_sampler
|
827 |
new_batch_sampler = SkipBatchSampler(batch_sampler, skip_batches=num_batches)
|
|
|
828 |
# We ignore all of those since they are all dealt with by our new_batch_sampler
|
829 |
ignore_kwargs = [
|
830 |
"batch_size",
|
|
|
833 |
"batch_sampler",
|
834 |
"drop_last",
|
835 |
]
|
|
|
836 |
kwargs = {
|
837 |
k: getattr(dataloader, k, _PYTORCH_DATALOADER_KWARGS[k])
|
838 |
for k in _PYTORCH_DATALOADER_KWARGS
|
839 |
if k not in ignore_kwargs
|
840 |
}
|
|
|
841 |
# Need to provide batch_size as batch_sampler is None for Iterable dataset
|
842 |
if new_batch_sampler is None:
|
843 |
kwargs["drop_last"] = dataloader.drop_last
|
844 |
kwargs["batch_size"] = dataloader.batch_size
|
|
|
845 |
if isinstance(dataloader, DataLoaderDispatcher):
|
846 |
if new_batch_sampler is None:
|
847 |
# Need to manually skip batches in the dataloader
|
|
|
875 |
dataloader = SkipDataLoader(dataset, skip_batches=num_batches, **kwargs)
|
876 |
else:
|
877 |
dataloader = DataLoader(dataset, batch_sampler=new_batch_sampler, **kwargs)
|
|
|
878 |
return dataloader
|
src/hooks.py
CHANGED
@@ -2,99 +2,76 @@ class ModelHook:
|
|
2 |
"""
|
3 |
A hook that contains callbacks to be executed just before and after the forward method of a model. The difference
|
4 |
with PyTorch existing hooks is that they get passed along the kwargs.
|
5 |
-
|
6 |
Class attribute:
|
7 |
- **no_grad** (`bool`, *optional*, defaults to `False`) -- Whether or not to execute the actual forward pass under
|
8 |
the `torch.no_grad()` context manager.
|
9 |
"""
|
10 |
no_grad = False
|
11 |
-
|
12 |
def init_hook(self, module):
|
13 |
"""
|
14 |
To be executed when the hook is attached to the module.
|
15 |
-
|
16 |
Args:
|
17 |
module (`torch.nn.Module`): The module attached to this hook.
|
18 |
"""
|
19 |
return module
|
20 |
-
|
21 |
def pre_forward(self, module, *args, **kwargs):
|
22 |
"""
|
23 |
To be executed just before the forward method of the model.
|
24 |
-
|
25 |
Args:
|
26 |
module (`torch.nn.Module`): The module whose forward pass will be executed just after this event.
|
27 |
args (`Tuple[Any]`): The positional arguments passed to the module.
|
28 |
kwargs (`Dict[Str, Any]`): The keyword arguments passed to the module.
|
29 |
-
|
30 |
Returns:
|
31 |
`Tuple[Tuple[Any], Dict[Str, Any]]`: A tuple with the treated `args` and `kwargs`.
|
32 |
"""
|
33 |
return args, kwargs
|
34 |
-
|
35 |
def post_forward(self, module, output):
|
36 |
"""
|
37 |
To be executed just after the forward method of the model.
|
38 |
-
|
39 |
Args:
|
40 |
module (`torch.nn.Module`): The module whose forward pass been executed just before this event.
|
41 |
output (`Any`): The output of the module.
|
42 |
-
|
43 |
Returns:
|
44 |
`Any`: The processed `output`.
|
45 |
"""
|
46 |
return output
|
47 |
-
|
48 |
def detach_hook(self, module):
|
49 |
"""
|
50 |
To be executed when the hook is detached from a module.
|
51 |
-
|
52 |
Args:
|
53 |
module (`torch.nn.Module`): The module detached from this hook.
|
54 |
"""
|
55 |
return module
|
56 |
-
|
57 |
-
|
58 |
class SequentialHook(ModelHook):
|
59 |
"""
|
60 |
A hook that can contain several hooks and iterates through them at each event.
|
61 |
"""
|
62 |
def __init__(self, *hooks):
|
63 |
self.hooks = hooks
|
64 |
-
|
65 |
def init_hook(self, module):
|
66 |
for hook in self.hooks:
|
67 |
module = hook.init_hook(module)
|
68 |
return module
|
69 |
-
|
70 |
def pre_forward(self, module, *args, **kwargs):
|
71 |
for hook in self.hooks:
|
72 |
args, kwargs = hook.pre_forward(module, *args, **kwargs)
|
73 |
return args, kwargs
|
74 |
-
|
75 |
def post_forward(self, module, output):
|
76 |
for hook in self.hooks:
|
77 |
output = hook.post_forward(module, output)
|
78 |
return output
|
79 |
-
|
80 |
def detach_hook(self, module):
|
81 |
for hook in self.hooks:
|
82 |
module = hook.detach_hook(module)
|
83 |
return module
|
84 |
-
|
85 |
-
|
86 |
def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False):
|
87 |
"""
|
88 |
Adds a hook to a given module. This will rewrite the `forward` method of the module to include the hook, to remove
|
89 |
this behavior and restore the original `forward` method, use `remove_hook_from_module`.
|
90 |
-
|
91 |
<Tip warning={true}>
|
92 |
-
|
93 |
If the module already contains a hook, this will replace it with the new hook passed by default. To chain two hooks
|
94 |
together, pass `append=True`, so it chains the current and new hook into an instance of the `SequentialHook` class.
|
95 |
-
|
96 |
</Tip>
|
97 |
-
|
98 |
Args:
|
99 |
module (`torch.nn.Module`):
|
100 |
The module to attach a hook to.
|
@@ -102,7 +79,6 @@ def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False)
|
|
102 |
The hook to attach.
|
103 |
append (`bool`, *optional*, defaults to `False`):
|
104 |
Whether the hook should be chained with an existing one (if module already contains a hook) or not.
|
105 |
-
|
106 |
Returns:
|
107 |
`torch.nn.Module`: The same module, with the hook attached (the module is modified in place, so the result can
|
108 |
be discarded).
|
@@ -111,17 +87,14 @@ def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False)
|
|
111 |
old_hook = module._hf_hook
|
112 |
remove_hook_from_module(module)
|
113 |
hook = SequentialHook(old_hook, hook)
|
114 |
-
|
115 |
if hasattr(module, "_hf_hook") and hasattr(module, "_old_forward"):
|
116 |
# If we already put some hook on this module, we replace it with the new one.
|
117 |
old_forward = module._old_forward
|
118 |
else:
|
119 |
old_forward = module.forward
|
120 |
module._old_forward = old_forward
|
121 |
-
|
122 |
module = hook.init_hook(module)
|
123 |
module._hf_hook = hook
|
124 |
-
|
125 |
def new_forward(module, *args, **kwargs):
|
126 |
args, kwargs = module._hf_hook.pre_forward(module, *args, **kwargs)
|
127 |
if module._hf_hook.no_grad:
|
@@ -130,20 +103,14 @@ def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False)
|
|
130 |
else:
|
131 |
output = module._old_forward(*args, **kwargs)
|
132 |
return module._hf_hook.post_forward(module, output)
|
133 |
-
|
134 |
module.forward = functools.update_wrapper(functools.partial(new_forward, module), old_forward)
|
135 |
-
|
136 |
return module
|
137 |
-
|
138 |
-
|
139 |
def remove_hook_from_module(module: nn.Module, recurse=False):
|
140 |
"""
|
141 |
Removes any hook attached to a module via `add_hook_to_module`.
|
142 |
-
|
143 |
Args:
|
144 |
module (`torch.nn.Module`): The module to attach a hook to.
|
145 |
recurse (`bool`, **optional**): Whether to remove the hooks recursively
|
146 |
-
|
147 |
Returns:
|
148 |
`torch.nn.Module`: The same module, with the hook detached (the module is modified in place, so the result can
|
149 |
be discarded).
|
@@ -151,23 +118,17 @@ def remove_hook_from_module(module: nn.Module, recurse=False):
|
|
151 |
if hasattr(module, "_hf_hook"):
|
152 |
module._hf_hook.detach_hook(module)
|
153 |
delattr(module, "_hf_hook")
|
154 |
-
|
155 |
if hasattr(module, "_old_forward"):
|
156 |
module.forward = module._old_forward
|
157 |
delattr(module, "_old_forward")
|
158 |
-
|
159 |
if recurse:
|
160 |
for child in module.children():
|
161 |
remove_hook_from_module(child, recurse)
|
162 |
-
|
163 |
return module
|
164 |
-
|
165 |
-
|
166 |
class AlignDevicesHook(ModelHook):
|
167 |
"""
|
168 |
A generic `ModelHook` that ensures inputs and model weights are on the same device for the forward pass of the
|
169 |
associated module, potentially offloading the weights after the forward pass.
|
170 |
-
|
171 |
Args:
|
172 |
execution_device (`torch.device`, *optional*):
|
173 |
The device on which inputs and model weights should be placed before the forward pass.
|
@@ -199,19 +160,16 @@ class AlignDevicesHook(ModelHook):
|
|
199 |
self.offload_buffers = offload_buffers
|
200 |
self.place_submodules = place_submodules
|
201 |
self.skip_keys = skip_keys
|
202 |
-
|
203 |
# Will contain the input device when `io_same_device=True`.
|
204 |
self.input_device = None
|
205 |
self.param_original_devices = {}
|
206 |
self.buffer_original_devices = {}
|
207 |
-
|
208 |
def __repr__(self):
|
209 |
return (
|
210 |
f"AlignDevicesHook(execution_device={self.execution_device}, offload={self.offload}, "
|
211 |
f"io_same_device={self.io_same_device}, offload_buffers={self.offload_buffers}, "
|
212 |
f"place_submodules={self.place_submodules}, skip_keys={repr(self.skip_keys)})"
|
213 |
)
|
214 |
-
|
215 |
def init_hook(self, module):
|
216 |
if not self.offload and self.execution_device is not None:
|
217 |
for name, _ in named_module_tensors(module, recurse=self.place_submodules):
|
@@ -237,9 +195,7 @@ class AlignDevicesHook(ModelHook):
|
|
237 |
elif self.offload_buffers and self.execution_device is not None:
|
238 |
for name in get_non_persistent_buffers(module, recurse=self.place_submodules):
|
239 |
set_module_tensor_to_device(module, name, self.execution_device)
|
240 |
-
|
241 |
return module
|
242 |
-
|
243 |
def pre_forward(self, module, *args, **kwargs):
|
244 |
if self.io_same_device:
|
245 |
self.input_device = find_device([args, kwargs])
|
@@ -257,11 +213,9 @@ class AlignDevicesHook(ModelHook):
|
|
257 |
set_module_tensor_to_device(
|
258 |
module, name, self.execution_device, value=self.weights_map[name], fp16_statistics=fp16_statistics
|
259 |
)
|
260 |
-
|
261 |
return send_to_device(args, self.execution_device), send_to_device(
|
262 |
kwargs, self.execution_device, skip_keys=self.skip_keys
|
263 |
)
|
264 |
-
|
265 |
def post_forward(self, module, output):
|
266 |
if self.offload:
|
267 |
for name, _ in named_module_tensors(
|
@@ -274,20 +228,15 @@ class AlignDevicesHook(ModelHook):
|
|
274 |
if type(module).__name__ == "Linear8bitLt":
|
275 |
module.state.SCB = None
|
276 |
module.state.CxB = None
|
277 |
-
|
278 |
if self.io_same_device and self.input_device is not None:
|
279 |
output = send_to_device(output, self.input_device, skip_keys=self.skip_keys)
|
280 |
-
|
281 |
return output
|
282 |
-
|
283 |
def detach_hook(self, module):
|
284 |
if self.offload:
|
285 |
for name, device in self.original_devices.items():
|
286 |
if device != torch.device("meta"):
|
287 |
set_module_tensor_to_device(module, name, device, value=self.weights_map.get(name, None))
|
288 |
return module
|
289 |
-
|
290 |
-
|
291 |
def attach_execution_device_hook(
|
292 |
module: torch.nn.Module,
|
293 |
execution_device: Union[int, str, torch.device],
|
@@ -297,7 +246,6 @@ def attach_execution_device_hook(
|
|
297 |
"""
|
298 |
Recursively attaches `AlignDevicesHook` to all submodules of a given model to make sure they have the right
|
299 |
execution device
|
300 |
-
|
301 |
Args:
|
302 |
module (`torch.nn.Module`):
|
303 |
The module where we want to attach the hooks.
|
@@ -313,15 +261,11 @@ def attach_execution_device_hook(
|
|
313 |
"""
|
314 |
if not hasattr(module, "_hf_hook") and len(module.state_dict()) > 0:
|
315 |
add_hook_to_module(module, AlignDevicesHook(execution_device, skip_keys=skip_keys))
|
316 |
-
|
317 |
# Break the recursion if we get to a preload module.
|
318 |
if preload_module_classes is not None and module.__class__.__name__ in preload_module_classes:
|
319 |
return
|
320 |
-
|
321 |
for child in module.children():
|
322 |
attach_execution_device_hook(child, execution_device)
|
323 |
-
|
324 |
-
|
325 |
def attach_align_device_hook(
|
326 |
module: torch.nn.Module,
|
327 |
execution_device: Optional[torch.device] = None,
|
@@ -335,7 +279,6 @@ def attach_align_device_hook(
|
|
335 |
"""
|
336 |
Recursively attaches `AlignDevicesHook` to all submodules of a given model that have direct parameters and/or
|
337 |
buffers.
|
338 |
-
|
339 |
Args:
|
340 |
module (`torch.nn.Module`):
|
341 |
The module where we want to attach the hooks.
|
@@ -362,7 +305,6 @@ def attach_align_device_hook(
|
|
362 |
full_offload = (
|
363 |
offload and preload_module_classes is not None and module.__class__.__name__ in preload_module_classes
|
364 |
)
|
365 |
-
|
366 |
if len(list(directs)) > 0 or full_offload:
|
367 |
if weights_map is not None:
|
368 |
prefix = f"{module_name}." if len(module_name) > 0 else ""
|
@@ -378,11 +320,9 @@ def attach_align_device_hook(
|
|
378 |
skip_keys=skip_keys,
|
379 |
)
|
380 |
add_hook_to_module(module, hook, append=True)
|
381 |
-
|
382 |
# We stop the recursion in case we hit the full offload.
|
383 |
if full_offload:
|
384 |
return
|
385 |
-
|
386 |
# Recurse on all children of the module.
|
387 |
for child_name, child in module.named_children():
|
388 |
child_name = f"{module_name}.{child_name}" if len(module_name) > 0 else child_name
|
@@ -396,20 +336,15 @@ def attach_align_device_hook(
|
|
396 |
preload_module_classes=preload_module_classes,
|
397 |
skip_keys=skip_keys,
|
398 |
)
|
399 |
-
|
400 |
-
|
401 |
def remove_hook_from_submodules(module: nn.Module):
|
402 |
"""
|
403 |
Recursively removes all hooks attached on the submodules of a given model.
|
404 |
-
|
405 |
Args:
|
406 |
module (`torch.nn.Module`): The module on which to remove all hooks.
|
407 |
"""
|
408 |
remove_hook_from_module(module)
|
409 |
for child in module.children():
|
410 |
remove_hook_from_submodules(child)
|
411 |
-
|
412 |
-
|
413 |
def attach_align_device_hook_on_blocks(
|
414 |
module: nn.Module,
|
415 |
execution_device: Optional[Union[torch.device, Dict[str, torch.device]]] = None,
|
@@ -422,7 +357,6 @@ def attach_align_device_hook_on_blocks(
|
|
422 |
):
|
423 |
"""
|
424 |
Attaches `AlignDevicesHook` to all blocks of a given model as needed.
|
425 |
-
|
426 |
Args:
|
427 |
module (`torch.nn.Module`):
|
428 |
The module where we want to attach the hooks.
|
@@ -464,12 +398,10 @@ def attach_align_device_hook_on_blocks(
|
|
464 |
skip_keys=skip_keys,
|
465 |
)
|
466 |
return
|
467 |
-
|
468 |
if not isinstance(execution_device, Mapping):
|
469 |
execution_device = {key: execution_device for key in offload.keys()}
|
470 |
if not isinstance(offload, Mapping):
|
471 |
offload = {key: offload for key in execution_device.keys()}
|
472 |
-
|
473 |
if module_name in execution_device and module_name in offload and not offload[module_name]:
|
474 |
hook = AlignDevicesHook(
|
475 |
execution_device=execution_device[module_name],
|
@@ -505,7 +437,6 @@ def attach_align_device_hook_on_blocks(
|
|
505 |
elif module_name == "":
|
506 |
hook = AlignDevicesHook(execution_device=execution_device.get(""), io_same_device=True, skip_keys=skip_keys)
|
507 |
add_hook_to_module(module, hook)
|
508 |
-
|
509 |
for child_name, child in module.named_children():
|
510 |
child_name = f"{module_name}.{child_name}" if len(module_name) > 0 else child_name
|
511 |
attach_align_device_hook_on_blocks(
|
@@ -518,13 +449,10 @@ def attach_align_device_hook_on_blocks(
|
|
518 |
preload_module_classes=preload_module_classes,
|
519 |
skip_keys=skip_keys,
|
520 |
)
|
521 |
-
|
522 |
-
|
523 |
class CpuOffload(ModelHook):
|
524 |
"""
|
525 |
Offloads a model on the CPU until its forward pass is called. The model will not be offloaded back to the CPU after
|
526 |
the forward, the user needs to call the `init_hook` method again for this.
|
527 |
-
|
528 |
Args:
|
529 |
execution_device(`str`, `int` or `torch.device`, *optional*):
|
530 |
The device on which the model should be executed. Will default to the MPS device if it's available, then
|
@@ -540,19 +468,14 @@ class CpuOffload(ModelHook):
|
|
540 |
prev_module_hook: Optional["UserCpuOffloadHook"] = None,
|
541 |
):
|
542 |
self.prev_module_hook = prev_module_hook
|
543 |
-
|
544 |
self.execution_device = execution_device if execution_device is not None else PartialState().default_device
|
545 |
-
|
546 |
def init_hook(self, module):
|
547 |
return module.to("cpu")
|
548 |
-
|
549 |
def pre_forward(self, module, *args, **kwargs):
|
550 |
if self.prev_module_hook is not None:
|
551 |
self.prev_module_hook.offload()
|
552 |
module.to(self.execution_device)
|
553 |
return send_to_device(args, self.execution_device), send_to_device(kwargs, self.execution_device)
|
554 |
-
|
555 |
-
|
556 |
class UserCpuOffloadHook:
|
557 |
"""
|
558 |
A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
|
@@ -561,9 +484,7 @@ class UserCpuOffloadHook:
|
|
561 |
def __init__(self, model, hook):
|
562 |
self.model = model
|
563 |
self.hook = hook
|
564 |
-
|
565 |
def offload(self):
|
566 |
self.hook.init_hook(self.model)
|
567 |
-
|
568 |
def remove(self):
|
569 |
remove_hook_from_module(self.model)
|
|
|
2 |
"""
|
3 |
A hook that contains callbacks to be executed just before and after the forward method of a model. The difference
|
4 |
with PyTorch existing hooks is that they get passed along the kwargs.
|
|
|
5 |
Class attribute:
|
6 |
- **no_grad** (`bool`, *optional*, defaults to `False`) -- Whether or not to execute the actual forward pass under
|
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.
|
|
|
13 |
Args:
|
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.
|
|
|
20 |
Args:
|
21 |
module (`torch.nn.Module`): The module whose forward pass will be executed just after this event.
|
22 |
args (`Tuple[Any]`): The positional arguments passed to the module.
|
23 |
kwargs (`Dict[Str, Any]`): The keyword arguments passed to the module.
|
|
|
24 |
Returns:
|
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.
|
|
|
31 |
Args:
|
32 |
module (`torch.nn.Module`): The module whose forward pass been executed just before this event.
|
33 |
output (`Any`): The output of the module.
|
|
|
34 |
Returns:
|
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.
|
|
|
41 |
Args:
|
42 |
module (`torch.nn.Module`): The module detached from this hook.
|
43 |
"""
|
44 |
return module
|
|
|
|
|
45 |
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)
|
66 |
return module
|
|
|
|
|
67 |
def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False):
|
68 |
"""
|
69 |
Adds a hook to a given module. This will rewrite the `forward` method of the module to include the hook, to remove
|
70 |
this behavior and restore the original `forward` method, use `remove_hook_from_module`.
|
|
|
71 |
<Tip warning={true}>
|
|
|
72 |
If the module already contains a hook, this will replace it with the new hook passed by default. To chain two hooks
|
73 |
together, pass `append=True`, so it chains the current and new hook into an instance of the `SequentialHook` class.
|
|
|
74 |
</Tip>
|
|
|
75 |
Args:
|
76 |
module (`torch.nn.Module`):
|
77 |
The module to attach a hook to.
|
|
|
79 |
The hook to attach.
|
80 |
append (`bool`, *optional*, defaults to `False`):
|
81 |
Whether the hook should be chained with an existing one (if module already contains a hook) or not.
|
|
|
82 |
Returns:
|
83 |
`torch.nn.Module`: The same module, with the hook attached (the module is modified in place, so the result can
|
84 |
be discarded).
|
|
|
87 |
old_hook = module._hf_hook
|
88 |
remove_hook_from_module(module)
|
89 |
hook = SequentialHook(old_hook, hook)
|
|
|
90 |
if hasattr(module, "_hf_hook") and hasattr(module, "_old_forward"):
|
91 |
# If we already put some hook on this module, we replace it with the new one.
|
92 |
old_forward = module._old_forward
|
93 |
else:
|
94 |
old_forward = module.forward
|
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:
|
|
|
103 |
else:
|
104 |
output = module._old_forward(*args, **kwargs)
|
105 |
return module._hf_hook.post_forward(module, output)
|
|
|
106 |
module.forward = functools.update_wrapper(functools.partial(new_forward, module), old_forward)
|
|
|
107 |
return module
|
|
|
|
|
108 |
def remove_hook_from_module(module: nn.Module, recurse=False):
|
109 |
"""
|
110 |
Removes any hook attached to a module via `add_hook_to_module`.
|
|
|
111 |
Args:
|
112 |
module (`torch.nn.Module`): The module to attach a hook to.
|
113 |
recurse (`bool`, **optional**): Whether to remove the hooks recursively
|
|
|
114 |
Returns:
|
115 |
`torch.nn.Module`: The same module, with the hook detached (the module is modified in place, so the result can
|
116 |
be discarded).
|
|
|
118 |
if hasattr(module, "_hf_hook"):
|
119 |
module._hf_hook.detach_hook(module)
|
120 |
delattr(module, "_hf_hook")
|
|
|
121 |
if hasattr(module, "_old_forward"):
|
122 |
module.forward = module._old_forward
|
123 |
delattr(module, "_old_forward")
|
|
|
124 |
if recurse:
|
125 |
for child in module.children():
|
126 |
remove_hook_from_module(child, recurse)
|
|
|
127 |
return module
|
|
|
|
|
128 |
class AlignDevicesHook(ModelHook):
|
129 |
"""
|
130 |
A generic `ModelHook` that ensures inputs and model weights are on the same device for the forward pass of the
|
131 |
associated module, potentially offloading the weights after the forward pass.
|
|
|
132 |
Args:
|
133 |
execution_device (`torch.device`, *optional*):
|
134 |
The device on which inputs and model weights should be placed before the forward pass.
|
|
|
160 |
self.offload_buffers = offload_buffers
|
161 |
self.place_submodules = place_submodules
|
162 |
self.skip_keys = skip_keys
|
|
|
163 |
# Will contain the input device when `io_same_device=True`.
|
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):
|
|
|
195 |
elif self.offload_buffers and self.execution_device is not None:
|
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])
|
|
|
213 |
set_module_tensor_to_device(
|
214 |
module, name, self.execution_device, value=self.weights_map[name], fp16_statistics=fp16_statistics
|
215 |
)
|
|
|
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(
|
|
|
228 |
if type(module).__name__ == "Linear8bitLt":
|
229 |
module.state.SCB = None
|
230 |
module.state.CxB = None
|
|
|
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():
|
237 |
if device != torch.device("meta"):
|
238 |
set_module_tensor_to_device(module, name, device, value=self.weights_map.get(name, None))
|
239 |
return module
|
|
|
|
|
240 |
def attach_execution_device_hook(
|
241 |
module: torch.nn.Module,
|
242 |
execution_device: Union[int, str, torch.device],
|
|
|
246 |
"""
|
247 |
Recursively attaches `AlignDevicesHook` to all submodules of a given model to make sure they have the right
|
248 |
execution device
|
|
|
249 |
Args:
|
250 |
module (`torch.nn.Module`):
|
251 |
The module where we want to attach the hooks.
|
|
|
261 |
"""
|
262 |
if not hasattr(module, "_hf_hook") and len(module.state_dict()) > 0:
|
263 |
add_hook_to_module(module, AlignDevicesHook(execution_device, skip_keys=skip_keys))
|
|
|
264 |
# Break the recursion if we get to a preload module.
|
265 |
if preload_module_classes is not None and module.__class__.__name__ in preload_module_classes:
|
266 |
return
|
|
|
267 |
for child in module.children():
|
268 |
attach_execution_device_hook(child, execution_device)
|
|
|
|
|
269 |
def attach_align_device_hook(
|
270 |
module: torch.nn.Module,
|
271 |
execution_device: Optional[torch.device] = None,
|
|
|
279 |
"""
|
280 |
Recursively attaches `AlignDevicesHook` to all submodules of a given model that have direct parameters and/or
|
281 |
buffers.
|
|
|
282 |
Args:
|
283 |
module (`torch.nn.Module`):
|
284 |
The module where we want to attach the hooks.
|
|
|
305 |
full_offload = (
|
306 |
offload and preload_module_classes is not None and module.__class__.__name__ in preload_module_classes
|
307 |
)
|
|
|
308 |
if len(list(directs)) > 0 or full_offload:
|
309 |
if weights_map is not None:
|
310 |
prefix = f"{module_name}." if len(module_name) > 0 else ""
|
|
|
320 |
skip_keys=skip_keys,
|
321 |
)
|
322 |
add_hook_to_module(module, hook, append=True)
|
|
|
323 |
# We stop the recursion in case we hit the full offload.
|
324 |
if full_offload:
|
325 |
return
|
|
|
326 |
# Recurse on all children of the module.
|
327 |
for child_name, child in module.named_children():
|
328 |
child_name = f"{module_name}.{child_name}" if len(module_name) > 0 else child_name
|
|
|
336 |
preload_module_classes=preload_module_classes,
|
337 |
skip_keys=skip_keys,
|
338 |
)
|
|
|
|
|
339 |
def remove_hook_from_submodules(module: nn.Module):
|
340 |
"""
|
341 |
Recursively removes all hooks attached on the submodules of a given model.
|
|
|
342 |
Args:
|
343 |
module (`torch.nn.Module`): The module on which to remove all hooks.
|
344 |
"""
|
345 |
remove_hook_from_module(module)
|
346 |
for child in module.children():
|
347 |
remove_hook_from_submodules(child)
|
|
|
|
|
348 |
def attach_align_device_hook_on_blocks(
|
349 |
module: nn.Module,
|
350 |
execution_device: Optional[Union[torch.device, Dict[str, torch.device]]] = None,
|
|
|
357 |
):
|
358 |
"""
|
359 |
Attaches `AlignDevicesHook` to all blocks of a given model as needed.
|
|
|
360 |
Args:
|
361 |
module (`torch.nn.Module`):
|
362 |
The module where we want to attach the hooks.
|
|
|
398 |
skip_keys=skip_keys,
|
399 |
)
|
400 |
return
|
|
|
401 |
if not isinstance(execution_device, Mapping):
|
402 |
execution_device = {key: execution_device for key in offload.keys()}
|
403 |
if not isinstance(offload, Mapping):
|
404 |
offload = {key: offload for key in execution_device.keys()}
|
|
|
405 |
if module_name in execution_device and module_name in offload and not offload[module_name]:
|
406 |
hook = AlignDevicesHook(
|
407 |
execution_device=execution_device[module_name],
|
|
|
437 |
elif module_name == "":
|
438 |
hook = AlignDevicesHook(execution_device=execution_device.get(""), io_same_device=True, skip_keys=skip_keys)
|
439 |
add_hook_to_module(module, hook)
|
|
|
440 |
for child_name, child in module.named_children():
|
441 |
child_name = f"{module_name}.{child_name}" if len(module_name) > 0 else child_name
|
442 |
attach_align_device_hook_on_blocks(
|
|
|
449 |
preload_module_classes=preload_module_classes,
|
450 |
skip_keys=skip_keys,
|
451 |
)
|
|
|
|
|
452 |
class CpuOffload(ModelHook):
|
453 |
"""
|
454 |
Offloads a model on the CPU until its forward pass is called. The model will not be offloaded back to the CPU after
|
455 |
the forward, the user needs to call the `init_hook` method again for this.
|
|
|
456 |
Args:
|
457 |
execution_device(`str`, `int` or `torch.device`, *optional*):
|
458 |
The device on which the model should be executed. Will default to the MPS device if it's available, then
|
|
|
468 |
prev_module_hook: Optional["UserCpuOffloadHook"] = None,
|
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()
|
477 |
module.to(self.execution_device)
|
478 |
return send_to_device(args, self.execution_device), send_to_device(kwargs, self.execution_device)
|
|
|
|
|
479 |
class UserCpuOffloadHook:
|
480 |
"""
|
481 |
A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
|
|
|
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)
|
src/launchers.py
CHANGED
@@ -1,8 +1,6 @@
|
|
1 |
def test_launch():
|
2 |
"Verify a `PartialState` can be initialized."
|
3 |
_ = PartialState()
|
4 |
-
|
5 |
-
|
6 |
def notebook_launcher(
|
7 |
function,
|
8 |
args=(),
|
@@ -16,17 +14,12 @@ def notebook_launcher(
|
|
16 |
"""
|
17 |
Launches a training function, using several processes or multiple nodes if it's possible in the current environment
|
18 |
(TPU with multiple cores for instance).
|
19 |
-
|
20 |
<Tip warning={true}>
|
21 |
-
|
22 |
To use this function absolutely zero calls to a CUDA device must be made in the notebook session before calling. If
|
23 |
any have been made, you will need to restart the notebook and make sure no cells use any CUDA capability.
|
24 |
-
|
25 |
Setting `ACCELERATE_DEBUG_MODE="1"` in your environment will run a test before truly launching to ensure that none
|
26 |
of those calls have been made.
|
27 |
-
|
28 |
</Tip>
|
29 |
-
|
30 |
Args:
|
31 |
function (`Callable`):
|
32 |
The training function to execute. If it accepts arguments, the first argument should be the index of the
|
@@ -46,19 +39,13 @@ def notebook_launcher(
|
|
46 |
The rank of the current node.
|
47 |
num_nodes (`int`, *optional*, defaults to 1):
|
48 |
The number of nodes to use for training.
|
49 |
-
|
50 |
Example:
|
51 |
-
|
52 |
```python
|
53 |
# Assume this is defined in a Jupyter Notebook on an instance with two GPUs
|
54 |
from accelerate import notebook_launcher
|
55 |
-
|
56 |
-
|
57 |
def train(*args):
|
58 |
# Your training function here
|
59 |
...
|
60 |
-
|
61 |
-
|
62 |
notebook_launcher(train, args=(arg1, arg2), num_processes=2, mixed_precision="fp16")
|
63 |
```
|
64 |
"""
|
@@ -69,18 +56,15 @@ def notebook_launcher(
|
|
69 |
in_kaggle = True
|
70 |
elif "IPython" in sys.modules:
|
71 |
in_colab = "google.colab" in str(sys.modules["IPython"].get_ipython())
|
72 |
-
|
73 |
try:
|
74 |
mixed_precision = PrecisionType(mixed_precision.lower())
|
75 |
except ValueError:
|
76 |
raise ValueError(
|
77 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
78 |
)
|
79 |
-
|
80 |
if (in_colab or in_kaggle) and (os.environ.get("TPU_NAME", None) is not None):
|
81 |
# TPU launch
|
82 |
import torch_xla.distributed.xla_multiprocessing as xmp
|
83 |
-
|
84 |
if len(AcceleratorState._shared_state) > 0:
|
85 |
raise ValueError(
|
86 |
"To train on TPU in Colab or Kaggle Kernel, the `Accelerator` should only be initialized inside "
|
@@ -89,7 +73,6 @@ def notebook_launcher(
|
|
89 |
)
|
90 |
if num_processes is None:
|
91 |
num_processes = 8
|
92 |
-
|
93 |
launcher = PrepareForLaunch(function, distributed_type="TPU")
|
94 |
print(f"Launching a training on {num_processes} TPU cores.")
|
95 |
xmp.spawn(launcher, args=args, nprocs=num_processes, start_method="fork")
|
@@ -111,7 +94,6 @@ def notebook_launcher(
|
|
111 |
# Multi-GPU launch
|
112 |
from torch.multiprocessing import start_processes
|
113 |
from torch.multiprocessing.spawn import ProcessRaisedException
|
114 |
-
|
115 |
if len(AcceleratorState._shared_state) > 0:
|
116 |
raise ValueError(
|
117 |
"To launch a multi-GPU training from your notebook, the `Accelerator` should only be initialized "
|
@@ -128,7 +110,6 @@ def notebook_launcher(
|
|
128 |
for lib_name in problematic_imports:
|
129 |
err += f"\n\t* `{lib_name}`"
|
130 |
raise RuntimeError(err)
|
131 |
-
|
132 |
patched_env = dict(
|
133 |
nproc=num_processes,
|
134 |
node_rank=node_rank,
|
@@ -137,12 +118,10 @@ def notebook_launcher(
|
|
137 |
master_port=use_port,
|
138 |
mixed_precision=mixed_precision,
|
139 |
)
|
140 |
-
|
141 |
# Check for CUDA P2P and IB issues
|
142 |
if not check_cuda_p2p_ib_support():
|
143 |
patched_env["nccl_p2p_disable"] = "1"
|
144 |
patched_env["nccl_ib_disable"] = "1"
|
145 |
-
|
146 |
# torch.distributed will expect a few environment variable to be here. We set the ones common to each
|
147 |
# process here (the other ones will be set be the launcher).
|
148 |
with patch_environment(**patched_env):
|
@@ -177,7 +156,6 @@ def notebook_launcher(
|
|
177 |
) from e
|
178 |
else:
|
179 |
raise RuntimeError(f"An issue was found when launching the training: {e}") from e
|
180 |
-
|
181 |
else:
|
182 |
# No need for a distributed launch otherwise as it's either CPU, GPU or MPS.
|
183 |
if is_mps_available():
|
@@ -188,19 +166,13 @@ def notebook_launcher(
|
|
188 |
else:
|
189 |
print("Launching training on CPU.")
|
190 |
function(*args)
|
191 |
-
|
192 |
-
|
193 |
def debug_launcher(function, args=(), num_processes=2):
|
194 |
"""
|
195 |
Launches a training function using several processes on CPU for debugging purposes.
|
196 |
-
|
197 |
<Tip warning={true}>
|
198 |
-
|
199 |
This function is provided for internal testing and debugging, but it's not intended for real trainings. It will
|
200 |
only use the CPU.
|
201 |
-
|
202 |
</Tip>
|
203 |
-
|
204 |
Args:
|
205 |
function (`Callable`):
|
206 |
The training function to execute.
|
@@ -210,7 +182,6 @@ def debug_launcher(function, args=(), num_processes=2):
|
|
210 |
The number of processes to use for training.
|
211 |
"""
|
212 |
from torch.multiprocessing import start_processes
|
213 |
-
|
214 |
with tempfile.NamedTemporaryFile() as tmp_file:
|
215 |
# torch.distributed will expect a few environment variable to be here. We set the ones common to each
|
216 |
# process here (the other ones will be set be the launcher).
|
|
|
1 |
def test_launch():
|
2 |
"Verify a `PartialState` can be initialized."
|
3 |
_ = PartialState()
|
|
|
|
|
4 |
def notebook_launcher(
|
5 |
function,
|
6 |
args=(),
|
|
|
14 |
"""
|
15 |
Launches a training function, using several processes or multiple nodes if it's possible in the current environment
|
16 |
(TPU with multiple cores for instance).
|
|
|
17 |
<Tip warning={true}>
|
|
|
18 |
To use this function absolutely zero calls to a CUDA device must be made in the notebook session before calling. If
|
19 |
any have been made, you will need to restart the notebook and make sure no cells use any CUDA capability.
|
|
|
20 |
Setting `ACCELERATE_DEBUG_MODE="1"` in your environment will run a test before truly launching to ensure that none
|
21 |
of those calls have been made.
|
|
|
22 |
</Tip>
|
|
|
23 |
Args:
|
24 |
function (`Callable`):
|
25 |
The training function to execute. If it accepts arguments, the first argument should be the index of the
|
|
|
39 |
The rank of the current node.
|
40 |
num_nodes (`int`, *optional*, defaults to 1):
|
41 |
The number of nodes to use for training.
|
|
|
42 |
Example:
|
|
|
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 |
...
|
|
|
|
|
49 |
notebook_launcher(train, args=(arg1, arg2), num_processes=2, mixed_precision="fp16")
|
50 |
```
|
51 |
"""
|
|
|
56 |
in_kaggle = True
|
57 |
elif "IPython" in sys.modules:
|
58 |
in_colab = "google.colab" in str(sys.modules["IPython"].get_ipython())
|
|
|
59 |
try:
|
60 |
mixed_precision = PrecisionType(mixed_precision.lower())
|
61 |
except ValueError:
|
62 |
raise ValueError(
|
63 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
64 |
)
|
|
|
65 |
if (in_colab or in_kaggle) and (os.environ.get("TPU_NAME", None) is not None):
|
66 |
# TPU launch
|
67 |
import torch_xla.distributed.xla_multiprocessing as xmp
|
|
|
68 |
if len(AcceleratorState._shared_state) > 0:
|
69 |
raise ValueError(
|
70 |
"To train on TPU in Colab or Kaggle Kernel, the `Accelerator` should only be initialized inside "
|
|
|
73 |
)
|
74 |
if num_processes is None:
|
75 |
num_processes = 8
|
|
|
76 |
launcher = PrepareForLaunch(function, distributed_type="TPU")
|
77 |
print(f"Launching a training on {num_processes} TPU cores.")
|
78 |
xmp.spawn(launcher, args=args, nprocs=num_processes, start_method="fork")
|
|
|
94 |
# Multi-GPU launch
|
95 |
from torch.multiprocessing import start_processes
|
96 |
from torch.multiprocessing.spawn import ProcessRaisedException
|
|
|
97 |
if len(AcceleratorState._shared_state) > 0:
|
98 |
raise ValueError(
|
99 |
"To launch a multi-GPU training from your notebook, the `Accelerator` should only be initialized "
|
|
|
110 |
for lib_name in problematic_imports:
|
111 |
err += f"\n\t* `{lib_name}`"
|
112 |
raise RuntimeError(err)
|
|
|
113 |
patched_env = dict(
|
114 |
nproc=num_processes,
|
115 |
node_rank=node_rank,
|
|
|
118 |
master_port=use_port,
|
119 |
mixed_precision=mixed_precision,
|
120 |
)
|
|
|
121 |
# Check for CUDA P2P and IB issues
|
122 |
if not check_cuda_p2p_ib_support():
|
123 |
patched_env["nccl_p2p_disable"] = "1"
|
124 |
patched_env["nccl_ib_disable"] = "1"
|
|
|
125 |
# torch.distributed will expect a few environment variable to be here. We set the ones common to each
|
126 |
# process here (the other ones will be set be the launcher).
|
127 |
with patch_environment(**patched_env):
|
|
|
156 |
) from e
|
157 |
else:
|
158 |
raise RuntimeError(f"An issue was found when launching the training: {e}") from e
|
|
|
159 |
else:
|
160 |
# No need for a distributed launch otherwise as it's either CPU, GPU or MPS.
|
161 |
if is_mps_available():
|
|
|
166 |
else:
|
167 |
print("Launching training on CPU.")
|
168 |
function(*args)
|
|
|
|
|
169 |
def debug_launcher(function, args=(), num_processes=2):
|
170 |
"""
|
171 |
Launches a training function using several processes on CPU for debugging purposes.
|
|
|
172 |
<Tip warning={true}>
|
|
|
173 |
This function is provided for internal testing and debugging, but it's not intended for real trainings. It will
|
174 |
only use the CPU.
|
|
|
175 |
</Tip>
|
|
|
176 |
Args:
|
177 |
function (`Callable`):
|
178 |
The training function to execute.
|
|
|
182 |
The number of processes to use for training.
|
183 |
"""
|
184 |
from torch.multiprocessing import start_processes
|
|
|
185 |
with tempfile.NamedTemporaryFile() as tmp_file:
|
186 |
# torch.distributed will expect a few environment variable to be here. We set the ones common to each
|
187 |
# process here (the other ones will be set be the launcher).
|
src/local_sgd.py
CHANGED
@@ -2,40 +2,29 @@ class LocalSGD:
|
|
2 |
"""
|
3 |
A helper class to support local SGD on top of Accelerator. It simply runs a given number of updates independently
|
4 |
on each device, and averages model weights every K synchronization step.
|
5 |
-
|
6 |
It should be used only in the multi-GPU (or multi-CPU) setup without extensions such as DeepSpeed. In particular,
|
7 |
this is a simple implementation that cannot support scenarios such as model parallelism.
|
8 |
-
|
9 |
-
|
10 |
Although we are not aware of the true origins of this simple approach, the idea of local SGD is quite old and goes
|
11 |
back to at least:
|
12 |
-
|
13 |
Zhang, J., De Sa, C., Mitliagkas, I., & Ré, C. (2016). [Parallel SGD: When does averaging help?. arXiv preprint
|
14 |
arXiv:1606.07365.](https://arxiv.org/abs/1606.07365)
|
15 |
-
|
16 |
We credit the term Local SGD to the following paper (but there might be earlier references we are not aware of).
|
17 |
-
|
18 |
Stich, Sebastian Urban. ["Local SGD Converges Fast and Communicates Little." ICLR 2019-International Conference on
|
19 |
Learning Representations. No. CONF. 2019.](https://arxiv.org/abs/1805.09767)
|
20 |
-
|
21 |
"""
|
22 |
def __enter__(self):
|
23 |
if self.enabled:
|
24 |
self.model_sync_obj = self.model.no_sync()
|
25 |
self.model_sync_obj.__enter__()
|
26 |
-
|
27 |
return self
|
28 |
-
|
29 |
def __exit__(self, type, value, tb):
|
30 |
if self.enabled:
|
31 |
# Average all models on exit
|
32 |
self._sync_and_avg_model_params()
|
33 |
self.model_sync_obj.__exit__(type, value, tb)
|
34 |
-
|
35 |
def __init__(self, accelerator: Accelerator, model: torch.nn.Module, local_sgd_steps: int, enabled: bool = True):
|
36 |
"""
|
37 |
Constructor.
|
38 |
-
|
39 |
Args:
|
40 |
model (`torch.nn.Module):
|
41 |
The model whose parameters we need to average.
|
@@ -58,7 +47,6 @@ class LocalSGD:
|
|
58 |
self.accelerator = accelerator
|
59 |
self.model = model
|
60 |
self.local_sgd_steps = local_sgd_steps
|
61 |
-
|
62 |
def step(self):
|
63 |
"""
|
64 |
This function makes a "step" and synchronizes model parameters if necessary.
|
@@ -66,10 +54,8 @@ class LocalSGD:
|
|
66 |
self.num_steps += 1
|
67 |
if not self.enabled:
|
68 |
return
|
69 |
-
|
70 |
if self.num_steps % self.local_sgd_steps == 0:
|
71 |
self._sync_and_avg_model_params()
|
72 |
-
|
73 |
def _sync_and_avg_model_params(self):
|
74 |
"""
|
75 |
Synchronize + Average model parameters across all GPUs
|
|
|
2 |
"""
|
3 |
A helper class to support local SGD on top of Accelerator. It simply runs a given number of updates independently
|
4 |
on each device, and averages model weights every K synchronization step.
|
|
|
5 |
It should be used only in the multi-GPU (or multi-CPU) setup without extensions such as DeepSpeed. In particular,
|
6 |
this is a simple implementation that cannot support scenarios such as model parallelism.
|
|
|
|
|
7 |
Although we are not aware of the true origins of this simple approach, the idea of local SGD is quite old and goes
|
8 |
back to at least:
|
|
|
9 |
Zhang, J., De Sa, C., Mitliagkas, I., & Ré, C. (2016). [Parallel SGD: When does averaging help?. arXiv preprint
|
10 |
arXiv:1606.07365.](https://arxiv.org/abs/1606.07365)
|
|
|
11 |
We credit the term Local SGD to the following paper (but there might be earlier references we are not aware of).
|
|
|
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.
|
|
|
28 |
Args:
|
29 |
model (`torch.nn.Module):
|
30 |
The model whose parameters we need to average.
|
|
|
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.
|
|
|
54 |
self.num_steps += 1
|
55 |
if not self.enabled:
|
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
|
src/logging.py
CHANGED
@@ -1,10 +1,8 @@
|
|
1 |
class MultiProcessAdapter(logging.LoggerAdapter):
|
2 |
"""
|
3 |
An adapter to assist with logging in multiprocess.
|
4 |
-
|
5 |
`log` takes in an additional `main_process_only` kwarg, which dictates whether it should be called on all processes
|
6 |
or only the main executed one. Default is `main_process_only=True`.
|
7 |
-
|
8 |
Does not require an `Accelerator` object to be created first.
|
9 |
"""
|
10 |
@staticmethod
|
@@ -12,18 +10,14 @@ class MultiProcessAdapter(logging.LoggerAdapter):
|
|
12 |
"Check if log should be performed"
|
13 |
state = PartialState()
|
14 |
return not main_process_only or (main_process_only and state.is_main_process)
|
15 |
-
|
16 |
def log(self, level, msg, *args, **kwargs):
|
17 |
"""
|
18 |
Delegates logger call after checking if we should log.
|
19 |
-
|
20 |
Accepts a new kwarg of `main_process_only`, which will dictate whether it will be logged across all processes
|
21 |
or only the main executed one. Default is `True` if not passed
|
22 |
-
|
23 |
Also accepts "in_order", which if `True` makes the processes log one by one, in order. This is much easier to
|
24 |
read, but comes at the cost of sometimes needing to wait for the other processes. Default is `False` to not
|
25 |
break with the previous behavior.
|
26 |
-
|
27 |
`in_order` is ignored if `main_process_only` is passed.
|
28 |
"""
|
29 |
if PartialState._shared_state == {}:
|
@@ -32,12 +26,10 @@ class MultiProcessAdapter(logging.LoggerAdapter):
|
|
32 |
)
|
33 |
main_process_only = kwargs.pop("main_process_only", True)
|
34 |
in_order = kwargs.pop("in_order", False)
|
35 |
-
|
36 |
if self.isEnabledFor(level):
|
37 |
if self._should_log(main_process_only):
|
38 |
msg, kwargs = self.process(msg, kwargs)
|
39 |
self.logger.log(level, msg, *args, **kwargs)
|
40 |
-
|
41 |
elif in_order:
|
42 |
state = PartialState()
|
43 |
for i in range(state.num_processes):
|
@@ -45,48 +37,36 @@ class MultiProcessAdapter(logging.LoggerAdapter):
|
|
45 |
msg, kwargs = self.process(msg, kwargs)
|
46 |
self.logger.log(level, msg, *args, **kwargs)
|
47 |
state.wait_for_everyone()
|
48 |
-
|
49 |
@functools.lru_cache(None)
|
50 |
def warning_once(self, *args, **kwargs):
|
51 |
"""
|
52 |
This method is identical to `logger.warning()`, but will emit the warning with the same message only once
|
53 |
-
|
54 |
Note: The cache is for the function arguments, so 2 different callers using the same arguments will hit the
|
55 |
cache. The assumption here is that all warning messages are unique across the code. If they aren't then need to
|
56 |
switch to another type of cache that includes the caller frame information in the hashing function.
|
57 |
"""
|
58 |
self.warning(*args, **kwargs)
|
59 |
-
|
60 |
-
|
61 |
def get_logger(name: str, log_level: str = None):
|
62 |
"""
|
63 |
Returns a `logging.Logger` for `name` that can handle multiprocessing.
|
64 |
-
|
65 |
If a log should be called on all processes, pass `main_process_only=False` If a log should be called on all
|
66 |
processes and in order, also pass `in_order=True`
|
67 |
-
|
68 |
Args:
|
69 |
name (`str`):
|
70 |
The name for the logger, such as `__file__`
|
71 |
log_level (`str`, *optional*):
|
72 |
The log level to use. If not passed, will default to the `LOG_LEVEL` environment variable, or `INFO` if not
|
73 |
-
|
74 |
Example:
|
75 |
-
|
76 |
```python
|
77 |
>>> from accelerate.logging import get_logger
|
78 |
>>> from accelerate import Accelerator
|
79 |
-
|
80 |
>>> logger = get_logger(__name__)
|
81 |
-
|
82 |
>>> accelerator = Accelerator()
|
83 |
>>> logger.info("My log", main_process_only=False)
|
84 |
>>> logger.debug("My log", main_process_only=True)
|
85 |
-
|
86 |
>>> logger = get_logger(__name__, log_level="DEBUG")
|
87 |
>>> logger.info("My log")
|
88 |
>>> logger.debug("My second log")
|
89 |
-
|
90 |
>>> array = ["a", "b", "c", "d"]
|
91 |
>>> letter_at_rank = array[accelerator.process_index]
|
92 |
>>> logger.info(letter_at_rank, in_order=True)
|
|
|
1 |
class MultiProcessAdapter(logging.LoggerAdapter):
|
2 |
"""
|
3 |
An adapter to assist with logging in multiprocess.
|
|
|
4 |
`log` takes in an additional `main_process_only` kwarg, which dictates whether it should be called on all processes
|
5 |
or only the main executed one. Default is `main_process_only=True`.
|
|
|
6 |
Does not require an `Accelerator` object to be created first.
|
7 |
"""
|
8 |
@staticmethod
|
|
|
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.
|
|
|
16 |
Accepts a new kwarg of `main_process_only`, which will dictate whether it will be logged across all processes
|
17 |
or only the main executed one. Default is `True` if not passed
|
|
|
18 |
Also accepts "in_order", which if `True` makes the processes log one by one, in order. This is much easier to
|
19 |
read, but comes at the cost of sometimes needing to wait for the other processes. Default is `False` to not
|
20 |
break with the previous behavior.
|
|
|
21 |
`in_order` is ignored if `main_process_only` is passed.
|
22 |
"""
|
23 |
if PartialState._shared_state == {}:
|
|
|
26 |
)
|
27 |
main_process_only = kwargs.pop("main_process_only", True)
|
28 |
in_order = kwargs.pop("in_order", False)
|
|
|
29 |
if self.isEnabledFor(level):
|
30 |
if self._should_log(main_process_only):
|
31 |
msg, kwargs = self.process(msg, kwargs)
|
32 |
self.logger.log(level, msg, *args, **kwargs)
|
|
|
33 |
elif in_order:
|
34 |
state = PartialState()
|
35 |
for i in range(state.num_processes):
|
|
|
37 |
msg, kwargs = self.process(msg, kwargs)
|
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
|
|
|
44 |
Note: The cache is for the function arguments, so 2 different callers using the same arguments will hit the
|
45 |
cache. The assumption here is that all warning messages are unique across the code. If they aren't then need to
|
46 |
switch to another type of cache that includes the caller frame information in the hashing function.
|
47 |
"""
|
48 |
self.warning(*args, **kwargs)
|
|
|
|
|
49 |
def get_logger(name: str, log_level: str = None):
|
50 |
"""
|
51 |
Returns a `logging.Logger` for `name` that can handle multiprocessing.
|
|
|
52 |
If a log should be called on all processes, pass `main_process_only=False` If a log should be called on all
|
53 |
processes and in order, also pass `in_order=True`
|
|
|
54 |
Args:
|
55 |
name (`str`):
|
56 |
The name for the logger, such as `__file__`
|
57 |
log_level (`str`, *optional*):
|
58 |
The log level to use. If not passed, will default to the `LOG_LEVEL` environment variable, or `INFO` if not
|
|
|
59 |
Example:
|
|
|
60 |
```python
|
61 |
>>> from accelerate.logging import get_logger
|
62 |
>>> from accelerate import Accelerator
|
|
|
63 |
>>> logger = get_logger(__name__)
|
|
|
64 |
>>> accelerator = Accelerator()
|
65 |
>>> logger.info("My log", main_process_only=False)
|
66 |
>>> logger.debug("My log", main_process_only=True)
|
|
|
67 |
>>> logger = get_logger(__name__, log_level="DEBUG")
|
68 |
>>> logger.info("My log")
|
69 |
>>> logger.debug("My second log")
|
|
|
70 |
>>> array = ["a", "b", "c", "d"]
|
71 |
>>> letter_at_rank = array[accelerator.process_index]
|
72 |
>>> logger.info(letter_at_rank, in_order=True)
|
src/optimizer.py
CHANGED
@@ -6,15 +6,11 @@ def move_to_device(state, device):
|
|
6 |
elif isinstance(state, torch.Tensor):
|
7 |
return state.to(device)
|
8 |
return state
|
9 |
-
|
10 |
-
|
11 |
class AcceleratedOptimizer(torch.optim.Optimizer):
|
12 |
"""
|
13 |
Internal wrapper around a torch optimizer.
|
14 |
-
|
15 |
Conditionally will perform `step` and `zero_grad` if gradients should be synchronized when performing gradient
|
16 |
accumulation.
|
17 |
-
|
18 |
Args:
|
19 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
20 |
The optimizer to wrap.
|
@@ -31,12 +27,10 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
31 |
self.gradient_state = GradientState()
|
32 |
self.device_placement = device_placement
|
33 |
self._is_overflow = False
|
34 |
-
|
35 |
if self.scaler is not None:
|
36 |
self._accelerate_step_called = False
|
37 |
self._optimizer_original_step_method = self.optimizer.step
|
38 |
self._optimizer_patched_step_method = patch_optimizer_step(self, self.optimizer.step)
|
39 |
-
|
40 |
# Handle device placement
|
41 |
if device_placement:
|
42 |
state_dict = self.optimizer.state_dict()
|
@@ -45,42 +39,32 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
45 |
else:
|
46 |
state_dict = move_to_device(state_dict, self.accelerator_state.device)
|
47 |
self.optimizer.load_state_dict(state_dict)
|
48 |
-
|
49 |
@property
|
50 |
def state(self):
|
51 |
return self.optimizer.state
|
52 |
-
|
53 |
@state.setter
|
54 |
def state(self, state):
|
55 |
self.optimizer.state = state
|
56 |
-
|
57 |
@property
|
58 |
def param_groups(self):
|
59 |
return self.optimizer.param_groups
|
60 |
-
|
61 |
@param_groups.setter
|
62 |
def param_groups(self, param_groups):
|
63 |
self.optimizer.param_groups = param_groups
|
64 |
-
|
65 |
@property
|
66 |
def defaults(self):
|
67 |
return self.optimizer.defaults
|
68 |
-
|
69 |
@defaults.setter
|
70 |
def defaults(self, defaults):
|
71 |
self.optimizer.defaults = defaults
|
72 |
-
|
73 |
def add_param_group(self, param_group):
|
74 |
self.optimizer.add_param_group(param_group)
|
75 |
-
|
76 |
def load_state_dict(self, state_dict):
|
77 |
if self.accelerator_state.distributed_type == DistributedType.TPU and self.device_placement:
|
78 |
xm.send_cpu_data_to_device(state_dict, self.accelerator_state.device)
|
79 |
self.optimizer.load_state_dict(state_dict)
|
80 |
-
|
81 |
def state_dict(self):
|
82 |
return self.optimizer.state_dict()
|
83 |
-
|
84 |
def zero_grad(self, set_to_none=None):
|
85 |
if self.gradient_state.sync_gradients:
|
86 |
accept_arg = "set_to_none" in inspect.signature(self.optimizer.zero_grad).parameters
|
@@ -92,7 +76,6 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
92 |
if set_to_none is not None:
|
93 |
raise ValueError("`set_to_none` for Optimizer.zero_grad` is not supported by this optimizer.")
|
94 |
self.optimizer.zero_grad()
|
95 |
-
|
96 |
def step(self, closure=None):
|
97 |
if self.gradient_state.sync_gradients:
|
98 |
if self.accelerator_state.distributed_type == DistributedType.TPU:
|
@@ -100,10 +83,8 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
100 |
xm.optimizer_step(self.optimizer, optimizer_args=optimizer_args)
|
101 |
elif self.scaler is not None:
|
102 |
self.optimizer.step = self._optimizer_patched_step_method
|
103 |
-
|
104 |
self.scaler.step(self.optimizer, closure)
|
105 |
self.scaler.update()
|
106 |
-
|
107 |
if not self._accelerate_step_called:
|
108 |
# If the optimizer step was skipped, gradient overflow was detected.
|
109 |
self._is_overflow = True
|
@@ -115,11 +96,9 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
115 |
self._accelerate_step_called = False
|
116 |
else:
|
117 |
self.optimizer.step(closure)
|
118 |
-
|
119 |
def _switch_parameters(self, parameters_map):
|
120 |
for param_group in self.optimizer.param_groups:
|
121 |
param_group["params"] = [parameters_map.get(p, p) for p in param_group["params"]]
|
122 |
-
|
123 |
@property
|
124 |
def is_overflow(self):
|
125 |
"""Whether or not the optimizer step was done, or skipped because of gradient overflow."""
|
@@ -129,12 +108,10 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
129 |
FutureWarning,
|
130 |
)
|
131 |
return self._is_overflow
|
132 |
-
|
133 |
@property
|
134 |
def step_was_skipped(self):
|
135 |
"""Whether or not the optimizer step was skipped."""
|
136 |
return self._is_overflow
|
137 |
-
|
138 |
def __getstate__(self):
|
139 |
_ignored_keys = [
|
140 |
"_accelerate_step_called",
|
@@ -142,18 +119,14 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
142 |
"_optimizer_patched_step_method",
|
143 |
]
|
144 |
return {k: v for k, v in self.__dict__.items() if k not in _ignored_keys}
|
145 |
-
|
146 |
def __setstate__(self, state):
|
147 |
self.__dict__.update(state)
|
148 |
if self.scaler is not None:
|
149 |
self._accelerate_step_called = False
|
150 |
self._optimizer_original_step_method = self.optimizer.step
|
151 |
self._optimizer_patched_step_method = patch_optimizer_step(self, self.optimizer.step)
|
152 |
-
|
153 |
-
|
154 |
def patch_optimizer_step(accelerated_optimizer: AcceleratedOptimizer, method):
|
155 |
def patched_step(*args, **kwargs):
|
156 |
accelerated_optimizer._accelerate_step_called = True
|
157 |
return method(*args, **kwargs)
|
158 |
-
|
159 |
return patched_step
|
|
|
6 |
elif isinstance(state, torch.Tensor):
|
7 |
return state.to(device)
|
8 |
return state
|
|
|
|
|
9 |
class AcceleratedOptimizer(torch.optim.Optimizer):
|
10 |
"""
|
11 |
Internal wrapper around a torch optimizer.
|
|
|
12 |
Conditionally will perform `step` and `zero_grad` if gradients should be synchronized when performing gradient
|
13 |
accumulation.
|
|
|
14 |
Args:
|
15 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
16 |
The optimizer to wrap.
|
|
|
27 |
self.gradient_state = GradientState()
|
28 |
self.device_placement = device_placement
|
29 |
self._is_overflow = False
|
|
|
30 |
if self.scaler is not None:
|
31 |
self._accelerate_step_called = False
|
32 |
self._optimizer_original_step_method = self.optimizer.step
|
33 |
self._optimizer_patched_step_method = patch_optimizer_step(self, self.optimizer.step)
|
|
|
34 |
# Handle device placement
|
35 |
if device_placement:
|
36 |
state_dict = self.optimizer.state_dict()
|
|
|
39 |
else:
|
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 |
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:
|
|
|
83 |
xm.optimizer_step(self.optimizer, optimizer_args=optimizer_args)
|
84 |
elif self.scaler is not None:
|
85 |
self.optimizer.step = self._optimizer_patched_step_method
|
|
|
86 |
self.scaler.step(self.optimizer, closure)
|
87 |
self.scaler.update()
|
|
|
88 |
if not self._accelerate_step_called:
|
89 |
# If the optimizer step was skipped, gradient overflow was detected.
|
90 |
self._is_overflow = True
|
|
|
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."""
|
|
|
108 |
FutureWarning,
|
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 |
"_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:
|
125 |
self._accelerate_step_called = False
|
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)
|
|
|
132 |
return patched_step
|
src/scheduler.py
CHANGED
@@ -1,16 +1,11 @@
|
|
1 |
-
|
2 |
-
|
3 |
# We ignore warnings about stepping the scheduler since we step it ourselves during gradient accumulation
|
4 |
-
|
5 |
class AcceleratedScheduler:
|
6 |
"""
|
7 |
A wrapper around a learning rate scheduler that will only step when the optimizer(s) have a training step. Useful
|
8 |
to avoid making a scheduler step too fast when gradients went overflow and there was no training step (in mixed
|
9 |
precision training)
|
10 |
-
|
11 |
When performing gradient accumulation scheduler lengths should not be changed accordingly, Accelerate will always
|
12 |
step the scheduler to account for it.
|
13 |
-
|
14 |
Args:
|
15 |
scheduler (`torch.optim.lr_scheduler._LRScheduler`):
|
16 |
The scheduler to wrap.
|
@@ -29,19 +24,16 @@ class AcceleratedScheduler:
|
|
29 |
self.split_batches = split_batches
|
30 |
self.step_with_optimizer = step_with_optimizer
|
31 |
self.gradient_state = GradientState()
|
32 |
-
|
33 |
def step(self, *args, **kwargs):
|
34 |
if not self.step_with_optimizer:
|
35 |
# No link between scheduler and optimizer -> just step
|
36 |
self.scheduler.step(*args, **kwargs)
|
37 |
return
|
38 |
-
|
39 |
# Otherwise, first make sure the optimizer was stepped.
|
40 |
if not self.gradient_state.sync_gradients:
|
41 |
if self.gradient_state.adjust_scheduler:
|
42 |
self.scheduler._step_count += 1
|
43 |
return
|
44 |
-
|
45 |
for opt in self.optimizers:
|
46 |
if opt.step_was_skipped:
|
47 |
return
|
@@ -59,19 +51,14 @@ class AcceleratedScheduler:
|
|
59 |
self.scheduler.step(*args, **kwargs)
|
60 |
else:
|
61 |
self.scheduler.step(*args, **kwargs)
|
62 |
-
|
63 |
# Passthroughs
|
64 |
def get_last_lr(self):
|
65 |
return self.scheduler.get_last_lr()
|
66 |
-
|
67 |
def state_dict(self):
|
68 |
return self.scheduler.state_dict()
|
69 |
-
|
70 |
def load_state_dict(self, state_dict):
|
71 |
self.scheduler.load_state_dict(state_dict)
|
72 |
-
|
73 |
def get_lr(self):
|
74 |
return self.scheduler.get_lr()
|
75 |
-
|
76 |
def print_lr(self, *args, **kwargs):
|
77 |
return self.scheduler.print_lr(*args, **kwargs)
|
|
|
|
|
|
|
1 |
# We ignore warnings about stepping the scheduler since we step it ourselves during gradient accumulation
|
|
|
2 |
class AcceleratedScheduler:
|
3 |
"""
|
4 |
A wrapper around a learning rate scheduler that will only step when the optimizer(s) have a training step. Useful
|
5 |
to avoid making a scheduler step too fast when gradients went overflow and there was no training step (in mixed
|
6 |
precision training)
|
|
|
7 |
When performing gradient accumulation scheduler lengths should not be changed accordingly, Accelerate will always
|
8 |
step the scheduler to account for it.
|
|
|
9 |
Args:
|
10 |
scheduler (`torch.optim.lr_scheduler._LRScheduler`):
|
11 |
The scheduler to wrap.
|
|
|
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
|
30 |
self.scheduler.step(*args, **kwargs)
|
31 |
return
|
|
|
32 |
# Otherwise, first make sure the optimizer was stepped.
|
33 |
if not self.gradient_state.sync_gradients:
|
34 |
if self.gradient_state.adjust_scheduler:
|
35 |
self.scheduler._step_count += 1
|
36 |
return
|
|
|
37 |
for opt in self.optimizers:
|
38 |
if opt.step_was_skipped:
|
39 |
return
|
|
|
51 |
self.scheduler.step(*args, **kwargs)
|
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)
|
src/state.py
CHANGED
@@ -1,58 +1,40 @@
|
|
1 |
logger = logging.getLogger(__name__)
|
2 |
-
|
3 |
-
|
4 |
def is_initialized() -> bool:
|
5 |
"""
|
6 |
Checks if the `AcceleratorState` has been initialized from `Accelerator`. Same as `AcceleratorState.initialized`,
|
7 |
but works as a module method.
|
8 |
"""
|
9 |
return AcceleratorState._shared_state != {}
|
10 |
-
|
11 |
-
|
12 |
# Lambda function that does nothing
|
13 |
def do_nothing(*args, **kwargs):
|
14 |
return None
|
15 |
-
|
16 |
-
|
17 |
class ThreadLocalSharedDict(threading.local):
|
18 |
"""
|
19 |
Descriptor that holds a dict shared between instances of a class in the same thread.
|
20 |
-
|
21 |
Note: Descriptors have slightly different semantics than just a dict field on its own.
|
22 |
`PartialState(...)._shared_state` and `PartialState._shared_state` (instance vs class) give the same value: the
|
23 |
underlying _storage dict. Likewise, `PartialState(...)._shared_state = {...}` overrides the _storage dict inside
|
24 |
the descriptor as you would expect. However, `PartialState._shared_state = {}` actually replaces the descriptor
|
25 |
object with a dict instead Thus, you should modify the _storage dict in-place (e.g. `_shared_state.clear()`).
|
26 |
-
|
27 |
See Python documentation for an explanation of descriptors: https://docs.python.org/3/howto/descriptor.html
|
28 |
-
|
29 |
This is required for using PyTorch/XLA with PJRT in multithreaded mode (required for TPU v2 and v3).
|
30 |
-
|
31 |
See https://github.com/pytorch/xla/blob/r2.0/docs/pjrt.md#multithreading-on-tpu-v2v3
|
32 |
"""
|
33 |
def __init__(self, thread_local: bool = False):
|
34 |
self._storage = {}
|
35 |
-
|
36 |
def __get__(self, obj, objtype=None):
|
37 |
return self._storage
|
38 |
-
|
39 |
def __set__(self, obj, value):
|
40 |
self._storage = value
|
41 |
-
|
42 |
-
|
43 |
# Prefer global shared dictionary, except when using TPU.
|
44 |
SharedDict = dict if not is_tpu_available(check_device=False) else ThreadLocalSharedDict
|
45 |
-
|
46 |
-
|
47 |
# Inspired by Alex Martelli's 'Borg'.
|
48 |
class PartialState:
|
49 |
"""
|
50 |
Singleton class that has information about the current training environment and functions to help with process
|
51 |
control. Designed to be used when only process control and device execution states are needed. Does *not* need to
|
52 |
be initialized from `Accelerator`.
|
53 |
-
|
54 |
**Available attributes:**
|
55 |
-
|
56 |
- **device** (`torch.device`) -- The device to use.
|
57 |
- **distributed_type** ([`~accelerate.state.DistributedType`]) -- The type of distributed environment currently
|
58 |
in use.
|
@@ -67,7 +49,6 @@ class PartialState:
|
|
67 |
- **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
|
68 |
"""
|
69 |
_shared_state = SharedDict()
|
70 |
-
|
71 |
def __init__(self, cpu: bool = False, **kwargs):
|
72 |
self.__dict__ = self._shared_state
|
73 |
if not self.initialized:
|
@@ -82,14 +63,12 @@ class PartialState:
|
|
82 |
os.environ.get("ACCELERATE_USE_SAGEMAKER", "false") == "true"
|
83 |
and os.environ.get("ACCELERATE_SAGEMAKER_DISTRIBUTED_TYPE") != SageMakerDistributedType.NO
|
84 |
)
|
85 |
-
|
86 |
if use_sagemaker_dp and not cpu:
|
87 |
if (
|
88 |
os.environ.get("ACCELERATE_SAGEMAKER_DISTRIBUTED_TYPE") == SageMakerDistributedType.DATA_PARALLEL
|
89 |
) or use_sagemaker_dp:
|
90 |
self.distributed_type = DistributedType.MULTI_GPU
|
91 |
import smdistributed.dataparallel.torch.torch_smddp # noqa
|
92 |
-
|
93 |
if not torch.distributed.is_initialized():
|
94 |
torch.distributed.init_process_group(backend="smddp")
|
95 |
self.backend = "smddp"
|
@@ -116,7 +95,6 @@ class PartialState:
|
|
116 |
self.distributed_type = DistributedType.DEEPSPEED
|
117 |
if not torch.distributed.is_initialized():
|
118 |
from deepspeed import comm as dist
|
119 |
-
|
120 |
# DeepSpeed always uses nccl
|
121 |
kwargs.pop("backend", None)
|
122 |
if is_xpu_available and is_ccl_available():
|
@@ -127,7 +105,6 @@ class PartialState:
|
|
127 |
else:
|
128 |
self.backend = "nccl"
|
129 |
dist.init_distributed(dist_backend=self.backend, auto_mpi_discovery=False, **kwargs)
|
130 |
-
|
131 |
self.num_processes = torch.distributed.get_world_size()
|
132 |
self.process_index = torch.distributed.get_rank()
|
133 |
self.local_process_index = int(os.environ.get("LOCAL_RANK", -1))
|
@@ -230,7 +207,6 @@ class PartialState:
|
|
230 |
and get_int_from_env(["OMP_NUM_THREADS", "MKL_NUM_THREADS"], 0) == 0
|
231 |
):
|
232 |
import psutil
|
233 |
-
|
234 |
num_cpu_threads_per_process = int(psutil.cpu_count(logical=False) / local_size)
|
235 |
if num_cpu_threads_per_process == 0:
|
236 |
num_cpu_threads_per_process = 1
|
@@ -261,12 +237,9 @@ class PartialState:
|
|
261 |
)
|
262 |
self.num_processes = 1
|
263 |
self.process_index = self.local_process_index = 0
|
264 |
-
|
265 |
if self.device is None:
|
266 |
self.device = torch.device("cpu") if cpu else self.default_device
|
267 |
-
|
268 |
self.fork_launched = parse_flag_from_env("FORK_LAUNCHED", 0)
|
269 |
-
|
270 |
def __repr__(self) -> str:
|
271 |
return (
|
272 |
f"Distributed environment: {self.distributed_type}{(' Backend: ' + self.backend) if self.backend else ''}\n"
|
@@ -275,36 +248,30 @@ class PartialState:
|
|
275 |
f"Local process index: {self.local_process_index}\n"
|
276 |
f"Device: {self.device}\n"
|
277 |
)
|
278 |
-
|
279 |
@staticmethod
|
280 |
def _reset_state():
|
281 |
"Resets `_shared_state`, is used internally and should not be called"
|
282 |
PartialState._shared_state.clear()
|
283 |
-
|
284 |
@property
|
285 |
def initialized(self) -> bool:
|
286 |
"Returns whether the `PartialState` has been initialized"
|
287 |
return self._shared_state != {}
|
288 |
-
|
289 |
@property
|
290 |
def use_distributed(self):
|
291 |
"""
|
292 |
Whether the Accelerator is configured for distributed training
|
293 |
"""
|
294 |
return self.distributed_type != DistributedType.NO and self.num_processes > 1
|
295 |
-
|
296 |
@property
|
297 |
def is_last_process(self) -> bool:
|
298 |
"Returns whether the current process is the last one"
|
299 |
return self.process_index == self.num_processes - 1
|
300 |
-
|
301 |
@property
|
302 |
def is_main_process(self) -> bool:
|
303 |
"Returns whether the current process is the main process"
|
304 |
return (
|
305 |
self.process_index == 0 if self.distributed_type != DistributedType.MEGATRON_LM else self.is_last_process
|
306 |
)
|
307 |
-
|
308 |
@property
|
309 |
def is_local_main_process(self) -> bool:
|
310 |
"Returns whether the current process is the main process on the local node"
|
@@ -313,19 +280,15 @@ class PartialState:
|
|
313 |
if self.distributed_type != DistributedType.MEGATRON_LM
|
314 |
else self.is_last_process
|
315 |
)
|
316 |
-
|
317 |
def wait_for_everyone(self):
|
318 |
"""
|
319 |
Will stop the execution of the current process until every other process has reached that point (so this does
|
320 |
nothing when the script is only run in one process). Useful to do before saving a model.
|
321 |
-
|
322 |
Example:
|
323 |
-
|
324 |
```python
|
325 |
>>> # Assuming two GPU processes
|
326 |
>>> import time
|
327 |
>>> from accelerate.state import PartialState
|
328 |
-
|
329 |
>>> state = PartialState()
|
330 |
>>> if state.is_main_process:
|
331 |
... time.sleep(2)
|
@@ -347,24 +310,18 @@ class PartialState:
|
|
347 |
torch.distributed.barrier()
|
348 |
elif self.distributed_type == DistributedType.TPU:
|
349 |
xm.rendezvous("accelerate.utils.wait_for_everyone")
|
350 |
-
|
351 |
def _goes_first(self, is_main: bool):
|
352 |
if not is_main:
|
353 |
self.wait_for_everyone()
|
354 |
-
|
355 |
yield
|
356 |
-
|
357 |
if is_main:
|
358 |
self.wait_for_everyone()
|
359 |
-
|
360 |
@contextmanager
|
361 |
def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
|
362 |
"""
|
363 |
Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
|
364 |
distributed inference, such as with different prompts.
|
365 |
-
|
366 |
Note that when using a `dict`, all keys need to have the same number of elements.
|
367 |
-
|
368 |
Args:
|
369 |
inputs (`list`, `tuple`, `torch.Tensor`, or `dict` of `list`/`tuple`/`torch.Tensor`):
|
370 |
The input to split between processes.
|
@@ -372,14 +329,10 @@ class PartialState:
|
|
372 |
Whether to apply padding by repeating the last element of the input so that all processes have the same
|
373 |
number of elements. Useful when trying to perform actions such as `gather()` on the outputs or passing
|
374 |
in less inputs than there are processes. If so, just remember to drop the padded elements afterwards.
|
375 |
-
|
376 |
-
|
377 |
Example:
|
378 |
-
|
379 |
```python
|
380 |
# Assume there are two processes
|
381 |
from accelerate import PartialState
|
382 |
-
|
383 |
state = PartialState()
|
384 |
with state.split_between_processes(["A", "B", "C"]) as inputs:
|
385 |
print(inputs)
|
@@ -387,7 +340,6 @@ class PartialState:
|
|
387 |
["A", "B"]
|
388 |
# Process 1
|
389 |
["C"]
|
390 |
-
|
391 |
with state.split_between_processes(["A", "B", "C"], apply_padding=True) as inputs:
|
392 |
print(inputs)
|
393 |
# Process 0
|
@@ -410,7 +362,6 @@ class PartialState:
|
|
410 |
end_index = start_index + num_samples_per_process
|
411 |
if (len(inputs) % self.num_processes != 0) and (self.process_index == self.num_processes - 1):
|
412 |
end_index = length
|
413 |
-
|
414 |
def _split_values(inputs, start_index, end_index):
|
415 |
if isinstance(inputs, (list, tuple, torch.Tensor)):
|
416 |
if start_index >= len(inputs):
|
@@ -420,7 +371,6 @@ class PartialState:
|
|
420 |
if apply_padding:
|
421 |
if isinstance(result, torch.Tensor):
|
422 |
from accelerate.utils import pad_across_processes, send_to_device
|
423 |
-
|
424 |
# The tensor needs to be on the device before we can pad it
|
425 |
tensorized_result = send_to_device(result, self.device)
|
426 |
result = pad_across_processes(tensorized_result, pad_index=inputs[-1])
|
@@ -433,21 +383,15 @@ class PartialState:
|
|
433 |
return inputs
|
434 |
else:
|
435 |
return inputs
|
436 |
-
|
437 |
yield _split_values(inputs, start_index, end_index)
|
438 |
-
|
439 |
@contextmanager
|
440 |
def main_process_first(self):
|
441 |
"""
|
442 |
Lets the main process go first inside a with block.
|
443 |
-
|
444 |
The other processes will enter the with block after the main process exits.
|
445 |
-
|
446 |
Example:
|
447 |
-
|
448 |
```python
|
449 |
>>> from accelerate import Accelerator
|
450 |
-
|
451 |
>>> accelerator = Accelerator()
|
452 |
>>> with accelerator.main_process_first():
|
453 |
... # This will be printed first by process 0 then in a seemingly
|
@@ -456,19 +400,14 @@ class PartialState:
|
|
456 |
```
|
457 |
"""
|
458 |
yield from self._goes_first(self.is_main_process)
|
459 |
-
|
460 |
@contextmanager
|
461 |
def local_main_process_first(self):
|
462 |
"""
|
463 |
Lets the local main process go inside a with block.
|
464 |
-
|
465 |
The other processes will enter the with block after the main process exits.
|
466 |
-
|
467 |
Example:
|
468 |
-
|
469 |
```python
|
470 |
>>> from accelerate.state import PartialState
|
471 |
-
|
472 |
>>> state = PartialState()
|
473 |
>>> with state.local_main_process_first():
|
474 |
... # This will be printed first by local process 0 then in a seemingly
|
@@ -477,27 +416,18 @@ class PartialState:
|
|
477 |
```
|
478 |
"""
|
479 |
yield from self._goes_first(self.is_local_main_process)
|
480 |
-
|
481 |
def on_main_process(self, function: Callable[..., Any] = None):
|
482 |
"""
|
483 |
Decorator that only runs the decorated function on the main process.
|
484 |
-
|
485 |
Args:
|
486 |
function (`Callable`): The function to decorate.
|
487 |
-
|
488 |
Example:
|
489 |
-
|
490 |
```python
|
491 |
>>> from accelerate.state import PartialState
|
492 |
-
|
493 |
>>> state = PartialState()
|
494 |
-
|
495 |
-
|
496 |
>>> @state.on_main_process
|
497 |
... def print_something():
|
498 |
... print("This will be printed by process 0 only.")
|
499 |
-
|
500 |
-
|
501 |
>>> print_something()
|
502 |
"This will be printed by process 0 only"
|
503 |
```
|
@@ -507,27 +437,19 @@ class PartialState:
|
|
507 |
if self.is_main_process or not self.use_distributed:
|
508 |
return function
|
509 |
return do_nothing
|
510 |
-
|
511 |
def on_local_main_process(self, function: Callable[..., Any] = None):
|
512 |
"""
|
513 |
Decorator that only runs the decorated function on the local main process.
|
514 |
-
|
515 |
Args:
|
516 |
function (`Callable`): The function to decorate.
|
517 |
-
|
518 |
Example:
|
519 |
```python
|
520 |
# Assume we have 2 servers with 4 processes each.
|
521 |
from accelerate.state import PartialState
|
522 |
-
|
523 |
state = PartialState()
|
524 |
-
|
525 |
-
|
526 |
@state.on_local_main_process
|
527 |
def print_something():
|
528 |
print("This will be printed by process 0 only on each server.")
|
529 |
-
|
530 |
-
|
531 |
print_something()
|
532 |
# On server 1:
|
533 |
"This will be printed by process 0 only"
|
@@ -538,27 +460,19 @@ class PartialState:
|
|
538 |
if self.is_local_main_process or not self.use_distributed:
|
539 |
return function
|
540 |
return do_nothing
|
541 |
-
|
542 |
def on_last_process(self, function: Callable[..., Any]):
|
543 |
"""
|
544 |
Decorator that only runs the decorated function on the last process.
|
545 |
-
|
546 |
Args:
|
547 |
function (`Callable`): The function to decorate.
|
548 |
-
|
549 |
Example:
|
550 |
```python
|
551 |
# Assume we have 4 processes.
|
552 |
from accelerate.state import PartialState
|
553 |
-
|
554 |
state = PartialState()
|
555 |
-
|
556 |
-
|
557 |
@state.on_last_process
|
558 |
def print_something():
|
559 |
print(f"Printed on process {state.process_index}")
|
560 |
-
|
561 |
-
|
562 |
print_something()
|
563 |
"Printed on process 3"
|
564 |
```
|
@@ -566,30 +480,22 @@ class PartialState:
|
|
566 |
if self.is_last_process or not self.use_distributed:
|
567 |
return function
|
568 |
return do_nothing
|
569 |
-
|
570 |
def on_process(self, function: Callable[..., Any] = None, process_index: int = None):
|
571 |
"""
|
572 |
Decorator that only runs the decorated function on the process with the given index.
|
573 |
-
|
574 |
Args:
|
575 |
function (`Callable`, `optional`):
|
576 |
The function to decorate.
|
577 |
process_index (`int`, `optional`):
|
578 |
The index of the process on which to run the function.
|
579 |
-
|
580 |
Example:
|
581 |
```python
|
582 |
# Assume we have 4 processes.
|
583 |
from accelerate.state import PartialState
|
584 |
-
|
585 |
state = PartialState()
|
586 |
-
|
587 |
-
|
588 |
@state.on_process(process_index=2)
|
589 |
def print_something():
|
590 |
print(f"Printed on process {state.process_index}")
|
591 |
-
|
592 |
-
|
593 |
print_something()
|
594 |
"Printed on process 2"
|
595 |
```
|
@@ -599,30 +505,22 @@ class PartialState:
|
|
599 |
if (self.process_index == process_index) or (not self.use_distributed):
|
600 |
return function
|
601 |
return do_nothing
|
602 |
-
|
603 |
def on_local_process(self, function: Callable[..., Any] = None, local_process_index: int = None):
|
604 |
"""
|
605 |
Decorator that only runs the decorated function on the process with the given index on the current node.
|
606 |
-
|
607 |
Args:
|
608 |
function (`Callable`, *optional*):
|
609 |
The function to decorate.
|
610 |
local_process_index (`int`, *optional*):
|
611 |
The index of the local process on which to run the function.
|
612 |
-
|
613 |
Example:
|
614 |
```python
|
615 |
# Assume we have 2 servers with 4 processes each.
|
616 |
from accelerate import Accelerator
|
617 |
-
|
618 |
accelerator = Accelerator()
|
619 |
-
|
620 |
-
|
621 |
@accelerator.on_local_process(local_process_index=2)
|
622 |
def print_something():
|
623 |
print(f"Printed on process {accelerator.local_process_index}")
|
624 |
-
|
625 |
-
|
626 |
print_something()
|
627 |
# On server 1:
|
628 |
"Printed on process 2"
|
@@ -635,11 +533,9 @@ class PartialState:
|
|
635 |
if (self.local_process_index == local_process_index) or (not self.use_distributed):
|
636 |
return function
|
637 |
return do_nothing
|
638 |
-
|
639 |
def print(self, *args, **kwargs):
|
640 |
if self.is_local_main_process:
|
641 |
print(*args, **kwargs)
|
642 |
-
|
643 |
@property
|
644 |
def default_device(self) -> torch.device:
|
645 |
"""
|
@@ -660,14 +556,10 @@ class PartialState:
|
|
660 |
return torch.device("npu")
|
661 |
else:
|
662 |
return torch.device("cpu")
|
663 |
-
|
664 |
-
|
665 |
class AcceleratorState:
|
666 |
"""
|
667 |
Singleton class that has information about the current training environment.
|
668 |
-
|
669 |
**Available attributes:**
|
670 |
-
|
671 |
- **device** (`torch.device`) -- The device to use.
|
672 |
- **distributed_type** ([`~accelerate.state.DistributedType`]) -- The type of distributed environment currently
|
673 |
in use.
|
@@ -683,7 +575,6 @@ class AcceleratorState:
|
|
683 |
- **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
|
684 |
"""
|
685 |
_shared_state = SharedDict()
|
686 |
-
|
687 |
def __init__(
|
688 |
self,
|
689 |
mixed_precision: str = None,
|
@@ -722,7 +613,6 @@ class AcceleratorState:
|
|
722 |
"or higher, compute capability of 8.9 or higher). Will use FP16 instead."
|
723 |
)
|
724 |
mixed_precision = "fp16"
|
725 |
-
|
726 |
self.dynamo_plugin = dynamo_plugin
|
727 |
if not _from_accelerator:
|
728 |
raise ValueError(
|
@@ -771,7 +661,6 @@ class AcceleratorState:
|
|
771 |
if self._mixed_precision != "no":
|
772 |
fsdp_plugin.set_mixed_precision(self._mixed_precision)
|
773 |
self.fsdp_plugin = fsdp_plugin
|
774 |
-
|
775 |
if (
|
776 |
self.dynamo_plugin.backend != DynamoBackend.NO
|
777 |
and self._mixed_precision == "no"
|
@@ -779,17 +668,14 @@ class AcceleratorState:
|
|
779 |
):
|
780 |
torch.backends.cuda.matmul.allow_tf32 = True
|
781 |
PartialState._shared_state["distributed_type"] = self.distributed_type
|
782 |
-
|
783 |
@property
|
784 |
def initialized(self) -> bool:
|
785 |
return self._shared_state != PartialState._shared_state
|
786 |
-
|
787 |
def __repr__(self):
|
788 |
repr = PartialState().__repr__() + f"\nMixed precision type: {self.mixed_precision}\n"
|
789 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
790 |
repr += f"ds_config: {self.deepspeed_plugin.deepspeed_config}\n"
|
791 |
return repr
|
792 |
-
|
793 |
def _check_initialized(self, mixed_precision=None, cpu=None):
|
794 |
"Checks if a modification is trying to be made and the `AcceleratorState` has already been initialized"
|
795 |
if self.initialized:
|
@@ -802,7 +688,6 @@ class AcceleratorState:
|
|
802 |
and self.distributed_type != DistributedType.DEEPSPEED
|
803 |
):
|
804 |
raise ValueError(err.format(flag=f"mixed_precision='{mixed_precision}'"))
|
805 |
-
|
806 |
# For backward compatibility
|
807 |
@property
|
808 |
def use_fp16(self):
|
@@ -812,7 +697,6 @@ class AcceleratorState:
|
|
812 |
FutureWarning,
|
813 |
)
|
814 |
return self._mixed_precision != "no"
|
815 |
-
|
816 |
@property
|
817 |
def mixed_precision(self):
|
818 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
@@ -826,47 +710,38 @@ class AcceleratorState:
|
|
826 |
else:
|
827 |
mixed_precision = self._mixed_precision
|
828 |
return mixed_precision
|
829 |
-
|
830 |
@staticmethod
|
831 |
def _reset_state(reset_partial_state: bool = False):
|
832 |
"Resets `_shared_state`, is used internally and should not be called"
|
833 |
AcceleratorState._shared_state.clear()
|
834 |
if reset_partial_state:
|
835 |
PartialState._reset_state()
|
836 |
-
|
837 |
@property
|
838 |
def use_distributed(self):
|
839 |
"""
|
840 |
Whether the Accelerator is configured for distributed training
|
841 |
"""
|
842 |
return PartialState().use_distributed
|
843 |
-
|
844 |
@property
|
845 |
def is_last_process(self) -> bool:
|
846 |
"Returns whether the current process is the last one"
|
847 |
return PartialState().is_last_process
|
848 |
-
|
849 |
@property
|
850 |
def is_main_process(self) -> bool:
|
851 |
"Returns whether the current process is the main process"
|
852 |
return PartialState().is_main_process
|
853 |
-
|
854 |
@property
|
855 |
def is_local_main_process(self) -> bool:
|
856 |
"Returns whether the current process is the main process on the local node"
|
857 |
return PartialState().is_local_main_process
|
858 |
-
|
859 |
def wait_for_everyone(self):
|
860 |
PartialState().wait_for_everyone()
|
861 |
-
|
862 |
@contextmanager
|
863 |
def split_between_processes(self, inputs: list | tuple | dict | torch.Tensor, apply_padding: bool = False):
|
864 |
"""
|
865 |
Splits `input` between `self.num_processes` quickly and can be then used on that process. Useful when doing
|
866 |
distributed inference, such as with different prompts.
|
867 |
-
|
868 |
Note that when using a `dict`, all keys need to have the same number of elements.
|
869 |
-
|
870 |
Args:
|
871 |
inputs (`list`, `tuple`, `torch.Tensor`, or `dict` of `list`/`tuple`/`torch.Tensor`):
|
872 |
The input to split between processes.
|
@@ -874,14 +749,10 @@ class AcceleratorState:
|
|
874 |
Whether to apply padding by repeating the last element of the input so that all processes have the same
|
875 |
number of elements. Useful when trying to perform actions such as `gather()` on the outputs or passing
|
876 |
in less inputs than there are processes. If so, just remember to drop the padded elements afterwards.
|
877 |
-
|
878 |
-
|
879 |
Example:
|
880 |
-
|
881 |
```python
|
882 |
# Assume there are two processes
|
883 |
from accelerate.state import AcceleratorState
|
884 |
-
|
885 |
state = AcceleratorState()
|
886 |
with state.split_between_processes(["A", "B", "C"]) as inputs:
|
887 |
print(inputs)
|
@@ -889,7 +760,6 @@ class AcceleratorState:
|
|
889 |
["A", "B"]
|
890 |
# Process 1
|
891 |
["C"]
|
892 |
-
|
893 |
with state.split_between_processes(["A", "B", "C"], apply_padding=True) as inputs:
|
894 |
print(inputs)
|
895 |
# Process 0
|
@@ -900,37 +770,28 @@ class AcceleratorState:
|
|
900 |
"""
|
901 |
with PartialState().split_between_processes(inputs, apply_padding=apply_padding) as inputs:
|
902 |
yield inputs
|
903 |
-
|
904 |
@contextmanager
|
905 |
def main_process_first(self):
|
906 |
"""
|
907 |
Lets the main process go first inside a with block.
|
908 |
-
|
909 |
The other processes will enter the with block after the main process exits.
|
910 |
"""
|
911 |
with PartialState().main_process_first():
|
912 |
yield
|
913 |
-
|
914 |
@contextmanager
|
915 |
def local_main_process_first(self):
|
916 |
"""
|
917 |
Lets the local main process go inside a with block.
|
918 |
-
|
919 |
The other processes will enter the with block after the main process exits.
|
920 |
"""
|
921 |
with PartialState().local_main_process_first():
|
922 |
yield
|
923 |
-
|
924 |
def print(self, *args, **kwargs):
|
925 |
PartialState().print(*args, **kwargs)
|
926 |
-
|
927 |
-
|
928 |
class GradientState:
|
929 |
"""
|
930 |
Singleton class that has information related to gradient synchronization for gradient accumulation
|
931 |
-
|
932 |
**Available attributes:**
|
933 |
-
|
934 |
- **end_of_dataloader** (`bool`) -- Whether we have reached the end the current dataloader
|
935 |
- **remainder** (`int`) -- The number of extra samples that were added from padding the dataloader
|
936 |
- **sync_gradients** (`bool`) -- Whether the gradients should be synced across all devices
|
@@ -944,7 +805,6 @@ class GradientState:
|
|
944 |
iteration and the number of total steps reset
|
945 |
"""
|
946 |
_shared_state = SharedDict()
|
947 |
-
|
948 |
def __init__(self, gradient_accumulation_plugin: Optional[GradientAccumulationPlugin] = None):
|
949 |
self.__dict__ = self._shared_state
|
950 |
if not self.initialized:
|
@@ -954,45 +814,37 @@ class GradientState:
|
|
954 |
self.plugin_kwargs = (
|
955 |
gradient_accumulation_plugin.to_kwargs() if gradient_accumulation_plugin is not None else {}
|
956 |
)
|
957 |
-
|
958 |
# Plugin args are different and can be updated
|
959 |
if gradient_accumulation_plugin is not None and self.plugin_kwargs != gradient_accumulation_plugin.to_kwargs():
|
960 |
self.plugin_kwargs = gradient_accumulation_plugin.to_kwargs()
|
961 |
-
|
962 |
@property
|
963 |
def num_steps(self) -> int:
|
964 |
"Returns the number of steps to accumulate over"
|
965 |
return self.plugin_kwargs.get("num_steps", 1)
|
966 |
-
|
967 |
@property
|
968 |
def adjust_scheduler(self) -> bool:
|
969 |
"Returns whether the scheduler should be adjusted"
|
970 |
return self.plugin_kwargs.get("adjust_scheduler", False)
|
971 |
-
|
972 |
@property
|
973 |
def sync_with_dataloader(self) -> bool:
|
974 |
"Returns whether the gradients should be synced at the end of the dataloader iteration and the number of total steps reset"
|
975 |
return self.plugin_kwargs.get("sync_with_dataloader", True)
|
976 |
-
|
977 |
@property
|
978 |
def initialized(self) -> bool:
|
979 |
"Returns whether the `GradientState` has been initialized"
|
980 |
return GradientState._shared_state != {}
|
981 |
-
|
982 |
@property
|
983 |
def end_of_dataloader(self) -> bool:
|
984 |
"Returns whether we have reached the end of the current dataloader"
|
985 |
if not self.in_dataloader:
|
986 |
return False
|
987 |
return self.active_dataloader.end_of_dataloader
|
988 |
-
|
989 |
@property
|
990 |
def remainder(self) -> int:
|
991 |
"Returns the number of extra samples that were added from padding the dataloader"
|
992 |
if not self.in_dataloader:
|
993 |
return -1
|
994 |
return self.active_dataloader.remainder
|
995 |
-
|
996 |
def __repr__(self):
|
997 |
return (
|
998 |
f"Sync Gradients: {self.sync_gradients}\n"
|
@@ -1000,26 +852,21 @@ class GradientState:
|
|
1000 |
f"Extra samples added: {self.remainder}\n"
|
1001 |
f"Gradient accumulation plugin: {self.plugin_kwargs}\n"
|
1002 |
)
|
1003 |
-
|
1004 |
def _set_sync_gradients(self, sync_gradients):
|
1005 |
"Private function that sets whether gradients should be synchronized. Users should not have to call this."
|
1006 |
self.sync_gradients = sync_gradients
|
1007 |
-
|
1008 |
def _add_dataloader(self, dataloader):
|
1009 |
"Private function that adds a dataloader to `self.dataloader_references` and sets `in_dataloader` to `True`. Users should not have to call this."
|
1010 |
self.active_dataloader = dataloader
|
1011 |
self.dataloader_references.append(self.active_dataloader)
|
1012 |
-
|
1013 |
def _remove_dataloader(self, dataloader):
|
1014 |
"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."
|
1015 |
self.dataloader_references.remove(dataloader)
|
1016 |
self.active_dataloader = self.dataloader_references[-1]
|
1017 |
-
|
1018 |
@property
|
1019 |
def in_dataloader(self) -> bool:
|
1020 |
"Returns whether the current process is in a dataloader"
|
1021 |
return self.active_dataloader is not None
|
1022 |
-
|
1023 |
@staticmethod
|
1024 |
def _reset_state():
|
1025 |
"Resets `_shared_state`, is used internally and should not be called"
|
|
|
1 |
logger = logging.getLogger(__name__)
|
|
|
|
|
2 |
def is_initialized() -> bool:
|
3 |
"""
|
4 |
Checks if the `AcceleratorState` has been initialized from `Accelerator`. Same as `AcceleratorState.initialized`,
|
5 |
but works as a module method.
|
6 |
"""
|
7 |
return AcceleratorState._shared_state != {}
|
|
|
|
|
8 |
# Lambda function that does nothing
|
9 |
def do_nothing(*args, **kwargs):
|
10 |
return None
|
|
|
|
|
11 |
class ThreadLocalSharedDict(threading.local):
|
12 |
"""
|
13 |
Descriptor that holds a dict shared between instances of a class in the same thread.
|
|
|
14 |
Note: Descriptors have slightly different semantics than just a dict field on its own.
|
15 |
`PartialState(...)._shared_state` and `PartialState._shared_state` (instance vs class) give the same value: the
|
16 |
underlying _storage dict. Likewise, `PartialState(...)._shared_state = {...}` overrides the _storage dict inside
|
17 |
the descriptor as you would expect. However, `PartialState._shared_state = {}` actually replaces the descriptor
|
18 |
object with a dict instead Thus, you should modify the _storage dict in-place (e.g. `_shared_state.clear()`).
|
|
|
19 |
See Python documentation for an explanation of descriptors: https://docs.python.org/3/howto/descriptor.html
|
|
|
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.
|
30 |
SharedDict = dict if not is_tpu_available(check_device=False) else ThreadLocalSharedDict
|
|
|
|
|
31 |
# Inspired by Alex Martelli's 'Borg'.
|
32 |
class PartialState:
|
33 |
"""
|
34 |
Singleton class that has information about the current training environment and functions to help with process
|
35 |
control. Designed to be used when only process control and device execution states are needed. Does *not* need to
|
36 |
be initialized from `Accelerator`.
|
|
|
37 |
**Available attributes:**
|
|
|
38 |
- **device** (`torch.device`) -- The device to use.
|
39 |
- **distributed_type** ([`~accelerate.state.DistributedType`]) -- The type of distributed environment currently
|
40 |
in use.
|
|
|
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:
|
|
|
63 |
os.environ.get("ACCELERATE_USE_SAGEMAKER", "false") == "true"
|
64 |
and os.environ.get("ACCELERATE_SAGEMAKER_DISTRIBUTED_TYPE") != SageMakerDistributedType.NO
|
65 |
)
|
|
|
66 |
if use_sagemaker_dp and not cpu:
|
67 |
if (
|
68 |
os.environ.get("ACCELERATE_SAGEMAKER_DISTRIBUTED_TYPE") == SageMakerDistributedType.DATA_PARALLEL
|
69 |
) or use_sagemaker_dp:
|
70 |
self.distributed_type = DistributedType.MULTI_GPU
|
71 |
import smdistributed.dataparallel.torch.torch_smddp # noqa
|
|
|
72 |
if not torch.distributed.is_initialized():
|
73 |
torch.distributed.init_process_group(backend="smddp")
|
74 |
self.backend = "smddp"
|
|
|
95 |
self.distributed_type = DistributedType.DEEPSPEED
|
96 |
if not torch.distributed.is_initialized():
|
97 |
from deepspeed import comm as dist
|
|
|
98 |
# DeepSpeed always uses nccl
|
99 |
kwargs.pop("backend", None)
|
100 |
if is_xpu_available and is_ccl_available():
|
|
|
105 |
else:
|
106 |
self.backend = "nccl"
|
107 |
dist.init_distributed(dist_backend=self.backend, auto_mpi_discovery=False, **kwargs)
|
|
|
108 |
self.num_processes = torch.distributed.get_world_size()
|
109 |
self.process_index = torch.distributed.get_rank()
|
110 |
self.local_process_index = int(os.environ.get("LOCAL_RANK", -1))
|
|
|
207 |
and get_int_from_env(["OMP_NUM_THREADS", "MKL_NUM_THREADS"], 0) == 0
|
208 |
):
|
209 |
import psutil
|
|
|
210 |
num_cpu_threads_per_process = int(psutil.cpu_count(logical=False) / local_size)
|
211 |
if num_cpu_threads_per_process == 0:
|
212 |
num_cpu_threads_per_process = 1
|
|
|
237 |
)
|
238 |
self.num_processes = 1
|
239 |
self.process_index = self.local_process_index = 0
|
|
|
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"
|
|
|
248 |
f"Local process index: {self.local_process_index}\n"
|
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"
|
|
|
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
|
286 |
nothing when the script is only run in one process). Useful to do before saving a model.
|
|
|
287 |
Example:
|
|
|
288 |
```python
|
289 |
>>> # Assuming two GPU processes
|
290 |
>>> import time
|
291 |
>>> from accelerate.state import PartialState
|
|
|
292 |
>>> state = PartialState()
|
293 |
>>> if state.is_main_process:
|
294 |
... time.sleep(2)
|
|
|
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()
|
|
|
316 |
yield
|
|
|
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
|
323 |
distributed inference, such as with different prompts.
|
|
|
324 |
Note that when using a `dict`, all keys need to have the same number of elements.
|
|
|
325 |
Args:
|
326 |
inputs (`list`, `tuple`, `torch.Tensor`, or `dict` of `list`/`tuple`/`torch.Tensor`):
|
327 |
The input to split between processes.
|
|
|
329 |
Whether to apply padding by repeating the last element of the input so that all processes have the same
|
330 |
number of elements. Useful when trying to perform actions such as `gather()` on the outputs or passing
|
331 |
in less inputs than there are processes. If so, just remember to drop the padded elements afterwards.
|
|
|
|
|
332 |
Example:
|
|
|
333 |
```python
|
334 |
# Assume there are two processes
|
335 |
from accelerate import PartialState
|
|
|
336 |
state = PartialState()
|
337 |
with state.split_between_processes(["A", "B", "C"]) as inputs:
|
338 |
print(inputs)
|
|
|
340 |
["A", "B"]
|
341 |
# Process 1
|
342 |
["C"]
|
|
|
343 |
with state.split_between_processes(["A", "B", "C"], apply_padding=True) as inputs:
|
344 |
print(inputs)
|
345 |
# Process 0
|
|
|
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):
|
|
|
371 |
if apply_padding:
|
372 |
if isinstance(result, torch.Tensor):
|
373 |
from accelerate.utils import pad_across_processes, send_to_device
|
|
|
374 |
# The tensor needs to be on the device before we can pad it
|
375 |
tensorized_result = send_to_device(result, self.device)
|
376 |
result = pad_across_processes(tensorized_result, pad_index=inputs[-1])
|
|
|
383 |
return inputs
|
384 |
else:
|
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.
|
|
|
391 |
The other processes will enter the with block after the main process exits.
|
|
|
392 |
Example:
|
|
|
393 |
```python
|
394 |
>>> from accelerate import Accelerator
|
|
|
395 |
>>> accelerator = Accelerator()
|
396 |
>>> with accelerator.main_process_first():
|
397 |
... # This will be printed first by process 0 then in a seemingly
|
|
|
400 |
```
|
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.
|
|
|
407 |
The other processes will enter the with block after the main process exits.
|
|
|
408 |
Example:
|
|
|
409 |
```python
|
410 |
>>> from accelerate.state import PartialState
|
|
|
411 |
>>> state = PartialState()
|
412 |
>>> with state.local_main_process_first():
|
413 |
... # This will be printed first by local process 0 then in a seemingly
|
|
|
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.
|
|
|
422 |
Args:
|
423 |
function (`Callable`): The function to decorate.
|
|
|
424 |
Example:
|
|
|
425 |
```python
|
426 |
>>> from accelerate.state import PartialState
|
|
|
427 |
>>> state = PartialState()
|
|
|
|
|
428 |
>>> @state.on_main_process
|
429 |
... def print_something():
|
430 |
... print("This will be printed by process 0 only.")
|
|
|
|
|
431 |
>>> print_something()
|
432 |
"This will be printed by process 0 only"
|
433 |
```
|
|
|
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.
|
|
|
443 |
Args:
|
444 |
function (`Callable`): The function to decorate.
|
|
|
445 |
Example:
|
446 |
```python
|
447 |
# Assume we have 2 servers with 4 processes each.
|
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:
|
455 |
"This will be printed by process 0 only"
|
|
|
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.
|
|
|
466 |
Args:
|
467 |
function (`Callable`): The function to decorate.
|
|
|
468 |
Example:
|
469 |
```python
|
470 |
# Assume we have 4 processes.
|
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"
|
478 |
```
|
|
|
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.
|
|
|
486 |
Args:
|
487 |
function (`Callable`, `optional`):
|
488 |
The function to decorate.
|
489 |
process_index (`int`, `optional`):
|
490 |
The index of the process on which to run the function.
|
|
|
491 |
Example:
|
492 |
```python
|
493 |
# Assume we have 4 processes.
|
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"
|
501 |
```
|
|
|
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.
|
|
|
511 |
Args:
|
512 |
function (`Callable`, *optional*):
|
513 |
The function to decorate.
|
514 |
local_process_index (`int`, *optional*):
|
515 |
The index of the local process on which to run the function.
|
|
|
516 |
Example:
|
517 |
```python
|
518 |
# Assume we have 2 servers with 4 processes each.
|
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:
|
526 |
"Printed on process 2"
|
|
|
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 |
"""
|
|
|
556 |
return torch.device("npu")
|
557 |
else:
|
558 |
return torch.device("cpu")
|
|
|
|
|
559 |
class AcceleratorState:
|
560 |
"""
|
561 |
Singleton class that has information about the current training environment.
|
|
|
562 |
**Available attributes:**
|
|
|
563 |
- **device** (`torch.device`) -- The device to use.
|
564 |
- **distributed_type** ([`~accelerate.state.DistributedType`]) -- The type of distributed environment currently
|
565 |
in use.
|
|
|
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,
|
|
|
613 |
"or higher, compute capability of 8.9 or higher). Will use FP16 instead."
|
614 |
)
|
615 |
mixed_precision = "fp16"
|
|
|
616 |
self.dynamo_plugin = dynamo_plugin
|
617 |
if not _from_accelerator:
|
618 |
raise ValueError(
|
|
|
661 |
if self._mixed_precision != "no":
|
662 |
fsdp_plugin.set_mixed_precision(self._mixed_precision)
|
663 |
self.fsdp_plugin = fsdp_plugin
|
|
|
664 |
if (
|
665 |
self.dynamo_plugin.backend != DynamoBackend.NO
|
666 |
and self._mixed_precision == "no"
|
|
|
668 |
):
|
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:
|
|
|
688 |
and self.distributed_type != DistributedType.DEEPSPEED
|
689 |
):
|
690 |
raise ValueError(err.format(flag=f"mixed_precision='{mixed_precision}'"))
|
|
|
691 |
# For backward compatibility
|
692 |
@property
|
693 |
def use_fp16(self):
|
|
|
697 |
FutureWarning,
|
698 |
)
|
699 |
return self._mixed_precision != "no"
|
|
|
700 |
@property
|
701 |
def mixed_precision(self):
|
702 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
|
|
710 |
else:
|
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
|
743 |
distributed inference, such as with different prompts.
|
|
|
744 |
Note that when using a `dict`, all keys need to have the same number of elements.
|
|
|
745 |
Args:
|
746 |
inputs (`list`, `tuple`, `torch.Tensor`, or `dict` of `list`/`tuple`/`torch.Tensor`):
|
747 |
The input to split between processes.
|
|
|
749 |
Whether to apply padding by repeating the last element of the input so that all processes have the same
|
750 |
number of elements. Useful when trying to perform actions such as `gather()` on the outputs or passing
|
751 |
in less inputs than there are processes. If so, just remember to drop the padded elements afterwards.
|
|
|
|
|
752 |
Example:
|
|
|
753 |
```python
|
754 |
# Assume there are two processes
|
755 |
from accelerate.state import AcceleratorState
|
|
|
756 |
state = AcceleratorState()
|
757 |
with state.split_between_processes(["A", "B", "C"]) as inputs:
|
758 |
print(inputs)
|
|
|
760 |
["A", "B"]
|
761 |
# Process 1
|
762 |
["C"]
|
|
|
763 |
with state.split_between_processes(["A", "B", "C"], apply_padding=True) as inputs:
|
764 |
print(inputs)
|
765 |
# Process 0
|
|
|
770 |
"""
|
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.
|
|
|
777 |
The other processes will enter the with block after the main process exits.
|
778 |
"""
|
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.
|
|
|
785 |
The other processes will enter the with block after the main process exits.
|
786 |
"""
|
787 |
with PartialState().local_main_process_first():
|
788 |
yield
|
|
|
789 |
def print(self, *args, **kwargs):
|
790 |
PartialState().print(*args, **kwargs)
|
|
|
|
|
791 |
class GradientState:
|
792 |
"""
|
793 |
Singleton class that has information related to gradient synchronization for gradient accumulation
|
|
|
794 |
**Available attributes:**
|
|
|
795 |
- **end_of_dataloader** (`bool`) -- Whether we have reached the end the current dataloader
|
796 |
- **remainder** (`int`) -- The number of extra samples that were added from padding the dataloader
|
797 |
- **sync_gradients** (`bool`) -- Whether the gradients should be synced across all devices
|
|
|
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:
|
|
|
814 |
self.plugin_kwargs = (
|
815 |
gradient_accumulation_plugin.to_kwargs() if gradient_accumulation_plugin is not None else {}
|
816 |
)
|
|
|
817 |
# Plugin args are different and can be updated
|
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 |
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"
|
src/tracking.py
CHANGED
@@ -1,39 +1,25 @@
|
|
1 |
-
|
2 |
-
|
3 |
# Expectation:
|
4 |
# Provide a project dir name, then each type of logger gets stored in project/{`logging_dir`}
|
5 |
-
|
6 |
_available_trackers = []
|
7 |
-
|
8 |
if is_tensorboard_available():
|
9 |
_available_trackers.append(LoggerType.TENSORBOARD)
|
10 |
-
|
11 |
if is_wandb_available():
|
12 |
_available_trackers.append(LoggerType.WANDB)
|
13 |
-
|
14 |
if is_comet_ml_available():
|
15 |
_available_trackers.append(LoggerType.COMETML)
|
16 |
-
|
17 |
if is_aim_available():
|
18 |
_available_trackers.append(LoggerType.AIM)
|
19 |
-
|
20 |
if is_mlflow_available():
|
21 |
_available_trackers.append(LoggerType.MLFLOW)
|
22 |
-
|
23 |
if is_clearml_available():
|
24 |
_available_trackers.append(LoggerType.CLEARML)
|
25 |
-
|
26 |
if is_dvclive_available():
|
27 |
_available_trackers.append(LoggerType.DVCLIVE)
|
28 |
-
|
29 |
logger = get_logger(__name__)
|
30 |
-
|
31 |
-
|
32 |
def on_main_process(function):
|
33 |
"""
|
34 |
Decorator to selectively run the decorated function on the main process only based on the `main_process_only`
|
35 |
attribute in a class.
|
36 |
-
|
37 |
Checks at function execution rather than initialization time, not triggering the initialization of the
|
38 |
`PartialState`.
|
39 |
"""
|
@@ -43,33 +29,23 @@ def on_main_process(function):
|
|
43 |
return PartialState().on_main_process(function)(self, *args, **kwargs)
|
44 |
else:
|
45 |
return function(self, *args, **kwargs)
|
46 |
-
|
47 |
return execute_on_main_process
|
48 |
-
|
49 |
-
|
50 |
def get_available_trackers():
|
51 |
"Returns a list of all supported available trackers in the system"
|
52 |
return _available_trackers
|
53 |
-
|
54 |
-
|
55 |
class GeneralTracker:
|
56 |
"""
|
57 |
A base Tracker class to be used for all logging integration implementations.
|
58 |
-
|
59 |
Each function should take in `**kwargs` that will automatically be passed in from a base dictionary provided to
|
60 |
[`Accelerator`].
|
61 |
-
|
62 |
Should implement `name`, `requires_logging_directory`, and `tracker` properties such that:
|
63 |
-
|
64 |
`name` (`str`): String representation of the tracker class name, such as "TensorBoard" `requires_logging_directory`
|
65 |
(`bool`): Whether the logger requires a directory to store their logs. `tracker` (`object`): Should return internal
|
66 |
tracking mechanism used by a tracker class (such as the `run` for wandb)
|
67 |
-
|
68 |
Implementations can also include a `main_process_only` (`bool`) attribute to toggle if relevent logging, init, and
|
69 |
other functions should occur on the main process or across all processes (by default will use `True`)
|
70 |
"""
|
71 |
main_process_only = True
|
72 |
-
|
73 |
def __init__(self, _blank=False):
|
74 |
if not _blank:
|
75 |
err = ""
|
@@ -79,7 +55,6 @@ class GeneralTracker:
|
|
79 |
if len(err) > 0:
|
80 |
err += ", "
|
81 |
err += "`requires_logging_directory`"
|
82 |
-
|
83 |
# as tracker is a @property that relies on post-init
|
84 |
if "tracker" not in dir(self):
|
85 |
if len(err) > 0:
|
@@ -91,24 +66,20 @@ class GeneralTracker:
|
|
91 |
f"required attributes. Please define them in the class definition: "
|
92 |
f"{err}"
|
93 |
)
|
94 |
-
|
95 |
def store_init_configuration(self, values: dict):
|
96 |
"""
|
97 |
Logs `values` as hyperparameters for the run. Implementations should use the experiment configuration
|
98 |
functionality of a tracking API.
|
99 |
-
|
100 |
Args:
|
101 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
102 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
103 |
`str`, `float`, `int`, or `None`.
|
104 |
"""
|
105 |
pass
|
106 |
-
|
107 |
def log(self, values: dict, step: Optional[int], **kwargs):
|
108 |
"""
|
109 |
Logs `values` to the current run. Base `log` implementations of a tracking API should go in here, along with
|
110 |
special behavior for the `step parameter.
|
111 |
-
|
112 |
Args:
|
113 |
values (Dictionary `str` to `str`, `float`, or `int`):
|
114 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, or `int`.
|
@@ -116,19 +87,15 @@ class GeneralTracker:
|
|
116 |
The run step. If included, the log will be affiliated with this step.
|
117 |
"""
|
118 |
pass
|
119 |
-
|
120 |
def finish(self):
|
121 |
"""
|
122 |
Should run any finalizing functions within the tracking API. If the API should not have one, just don't
|
123 |
overwrite that method.
|
124 |
"""
|
125 |
pass
|
126 |
-
|
127 |
-
|
128 |
class TensorBoardTracker(GeneralTracker):
|
129 |
"""
|
130 |
A `Tracker` class that supports `tensorboard`. Should be initialized at the start of your script.
|
131 |
-
|
132 |
Args:
|
133 |
run_name (`str`):
|
134 |
The name of the experiment run
|
@@ -139,7 +106,6 @@ class TensorBoardTracker(GeneralTracker):
|
|
139 |
"""
|
140 |
name = "tensorboard"
|
141 |
requires_logging_directory = True
|
142 |
-
|
143 |
@on_main_process
|
144 |
def __init__(self, run_name: str, logging_dir: Union[str, os.PathLike], **kwargs):
|
145 |
try:
|
@@ -154,17 +120,14 @@ class TensorBoardTracker(GeneralTracker):
|
|
154 |
logger.debug(
|
155 |
"Make sure to log any initial configurations with `self.store_init_configuration` before training!"
|
156 |
)
|
157 |
-
|
158 |
@property
|
159 |
def tracker(self):
|
160 |
return self.writer
|
161 |
-
|
162 |
@on_main_process
|
163 |
def store_init_configuration(self, values: dict):
|
164 |
"""
|
165 |
Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment. Stores the
|
166 |
hyperparameters in a yaml file for future use.
|
167 |
-
|
168 |
Args:
|
169 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
170 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
@@ -182,12 +145,10 @@ class TensorBoardTracker(GeneralTracker):
|
|
182 |
logger.error("Serialization to store hyperparameters failed")
|
183 |
raise
|
184 |
logger.debug("Stored initial configuration hyperparameters to TensorBoard and hparams yaml file")
|
185 |
-
|
186 |
@on_main_process
|
187 |
def log(self, values: dict, step: Optional[int] = None, **kwargs):
|
188 |
"""
|
189 |
Logs `values` to the current run.
|
190 |
-
|
191 |
Args:
|
192 |
values (Dictionary `str` to `str`, `float`, `int` or `dict` of `str` to `float`/`int`):
|
193 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, `int` or `dict` of
|
@@ -208,12 +169,10 @@ class TensorBoardTracker(GeneralTracker):
|
|
208 |
self.writer.add_scalars(k, v, global_step=step, **kwargs)
|
209 |
self.writer.flush()
|
210 |
logger.debug("Successfully logged to TensorBoard")
|
211 |
-
|
212 |
@on_main_process
|
213 |
def log_images(self, values: dict, step: Optional[int], **kwargs):
|
214 |
"""
|
215 |
Logs `images` to the current run.
|
216 |
-
|
217 |
Args:
|
218 |
values (Dictionary `str` to `List` of `np.ndarray` or `PIL.Image`):
|
219 |
Values to be logged as key-value pairs. The values need to have type `List` of `np.ndarray` or
|
@@ -225,7 +184,6 @@ class TensorBoardTracker(GeneralTracker):
|
|
225 |
for k, v in values.items():
|
226 |
self.writer.add_images(k, v, global_step=step, **kwargs)
|
227 |
logger.debug("Successfully logged images to TensorBoard")
|
228 |
-
|
229 |
@on_main_process
|
230 |
def finish(self):
|
231 |
"""
|
@@ -233,12 +191,9 @@ class TensorBoardTracker(GeneralTracker):
|
|
233 |
"""
|
234 |
self.writer.close()
|
235 |
logger.debug("TensorBoard writer closed")
|
236 |
-
|
237 |
-
|
238 |
class WandBTracker(GeneralTracker):
|
239 |
"""
|
240 |
A `Tracker` class that supports `wandb`. Should be initialized at the start of your script.
|
241 |
-
|
242 |
Args:
|
243 |
run_name (`str`):
|
244 |
The name of the experiment run.
|
@@ -248,44 +203,35 @@ class WandBTracker(GeneralTracker):
|
|
248 |
name = "wandb"
|
249 |
requires_logging_directory = False
|
250 |
main_process_only = False
|
251 |
-
|
252 |
@on_main_process
|
253 |
def __init__(self, run_name: str, **kwargs):
|
254 |
super().__init__()
|
255 |
self.run_name = run_name
|
256 |
-
|
257 |
import wandb
|
258 |
-
|
259 |
self.run = wandb.init(project=self.run_name, **kwargs)
|
260 |
logger.debug(f"Initialized WandB project {self.run_name}")
|
261 |
logger.debug(
|
262 |
"Make sure to log any initial configurations with `self.store_init_configuration` before training!"
|
263 |
)
|
264 |
-
|
265 |
@property
|
266 |
def tracker(self):
|
267 |
return self.run
|
268 |
-
|
269 |
@on_main_process
|
270 |
def store_init_configuration(self, values: dict):
|
271 |
"""
|
272 |
Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
|
273 |
-
|
274 |
Args:
|
275 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
276 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
277 |
`str`, `float`, `int`, or `None`.
|
278 |
"""
|
279 |
import wandb
|
280 |
-
|
281 |
wandb.config.update(values, allow_val_change=True)
|
282 |
logger.debug("Stored initial configuration hyperparameters to WandB")
|
283 |
-
|
284 |
@on_main_process
|
285 |
def log(self, values: dict, step: Optional[int] = None, **kwargs):
|
286 |
"""
|
287 |
Logs `values` to the current run.
|
288 |
-
|
289 |
Args:
|
290 |
values (Dictionary `str` to `str`, `float`, `int` or `dict` of `str` to `float`/`int`):
|
291 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, `int` or `dict` of
|
@@ -297,12 +243,10 @@ class WandBTracker(GeneralTracker):
|
|
297 |
"""
|
298 |
self.run.log(values, step=step, **kwargs)
|
299 |
logger.debug("Successfully logged to WandB")
|
300 |
-
|
301 |
@on_main_process
|
302 |
def log_images(self, values: dict, step: Optional[int] = None, **kwargs):
|
303 |
"""
|
304 |
Logs `images` to the current run.
|
305 |
-
|
306 |
Args:
|
307 |
values (Dictionary `str` to `List` of `np.ndarray` or `PIL.Image`):
|
308 |
Values to be logged as key-value pairs. The values need to have type `List` of `np.ndarray` or
|
@@ -312,11 +256,9 @@ class WandBTracker(GeneralTracker):
|
|
312 |
Additional key word arguments passed along to the `wandb.log` method.
|
313 |
"""
|
314 |
import wandb
|
315 |
-
|
316 |
for k, v in values.items():
|
317 |
self.log({k: [wandb.Image(image) for image in v]}, step=step, **kwargs)
|
318 |
logger.debug("Successfully logged images to WandB")
|
319 |
-
|
320 |
@on_main_process
|
321 |
def log_table(
|
322 |
self,
|
@@ -330,7 +272,6 @@ class WandBTracker(GeneralTracker):
|
|
330 |
"""
|
331 |
Log a Table containing any object type (text, image, audio, video, molecule, html, etc). Can be defined either
|
332 |
with `columns` and `data` or with `dataframe`.
|
333 |
-
|
334 |
Args:
|
335 |
table_name (`str`):
|
336 |
The name to give to the logged table on the wandb workspace
|
@@ -344,10 +285,8 @@ class WandBTracker(GeneralTracker):
|
|
344 |
The run step. If included, the log will be affiliated with this step.
|
345 |
"""
|
346 |
import wandb
|
347 |
-
|
348 |
values = {table_name: wandb.Table(columns=columns, data=data, dataframe=dataframe)}
|
349 |
self.log(values, step=step, **kwargs)
|
350 |
-
|
351 |
@on_main_process
|
352 |
def finish(self):
|
353 |
"""
|
@@ -355,14 +294,10 @@ class WandBTracker(GeneralTracker):
|
|
355 |
"""
|
356 |
self.run.finish()
|
357 |
logger.debug("WandB run closed")
|
358 |
-
|
359 |
-
|
360 |
class CometMLTracker(GeneralTracker):
|
361 |
"""
|
362 |
A `Tracker` class that supports `comet_ml`. Should be initialized at the start of your script.
|
363 |
-
|
364 |
API keys must be stored in a Comet config file.
|
365 |
-
|
366 |
Args:
|
367 |
run_name (`str`):
|
368 |
The name of the experiment run.
|
@@ -371,29 +306,23 @@ class CometMLTracker(GeneralTracker):
|
|
371 |
"""
|
372 |
name = "comet_ml"
|
373 |
requires_logging_directory = False
|
374 |
-
|
375 |
@on_main_process
|
376 |
def __init__(self, run_name: str, **kwargs):
|
377 |
super().__init__()
|
378 |
self.run_name = run_name
|
379 |
-
|
380 |
from comet_ml import Experiment
|
381 |
-
|
382 |
self.writer = Experiment(project_name=run_name, **kwargs)
|
383 |
logger.debug(f"Initialized CometML project {self.run_name}")
|
384 |
logger.debug(
|
385 |
"Make sure to log any initial configurations with `self.store_init_configuration` before training!"
|
386 |
)
|
387 |
-
|
388 |
@property
|
389 |
def tracker(self):
|
390 |
return self.writer
|
391 |
-
|
392 |
@on_main_process
|
393 |
def store_init_configuration(self, values: dict):
|
394 |
"""
|
395 |
Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
|
396 |
-
|
397 |
Args:
|
398 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
399 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
@@ -401,12 +330,10 @@ class CometMLTracker(GeneralTracker):
|
|
401 |
"""
|
402 |
self.writer.log_parameters(values)
|
403 |
logger.debug("Stored initial configuration hyperparameters to CometML")
|
404 |
-
|
405 |
@on_main_process
|
406 |
def log(self, values: dict, step: Optional[int] = None, **kwargs):
|
407 |
"""
|
408 |
Logs `values` to the current run.
|
409 |
-
|
410 |
Args:
|
411 |
values (Dictionary `str` to `str`, `float`, `int` or `dict` of `str` to `float`/`int`):
|
412 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, `int` or `dict` of
|
@@ -427,7 +354,6 @@ class CometMLTracker(GeneralTracker):
|
|
427 |
elif isinstance(v, dict):
|
428 |
self.writer.log_metrics(v, step=step, **kwargs)
|
429 |
logger.debug("Successfully logged to CometML")
|
430 |
-
|
431 |
@on_main_process
|
432 |
def finish(self):
|
433 |
"""
|
@@ -435,12 +361,9 @@ class CometMLTracker(GeneralTracker):
|
|
435 |
"""
|
436 |
self.writer.end()
|
437 |
logger.debug("CometML run closed")
|
438 |
-
|
439 |
-
|
440 |
class AimTracker(GeneralTracker):
|
441 |
"""
|
442 |
A `Tracker` class that supports `aim`. Should be initialized at the start of your script.
|
443 |
-
|
444 |
Args:
|
445 |
run_name (`str`):
|
446 |
The name of the experiment run.
|
@@ -449,40 +372,32 @@ class AimTracker(GeneralTracker):
|
|
449 |
"""
|
450 |
name = "aim"
|
451 |
requires_logging_directory = True
|
452 |
-
|
453 |
@on_main_process
|
454 |
def __init__(self, run_name: str, logging_dir: Optional[Union[str, os.PathLike]] = ".", **kwargs):
|
455 |
self.run_name = run_name
|
456 |
-
|
457 |
from aim import Run
|
458 |
-
|
459 |
self.writer = Run(repo=logging_dir, **kwargs)
|
460 |
self.writer.name = self.run_name
|
461 |
logger.debug(f"Initialized Aim project {self.run_name}")
|
462 |
logger.debug(
|
463 |
"Make sure to log any initial configurations with `self.store_init_configuration` before training!"
|
464 |
)
|
465 |
-
|
466 |
@property
|
467 |
def tracker(self):
|
468 |
return self.writer
|
469 |
-
|
470 |
@on_main_process
|
471 |
def store_init_configuration(self, values: dict):
|
472 |
"""
|
473 |
Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
|
474 |
-
|
475 |
Args:
|
476 |
values (`dict`):
|
477 |
Values to be stored as initial hyperparameters as key-value pairs.
|
478 |
"""
|
479 |
self.writer["hparams"] = values
|
480 |
-
|
481 |
@on_main_process
|
482 |
def log(self, values: dict, step: Optional[int], **kwargs):
|
483 |
"""
|
484 |
Logs `values` to the current run.
|
485 |
-
|
486 |
Args:
|
487 |
values (`dict`):
|
488 |
Values to be logged as key-value pairs.
|
@@ -494,12 +409,10 @@ class AimTracker(GeneralTracker):
|
|
494 |
# Note: replace this with the dictionary support when merged
|
495 |
for key, value in values.items():
|
496 |
self.writer.track(value, name=key, step=step, **kwargs)
|
497 |
-
|
498 |
@on_main_process
|
499 |
def log_images(self, values: dict, step: Optional[int] = None, kwargs: Optional[Dict[str, dict]] = None):
|
500 |
"""
|
501 |
Logs `images` to the current run.
|
502 |
-
|
503 |
Args:
|
504 |
values (`Dict[str, Union[np.ndarray, PIL.Image, Tuple[np.ndarray, str], Tuple[PIL.Image, str]]]`):
|
505 |
Values to be logged as key-value pairs. The values need to have type `np.ndarray` or PIL.Image. If a
|
@@ -511,14 +424,11 @@ class AimTracker(GeneralTracker):
|
|
511 |
keys `aim_image` and `track`, respectively.
|
512 |
"""
|
513 |
import aim
|
514 |
-
|
515 |
aim_image_kw = {}
|
516 |
track_kw = {}
|
517 |
-
|
518 |
if kwargs is not None:
|
519 |
aim_image_kw = kwargs.get("aim_image", {})
|
520 |
track_kw = kwargs.get("track", {})
|
521 |
-
|
522 |
for key, value in values.items():
|
523 |
if isinstance(value, tuple):
|
524 |
img, caption = value
|
@@ -526,19 +436,15 @@ class AimTracker(GeneralTracker):
|
|
526 |
img, caption = value, ""
|
527 |
aim_image = aim.Image(img, caption=caption, **aim_image_kw)
|
528 |
self.writer.track(aim_image, name=key, step=step, **track_kw)
|
529 |
-
|
530 |
@on_main_process
|
531 |
def finish(self):
|
532 |
"""
|
533 |
Closes `aim` writer
|
534 |
"""
|
535 |
self.writer.close()
|
536 |
-
|
537 |
-
|
538 |
class MLflowTracker(GeneralTracker):
|
539 |
"""
|
540 |
A `Tracker` class that supports `mlflow`. Should be initialized at the start of your script.
|
541 |
-
|
542 |
Args:
|
543 |
experiment_name (`str`, *optional*):
|
544 |
Name of the experiment. Environment variable MLFLOW_EXPERIMENT_NAME has priority over this argument.
|
@@ -564,7 +470,6 @@ class MLflowTracker(GeneralTracker):
|
|
564 |
"""
|
565 |
name = "mlflow"
|
566 |
requires_logging_directory = False
|
567 |
-
|
568 |
@on_main_process
|
569 |
def __init__(
|
570 |
self,
|
@@ -581,11 +486,8 @@ class MLflowTracker(GeneralTracker):
|
|
581 |
tags = os.getenv("MLFLOW_TAGS", tags)
|
582 |
if isinstance(tags, str):
|
583 |
tags = json.loads(tags)
|
584 |
-
|
585 |
nested_run = os.getenv("MLFLOW_NESTED_RUN", nested_run)
|
586 |
-
|
587 |
import mlflow
|
588 |
-
|
589 |
exps = mlflow.search_experiments(filter_string=f"name = '{experiment_name}'")
|
590 |
if len(exps) > 0:
|
591 |
if len(exps) > 1:
|
@@ -597,7 +499,6 @@ class MLflowTracker(GeneralTracker):
|
|
597 |
artifact_location=logging_dir,
|
598 |
tags=tags,
|
599 |
)
|
600 |
-
|
601 |
self.active_run = mlflow.start_run(
|
602 |
run_id=run_id,
|
603 |
experiment_id=experiment_id,
|
@@ -606,27 +507,22 @@ class MLflowTracker(GeneralTracker):
|
|
606 |
tags=tags,
|
607 |
description=description,
|
608 |
)
|
609 |
-
|
610 |
logger.debug(f"Initialized mlflow experiment {experiment_name}")
|
611 |
logger.debug(
|
612 |
"Make sure to log any initial configurations with `self.store_init_configuration` before training!"
|
613 |
)
|
614 |
-
|
615 |
@property
|
616 |
def tracker(self):
|
617 |
return self.active_run
|
618 |
-
|
619 |
@on_main_process
|
620 |
def store_init_configuration(self, values: dict):
|
621 |
"""
|
622 |
Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment.
|
623 |
-
|
624 |
Args:
|
625 |
values (`dict`):
|
626 |
Values to be stored as initial hyperparameters as key-value pairs.
|
627 |
"""
|
628 |
import mlflow
|
629 |
-
|
630 |
for name, value in list(values.items()):
|
631 |
# internally, all values are converted to str in MLflow
|
632 |
if len(str(value)) > mlflow.utils.validation.MAX_PARAM_VAL_LENGTH:
|
@@ -635,20 +531,15 @@ class MLflowTracker(GeneralTracker):
|
|
635 |
f" log_param() only accepts values no longer than {mlflow.utils.validation.MAX_PARAM_VAL_LENGTH} characters so we dropped this attribute."
|
636 |
)
|
637 |
del values[name]
|
638 |
-
|
639 |
values_list = list(values.items())
|
640 |
-
|
641 |
# MLflow cannot log more than 100 values in one go, so we have to split it
|
642 |
for i in range(0, len(values_list), mlflow.utils.validation.MAX_PARAMS_TAGS_PER_BATCH):
|
643 |
mlflow.log_params(dict(values_list[i : i + mlflow.utils.validation.MAX_PARAMS_TAGS_PER_BATCH]))
|
644 |
-
|
645 |
logger.debug("Stored initial configuration hyperparameters to MLflow")
|
646 |
-
|
647 |
@on_main_process
|
648 |
def log(self, values: dict, step: Optional[int]):
|
649 |
"""
|
650 |
Logs `values` to the current run.
|
651 |
-
|
652 |
Args:
|
653 |
values (`dict`):
|
654 |
Values to be logged as key-value pairs.
|
@@ -665,24 +556,18 @@ class MLflowTracker(GeneralTracker):
|
|
665 |
"MLflow's log_metric() only accepts float and int types so we dropped this attribute."
|
666 |
)
|
667 |
import mlflow
|
668 |
-
|
669 |
mlflow.log_metrics(metrics, step=step)
|
670 |
logger.debug("Successfully logged to mlflow")
|
671 |
-
|
672 |
@on_main_process
|
673 |
def finish(self):
|
674 |
"""
|
675 |
End the active MLflow run.
|
676 |
"""
|
677 |
import mlflow
|
678 |
-
|
679 |
mlflow.end_run()
|
680 |
-
|
681 |
-
|
682 |
class ClearMLTracker(GeneralTracker):
|
683 |
"""
|
684 |
A `Tracker` class that supports `clearml`. Should be initialized at the start of your script.
|
685 |
-
|
686 |
Args:
|
687 |
run_name (`str`, *optional*):
|
688 |
Name of the experiment. Environment variables `CLEARML_PROJECT` and `CLEARML_TASK` have priority over this
|
@@ -692,43 +577,35 @@ class ClearMLTracker(GeneralTracker):
|
|
692 |
"""
|
693 |
name = "clearml"
|
694 |
requires_logging_directory = False
|
695 |
-
|
696 |
@on_main_process
|
697 |
def __init__(self, run_name: str = None, **kwargs):
|
698 |
from clearml import Task
|
699 |
-
|
700 |
current_task = Task.current_task()
|
701 |
self._initialized_externally = False
|
702 |
if current_task:
|
703 |
self._initialized_externally = True
|
704 |
self.task = current_task
|
705 |
return
|
706 |
-
|
707 |
kwargs.setdefault("project_name", os.environ.get("CLEARML_PROJECT", run_name))
|
708 |
kwargs.setdefault("task_name", os.environ.get("CLEARML_TASK", run_name))
|
709 |
self.task = Task.init(**kwargs)
|
710 |
-
|
711 |
@property
|
712 |
def tracker(self):
|
713 |
return self.task
|
714 |
-
|
715 |
@on_main_process
|
716 |
def store_init_configuration(self, values: dict):
|
717 |
"""
|
718 |
Connect configuration dictionary to the Task object. Should be run at the beginning of your experiment.
|
719 |
-
|
720 |
Args:
|
721 |
values (`dict`):
|
722 |
Values to be stored as initial hyperparameters as key-value pairs.
|
723 |
"""
|
724 |
return self.task.connect_configuration(values)
|
725 |
-
|
726 |
@on_main_process
|
727 |
def log(self, values: Dict[str, Union[int, float]], step: Optional[int] = None, **kwargs):
|
728 |
"""
|
729 |
Logs `values` dictionary to the current run. The dictionary keys must be strings. The dictionary values must be
|
730 |
ints or floats
|
731 |
-
|
732 |
Args:
|
733 |
values (`Dict[str, Union[int, float]]`):
|
734 |
Values to be logged as key-value pairs. If the key starts with 'eval_'/'test_'/'train_', the value will
|
@@ -756,12 +633,10 @@ class ClearMLTracker(GeneralTracker):
|
|
756 |
continue
|
757 |
title, series = ClearMLTracker._get_title_series(k)
|
758 |
clearml_logger.report_scalar(title=title, series=series, value=v, iteration=step, **kwargs)
|
759 |
-
|
760 |
@on_main_process
|
761 |
def log_images(self, values: dict, step: Optional[int] = None, **kwargs):
|
762 |
"""
|
763 |
Logs `images` to the current run.
|
764 |
-
|
765 |
Args:
|
766 |
values (`Dict[str, List[Union[np.ndarray, PIL.Image]]`):
|
767 |
Values to be logged as key-value pairs. The values need to have type `List` of `np.ndarray` or
|
@@ -774,7 +649,6 @@ class ClearMLTracker(GeneralTracker):
|
|
774 |
for k, v in values.items():
|
775 |
title, series = ClearMLTracker._get_title_series(k)
|
776 |
clearml_logger.report_image(title=title, series=series, iteration=step, image=v, **kwargs)
|
777 |
-
|
778 |
@on_main_process
|
779 |
def log_table(
|
780 |
self,
|
@@ -787,7 +661,6 @@ class ClearMLTracker(GeneralTracker):
|
|
787 |
):
|
788 |
"""
|
789 |
Log a Table to the task. Can be defined eitherwith `columns` and `data` or with `dataframe`.
|
790 |
-
|
791 |
Args:
|
792 |
table_name (`str`):
|
793 |
The name of the table
|
@@ -812,7 +685,6 @@ class ClearMLTracker(GeneralTracker):
|
|
812 |
to_report = [columns] + data if columns else data
|
813 |
title, series = ClearMLTracker._get_title_series(table_name)
|
814 |
self.task.get_logger().report_table(title=title, series=series, table_plot=to_report, iteration=step, **kwargs)
|
815 |
-
|
816 |
@on_main_process
|
817 |
def finish(self):
|
818 |
"""
|
@@ -821,66 +693,52 @@ class ClearMLTracker(GeneralTracker):
|
|
821 |
"""
|
822 |
if self.task and not self._initialized_externally:
|
823 |
self.task.close()
|
824 |
-
|
825 |
@staticmethod
|
826 |
def _get_title_series(name):
|
827 |
for prefix in ["eval", "test", "train"]:
|
828 |
if name.startswith(prefix + "_"):
|
829 |
return name[len(prefix) + 1 :], prefix
|
830 |
return name, "train"
|
831 |
-
|
832 |
-
|
833 |
class DVCLiveTracker(GeneralTracker):
|
834 |
"""
|
835 |
A `Tracker` class that supports `dvclive`. Should be initialized at the start of your script.
|
836 |
-
|
837 |
Args:
|
838 |
run_name (`str`, *optional*):
|
839 |
Ignored for dvclive. See `kwargs` instead.
|
840 |
kwargs:
|
841 |
Additional key word arguments passed along to [`dvclive.Live()`](https://dvc.org/doc/dvclive/live).
|
842 |
-
|
843 |
Example:
|
844 |
-
|
845 |
```py
|
846 |
from accelerate import Accelerator
|
847 |
-
|
848 |
accelerator = Accelerator(log_with="dvclive")
|
849 |
accelerator.init_trackers(project_name="my_project", init_kwargs={"dvclive": {"dir": "my_directory"}})
|
850 |
```
|
851 |
"""
|
852 |
name = "dvclive"
|
853 |
requires_logging_directory = False
|
854 |
-
|
855 |
@on_main_process
|
856 |
def __init__(self, run_name: Optional[str] = None, live: Optional[Any] = None, **kwargs):
|
857 |
from dvclive import Live
|
858 |
-
|
859 |
super().__init__()
|
860 |
self.live = live if live is not None else Live(**kwargs)
|
861 |
-
|
862 |
@property
|
863 |
def tracker(self):
|
864 |
return self.live
|
865 |
-
|
866 |
@on_main_process
|
867 |
def store_init_configuration(self, values: dict):
|
868 |
"""
|
869 |
Logs `values` as hyperparameters for the run. Should be run at the beginning of your experiment. Stores the
|
870 |
hyperparameters in a yaml file for future use.
|
871 |
-
|
872 |
Args:
|
873 |
values (Dictionary `str` to `bool`, `str`, `float`, `int`, or a List or Dict of those types):
|
874 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
875 |
`str`, `float`, or `int`.
|
876 |
"""
|
877 |
self.live.log_params(values)
|
878 |
-
|
879 |
@on_main_process
|
880 |
def log(self, values: dict, step: Optional[int] = None, **kwargs):
|
881 |
"""
|
882 |
Logs `values` to the current run.
|
883 |
-
|
884 |
Args:
|
885 |
values (Dictionary `str` to `str`, `float`, or `int`):
|
886 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, or `int`.
|
@@ -890,7 +748,6 @@ class DVCLiveTracker(GeneralTracker):
|
|
890 |
Additional key word arguments passed along to `dvclive.Live.log_metric()`.
|
891 |
"""
|
892 |
from dvclive.plots import Metric
|
893 |
-
|
894 |
if step is not None:
|
895 |
self.live.step = step
|
896 |
for k, v in values.items():
|
@@ -903,15 +760,12 @@ class DVCLiveTracker(GeneralTracker):
|
|
903 |
"This invocation of DVCLive's Live.log_metric() "
|
904 |
"is incorrect so we dropped this attribute."
|
905 |
)
|
906 |
-
|
907 |
@on_main_process
|
908 |
def finish(self):
|
909 |
"""
|
910 |
Closes `dvclive.Live()`.
|
911 |
"""
|
912 |
self.live.end()
|
913 |
-
|
914 |
-
|
915 |
LOGGER_TYPE_TO_CLASS = {
|
916 |
"aim": AimTracker,
|
917 |
"comet_ml": CometMLTracker,
|
@@ -921,8 +775,6 @@ LOGGER_TYPE_TO_CLASS = {
|
|
921 |
"clearml": ClearMLTracker,
|
922 |
"dvclive": DVCLiveTracker,
|
923 |
}
|
924 |
-
|
925 |
-
|
926 |
def filter_trackers(
|
927 |
log_with: List[Union[str, LoggerType, GeneralTracker]],
|
928 |
logging_dir: Union[str, os.PathLike] = None,
|
@@ -933,11 +785,9 @@ def filter_trackers(
|
|
933 |
- Filters out repeats of tracker types
|
934 |
- If `all` is in `log_with`, will return all trackers in the environment
|
935 |
- If a tracker requires a `logging_dir`, ensures that `logging_dir` is not `None`
|
936 |
-
|
937 |
Args:
|
938 |
log_with (list of `str`, [`~utils.LoggerType`] or [`~tracking.GeneralTracker`], *optional*):
|
939 |
A list of loggers to be setup for experiment tracking. Should be one or several of:
|
940 |
-
|
941 |
- `"all"`
|
942 |
- `"tensorboard"`
|
943 |
- `"wandb"`
|
@@ -974,5 +824,4 @@ def filter_trackers(
|
|
974 |
loggers.append(log_type)
|
975 |
else:
|
976 |
logger.debug(f"Tried adding logger {log_type}, but package is unavailable in the system.")
|
977 |
-
|
978 |
return loggers
|
|
|
|
|
|
|
1 |
# Expectation:
|
2 |
# Provide a project dir name, then each type of logger gets stored in project/{`logging_dir`}
|
|
|
3 |
_available_trackers = []
|
|
|
4 |
if is_tensorboard_available():
|
5 |
_available_trackers.append(LoggerType.TENSORBOARD)
|
|
|
6 |
if is_wandb_available():
|
7 |
_available_trackers.append(LoggerType.WANDB)
|
|
|
8 |
if is_comet_ml_available():
|
9 |
_available_trackers.append(LoggerType.COMETML)
|
|
|
10 |
if is_aim_available():
|
11 |
_available_trackers.append(LoggerType.AIM)
|
|
|
12 |
if is_mlflow_available():
|
13 |
_available_trackers.append(LoggerType.MLFLOW)
|
|
|
14 |
if is_clearml_available():
|
15 |
_available_trackers.append(LoggerType.CLEARML)
|
|
|
16 |
if is_dvclive_available():
|
17 |
_available_trackers.append(LoggerType.DVCLIVE)
|
|
|
18 |
logger = get_logger(__name__)
|
|
|
|
|
19 |
def on_main_process(function):
|
20 |
"""
|
21 |
Decorator to selectively run the decorated function on the main process only based on the `main_process_only`
|
22 |
attribute in a class.
|
|
|
23 |
Checks at function execution rather than initialization time, not triggering the initialization of the
|
24 |
`PartialState`.
|
25 |
"""
|
|
|
29 |
return PartialState().on_main_process(function)(self, *args, **kwargs)
|
30 |
else:
|
31 |
return function(self, *args, **kwargs)
|
|
|
32 |
return execute_on_main_process
|
|
|
|
|
33 |
def get_available_trackers():
|
34 |
"Returns a list of all supported available trackers in the system"
|
35 |
return _available_trackers
|
|
|
|
|
36 |
class GeneralTracker:
|
37 |
"""
|
38 |
A base Tracker class to be used for all logging integration implementations.
|
|
|
39 |
Each function should take in `**kwargs` that will automatically be passed in from a base dictionary provided to
|
40 |
[`Accelerator`].
|
|
|
41 |
Should implement `name`, `requires_logging_directory`, and `tracker` properties such that:
|
|
|
42 |
`name` (`str`): String representation of the tracker class name, such as "TensorBoard" `requires_logging_directory`
|
43 |
(`bool`): Whether the logger requires a directory to store their logs. `tracker` (`object`): Should return internal
|
44 |
tracking mechanism used by a tracker class (such as the `run` for wandb)
|
|
|
45 |
Implementations can also include a `main_process_only` (`bool`) attribute to toggle if relevent logging, init, and
|
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 = ""
|
|
|
55 |
if len(err) > 0:
|
56 |
err += ", "
|
57 |
err += "`requires_logging_directory`"
|
|
|
58 |
# as tracker is a @property that relies on post-init
|
59 |
if "tracker" not in dir(self):
|
60 |
if len(err) > 0:
|
|
|
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
|
72 |
functionality of a tracking API.
|
|
|
73 |
Args:
|
74 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
75 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
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
|
82 |
special behavior for the `step parameter.
|
|
|
83 |
Args:
|
84 |
values (Dictionary `str` to `str`, `float`, or `int`):
|
85 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, or `int`.
|
|
|
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
|
93 |
overwrite that method.
|
94 |
"""
|
95 |
pass
|
|
|
|
|
96 |
class TensorBoardTracker(GeneralTracker):
|
97 |
"""
|
98 |
A `Tracker` class that supports `tensorboard`. Should be initialized at the start of your script.
|
|
|
99 |
Args:
|
100 |
run_name (`str`):
|
101 |
The name of the experiment run
|
|
|
106 |
"""
|
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:
|
|
|
120 |
logger.debug(
|
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
|
130 |
hyperparameters in a yaml file for future use.
|
|
|
131 |
Args:
|
132 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
133 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
|
|
145 |
logger.error("Serialization to store hyperparameters failed")
|
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.
|
|
|
152 |
Args:
|
153 |
values (Dictionary `str` to `str`, `float`, `int` or `dict` of `str` to `float`/`int`):
|
154 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, `int` or `dict` of
|
|
|
169 |
self.writer.add_scalars(k, v, global_step=step, **kwargs)
|
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.
|
|
|
176 |
Args:
|
177 |
values (Dictionary `str` to `List` of `np.ndarray` or `PIL.Image`):
|
178 |
Values to be logged as key-value pairs. The values need to have type `List` of `np.ndarray` or
|
|
|
184 |
for k, v in values.items():
|
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 |
"""
|
|
|
191 |
"""
|
192 |
self.writer.close()
|
193 |
logger.debug("TensorBoard writer closed")
|
|
|
|
|
194 |
class WandBTracker(GeneralTracker):
|
195 |
"""
|
196 |
A `Tracker` class that supports `wandb`. Should be initialized at the start of your script.
|
|
|
197 |
Args:
|
198 |
run_name (`str`):
|
199 |
The name of the experiment run.
|
|
|
203 |
name = "wandb"
|
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
|
|
|
210 |
import wandb
|
|
|
211 |
self.run = wandb.init(project=self.run_name, **kwargs)
|
212 |
logger.debug(f"Initialized WandB project {self.run_name}")
|
213 |
logger.debug(
|
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.
|
|
|
223 |
Args:
|
224 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
225 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
226 |
`str`, `float`, `int`, or `None`.
|
227 |
"""
|
228 |
import wandb
|
|
|
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.
|
|
|
235 |
Args:
|
236 |
values (Dictionary `str` to `str`, `float`, `int` or `dict` of `str` to `float`/`int`):
|
237 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, `int` or `dict` of
|
|
|
243 |
"""
|
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.
|
|
|
250 |
Args:
|
251 |
values (Dictionary `str` to `List` of `np.ndarray` or `PIL.Image`):
|
252 |
Values to be logged as key-value pairs. The values need to have type `List` of `np.ndarray` or
|
|
|
256 |
Additional key word arguments passed along to the `wandb.log` method.
|
257 |
"""
|
258 |
import wandb
|
|
|
259 |
for k, v in values.items():
|
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,
|
|
|
272 |
"""
|
273 |
Log a Table containing any object type (text, image, audio, video, molecule, html, etc). Can be defined either
|
274 |
with `columns` and `data` or with `dataframe`.
|
|
|
275 |
Args:
|
276 |
table_name (`str`):
|
277 |
The name to give to the logged table on the wandb workspace
|
|
|
285 |
The run step. If included, the log will be affiliated with this step.
|
286 |
"""
|
287 |
import wandb
|
|
|
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 |
"""
|
|
|
294 |
"""
|
295 |
self.run.finish()
|
296 |
logger.debug("WandB run closed")
|
|
|
|
|
297 |
class CometMLTracker(GeneralTracker):
|
298 |
"""
|
299 |
A `Tracker` class that supports `comet_ml`. Should be initialized at the start of your script.
|
|
|
300 |
API keys must be stored in a Comet config file.
|
|
|
301 |
Args:
|
302 |
run_name (`str`):
|
303 |
The name of the experiment run.
|
|
|
306 |
"""
|
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
|
|
|
313 |
from comet_ml import Experiment
|
|
|
314 |
self.writer = Experiment(project_name=run_name, **kwargs)
|
315 |
logger.debug(f"Initialized CometML project {self.run_name}")
|
316 |
logger.debug(
|
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.
|
|
|
326 |
Args:
|
327 |
values (Dictionary `str` to `bool`, `str`, `float` or `int`):
|
328 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
|
|
330 |
"""
|
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.
|
|
|
337 |
Args:
|
338 |
values (Dictionary `str` to `str`, `float`, `int` or `dict` of `str` to `float`/`int`):
|
339 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, `int` or `dict` of
|
|
|
354 |
elif isinstance(v, dict):
|
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 |
"""
|
|
|
361 |
"""
|
362 |
self.writer.end()
|
363 |
logger.debug("CometML run closed")
|
|
|
|
|
364 |
class AimTracker(GeneralTracker):
|
365 |
"""
|
366 |
A `Tracker` class that supports `aim`. Should be initialized at the start of your script.
|
|
|
367 |
Args:
|
368 |
run_name (`str`):
|
369 |
The name of the experiment run.
|
|
|
372 |
"""
|
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
|
|
|
379 |
self.writer = Run(repo=logging_dir, **kwargs)
|
380 |
self.writer.name = self.run_name
|
381 |
logger.debug(f"Initialized Aim project {self.run_name}")
|
382 |
logger.debug(
|
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.
|
|
|
392 |
Args:
|
393 |
values (`dict`):
|
394 |
Values to be stored as initial hyperparameters as key-value pairs.
|
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.
|
|
|
401 |
Args:
|
402 |
values (`dict`):
|
403 |
Values to be logged as key-value pairs.
|
|
|
409 |
# Note: replace this with the dictionary support when merged
|
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.
|
|
|
416 |
Args:
|
417 |
values (`Dict[str, Union[np.ndarray, PIL.Image, Tuple[np.ndarray, str], Tuple[PIL.Image, str]]]`):
|
418 |
Values to be logged as key-value pairs. The values need to have type `np.ndarray` or PIL.Image. If a
|
|
|
424 |
keys `aim_image` and `track`, respectively.
|
425 |
"""
|
426 |
import aim
|
|
|
427 |
aim_image_kw = {}
|
428 |
track_kw = {}
|
|
|
429 |
if kwargs is not None:
|
430 |
aim_image_kw = kwargs.get("aim_image", {})
|
431 |
track_kw = kwargs.get("track", {})
|
|
|
432 |
for key, value in values.items():
|
433 |
if isinstance(value, tuple):
|
434 |
img, caption = value
|
|
|
436 |
img, caption = value, ""
|
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
|
443 |
"""
|
444 |
self.writer.close()
|
|
|
|
|
445 |
class MLflowTracker(GeneralTracker):
|
446 |
"""
|
447 |
A `Tracker` class that supports `mlflow`. Should be initialized at the start of your script.
|
|
|
448 |
Args:
|
449 |
experiment_name (`str`, *optional*):
|
450 |
Name of the experiment. Environment variable MLFLOW_EXPERIMENT_NAME has priority over this argument.
|
|
|
470 |
"""
|
471 |
name = "mlflow"
|
472 |
requires_logging_directory = False
|
|
|
473 |
@on_main_process
|
474 |
def __init__(
|
475 |
self,
|
|
|
486 |
tags = os.getenv("MLFLOW_TAGS", tags)
|
487 |
if isinstance(tags, str):
|
488 |
tags = json.loads(tags)
|
|
|
489 |
nested_run = os.getenv("MLFLOW_NESTED_RUN", nested_run)
|
|
|
490 |
import mlflow
|
|
|
491 |
exps = mlflow.search_experiments(filter_string=f"name = '{experiment_name}'")
|
492 |
if len(exps) > 0:
|
493 |
if len(exps) > 1:
|
|
|
499 |
artifact_location=logging_dir,
|
500 |
tags=tags,
|
501 |
)
|
|
|
502 |
self.active_run = mlflow.start_run(
|
503 |
run_id=run_id,
|
504 |
experiment_id=experiment_id,
|
|
|
507 |
tags=tags,
|
508 |
description=description,
|
509 |
)
|
|
|
510 |
logger.debug(f"Initialized mlflow experiment {experiment_name}")
|
511 |
logger.debug(
|
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.
|
|
|
521 |
Args:
|
522 |
values (`dict`):
|
523 |
Values to be stored as initial hyperparameters as key-value pairs.
|
524 |
"""
|
525 |
import mlflow
|
|
|
526 |
for name, value in list(values.items()):
|
527 |
# internally, all values are converted to str in MLflow
|
528 |
if len(str(value)) > mlflow.utils.validation.MAX_PARAM_VAL_LENGTH:
|
|
|
531 |
f" log_param() only accepts values no longer than {mlflow.utils.validation.MAX_PARAM_VAL_LENGTH} characters so we dropped this attribute."
|
532 |
)
|
533 |
del values[name]
|
|
|
534 |
values_list = list(values.items())
|
|
|
535 |
# MLflow cannot log more than 100 values in one go, so we have to split it
|
536 |
for i in range(0, len(values_list), mlflow.utils.validation.MAX_PARAMS_TAGS_PER_BATCH):
|
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.
|
|
|
543 |
Args:
|
544 |
values (`dict`):
|
545 |
Values to be logged as key-value pairs.
|
|
|
556 |
"MLflow's log_metric() only accepts float and int types so we dropped this attribute."
|
557 |
)
|
558 |
import mlflow
|
|
|
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.
|
565 |
"""
|
566 |
import mlflow
|
|
|
567 |
mlflow.end_run()
|
|
|
|
|
568 |
class ClearMLTracker(GeneralTracker):
|
569 |
"""
|
570 |
A `Tracker` class that supports `clearml`. Should be initialized at the start of your script.
|
|
|
571 |
Args:
|
572 |
run_name (`str`, *optional*):
|
573 |
Name of the experiment. Environment variables `CLEARML_PROJECT` and `CLEARML_TASK` have priority over this
|
|
|
577 |
"""
|
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()
|
584 |
self._initialized_externally = False
|
585 |
if current_task:
|
586 |
self._initialized_externally = True
|
587 |
self.task = current_task
|
588 |
return
|
|
|
589 |
kwargs.setdefault("project_name", os.environ.get("CLEARML_PROJECT", run_name))
|
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.
|
|
|
599 |
Args:
|
600 |
values (`dict`):
|
601 |
Values to be stored as initial hyperparameters as key-value pairs.
|
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
|
608 |
ints or floats
|
|
|
609 |
Args:
|
610 |
values (`Dict[str, Union[int, float]]`):
|
611 |
Values to be logged as key-value pairs. If the key starts with 'eval_'/'test_'/'train_', the value will
|
|
|
633 |
continue
|
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.
|
|
|
640 |
Args:
|
641 |
values (`Dict[str, List[Union[np.ndarray, PIL.Image]]`):
|
642 |
Values to be logged as key-value pairs. The values need to have type `List` of `np.ndarray` or
|
|
|
649 |
for k, v in values.items():
|
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,
|
|
|
661 |
):
|
662 |
"""
|
663 |
Log a Table to the task. Can be defined eitherwith `columns` and `data` or with `dataframe`.
|
|
|
664 |
Args:
|
665 |
table_name (`str`):
|
666 |
The name of the table
|
|
|
685 |
to_report = [columns] + data if columns else data
|
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 |
"""
|
|
|
693 |
"""
|
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 + "_"):
|
700 |
return name[len(prefix) + 1 :], prefix
|
701 |
return name, "train"
|
|
|
|
|
702 |
class DVCLiveTracker(GeneralTracker):
|
703 |
"""
|
704 |
A `Tracker` class that supports `dvclive`. Should be initialized at the start of your script.
|
|
|
705 |
Args:
|
706 |
run_name (`str`, *optional*):
|
707 |
Ignored for dvclive. See `kwargs` instead.
|
708 |
kwargs:
|
709 |
Additional key word arguments passed along to [`dvclive.Live()`](https://dvc.org/doc/dvclive/live).
|
|
|
710 |
Example:
|
|
|
711 |
```py
|
712 |
from accelerate import Accelerator
|
|
|
713 |
accelerator = Accelerator(log_with="dvclive")
|
714 |
accelerator.init_trackers(project_name="my_project", init_kwargs={"dvclive": {"dir": "my_directory"}})
|
715 |
```
|
716 |
"""
|
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
|
731 |
hyperparameters in a yaml file for future use.
|
|
|
732 |
Args:
|
733 |
values (Dictionary `str` to `bool`, `str`, `float`, `int`, or a List or Dict of those types):
|
734 |
Values to be stored as initial hyperparameters as key-value pairs. The values need to have type `bool`,
|
735 |
`str`, `float`, or `int`.
|
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.
|
|
|
742 |
Args:
|
743 |
values (Dictionary `str` to `str`, `float`, or `int`):
|
744 |
Values to be logged as key-value pairs. The values need to have type `str`, `float`, or `int`.
|
|
|
748 |
Additional key word arguments passed along to `dvclive.Live.log_metric()`.
|
749 |
"""
|
750 |
from dvclive.plots import Metric
|
|
|
751 |
if step is not None:
|
752 |
self.live.step = step
|
753 |
for k, v in values.items():
|
|
|
760 |
"This invocation of DVCLive's Live.log_metric() "
|
761 |
"is incorrect so we dropped this attribute."
|
762 |
)
|
|
|
763 |
@on_main_process
|
764 |
def finish(self):
|
765 |
"""
|
766 |
Closes `dvclive.Live()`.
|
767 |
"""
|
768 |
self.live.end()
|
|
|
|
|
769 |
LOGGER_TYPE_TO_CLASS = {
|
770 |
"aim": AimTracker,
|
771 |
"comet_ml": CometMLTracker,
|
|
|
775 |
"clearml": ClearMLTracker,
|
776 |
"dvclive": DVCLiveTracker,
|
777 |
}
|
|
|
|
|
778 |
def filter_trackers(
|
779 |
log_with: List[Union[str, LoggerType, GeneralTracker]],
|
780 |
logging_dir: Union[str, os.PathLike] = None,
|
|
|
785 |
- Filters out repeats of tracker types
|
786 |
- If `all` is in `log_with`, will return all trackers in the environment
|
787 |
- If a tracker requires a `logging_dir`, ensures that `logging_dir` is not `None`
|
|
|
788 |
Args:
|
789 |
log_with (list of `str`, [`~utils.LoggerType`] or [`~tracking.GeneralTracker`], *optional*):
|
790 |
A list of loggers to be setup for experiment tracking. Should be one or several of:
|
|
|
791 |
- `"all"`
|
792 |
- `"tensorboard"`
|
793 |
- `"wandb"`
|
|
|
824 |
loggers.append(log_type)
|
825 |
else:
|
826 |
logger.debug(f"Tried adding logger {log_type}, but package is unavailable in the system.")
|
|
|
827 |
return loggers
|
src/utils/bnb.py
CHANGED
@@ -1,6 +1,4 @@
|
|
1 |
logger = logging.getLogger(__name__)
|
2 |
-
|
3 |
-
|
4 |
def load_and_quantize_model(
|
5 |
model: torch.nn.Module,
|
6 |
bnb_quantization_config: BnbQuantizationConfig,
|
@@ -15,7 +13,6 @@ def load_and_quantize_model(
|
|
15 |
This function will quantize the input model with the associated config passed in `bnb_quantization_config`. If the
|
16 |
model is in the meta device, we will load and dispatch the weights according to the `device_map` passed. If the
|
17 |
model is already loaded, we will quantize the model and put the model on the GPU,
|
18 |
-
|
19 |
Args:
|
20 |
model (`torch.nn.Module`):
|
21 |
Input model. The model can be already loaded or on the meta device
|
@@ -40,13 +37,11 @@ def load_and_quantize_model(
|
|
40 |
offload_state_dict (`bool`, *optional*, defaults to `False`):
|
41 |
If `True`, will temporarily offload the CPU state dict on the hard drive to avoid getting out of CPU RAM if
|
42 |
the weight of the CPU state dict + the biggest shard does not fit.
|
43 |
-
|
44 |
Returns:
|
45 |
`torch.nn.Module`: The quantized model
|
46 |
"""
|
47 |
load_in_4bit = bnb_quantization_config.load_in_4bit
|
48 |
load_in_8bit = bnb_quantization_config.load_in_8bit
|
49 |
-
|
50 |
if load_in_8bit and not is_8bit_bnb_available():
|
51 |
raise ImportError(
|
52 |
"You have a version of `bitsandbytes` that is not compatible with 8bit quantization,"
|
@@ -57,31 +52,25 @@ def load_and_quantize_model(
|
|
57 |
"You have a version of `bitsandbytes` that is not compatible with 4bit quantization,"
|
58 |
"make sure you have the latest version of `bitsandbytes` installed."
|
59 |
)
|
60 |
-
|
61 |
modules_on_cpu = []
|
62 |
# custom device map
|
63 |
if isinstance(device_map, dict) and len(device_map.keys()) > 1:
|
64 |
modules_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]]
|
65 |
-
|
66 |
# We keep some modules such as the lm_head in their original dtype for numerical stability reasons
|
67 |
if bnb_quantization_config.skip_modules is None:
|
68 |
bnb_quantization_config.skip_modules = get_keys_to_not_convert(model)
|
69 |
-
|
70 |
# add cpu modules to skip modules only for 4-bit modules
|
71 |
if load_in_4bit:
|
72 |
bnb_quantization_config.skip_modules.extend(modules_on_cpu)
|
73 |
modules_to_not_convert = bnb_quantization_config.skip_modules
|
74 |
-
|
75 |
# We add the modules we want to keep in full precision
|
76 |
if bnb_quantization_config.keep_in_fp32_modules is None:
|
77 |
bnb_quantization_config.keep_in_fp32_modules = []
|
78 |
keep_in_fp32_modules = bnb_quantization_config.keep_in_fp32_modules
|
79 |
modules_to_not_convert.extend(keep_in_fp32_modules)
|
80 |
-
|
81 |
# compatibility with peft
|
82 |
model.is_loaded_in_4bit = load_in_4bit
|
83 |
model.is_loaded_in_8bit = load_in_8bit
|
84 |
-
|
85 |
model_device = get_parameter_device(model)
|
86 |
if model_device.type != "meta":
|
87 |
# quantization of an already loaded model
|
@@ -115,18 +104,15 @@ def load_and_quantize_model(
|
|
115 |
"We move the model to cuda."
|
116 |
)
|
117 |
return model
|
118 |
-
|
119 |
elif weights_location is None:
|
120 |
raise RuntimeError(
|
121 |
f"`weights_location` needs to be the folder path containing the weights of the model, but we found {weights_location} "
|
122 |
)
|
123 |
-
|
124 |
else:
|
125 |
with init_empty_weights():
|
126 |
model = replace_with_bnb_layers(
|
127 |
model, bnb_quantization_config, modules_to_not_convert=modules_to_not_convert
|
128 |
)
|
129 |
-
|
130 |
device_map = get_quantized_model_device_map(
|
131 |
model,
|
132 |
bnb_quantization_config,
|
@@ -136,9 +122,7 @@ def load_and_quantize_model(
|
|
136 |
)
|
137 |
if offload_state_dict is None and device_map is not None and "disk" in device_map.values():
|
138 |
offload_state_dict = True
|
139 |
-
|
140 |
offload = any(x in list(device_map.values()) for x in ["cpu", "disk"])
|
141 |
-
|
142 |
load_checkpoint_in_model(
|
143 |
model,
|
144 |
weights_location,
|
@@ -150,8 +134,6 @@ def load_and_quantize_model(
|
|
150 |
offload_8bit_bnb=load_in_8bit and offload,
|
151 |
)
|
152 |
return dispatch_model(model, device_map=device_map, offload_dir=offload_folder)
|
153 |
-
|
154 |
-
|
155 |
def get_quantized_model_device_map(
|
156 |
model, bnb_quantization_config, device_map=None, max_memory=None, no_split_module_classes=None
|
157 |
):
|
@@ -161,14 +143,12 @@ def get_quantized_model_device_map(
|
|
161 |
else:
|
162 |
raise RuntimeError("No GPU found. A GPU is needed for quantization.")
|
163 |
logger.info("The device_map was not initialized." "Setting device_map to `{'':torch.cuda.current_device()}`.")
|
164 |
-
|
165 |
if isinstance(device_map, str):
|
166 |
if device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
|
167 |
raise ValueError(
|
168 |
"If passing a string for `device_map`, please choose 'auto', 'balanced', 'balanced_low_0' or "
|
169 |
"'sequential'."
|
170 |
)
|
171 |
-
|
172 |
special_dtypes = {}
|
173 |
special_dtypes.update(
|
174 |
{
|
@@ -184,12 +164,10 @@ def get_quantized_model_device_map(
|
|
184 |
if any(m in name for m in bnb_quantization_config.keep_in_fp32_modules)
|
185 |
}
|
186 |
)
|
187 |
-
|
188 |
kwargs = {}
|
189 |
kwargs["special_dtypes"] = special_dtypes
|
190 |
kwargs["no_split_module_classes"] = no_split_module_classes
|
191 |
kwargs["dtype"] = bnb_quantization_config.target_dtype
|
192 |
-
|
193 |
# get max_memory for each device.
|
194 |
if device_map != "sequential":
|
195 |
max_memory = get_balanced_memory(
|
@@ -198,14 +176,11 @@ def get_quantized_model_device_map(
|
|
198 |
max_memory=max_memory,
|
199 |
**kwargs,
|
200 |
)
|
201 |
-
|
202 |
kwargs["max_memory"] = max_memory
|
203 |
device_map = infer_auto_device_map(model, **kwargs)
|
204 |
-
|
205 |
if isinstance(device_map, dict):
|
206 |
# check if don't have any quantized module on the cpu
|
207 |
modules_not_to_convert = bnb_quantization_config.skip_modules + bnb_quantization_config.keep_in_fp32_modules
|
208 |
-
|
209 |
device_map_without_some_modules = {
|
210 |
key: device_map[key] for key in device_map.keys() if key not in modules_not_to_convert
|
211 |
}
|
@@ -228,13 +203,10 @@ def get_quantized_model_device_map(
|
|
228 |
)
|
229 |
del device_map_without_some_modules
|
230 |
return device_map
|
231 |
-
|
232 |
-
|
233 |
def replace_with_bnb_layers(model, bnb_quantization_config, modules_to_not_convert=None, current_key_name=None):
|
234 |
"""
|
235 |
A helper function to replace all `torch.nn.Linear` modules by `bnb.nn.Linear8bit` modules or by `bnb.nn.Linear4bit`
|
236 |
modules from the `bitsandbytes`library. The function will be run recursively and replace `torch.nn.Linear` modules.
|
237 |
-
|
238 |
Parameters:
|
239 |
model (`torch.nn.Module`):
|
240 |
Input model or `torch.nn.Module` as the function is run recursively.
|
@@ -247,7 +219,6 @@ def replace_with_bnb_layers(model, bnb_quantization_config, modules_to_not_conve
|
|
247 |
"""
|
248 |
if modules_to_not_convert is None:
|
249 |
modules_to_not_convert = []
|
250 |
-
|
251 |
model, has_been_replaced = _replace_with_bnb_layers(
|
252 |
model, bnb_quantization_config, modules_to_not_convert, current_key_name
|
253 |
)
|
@@ -259,8 +230,6 @@ def replace_with_bnb_layers(model, bnb_quantization_config, modules_to_not_conve
|
|
259 |
" a bug."
|
260 |
)
|
261 |
return model
|
262 |
-
|
263 |
-
|
264 |
def _replace_with_bnb_layers(
|
265 |
model,
|
266 |
bnb_quantization_config,
|
@@ -269,12 +238,10 @@ def _replace_with_bnb_layers(
|
|
269 |
):
|
270 |
"""
|
271 |
Private method that wraps the recursion for module replacement.
|
272 |
-
|
273 |
Returns the converted model and a boolean that indicates if the conversion has been successfull or not.
|
274 |
"""
|
275 |
# bitsandbytes will initialize CUDA on import, so it needs to be imported lazily
|
276 |
import bitsandbytes as bnb
|
277 |
-
|
278 |
has_been_replaced = False
|
279 |
for name, module in model.named_children():
|
280 |
if current_key_name is None:
|
@@ -325,15 +292,12 @@ def _replace_with_bnb_layers(
|
|
325 |
# Remove the last key for recursion
|
326 |
current_key_name.pop(-1)
|
327 |
return model, has_been_replaced
|
328 |
-
|
329 |
-
|
330 |
def get_keys_to_not_convert(model):
|
331 |
r"""
|
332 |
An utility function to get the key of the module to keep in full precision if any For example for CausalLM modules
|
333 |
we may want to keep the lm_head in full precision for numerical stability reasons. For other architectures, we want
|
334 |
to keep the tied weights of the model. The function will return a list of the keys of the modules to not convert in
|
335 |
int8.
|
336 |
-
|
337 |
Parameters:
|
338 |
model (`torch.nn.Module`):
|
339 |
Input model
|
@@ -341,7 +305,6 @@ def get_keys_to_not_convert(model):
|
|
341 |
# Create a copy of the model
|
342 |
with init_empty_weights():
|
343 |
tied_model = deepcopy(model) # this has 0 cost since it is done inside `init_empty_weights` context manager`
|
344 |
-
|
345 |
tied_params = find_tied_parameters(tied_model)
|
346 |
# For compatibility with Accelerate < 0.18
|
347 |
if isinstance(tied_params, dict):
|
@@ -349,24 +312,19 @@ def get_keys_to_not_convert(model):
|
|
349 |
else:
|
350 |
tied_keys = sum(tied_params, [])
|
351 |
has_tied_params = len(tied_keys) > 0
|
352 |
-
|
353 |
# Check if it is a base model
|
354 |
is_base_model = False
|
355 |
if hasattr(model, "base_model_prefix"):
|
356 |
is_base_model = not hasattr(model, model.base_model_prefix)
|
357 |
-
|
358 |
# Ignore this for base models (BertModel, GPT2Model, etc.)
|
359 |
if (not has_tied_params) and is_base_model:
|
360 |
return []
|
361 |
-
|
362 |
# otherwise they have an attached head
|
363 |
list_modules = list(model.named_children())
|
364 |
list_last_module = [list_modules[-1][0]]
|
365 |
-
|
366 |
# add last module together with tied weights
|
367 |
intersection = set(list_last_module) - set(tied_keys)
|
368 |
list_untouched = list(set(tied_keys)) + list(intersection)
|
369 |
-
|
370 |
# remove ".weight" from the keys
|
371 |
names_to_remove = [".weight", ".bias"]
|
372 |
filtered_module_names = []
|
@@ -375,25 +333,17 @@ def get_keys_to_not_convert(model):
|
|
375 |
if name_to_remove in name:
|
376 |
name = name.replace(name_to_remove, "")
|
377 |
filtered_module_names.append(name)
|
378 |
-
|
379 |
return filtered_module_names
|
380 |
-
|
381 |
-
|
382 |
def has_4bit_bnb_layers(model):
|
383 |
"""Check if we have `bnb.nn.Linear4bit` or `bnb.nn.Linear8bitLt` layers inside our model"""
|
384 |
# bitsandbytes will initialize CUDA on import, so it needs to be imported lazily
|
385 |
import bitsandbytes as bnb
|
386 |
-
|
387 |
for m in model.modules():
|
388 |
if isinstance(m, bnb.nn.Linear4bit):
|
389 |
return True
|
390 |
return False
|
391 |
-
|
392 |
-
|
393 |
def get_parameter_device(parameter: nn.Module):
|
394 |
return next(parameter.parameters()).device
|
395 |
-
|
396 |
-
|
397 |
def quantize_and_offload_8bit(model, param, param_name, new_dtype, offload_folder, offload_index, fp16_statistics):
|
398 |
# if it is not quantized, we quantize and offload the quantized weights and the SCB stats
|
399 |
if fp16_statistics is None:
|
@@ -421,5 +371,4 @@ def quantize_and_offload_8bit(model, param, param_name, new_dtype, offload_folde
|
|
421 |
else:
|
422 |
offload_weight(param, param_name, offload_folder, index=offload_index)
|
423 |
offload_weight(fp16_statistics, param_name.replace("weight", "SCB"), offload_folder, index=offload_index)
|
424 |
-
|
425 |
set_module_tensor_to_device(model, param_name, "meta", dtype=new_dtype, value=torch.empty(*param.size()))
|
|
|
1 |
logger = logging.getLogger(__name__)
|
|
|
|
|
2 |
def load_and_quantize_model(
|
3 |
model: torch.nn.Module,
|
4 |
bnb_quantization_config: BnbQuantizationConfig,
|
|
|
13 |
This function will quantize the input model with the associated config passed in `bnb_quantization_config`. If the
|
14 |
model is in the meta device, we will load and dispatch the weights according to the `device_map` passed. If the
|
15 |
model is already loaded, we will quantize the model and put the model on the GPU,
|
|
|
16 |
Args:
|
17 |
model (`torch.nn.Module`):
|
18 |
Input model. The model can be already loaded or on the meta device
|
|
|
37 |
offload_state_dict (`bool`, *optional*, defaults to `False`):
|
38 |
If `True`, will temporarily offload the CPU state dict on the hard drive to avoid getting out of CPU RAM if
|
39 |
the weight of the CPU state dict + the biggest shard does not fit.
|
|
|
40 |
Returns:
|
41 |
`torch.nn.Module`: The quantized model
|
42 |
"""
|
43 |
load_in_4bit = bnb_quantization_config.load_in_4bit
|
44 |
load_in_8bit = bnb_quantization_config.load_in_8bit
|
|
|
45 |
if load_in_8bit and not is_8bit_bnb_available():
|
46 |
raise ImportError(
|
47 |
"You have a version of `bitsandbytes` that is not compatible with 8bit quantization,"
|
|
|
52 |
"You have a version of `bitsandbytes` that is not compatible with 4bit quantization,"
|
53 |
"make sure you have the latest version of `bitsandbytes` installed."
|
54 |
)
|
|
|
55 |
modules_on_cpu = []
|
56 |
# custom device map
|
57 |
if isinstance(device_map, dict) and len(device_map.keys()) > 1:
|
58 |
modules_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]]
|
|
|
59 |
# We keep some modules such as the lm_head in their original dtype for numerical stability reasons
|
60 |
if bnb_quantization_config.skip_modules is None:
|
61 |
bnb_quantization_config.skip_modules = get_keys_to_not_convert(model)
|
|
|
62 |
# add cpu modules to skip modules only for 4-bit modules
|
63 |
if load_in_4bit:
|
64 |
bnb_quantization_config.skip_modules.extend(modules_on_cpu)
|
65 |
modules_to_not_convert = bnb_quantization_config.skip_modules
|
|
|
66 |
# We add the modules we want to keep in full precision
|
67 |
if bnb_quantization_config.keep_in_fp32_modules is None:
|
68 |
bnb_quantization_config.keep_in_fp32_modules = []
|
69 |
keep_in_fp32_modules = bnb_quantization_config.keep_in_fp32_modules
|
70 |
modules_to_not_convert.extend(keep_in_fp32_modules)
|
|
|
71 |
# compatibility with peft
|
72 |
model.is_loaded_in_4bit = load_in_4bit
|
73 |
model.is_loaded_in_8bit = load_in_8bit
|
|
|
74 |
model_device = get_parameter_device(model)
|
75 |
if model_device.type != "meta":
|
76 |
# quantization of an already loaded model
|
|
|
104 |
"We move the model to cuda."
|
105 |
)
|
106 |
return model
|
|
|
107 |
elif weights_location is None:
|
108 |
raise RuntimeError(
|
109 |
f"`weights_location` needs to be the folder path containing the weights of the model, but we found {weights_location} "
|
110 |
)
|
|
|
111 |
else:
|
112 |
with init_empty_weights():
|
113 |
model = replace_with_bnb_layers(
|
114 |
model, bnb_quantization_config, modules_to_not_convert=modules_to_not_convert
|
115 |
)
|
|
|
116 |
device_map = get_quantized_model_device_map(
|
117 |
model,
|
118 |
bnb_quantization_config,
|
|
|
122 |
)
|
123 |
if offload_state_dict is None and device_map is not None and "disk" in device_map.values():
|
124 |
offload_state_dict = True
|
|
|
125 |
offload = any(x in list(device_map.values()) for x in ["cpu", "disk"])
|
|
|
126 |
load_checkpoint_in_model(
|
127 |
model,
|
128 |
weights_location,
|
|
|
134 |
offload_8bit_bnb=load_in_8bit and offload,
|
135 |
)
|
136 |
return dispatch_model(model, device_map=device_map, offload_dir=offload_folder)
|
|
|
|
|
137 |
def get_quantized_model_device_map(
|
138 |
model, bnb_quantization_config, device_map=None, max_memory=None, no_split_module_classes=None
|
139 |
):
|
|
|
143 |
else:
|
144 |
raise RuntimeError("No GPU found. A GPU is needed for quantization.")
|
145 |
logger.info("The device_map was not initialized." "Setting device_map to `{'':torch.cuda.current_device()}`.")
|
|
|
146 |
if isinstance(device_map, str):
|
147 |
if device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
|
148 |
raise ValueError(
|
149 |
"If passing a string for `device_map`, please choose 'auto', 'balanced', 'balanced_low_0' or "
|
150 |
"'sequential'."
|
151 |
)
|
|
|
152 |
special_dtypes = {}
|
153 |
special_dtypes.update(
|
154 |
{
|
|
|
164 |
if any(m in name for m in bnb_quantization_config.keep_in_fp32_modules)
|
165 |
}
|
166 |
)
|
|
|
167 |
kwargs = {}
|
168 |
kwargs["special_dtypes"] = special_dtypes
|
169 |
kwargs["no_split_module_classes"] = no_split_module_classes
|
170 |
kwargs["dtype"] = bnb_quantization_config.target_dtype
|
|
|
171 |
# get max_memory for each device.
|
172 |
if device_map != "sequential":
|
173 |
max_memory = get_balanced_memory(
|
|
|
176 |
max_memory=max_memory,
|
177 |
**kwargs,
|
178 |
)
|
|
|
179 |
kwargs["max_memory"] = max_memory
|
180 |
device_map = infer_auto_device_map(model, **kwargs)
|
|
|
181 |
if isinstance(device_map, dict):
|
182 |
# check if don't have any quantized module on the cpu
|
183 |
modules_not_to_convert = bnb_quantization_config.skip_modules + bnb_quantization_config.keep_in_fp32_modules
|
|
|
184 |
device_map_without_some_modules = {
|
185 |
key: device_map[key] for key in device_map.keys() if key not in modules_not_to_convert
|
186 |
}
|
|
|
203 |
)
|
204 |
del device_map_without_some_modules
|
205 |
return device_map
|
|
|
|
|
206 |
def replace_with_bnb_layers(model, bnb_quantization_config, modules_to_not_convert=None, current_key_name=None):
|
207 |
"""
|
208 |
A helper function to replace all `torch.nn.Linear` modules by `bnb.nn.Linear8bit` modules or by `bnb.nn.Linear4bit`
|
209 |
modules from the `bitsandbytes`library. The function will be run recursively and replace `torch.nn.Linear` modules.
|
|
|
210 |
Parameters:
|
211 |
model (`torch.nn.Module`):
|
212 |
Input model or `torch.nn.Module` as the function is run recursively.
|
|
|
219 |
"""
|
220 |
if modules_to_not_convert is None:
|
221 |
modules_to_not_convert = []
|
|
|
222 |
model, has_been_replaced = _replace_with_bnb_layers(
|
223 |
model, bnb_quantization_config, modules_to_not_convert, current_key_name
|
224 |
)
|
|
|
230 |
" a bug."
|
231 |
)
|
232 |
return model
|
|
|
|
|
233 |
def _replace_with_bnb_layers(
|
234 |
model,
|
235 |
bnb_quantization_config,
|
|
|
238 |
):
|
239 |
"""
|
240 |
Private method that wraps the recursion for module replacement.
|
|
|
241 |
Returns the converted model and a boolean that indicates if the conversion has been successfull or not.
|
242 |
"""
|
243 |
# bitsandbytes will initialize CUDA on import, so it needs to be imported lazily
|
244 |
import bitsandbytes as bnb
|
|
|
245 |
has_been_replaced = False
|
246 |
for name, module in model.named_children():
|
247 |
if current_key_name is None:
|
|
|
292 |
# Remove the last key for recursion
|
293 |
current_key_name.pop(-1)
|
294 |
return model, has_been_replaced
|
|
|
|
|
295 |
def get_keys_to_not_convert(model):
|
296 |
r"""
|
297 |
An utility function to get the key of the module to keep in full precision if any For example for CausalLM modules
|
298 |
we may want to keep the lm_head in full precision for numerical stability reasons. For other architectures, we want
|
299 |
to keep the tied weights of the model. The function will return a list of the keys of the modules to not convert in
|
300 |
int8.
|
|
|
301 |
Parameters:
|
302 |
model (`torch.nn.Module`):
|
303 |
Input model
|
|
|
305 |
# Create a copy of the model
|
306 |
with init_empty_weights():
|
307 |
tied_model = deepcopy(model) # this has 0 cost since it is done inside `init_empty_weights` context manager`
|
|
|
308 |
tied_params = find_tied_parameters(tied_model)
|
309 |
# For compatibility with Accelerate < 0.18
|
310 |
if isinstance(tied_params, dict):
|
|
|
312 |
else:
|
313 |
tied_keys = sum(tied_params, [])
|
314 |
has_tied_params = len(tied_keys) > 0
|
|
|
315 |
# Check if it is a base model
|
316 |
is_base_model = False
|
317 |
if hasattr(model, "base_model_prefix"):
|
318 |
is_base_model = not hasattr(model, model.base_model_prefix)
|
|
|
319 |
# Ignore this for base models (BertModel, GPT2Model, etc.)
|
320 |
if (not has_tied_params) and is_base_model:
|
321 |
return []
|
|
|
322 |
# otherwise they have an attached head
|
323 |
list_modules = list(model.named_children())
|
324 |
list_last_module = [list_modules[-1][0]]
|
|
|
325 |
# add last module together with tied weights
|
326 |
intersection = set(list_last_module) - set(tied_keys)
|
327 |
list_untouched = list(set(tied_keys)) + list(intersection)
|
|
|
328 |
# remove ".weight" from the keys
|
329 |
names_to_remove = [".weight", ".bias"]
|
330 |
filtered_module_names = []
|
|
|
333 |
if name_to_remove in name:
|
334 |
name = name.replace(name_to_remove, "")
|
335 |
filtered_module_names.append(name)
|
|
|
336 |
return filtered_module_names
|
|
|
|
|
337 |
def has_4bit_bnb_layers(model):
|
338 |
"""Check if we have `bnb.nn.Linear4bit` or `bnb.nn.Linear8bitLt` layers inside our model"""
|
339 |
# bitsandbytes will initialize CUDA on import, so it needs to be imported lazily
|
340 |
import bitsandbytes as bnb
|
|
|
341 |
for m in model.modules():
|
342 |
if isinstance(m, bnb.nn.Linear4bit):
|
343 |
return True
|
344 |
return False
|
|
|
|
|
345 |
def get_parameter_device(parameter: nn.Module):
|
346 |
return next(parameter.parameters()).device
|
|
|
|
|
347 |
def quantize_and_offload_8bit(model, param, param_name, new_dtype, offload_folder, offload_index, fp16_statistics):
|
348 |
# if it is not quantized, we quantize and offload the quantized weights and the SCB stats
|
349 |
if fp16_statistics is None:
|
|
|
371 |
else:
|
372 |
offload_weight(param, param_name, offload_folder, index=offload_index)
|
373 |
offload_weight(fp16_statistics, param_name.replace("weight", "SCB"), offload_folder, index=offload_index)
|
|
|
374 |
set_module_tensor_to_device(model, param_name, "meta", dtype=new_dtype, value=torch.empty(*param.size()))
|
src/utils/constants.py
CHANGED
@@ -21,9 +21,7 @@ FSDP_PYTORCH_VERSION = "2.1.0"
|
|
21 |
FSDP_MODEL_NAME = "pytorch_model_fsdp"
|
22 |
DEEPSPEED_MULTINODE_LAUNCHERS = ["pdsh", "standard", "openmpi", "mvapich", "mpich"]
|
23 |
TORCH_DYNAMO_MODES = ["default", "reduce-overhead", "max-autotune"]
|
24 |
-
|
25 |
STR_OPERATION_TO_FUNC = {">": op.gt, ">=": op.ge, "==": op.eq, "!=": op.ne, "<=": op.le, "<": op.lt}
|
26 |
-
|
27 |
# These are the args for `torch.distributed.launch` for pytorch < 1.9
|
28 |
TORCH_LAUNCH_PARAMS = [
|
29 |
"nnodes",
|
@@ -50,6 +48,5 @@ TORCH_LAUNCH_PARAMS = [
|
|
50 |
"master_addr",
|
51 |
"master_port",
|
52 |
]
|
53 |
-
|
54 |
CUDA_DISTRIBUTED_TYPES = ["DEEPSPEED", "MULTI_GPU", "FSDP", "MEGATRON_LM"]
|
55 |
TORCH_DISTRIBUTED_OPERATION_TYPES = CUDA_DISTRIBUTED_TYPES + ["MULTI_NPU", "MULTI_XPU", "MULTI_CPU"]
|
|
|
21 |
FSDP_MODEL_NAME = "pytorch_model_fsdp"
|
22 |
DEEPSPEED_MULTINODE_LAUNCHERS = ["pdsh", "standard", "openmpi", "mvapich", "mpich"]
|
23 |
TORCH_DYNAMO_MODES = ["default", "reduce-overhead", "max-autotune"]
|
|
|
24 |
STR_OPERATION_TO_FUNC = {">": op.gt, ">=": op.ge, "==": op.eq, "!=": op.ne, "<=": op.le, "<": op.lt}
|
|
|
25 |
# These are the args for `torch.distributed.launch` for pytorch < 1.9
|
26 |
TORCH_LAUNCH_PARAMS = [
|
27 |
"nnodes",
|
|
|
48 |
"master_addr",
|
49 |
"master_port",
|
50 |
]
|
|
|
51 |
CUDA_DISTRIBUTED_TYPES = ["DEEPSPEED", "MULTI_GPU", "FSDP", "MEGATRON_LM"]
|
52 |
TORCH_DISTRIBUTED_OPERATION_TYPES = CUDA_DISTRIBUTED_TYPES + ["MULTI_NPU", "MULTI_XPU", "MULTI_CPU"]
|
src/utils/dataclasses.py
CHANGED
@@ -7,41 +7,32 @@ class KwargsHandler:
|
|
7 |
"""
|
8 |
def to_dict(self):
|
9 |
return copy.deepcopy(self.__dict__)
|
10 |
-
|
11 |
def to_kwargs(self):
|
12 |
"""
|
13 |
Returns a dictionary containing the attributes with values different from the default of this class.
|
14 |
"""
|
15 |
# import clear_environment here to avoid circular import problem
|
16 |
from .other import clear_environment
|
17 |
-
|
18 |
with clear_environment():
|
19 |
default_dict = self.__class__().to_dict()
|
20 |
this_dict = self.to_dict()
|
21 |
return {k: v for k, v in this_dict.items() if default_dict[k] != v}
|
22 |
-
|
23 |
-
|
24 |
@dataclass
|
25 |
class AutocastKwargs(KwargsHandler):
|
26 |
"""
|
27 |
Use this object in your [`Accelerator`] to customize how `torch.autocast` behaves. Please refer to the
|
28 |
documentation of this [context manager](https://pytorch.org/docs/stable/amp.html#torch.autocast) for more
|
29 |
information on each argument.
|
30 |
-
|
31 |
Example:
|
32 |
-
|
33 |
```python
|
34 |
from accelerate import Accelerator
|
35 |
from accelerate.utils import AutocastKwargs
|
36 |
-
|
37 |
kwargs = AutocastKwargs(cache_enabled=True)
|
38 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
39 |
```
|
40 |
"""
|
41 |
enabled: bool = True
|
42 |
cache_enabled: bool = None
|
43 |
-
|
44 |
-
|
45 |
@dataclass
|
46 |
class DistributedDataParallelKwargs(KwargsHandler):
|
47 |
"""
|
@@ -49,21 +40,14 @@ class DistributedDataParallelKwargs(KwargsHandler):
|
|
49 |
`torch.nn.parallel.DistributedDataParallel`. Please refer to the documentation of this
|
50 |
[wrapper](https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html) for more
|
51 |
information on each argument.
|
52 |
-
|
53 |
<Tip warning={true}>
|
54 |
-
|
55 |
`gradient_as_bucket_view` is only available in PyTorch 1.7.0 and later versions.
|
56 |
-
|
57 |
`static_graph` is only available in PyTorch 1.11.0 and later versions.
|
58 |
-
|
59 |
</Tip>
|
60 |
-
|
61 |
Example:
|
62 |
-
|
63 |
```python
|
64 |
from accelerate import Accelerator
|
65 |
from accelerate.utils import DistributedDataParallelKwargs
|
66 |
-
|
67 |
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
|
68 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
69 |
```
|
@@ -75,27 +59,19 @@ class DistributedDataParallelKwargs(KwargsHandler):
|
|
75 |
check_reduction: bool = False
|
76 |
gradient_as_bucket_view: bool = False
|
77 |
static_graph: bool = False
|
78 |
-
|
79 |
-
|
80 |
@dataclass
|
81 |
class GradScalerKwargs(KwargsHandler):
|
82 |
"""
|
83 |
Use this object in your [`Accelerator`] to customize the behavior of mixed precision, specifically how the
|
84 |
`torch.cuda.amp.GradScaler` used is created. Please refer to the documentation of this
|
85 |
[scaler](https://pytorch.org/docs/stable/amp.html?highlight=gradscaler) for more information on each argument.
|
86 |
-
|
87 |
<Tip warning={true}>
|
88 |
-
|
89 |
`GradScaler` is only available in PyTorch 1.5.0 and later versions.
|
90 |
-
|
91 |
</Tip>
|
92 |
-
|
93 |
Example:
|
94 |
-
|
95 |
```python
|
96 |
from accelerate import Accelerator
|
97 |
from accelerate.utils import GradScalerKwargs
|
98 |
-
|
99 |
kwargs = GradScalerKwargs(backoff_filter=0.25)
|
100 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
101 |
```
|
@@ -105,8 +81,6 @@ class GradScalerKwargs(KwargsHandler):
|
|
105 |
backoff_factor: float = 0.5
|
106 |
growth_interval: int = 2000
|
107 |
enabled: bool = True
|
108 |
-
|
109 |
-
|
110 |
@dataclass
|
111 |
class InitProcessGroupKwargs(KwargsHandler):
|
112 |
"""
|
@@ -114,12 +88,10 @@ class InitProcessGroupKwargs(KwargsHandler):
|
|
114 |
to the documentation of this
|
115 |
[method](https://pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group) for more
|
116 |
information on each argument.
|
117 |
-
|
118 |
```python
|
119 |
from datetime import timedelta
|
120 |
from accelerate import Accelerator
|
121 |
from accelerate.utils import InitProcessGroupKwargs
|
122 |
-
|
123 |
kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=800))
|
124 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
125 |
```
|
@@ -127,45 +99,32 @@ class InitProcessGroupKwargs(KwargsHandler):
|
|
127 |
backend: Optional[str] = "nccl"
|
128 |
init_method: Optional[str] = None
|
129 |
timeout: timedelta = timedelta(seconds=1800)
|
130 |
-
|
131 |
-
|
132 |
# Literals
|
133 |
Backend = Literal["msamp", "te"]
|
134 |
OptLevel = Literal["O1", "O2"]
|
135 |
FP8Format = Literal["E4M3", "HYBRID"]
|
136 |
AmaxComputeAlgorithm = Literal["max", "most_recent"]
|
137 |
-
|
138 |
-
|
139 |
@dataclass
|
140 |
class FP8RecipeKwargs(KwargsHandler):
|
141 |
"""
|
142 |
Use this object in your [`Accelerator`] to customize the initialization of the recipe for FP8 mixed precision
|
143 |
training with `transformer-engine` or `ms-amp`.
|
144 |
-
|
145 |
<Tip>
|
146 |
-
|
147 |
For more information on `transformer-engine` args, please refer to the API
|
148 |
[documentation](https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/api/common.html).
|
149 |
-
|
150 |
For more information on the `ms-amp` args, please refer to the Optimization Level
|
151 |
[documentation](https://azure.github.io/MS-AMP/docs/user-tutorial/optimization-level).
|
152 |
-
|
153 |
</Tip>
|
154 |
-
|
155 |
```python
|
156 |
from accelerate import Accelerator
|
157 |
from accelerate.utils import FP8RecipeKwargs
|
158 |
-
|
159 |
kwargs = FP8RecipeKwargs(backend="te", fp8_format="HYBRID")
|
160 |
accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=[kwargs])
|
161 |
```
|
162 |
-
|
163 |
To use MS-AMP as an engine, pass `backend="msamp"` and the `optimization_level`:
|
164 |
-
|
165 |
```python
|
166 |
kwargs = FP8RecipeKwargs(backend="msamp", optimization_level="02")
|
167 |
```
|
168 |
-
|
169 |
Args:
|
170 |
backend (`str`, *optional*, defaults to "msamp"):
|
171 |
Which FP8 engine to use. Must be one of `"msamp"` (MS-AMP) or `"te"` (TransformerEngine).
|
@@ -200,7 +159,6 @@ class FP8RecipeKwargs(KwargsHandler):
|
|
200 |
amax_history_len: int = 1
|
201 |
amax_compute_algo: AmaxComputeAlgorithm = "most_recent"
|
202 |
override_linear_precision: Tuple[bool, bool, bool] = (False, False, False)
|
203 |
-
|
204 |
def __post_init__(self):
|
205 |
self.backend = self.backend.upper()
|
206 |
if self.backend not in get_args(Backend):
|
@@ -215,37 +173,26 @@ class FP8RecipeKwargs(KwargsHandler):
|
|
215 |
elif self.backend == "MSAMP":
|
216 |
if self.opt_level not in get_args(OptLevel):
|
217 |
raise ValueError(f"`optimization_level` must be one of {' or '.join(get_args(OptLevel))}")
|
218 |
-
|
219 |
-
|
220 |
class EnumWithContains(enum.EnumMeta):
|
221 |
"A metaclass that adds the ability to check if `self` contains an item with the `in` operator"
|
222 |
-
|
223 |
def __contains__(cls, item):
|
224 |
try:
|
225 |
cls(item)
|
226 |
except ValueError:
|
227 |
return False
|
228 |
return True
|
229 |
-
|
230 |
-
|
231 |
class BaseEnum(enum.Enum, metaclass=EnumWithContains):
|
232 |
"An enum class that can get the value of an item with `str(Enum.key)`"
|
233 |
-
|
234 |
def __str__(self):
|
235 |
return self.value
|
236 |
-
|
237 |
@classmethod
|
238 |
def list(cls):
|
239 |
"Method to list all the possible items in `cls`"
|
240 |
return list(map(str, cls))
|
241 |
-
|
242 |
-
|
243 |
class DistributedType(str, enum.Enum):
|
244 |
"""
|
245 |
Represents a type of distributed environment.
|
246 |
-
|
247 |
Values:
|
248 |
-
|
249 |
- **NO** -- Not a distributed environment, just a single process.
|
250 |
- **MULTI_CPU** -- Distributed on multiple CPU nodes.
|
251 |
- **MULTI_GPU** -- Distributed on multiple GPUs.
|
@@ -264,14 +211,10 @@ class DistributedType(str, enum.Enum):
|
|
264 |
FSDP = "FSDP"
|
265 |
TPU = "TPU"
|
266 |
MEGATRON_LM = "MEGATRON_LM"
|
267 |
-
|
268 |
-
|
269 |
class SageMakerDistributedType(str, enum.Enum):
|
270 |
"""
|
271 |
Represents a type of distributed environment.
|
272 |
-
|
273 |
Values:
|
274 |
-
|
275 |
- **NO** -- Not a distributed environment, just a single process.
|
276 |
- **DATA_PARALLEL** -- using sagemaker distributed data parallelism.
|
277 |
- **MODEL_PARALLEL** -- using sagemaker distributed model parallelism.
|
@@ -280,28 +223,20 @@ class SageMakerDistributedType(str, enum.Enum):
|
|
280 |
NO = "NO"
|
281 |
DATA_PARALLEL = "DATA_PARALLEL"
|
282 |
MODEL_PARALLEL = "MODEL_PARALLEL"
|
283 |
-
|
284 |
-
|
285 |
class ComputeEnvironment(str, enum.Enum):
|
286 |
"""
|
287 |
Represents a type of the compute environment.
|
288 |
-
|
289 |
Values:
|
290 |
-
|
291 |
- **LOCAL_MACHINE** -- private/custom cluster hardware.
|
292 |
- **AMAZON_SAGEMAKER** -- Amazon SageMaker as compute environment.
|
293 |
"""
|
294 |
# Subclassing str as well as Enum allows the `ComputeEnvironment` to be JSON-serializable out of the box.
|
295 |
LOCAL_MACHINE = "LOCAL_MACHINE"
|
296 |
AMAZON_SAGEMAKER = "AMAZON_SAGEMAKER"
|
297 |
-
|
298 |
-
|
299 |
class DynamoBackend(str, BaseEnum):
|
300 |
"""
|
301 |
Represents a dynamo backend (see https://github.com/pytorch/torchdynamo).
|
302 |
-
|
303 |
Values:
|
304 |
-
|
305 |
- **NO** -- Do not use torch dynamo.
|
306 |
- **EAGER** -- Uses PyTorch to run the extracted GraphModule. This is quite useful in debugging TorchDynamo
|
307 |
issues.
|
@@ -325,7 +260,6 @@ class DynamoBackend(str, BaseEnum):
|
|
325 |
- **IPEX** -- Uses IPEX for inference on CPU. Inference only. [Read
|
326 |
more](https://github.com/intel/intel-extension-for-pytorch).
|
327 |
- **TVM** -- Uses Apach TVM for inference optimizations. [Read more](https://tvm.apache.org/)
|
328 |
-
|
329 |
"""
|
330 |
# Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
|
331 |
NO = "NO"
|
@@ -341,13 +275,9 @@ class DynamoBackend(str, BaseEnum):
|
|
341 |
TENSORRT = "TENSORRT"
|
342 |
IPEX = "IPEX"
|
343 |
TVM = "TVM"
|
344 |
-
|
345 |
-
|
346 |
class LoggerType(BaseEnum):
|
347 |
"""Represents a type of supported experiment tracker
|
348 |
-
|
349 |
Values:
|
350 |
-
|
351 |
- **ALL** -- all available trackers in the environment that are supported
|
352 |
- **TENSORBOARD** -- TensorBoard as an experiment tracker
|
353 |
- **WANDB** -- wandb as an experiment tracker
|
@@ -362,13 +292,9 @@ class LoggerType(BaseEnum):
|
|
362 |
MLFLOW = "mlflow"
|
363 |
CLEARML = "clearml"
|
364 |
DVCLIVE = "dvclive"
|
365 |
-
|
366 |
-
|
367 |
class PrecisionType(BaseEnum):
|
368 |
"""Represents a type of precision used on floating point values
|
369 |
-
|
370 |
Values:
|
371 |
-
|
372 |
- **NO** -- using full precision (FP32)
|
373 |
- **FP16** -- using half precision
|
374 |
- **BF16** -- using brain floating point precision
|
@@ -377,8 +303,6 @@ class PrecisionType(BaseEnum):
|
|
377 |
FP8 = "fp8"
|
378 |
FP16 = "fp16"
|
379 |
BF16 = "bf16"
|
380 |
-
|
381 |
-
|
382 |
class RNGType(BaseEnum):
|
383 |
TORCH = "torch"
|
384 |
CUDA = "cuda"
|
@@ -386,25 +310,17 @@ class RNGType(BaseEnum):
|
|
386 |
XLA = "xla"
|
387 |
XPU = "xpu"
|
388 |
GENERATOR = "generator"
|
389 |
-
|
390 |
-
|
391 |
class CustomDtype(enum.Enum):
|
392 |
r"""
|
393 |
An enum that contains multiple custom dtypes that can be used for `infer_auto_device_map`.
|
394 |
"""
|
395 |
FP8 = "fp8"
|
396 |
INT4 = "int4"
|
397 |
-
|
398 |
-
|
399 |
# data classes
|
400 |
-
|
401 |
-
|
402 |
@dataclass
|
403 |
class TensorInformation:
|
404 |
shape: torch.Size
|
405 |
dtype: torch.dtype
|
406 |
-
|
407 |
-
|
408 |
@dataclass
|
409 |
class ProjectConfiguration:
|
410 |
"""
|
@@ -421,17 +337,14 @@ class ProjectConfiguration:
|
|
421 |
default=False,
|
422 |
metadata={"help": "Whether saved states should be automatically iteratively named."},
|
423 |
)
|
424 |
-
|
425 |
total_limit: int = field(
|
426 |
default=None,
|
427 |
metadata={"help": "The maximum number of total saved states to keep."},
|
428 |
)
|
429 |
-
|
430 |
iteration: int = field(
|
431 |
default=0,
|
432 |
metadata={"help": "The current save iteration."},
|
433 |
)
|
434 |
-
|
435 |
save_on_each_node: bool = field(
|
436 |
default=False,
|
437 |
metadata={
|
@@ -441,17 +354,13 @@ class ProjectConfiguration:
|
|
441 |
)
|
442 |
},
|
443 |
)
|
444 |
-
|
445 |
def set_directories(self, project_dir: str = None):
|
446 |
"Sets `self.project_dir` and `self.logging_dir` to the appropriate values."
|
447 |
self.project_dir = project_dir
|
448 |
if self.logging_dir is None:
|
449 |
self.logging_dir = project_dir
|
450 |
-
|
451 |
def __post_init__(self):
|
452 |
self.set_directories(self.project_dir)
|
453 |
-
|
454 |
-
|
455 |
@dataclass
|
456 |
class GradientAccumulationPlugin(KwargsHandler):
|
457 |
"""
|
@@ -470,8 +379,6 @@ class GradientAccumulationPlugin(KwargsHandler):
|
|
470 |
"help": "Whether to synchronize setting the gradients when at the end of the dataloader. Should only be set to `False` if you know what you're doing."
|
471 |
},
|
472 |
)
|
473 |
-
|
474 |
-
|
475 |
@dataclass
|
476 |
class TorchDynamoPlugin(KwargsHandler):
|
477 |
"""
|
@@ -488,7 +395,6 @@ class TorchDynamoPlugin(KwargsHandler):
|
|
488 |
dynamic: bool = field(default=None, metadata={"help": "Whether to use dynamic shape for tracing"})
|
489 |
options: Any = field(default=None, metadata={"help": "A dictionary of options to pass to the backend."})
|
490 |
disable: bool = field(default=False, metadata={"help": "Turn torch.compile() into a no-op for testing"})
|
491 |
-
|
492 |
def __post_init__(self):
|
493 |
prefix = "ACCELERATE_DYNAMO_"
|
494 |
if self.backend is None:
|
@@ -500,13 +406,10 @@ class TorchDynamoPlugin(KwargsHandler):
|
|
500 |
self.fullgraph = str_to_bool(os.environ.get(prefix + "USE_FULLGRAPH", "False")) == 1
|
501 |
if self.dynamic is None:
|
502 |
self.dynamic = str_to_bool(os.environ.get(prefix + "USE_DYNAMIC", "False")) == 1
|
503 |
-
|
504 |
def to_dict(self):
|
505 |
dynamo_config = copy.deepcopy(self.__dict__)
|
506 |
dynamo_config["backend"] = dynamo_config["backend"].value.lower()
|
507 |
return dynamo_config
|
508 |
-
|
509 |
-
|
510 |
@dataclass
|
511 |
class DeepSpeedPlugin:
|
512 |
"""
|
@@ -560,41 +463,31 @@ class DeepSpeedPlugin:
|
|
560 |
default=None,
|
561 |
metadata={"help": "Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3."},
|
562 |
)
|
563 |
-
|
564 |
def __post_init__(self):
|
565 |
from .deepspeed import HfDeepSpeedConfig
|
566 |
-
|
567 |
if self.gradient_accumulation_steps is None:
|
568 |
gas = os.environ.get("ACCELERATE_GRADIENT_ACCUMULATION_STEPS", "auto")
|
569 |
self.gradient_accumulation_steps = int(gas) if gas.isdigit() else gas
|
570 |
-
|
571 |
if self.gradient_clipping is None:
|
572 |
gradient_clipping = os.environ.get("ACCELERATE_GRADIENT_CLIPPING", "none")
|
573 |
if gradient_clipping != "none":
|
574 |
self.gradient_clipping = float(gradient_clipping)
|
575 |
-
|
576 |
if self.zero_stage is None:
|
577 |
self.zero_stage = int(os.environ.get("ACCELERATE_DEEPSPEED_ZERO_STAGE", 2))
|
578 |
-
|
579 |
if self.offload_optimizer_device is None:
|
580 |
self.offload_optimizer_device = os.environ.get("ACCELERATE_DEEPSPEED_OFFLOAD_OPTIMIZER_DEVICE", "none")
|
581 |
-
|
582 |
if self.offload_param_device is None:
|
583 |
self.offload_param_device = os.environ.get("ACCELERATE_DEEPSPEED_OFFLOAD_PARAM_DEVICE", "none")
|
584 |
-
|
585 |
if self.offload_optimizer_nvme_path is None:
|
586 |
self.offload_optimizer_nvme_path = os.environ.get(
|
587 |
"ACCELERATE_DEEPSPEED_OFFLOAD_OPTIMIZER_NVME_PATH", "none"
|
588 |
)
|
589 |
-
|
590 |
if self.offload_param_nvme_path is None:
|
591 |
self.offload_param_nvme_path = os.environ.get("ACCELERATE_DEEPSPEED_OFFLOAD_PARAM_NVME_PATH", "none")
|
592 |
-
|
593 |
if self.zero3_save_16bit_model is None:
|
594 |
self.zero3_save_16bit_model = (
|
595 |
os.environ.get("ACCELERATE_DEEPSPEED_ZERO3_SAVE_16BIT_MODEL", "false") == "true"
|
596 |
)
|
597 |
-
|
598 |
if self.hf_ds_config is None:
|
599 |
self.hf_ds_config = os.environ.get("ACCELERATE_DEEPSPEED_CONFIG_FILE", "none")
|
600 |
if (
|
@@ -608,7 +501,6 @@ class DeepSpeedPlugin:
|
|
608 |
self.hf_ds_config.config["gradient_accumulation_steps"] = 1
|
609 |
if "zero_optimization" not in self.hf_ds_config.config:
|
610 |
raise ValueError("Please specify the ZeRO optimization config in the DeepSpeed config.")
|
611 |
-
|
612 |
self._deepspeed_config_checks()
|
613 |
plugin_to_config_mapping = {
|
614 |
"gradient_accumulation_steps": "gradient_accumulation_steps",
|
@@ -624,7 +516,6 @@ class DeepSpeedPlugin:
|
|
624 |
for key in kwargs.keys():
|
625 |
self.fill_match(key, **kwargs, must_match=False)
|
626 |
self.hf_ds_config.set_stage_and_offload()
|
627 |
-
|
628 |
# filling the missing values in the class attributes from the DeepSpeed config
|
629 |
# when using the DeepSpeed config file.
|
630 |
for key, value in plugin_to_config_mapping.items():
|
@@ -654,7 +545,6 @@ class DeepSpeedPlugin:
|
|
654 |
if self.gradient_clipping:
|
655 |
config["gradient_clipping"] = self.gradient_clipping
|
656 |
self.hf_ds_config = HfDeepSpeedConfig(config)
|
657 |
-
|
658 |
self.deepspeed_config = self.hf_ds_config.config
|
659 |
self.deepspeed_config["steps_per_print"] = float("inf") # this will stop deepspeed from logging @ stdout
|
660 |
if self.zero3_init_flag is None:
|
@@ -664,13 +554,11 @@ class DeepSpeedPlugin:
|
|
664 |
if self.zero3_init_flag and not self.hf_ds_config.is_zero3():
|
665 |
warnings.warn("DeepSpeed Zero3 Init flag is only applicable for ZeRO Stage 3. Setting it to False.")
|
666 |
self.zero3_init_flag = False
|
667 |
-
|
668 |
def fill_match(self, ds_key_long, mismatches=None, must_match=True, **kwargs):
|
669 |
mismatches = [] if mismatches is None else mismatches
|
670 |
config, ds_key = self.hf_ds_config.find_config_node(ds_key_long)
|
671 |
if config is None:
|
672 |
return
|
673 |
-
|
674 |
if config.get(ds_key) == "auto":
|
675 |
if ds_key_long in kwargs:
|
676 |
config[ds_key] = kwargs[ds_key_long]
|
@@ -681,15 +569,12 @@ class DeepSpeedPlugin:
|
|
681 |
f"Please specify `{ds_key_long}` without `auto`(set to correct value) in the DeepSpeed config file or "
|
682 |
"pass it in kwargs."
|
683 |
)
|
684 |
-
|
685 |
if not must_match:
|
686 |
return
|
687 |
-
|
688 |
ds_val = config.get(ds_key)
|
689 |
if ds_val is not None and ds_key_long in kwargs:
|
690 |
if ds_val != kwargs[ds_key_long]:
|
691 |
mismatches.append(f"- ds {ds_key_long}={ds_val} vs arg {ds_key_long}={kwargs[ds_key_long]}")
|
692 |
-
|
693 |
def deepspeed_config_process(self, prefix="", mismatches=None, config=None, must_match=True, **kwargs):
|
694 |
"""Process the DeepSpeed config with the values from the kwargs."""
|
695 |
mismatches = [] if mismatches is None else mismatches
|
@@ -708,7 +593,6 @@ class DeepSpeedPlugin:
|
|
708 |
"Please correct the following DeepSpeed config values that mismatch kwargs "
|
709 |
f" values:\n{mismatches_msg}\nThe easiest method is to set these DeepSpeed config values to 'auto'."
|
710 |
)
|
711 |
-
|
712 |
def set_mixed_precision(self, mixed_precision):
|
713 |
ds_config = self.deepspeed_config
|
714 |
kwargs = {
|
@@ -721,7 +605,6 @@ class DeepSpeedPlugin:
|
|
721 |
elif mixed_precision == "bf16":
|
722 |
if "bf16" not in ds_config:
|
723 |
ds_config["bf16"] = {"enabled": True}
|
724 |
-
|
725 |
if mixed_precision != "no":
|
726 |
diff_dtype = "bf16" if mixed_precision == "fp16" else "fp16"
|
727 |
if str(ds_config.get(diff_dtype, {}).get("enabled", "False")).lower() == "true":
|
@@ -733,10 +616,8 @@ class DeepSpeedPlugin:
|
|
733 |
ds_config[dtype] = {"enabled": False}
|
734 |
self.fill_match("fp16.enabled", must_match=False, **kwargs)
|
735 |
self.fill_match("bf16.enabled", must_match=False, **kwargs)
|
736 |
-
|
737 |
def set_deepspeed_weakref(self):
|
738 |
from .imports import is_transformers_available
|
739 |
-
|
740 |
if self.zero3_init_flag:
|
741 |
if not is_transformers_available():
|
742 |
raise Exception(
|
@@ -753,17 +634,13 @@ class DeepSpeedPlugin:
|
|
753 |
ds_config["train_micro_batch_size_per_gpu"] = 1
|
754 |
if ds_config.get("train_batch_size", None) == "auto":
|
755 |
del ds_config["train_batch_size"]
|
756 |
-
|
757 |
if compare_versions("transformers", "<", "4.33"):
|
758 |
from transformers.deepspeed import HfDeepSpeedConfig
|
759 |
else:
|
760 |
from transformers.integrations import HfDeepSpeedConfig
|
761 |
-
|
762 |
self.dschf = HfDeepSpeedConfig(ds_config) # keep this object alive # noqa
|
763 |
-
|
764 |
def is_zero3_init_enabled(self):
|
765 |
return self.zero3_init_flag
|
766 |
-
|
767 |
@contextmanager
|
768 |
def zero3_init_context_manager(self, enable=False):
|
769 |
old = self.zero3_init_flag
|
@@ -777,7 +654,6 @@ class DeepSpeedPlugin:
|
|
777 |
self.zero3_init_flag = old
|
778 |
self.dschf = None
|
779 |
self.set_deepspeed_weakref()
|
780 |
-
|
781 |
def _deepspeed_config_checks(self):
|
782 |
env_variable_names_to_ignore = [
|
783 |
"ACCELERATE_GRADIENT_ACCUMULATION_STEPS",
|
@@ -793,9 +669,7 @@ class DeepSpeedPlugin:
|
|
793 |
env_variable_names_to_ignore = [
|
794 |
name.replace("ACCELERATE_", "").replace("DEEPSPEED_", "").lower() for name in env_variable_names_to_ignore
|
795 |
]
|
796 |
-
|
797 |
deepspeed_fields_from_accelerate_config = os.environ.get("ACCELERATE_CONFIG_DS_FIELDS", "").split(",")
|
798 |
-
|
799 |
if any(name in env_variable_names_to_ignore for name in deepspeed_fields_from_accelerate_config):
|
800 |
raise ValueError(
|
801 |
f"When using `deepspeed_config_file`, the following accelerate config variables will be ignored: {env_variable_names_to_ignore}.\n"
|
@@ -804,8 +678,6 @@ class DeepSpeedPlugin:
|
|
804 |
"The easiest method is to create a new config following the questionnaire via `accelerate config`.\n"
|
805 |
"It will only ask for the necessary config variables when using `deepspeed_config_file`."
|
806 |
)
|
807 |
-
|
808 |
-
|
809 |
@dataclass
|
810 |
class FullyShardedDataParallelPlugin:
|
811 |
"""
|
@@ -912,10 +784,8 @@ class FullyShardedDataParallelPlugin:
|
|
912 |
"for reduced memory usage."
|
913 |
},
|
914 |
)
|
915 |
-
|
916 |
def __post_init__(self):
|
917 |
from torch.distributed.fsdp.fully_sharded_data_parallel import BackwardPrefetch, CPUOffload, ShardingStrategy
|
918 |
-
|
919 |
prefix = "FSDP_"
|
920 |
if self.sharding_strategy is None:
|
921 |
sharding_strategy = os.environ.get(prefix + "SHARDING_STRATEGY", "FULL_SHARD")
|
@@ -925,18 +795,15 @@ class FullyShardedDataParallelPlugin:
|
|
925 |
else int(sharding_strategy)
|
926 |
)
|
927 |
self.sharding_strategy = ShardingStrategy(sharding_strategy)
|
928 |
-
|
929 |
if self.cpu_offload is None:
|
930 |
if str_to_bool(os.environ.get(prefix + "OFFLOAD_PARAMS", "False")) == 1:
|
931 |
self.cpu_offload = CPUOffload(offload_params=True)
|
932 |
else:
|
933 |
self.cpu_offload = CPUOffload(offload_params=False)
|
934 |
-
|
935 |
if self.backward_prefetch is None:
|
936 |
prefetch_policy = os.environ.get(prefix + "BACKWARD_PREFETCH", "NO_PREFETCH")
|
937 |
if prefetch_policy != FSDP_BACKWARD_PREFETCH[-1]:
|
938 |
self.backward_prefetch = BackwardPrefetch(FSDP_BACKWARD_PREFETCH.index(prefetch_policy) + 1)
|
939 |
-
|
940 |
if self.state_dict_type is None:
|
941 |
state_dict_type_policy = os.environ.get(prefix + "STATE_DICT_TYPE", "FULL_STATE_DICT")
|
942 |
self.set_state_dict_type(state_dict_type_policy)
|
@@ -944,7 +811,6 @@ class FullyShardedDataParallelPlugin:
|
|
944 |
self.sync_module_states = str_to_bool(os.environ.get(prefix + "SYNC_MODULE_STATES", "True")) == 1
|
945 |
self.forward_prefetch = str_to_bool(os.environ.get(prefix + "FORWARD_PREFETCH", "False")) == 1
|
946 |
self.activation_checkpointing = str_to_bool(os.environ.get(prefix + "ACTIVATION_CHECKPOINTING", "False")) == 1
|
947 |
-
|
948 |
if self.sync_module_states:
|
949 |
if is_npu_available():
|
950 |
device = torch.npu.current_device()
|
@@ -957,12 +823,10 @@ class FullyShardedDataParallelPlugin:
|
|
957 |
"There are currently no available devices found, must be one of 'XPU', 'CUDA', or 'NPU'."
|
958 |
)
|
959 |
self.param_init_fn = lambda x: x.to_empty(device=device, recurse=False)
|
960 |
-
|
961 |
@staticmethod
|
962 |
def get_module_class_from_name(module, name):
|
963 |
"""
|
964 |
Gets a class from a module by its name.
|
965 |
-
|
966 |
Args:
|
967 |
module (`torch.nn.Module`): The module to get the class from.
|
968 |
name (`str`): The name of the class.
|
@@ -977,10 +841,8 @@ class FullyShardedDataParallelPlugin:
|
|
977 |
module_class = FullyShardedDataParallelPlugin.get_module_class_from_name(child_module, name)
|
978 |
if module_class is not None:
|
979 |
return module_class
|
980 |
-
|
981 |
def set_auto_wrap_policy(self, model):
|
982 |
from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy, transformer_auto_wrap_policy
|
983 |
-
|
984 |
default_transformer_cls_names_to_wrap = (
|
985 |
",".join(model._no_split_modules) if getattr(model, "_no_split_modules", None) is not None else ""
|
986 |
)
|
@@ -997,7 +859,6 @@ class FullyShardedDataParallelPlugin:
|
|
997 |
raise Exception("Could not find the transformer layer class to wrap in the model.")
|
998 |
else:
|
999 |
transformer_cls_to_wrap.add(transformer_cls)
|
1000 |
-
|
1001 |
self.auto_wrap_policy = functools.partial(
|
1002 |
transformer_auto_wrap_policy,
|
1003 |
# Transformer layer class to wrap
|
@@ -1009,7 +870,6 @@ class FullyShardedDataParallelPlugin:
|
|
1009 |
self.auto_wrap_policy = functools.partial(
|
1010 |
size_based_auto_wrap_policy, min_num_params=min_num_params
|
1011 |
)
|
1012 |
-
|
1013 |
def set_mixed_precision(self, mixed_precision):
|
1014 |
if mixed_precision == "fp16":
|
1015 |
dtype = torch.float16
|
@@ -1018,26 +878,20 @@ class FullyShardedDataParallelPlugin:
|
|
1018 |
else:
|
1019 |
raise ValueError(f"Unknown mixed precision value: {mixed_precision}")
|
1020 |
from torch.distributed.fsdp.fully_sharded_data_parallel import MixedPrecision
|
1021 |
-
|
1022 |
if self.mixed_precision_policy is None:
|
1023 |
self.mixed_precision_policy = MixedPrecision(param_dtype=dtype, reduce_dtype=dtype, buffer_dtype=dtype)
|
1024 |
-
|
1025 |
def set_state_dict_type(self, state_dict_type_policy):
|
1026 |
from torch.distributed.fsdp.fully_sharded_data_parallel import (
|
1027 |
FullOptimStateDictConfig,
|
1028 |
FullStateDictConfig,
|
1029 |
StateDictType,
|
1030 |
)
|
1031 |
-
|
1032 |
self.state_dict_type = StateDictType(FSDP_STATE_DICT_TYPE.index(state_dict_type_policy) + 1)
|
1033 |
-
|
1034 |
if self.state_dict_type == StateDictType.FULL_STATE_DICT:
|
1035 |
if self.state_dict_config is None:
|
1036 |
self.state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
|
1037 |
if self.optim_state_dict_config is None:
|
1038 |
self.optim_state_dict_config = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
|
1039 |
-
|
1040 |
-
|
1041 |
@dataclass
|
1042 |
class MegatronLMPlugin:
|
1043 |
"""
|
@@ -1169,7 +1023,6 @@ class MegatronLMPlugin:
|
|
1169 |
default=False,
|
1170 |
metadata={"help": "Whether to return logits from the model."},
|
1171 |
)
|
1172 |
-
|
1173 |
# custom train step args
|
1174 |
custom_train_step_class: Optional[Any] = field(
|
1175 |
default=None,
|
@@ -1179,7 +1032,6 @@ class MegatronLMPlugin:
|
|
1179 |
default=None,
|
1180 |
metadata={"help": "Custom train step kwargs."},
|
1181 |
)
|
1182 |
-
|
1183 |
# custom model args
|
1184 |
custom_model_provider_function: Optional[Callable] = field(
|
1185 |
default=None,
|
@@ -1189,14 +1041,12 @@ class MegatronLMPlugin:
|
|
1189 |
default=None,
|
1190 |
metadata={"help": "Custom prepare model function."},
|
1191 |
)
|
1192 |
-
|
1193 |
# remaining args such as enabling Alibi/ROPE positional embeddings,
|
1194 |
# wandb logging, Multi-Query Attention, etc.
|
1195 |
other_megatron_args: Optional[Dict[str, Any]] = field(
|
1196 |
default=None,
|
1197 |
metadata={"help": "Other Megatron-LM arguments. Please refer Megatron-LM"},
|
1198 |
)
|
1199 |
-
|
1200 |
def __post_init__(self):
|
1201 |
prefix = "MEGATRON_LM_"
|
1202 |
if self.tp_degree is None:
|
@@ -1215,18 +1065,15 @@ class MegatronLMPlugin:
|
|
1215 |
)
|
1216 |
if self.sequence_parallelism is None:
|
1217 |
self.sequence_parallelism = str_to_bool(os.environ.get(prefix + "SEQUENCE_PARALLELISM", "False")) == 1
|
1218 |
-
|
1219 |
if self.pp_degree > 1 or self.use_distributed_optimizer:
|
1220 |
self.DDP_impl = "local"
|
1221 |
else:
|
1222 |
self.DDP_impl = "torch"
|
1223 |
-
|
1224 |
if self.consumed_samples is not None:
|
1225 |
if len(self.consumed_samples) == 1:
|
1226 |
self.consumed_samples.extend([0, 0])
|
1227 |
elif len(self.consumed_samples) == 2:
|
1228 |
self.consumed_samples.append(0)
|
1229 |
-
|
1230 |
self.megatron_lm_default_args = {
|
1231 |
"tensor_model_parallel_size": self.tp_degree,
|
1232 |
"pipeline_model_parallel_size": self.pp_degree,
|
@@ -1253,7 +1100,6 @@ class MegatronLMPlugin:
|
|
1253 |
self.set_tensorboard_logging_options()
|
1254 |
if self.other_megatron_args is not None:
|
1255 |
self.megatron_lm_default_args.update(self.other_megatron_args)
|
1256 |
-
|
1257 |
def set_network_size_args(self, model, batch_data=None):
|
1258 |
# Check if the model is either BERT, GPT or T5 else raise error
|
1259 |
# set 'num_layers', 'hidden_size', 'num_attention_heads', 'max_position_embeddings'
|
@@ -1317,7 +1163,6 @@ class MegatronLMPlugin:
|
|
1317 |
self.decoder_seq_length = batch_data["labels"].shape[1]
|
1318 |
else:
|
1319 |
self.decoder_seq_length = max_position_embeddings
|
1320 |
-
|
1321 |
self.megatron_lm_default_args["encoder_seq_length"] = self.encoder_seq_length
|
1322 |
self.megatron_lm_default_args["decoder_seq_length"] = self.decoder_seq_length
|
1323 |
else:
|
@@ -1325,7 +1170,6 @@ class MegatronLMPlugin:
|
|
1325 |
"🤗 Accelerate Megatron-LM integration supports only BERT, GPT and T5 model. "
|
1326 |
"Please check the model you are using is one of those."
|
1327 |
)
|
1328 |
-
|
1329 |
self.megatron_lm_default_args["model_type_name"] = model_type_name
|
1330 |
self.megatron_lm_default_args["num_layers"] = num_layers
|
1331 |
self.megatron_lm_default_args["hidden_size"] = hidden_size
|
@@ -1336,7 +1180,6 @@ class MegatronLMPlugin:
|
|
1336 |
self.megatron_lm_default_args["model_return_dict"] = model.config.return_dict
|
1337 |
if model_type_name == "bert":
|
1338 |
self.megatron_lm_default_args["num_labels"] = num_labels
|
1339 |
-
|
1340 |
def set_mixed_precision(self, mixed_precision):
|
1341 |
if mixed_precision == "fp16":
|
1342 |
self.megatron_lm_default_args["fp16"] = True
|
@@ -1344,7 +1187,6 @@ class MegatronLMPlugin:
|
|
1344 |
self.megatron_lm_default_args["bf16"] = True
|
1345 |
self.DDP_impl = "local"
|
1346 |
self.megatron_lm_default_args["DDP_impl"] = self.DDP_impl
|
1347 |
-
|
1348 |
def set_training_args(self, micro_batch_size, dp_degree):
|
1349 |
self.data_parallel_size = dp_degree
|
1350 |
self.micro_batch_size = micro_batch_size
|
@@ -1352,7 +1194,6 @@ class MegatronLMPlugin:
|
|
1352 |
self.megatron_lm_default_args["data_parallel_size"] = self.data_parallel_size
|
1353 |
self.megatron_lm_default_args["micro_batch_size"] = self.micro_batch_size
|
1354 |
self.megatron_lm_default_args["global_batch_size"] = self.global_batch_size
|
1355 |
-
|
1356 |
def set_optimizer_type(self, optimizer):
|
1357 |
optimizer_name = optimizer.__class__.__name__.lower()
|
1358 |
if "adam" in optimizer_name:
|
@@ -1365,10 +1206,8 @@ class MegatronLMPlugin:
|
|
1365 |
self.megatron_lm_default_args["sgd_momentum"] = optimizer.defaults["momentum"]
|
1366 |
else:
|
1367 |
raise ValueError(f"Optimizer {optimizer_name} is not supported by Megatron-LM")
|
1368 |
-
|
1369 |
self.megatron_lm_default_args["lr"] = optimizer.defaults["lr"]
|
1370 |
self.megatron_lm_default_args["weight_decay"] = optimizer.defaults["weight_decay"]
|
1371 |
-
|
1372 |
def set_scheduler_args(self, scheduler):
|
1373 |
if self.train_iters is None:
|
1374 |
self.train_iters = scheduler.total_num_steps // self.megatron_lm_default_args["data_parallel_size"]
|
@@ -1384,7 +1223,6 @@ class MegatronLMPlugin:
|
|
1384 |
"Ignoring `lr_warmup_samples` as `lr_warmup_iters` based on scheduler is being used for training."
|
1385 |
)
|
1386 |
self.lr_warmup_samples = 0
|
1387 |
-
|
1388 |
self.megatron_lm_default_args["train_iters"] = self.train_iters
|
1389 |
self.megatron_lm_default_args["lr_warmup_iters"] = self.lr_warmup_iters
|
1390 |
self.megatron_lm_default_args["train_samples"] = self.train_samples
|
@@ -1397,10 +1235,8 @@ class MegatronLMPlugin:
|
|
1397 |
self.megatron_lm_default_args["start_weight_decay"] = self.start_weight_decay
|
1398 |
self.megatron_lm_default_args["end_weight_decay"] = self.end_weight_decay
|
1399 |
self.megatron_lm_default_args["min_lr"] = self.min_lr
|
1400 |
-
|
1401 |
def set_tensorboard_logging_options(self):
|
1402 |
from megatron.arguments import _add_logging_args
|
1403 |
-
|
1404 |
parser = argparse.ArgumentParser()
|
1405 |
parser = _add_logging_args(parser)
|
1406 |
logging_args = parser.parse_known_args()
|
@@ -1410,35 +1246,28 @@ class MegatronLMPlugin:
|
|
1410 |
self.megatron_lm_default_args[key] = True
|
1411 |
elif key.startswith("no_log_"):
|
1412 |
self.megatron_lm_default_args[key.replace("no_", "")] = True
|
1413 |
-
|
1414 |
-
|
1415 |
@dataclass
|
1416 |
class BnbQuantizationConfig:
|
1417 |
"""
|
1418 |
A plugin to enable BitsAndBytes 4bit and 8bit quantization
|
1419 |
"""
|
1420 |
load_in_8bit: bool = field(default=False, metadata={"help": "enable 8bit quantization."})
|
1421 |
-
|
1422 |
llm_int8_threshold: float = field(
|
1423 |
default=6.0, metadata={"help": "value of the outliner threshold. only relevant when load_in_8bit=True"}
|
1424 |
)
|
1425 |
-
|
1426 |
load_in_4bit: bool = field(default=False, metadata={"help": "enable 4bit quantization."})
|
1427 |
-
|
1428 |
bnb_4bit_quant_type: str = field(
|
1429 |
default="fp4",
|
1430 |
metadata={
|
1431 |
"help": "set the quantization data type in the `bnb.nn.Linear4Bit` layers. Options are {'fp4','np4'}."
|
1432 |
},
|
1433 |
)
|
1434 |
-
|
1435 |
bnb_4bit_use_double_quant: bool = field(
|
1436 |
default=False,
|
1437 |
metadata={
|
1438 |
"help": "enable nested quantization where the quantization constants from the first quantization are quantized again."
|
1439 |
},
|
1440 |
)
|
1441 |
-
|
1442 |
bnb_4bit_compute_dtype: bool = field(
|
1443 |
default="fp16",
|
1444 |
metadata={
|
@@ -1446,7 +1275,6 @@ class BnbQuantizationConfig:
|
|
1446 |
"fp32, but computation can be set to bf16 for speedups. Options are {'fp32','fp16','bf16'}."
|
1447 |
},
|
1448 |
)
|
1449 |
-
|
1450 |
torch_dtype: torch.dtype = field(
|
1451 |
default=None,
|
1452 |
metadata={
|
@@ -1454,46 +1282,36 @@ class BnbQuantizationConfig:
|
|
1454 |
"to `torch.float16` for 8 bit model and use the same dtype as the compute dtype for 4 bit model "
|
1455 |
},
|
1456 |
)
|
1457 |
-
|
1458 |
skip_modules: List[str] = field(
|
1459 |
default=None,
|
1460 |
metadata={
|
1461 |
"help": "an explicit list of the modules that we don't quantize. The dtype of these modules will be `torch_dtype`."
|
1462 |
},
|
1463 |
)
|
1464 |
-
|
1465 |
keep_in_fp32_modules: List[str] = field(
|
1466 |
default=None,
|
1467 |
metadata={"help": "an explicit list of the modules that we don't quantize. We keep them in `torch.float32`."},
|
1468 |
)
|
1469 |
-
|
1470 |
def __post_init__(self):
|
1471 |
"""
|
1472 |
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
|
1473 |
"""
|
1474 |
if not isinstance(self.load_in_8bit, bool):
|
1475 |
raise ValueError("load_in_8bit must be a boolean")
|
1476 |
-
|
1477 |
if not isinstance(self.load_in_4bit, bool):
|
1478 |
raise ValueError("load_in_4bit must be a boolean")
|
1479 |
-
|
1480 |
if self.load_in_4bit and self.load_in_8bit:
|
1481 |
raise ValueError("load_in_4bit and load_in_8 can't be both True")
|
1482 |
-
|
1483 |
if not self.load_in_4bit and not self.load_in_8bit:
|
1484 |
raise ValueError("load_in_4bit and load_in_8 can't be both False")
|
1485 |
-
|
1486 |
if not isinstance(self.llm_int8_threshold, (int, float)):
|
1487 |
raise ValueError("llm_int8_threshold must be a float or an int")
|
1488 |
-
|
1489 |
if not isinstance(self.bnb_4bit_quant_type, str):
|
1490 |
raise ValueError("bnb_4bit_quant_type must be a string")
|
1491 |
elif self.bnb_4bit_quant_type not in ["fp4", "nf4"]:
|
1492 |
raise ValueError(f"bnb_4bit_quant_type must be in ['fp4','nf4'] but found {self.bnb_4bit_quant_type}")
|
1493 |
-
|
1494 |
if not isinstance(self.bnb_4bit_use_double_quant, bool):
|
1495 |
raise ValueError("bnb_4bit_use_double_quant must be a boolean")
|
1496 |
-
|
1497 |
if isinstance(self.bnb_4bit_compute_dtype, str):
|
1498 |
if self.bnb_4bit_compute_dtype == "fp32":
|
1499 |
self.bnb_4bit_compute_dtype = torch.float32
|
@@ -1507,22 +1325,16 @@ class BnbQuantizationConfig:
|
|
1507 |
)
|
1508 |
elif not isinstance(self.bnb_4bit_compute_dtype, torch.dtype):
|
1509 |
raise ValueError("bnb_4bit_compute_dtype must be a string or a torch.dtype")
|
1510 |
-
|
1511 |
if self.skip_modules is not None and not isinstance(self.skip_modules, list):
|
1512 |
raise ValueError("skip_modules must be a list of strings")
|
1513 |
-
|
1514 |
if self.keep_in_fp32_modules is not None and not isinstance(self.keep_in_fp32_modules, list):
|
1515 |
raise ValueError("keep_in_fp_32_modules must be a list of strings")
|
1516 |
-
|
1517 |
if self.load_in_4bit:
|
1518 |
self.target_dtype = CustomDtype.INT4
|
1519 |
-
|
1520 |
if self.load_in_8bit:
|
1521 |
self.target_dtype = torch.int8
|
1522 |
-
|
1523 |
if self.load_in_4bit and self.llm_int8_threshold != 6.0:
|
1524 |
warnings.warn("llm_int8_threshold can only be used for model loaded in 8bit")
|
1525 |
-
|
1526 |
if isinstance(self.torch_dtype, str):
|
1527 |
if self.torch_dtype == "fp32":
|
1528 |
self.torch_dtype = torch.float32
|
@@ -1534,9 +1346,7 @@ class BnbQuantizationConfig:
|
|
1534 |
raise ValueError(f"torch_dtype must be in ['fp32','fp16','bf16'] but found {self.torch_dtype}")
|
1535 |
if self.load_in_8bit and self.torch_dtype is None:
|
1536 |
self.torch_dtype = torch.float16
|
1537 |
-
|
1538 |
if self.load_in_4bit and self.torch_dtype is None:
|
1539 |
self.torch_dtype = self.bnb_4bit_compute_dtype
|
1540 |
-
|
1541 |
if not isinstance(self.torch_dtype, torch.dtype):
|
1542 |
raise ValueError("torch_dtype must be a torch.dtype")
|
|
|
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.
|
13 |
"""
|
14 |
# import clear_environment here to avoid circular import problem
|
15 |
from .other import clear_environment
|
|
|
16 |
with clear_environment():
|
17 |
default_dict = self.__class__().to_dict()
|
18 |
this_dict = self.to_dict()
|
19 |
return {k: v for k, v in this_dict.items() if default_dict[k] != v}
|
|
|
|
|
20 |
@dataclass
|
21 |
class AutocastKwargs(KwargsHandler):
|
22 |
"""
|
23 |
Use this object in your [`Accelerator`] to customize how `torch.autocast` behaves. Please refer to the
|
24 |
documentation of this [context manager](https://pytorch.org/docs/stable/amp.html#torch.autocast) for more
|
25 |
information on each argument.
|
|
|
26 |
Example:
|
|
|
27 |
```python
|
28 |
from accelerate import Accelerator
|
29 |
from accelerate.utils import AutocastKwargs
|
|
|
30 |
kwargs = AutocastKwargs(cache_enabled=True)
|
31 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
32 |
```
|
33 |
"""
|
34 |
enabled: bool = True
|
35 |
cache_enabled: bool = None
|
|
|
|
|
36 |
@dataclass
|
37 |
class DistributedDataParallelKwargs(KwargsHandler):
|
38 |
"""
|
|
|
40 |
`torch.nn.parallel.DistributedDataParallel`. Please refer to the documentation of this
|
41 |
[wrapper](https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html) for more
|
42 |
information on each argument.
|
|
|
43 |
<Tip warning={true}>
|
|
|
44 |
`gradient_as_bucket_view` is only available in PyTorch 1.7.0 and later versions.
|
|
|
45 |
`static_graph` is only available in PyTorch 1.11.0 and later versions.
|
|
|
46 |
</Tip>
|
|
|
47 |
Example:
|
|
|
48 |
```python
|
49 |
from accelerate import Accelerator
|
50 |
from accelerate.utils import DistributedDataParallelKwargs
|
|
|
51 |
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
|
52 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
53 |
```
|
|
|
59 |
check_reduction: bool = False
|
60 |
gradient_as_bucket_view: bool = False
|
61 |
static_graph: bool = False
|
|
|
|
|
62 |
@dataclass
|
63 |
class GradScalerKwargs(KwargsHandler):
|
64 |
"""
|
65 |
Use this object in your [`Accelerator`] to customize the behavior of mixed precision, specifically how the
|
66 |
`torch.cuda.amp.GradScaler` used is created. Please refer to the documentation of this
|
67 |
[scaler](https://pytorch.org/docs/stable/amp.html?highlight=gradscaler) for more information on each argument.
|
|
|
68 |
<Tip warning={true}>
|
|
|
69 |
`GradScaler` is only available in PyTorch 1.5.0 and later versions.
|
|
|
70 |
</Tip>
|
|
|
71 |
Example:
|
|
|
72 |
```python
|
73 |
from accelerate import Accelerator
|
74 |
from accelerate.utils import GradScalerKwargs
|
|
|
75 |
kwargs = GradScalerKwargs(backoff_filter=0.25)
|
76 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
77 |
```
|
|
|
81 |
backoff_factor: float = 0.5
|
82 |
growth_interval: int = 2000
|
83 |
enabled: bool = True
|
|
|
|
|
84 |
@dataclass
|
85 |
class InitProcessGroupKwargs(KwargsHandler):
|
86 |
"""
|
|
|
88 |
to the documentation of this
|
89 |
[method](https://pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group) for more
|
90 |
information on each argument.
|
|
|
91 |
```python
|
92 |
from datetime import timedelta
|
93 |
from accelerate import Accelerator
|
94 |
from accelerate.utils import InitProcessGroupKwargs
|
|
|
95 |
kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=800))
|
96 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
97 |
```
|
|
|
99 |
backend: Optional[str] = "nccl"
|
100 |
init_method: Optional[str] = None
|
101 |
timeout: timedelta = timedelta(seconds=1800)
|
|
|
|
|
102 |
# Literals
|
103 |
Backend = Literal["msamp", "te"]
|
104 |
OptLevel = Literal["O1", "O2"]
|
105 |
FP8Format = Literal["E4M3", "HYBRID"]
|
106 |
AmaxComputeAlgorithm = Literal["max", "most_recent"]
|
|
|
|
|
107 |
@dataclass
|
108 |
class FP8RecipeKwargs(KwargsHandler):
|
109 |
"""
|
110 |
Use this object in your [`Accelerator`] to customize the initialization of the recipe for FP8 mixed precision
|
111 |
training with `transformer-engine` or `ms-amp`.
|
|
|
112 |
<Tip>
|
|
|
113 |
For more information on `transformer-engine` args, please refer to the API
|
114 |
[documentation](https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/api/common.html).
|
|
|
115 |
For more information on the `ms-amp` args, please refer to the Optimization Level
|
116 |
[documentation](https://azure.github.io/MS-AMP/docs/user-tutorial/optimization-level).
|
|
|
117 |
</Tip>
|
|
|
118 |
```python
|
119 |
from accelerate import Accelerator
|
120 |
from accelerate.utils import FP8RecipeKwargs
|
|
|
121 |
kwargs = FP8RecipeKwargs(backend="te", fp8_format="HYBRID")
|
122 |
accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=[kwargs])
|
123 |
```
|
|
|
124 |
To use MS-AMP as an engine, pass `backend="msamp"` and the `optimization_level`:
|
|
|
125 |
```python
|
126 |
kwargs = FP8RecipeKwargs(backend="msamp", optimization_level="02")
|
127 |
```
|
|
|
128 |
Args:
|
129 |
backend (`str`, *optional*, defaults to "msamp"):
|
130 |
Which FP8 engine to use. Must be one of `"msamp"` (MS-AMP) or `"te"` (TransformerEngine).
|
|
|
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):
|
|
|
173 |
elif self.backend == "MSAMP":
|
174 |
if self.opt_level not in get_args(OptLevel):
|
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)
|
181 |
except ValueError:
|
182 |
return False
|
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))
|
|
|
|
|
192 |
class DistributedType(str, enum.Enum):
|
193 |
"""
|
194 |
Represents a type of distributed environment.
|
|
|
195 |
Values:
|
|
|
196 |
- **NO** -- Not a distributed environment, just a single process.
|
197 |
- **MULTI_CPU** -- Distributed on multiple CPU nodes.
|
198 |
- **MULTI_GPU** -- Distributed on multiple GPUs.
|
|
|
211 |
FSDP = "FSDP"
|
212 |
TPU = "TPU"
|
213 |
MEGATRON_LM = "MEGATRON_LM"
|
|
|
|
|
214 |
class SageMakerDistributedType(str, enum.Enum):
|
215 |
"""
|
216 |
Represents a type of distributed environment.
|
|
|
217 |
Values:
|
|
|
218 |
- **NO** -- Not a distributed environment, just a single process.
|
219 |
- **DATA_PARALLEL** -- using sagemaker distributed data parallelism.
|
220 |
- **MODEL_PARALLEL** -- using sagemaker distributed model parallelism.
|
|
|
223 |
NO = "NO"
|
224 |
DATA_PARALLEL = "DATA_PARALLEL"
|
225 |
MODEL_PARALLEL = "MODEL_PARALLEL"
|
|
|
|
|
226 |
class ComputeEnvironment(str, enum.Enum):
|
227 |
"""
|
228 |
Represents a type of the compute environment.
|
|
|
229 |
Values:
|
|
|
230 |
- **LOCAL_MACHINE** -- private/custom cluster hardware.
|
231 |
- **AMAZON_SAGEMAKER** -- Amazon SageMaker as compute environment.
|
232 |
"""
|
233 |
# Subclassing str as well as Enum allows the `ComputeEnvironment` to be JSON-serializable out of the box.
|
234 |
LOCAL_MACHINE = "LOCAL_MACHINE"
|
235 |
AMAZON_SAGEMAKER = "AMAZON_SAGEMAKER"
|
|
|
|
|
236 |
class DynamoBackend(str, BaseEnum):
|
237 |
"""
|
238 |
Represents a dynamo backend (see https://github.com/pytorch/torchdynamo).
|
|
|
239 |
Values:
|
|
|
240 |
- **NO** -- Do not use torch dynamo.
|
241 |
- **EAGER** -- Uses PyTorch to run the extracted GraphModule. This is quite useful in debugging TorchDynamo
|
242 |
issues.
|
|
|
260 |
- **IPEX** -- Uses IPEX for inference on CPU. Inference only. [Read
|
261 |
more](https://github.com/intel/intel-extension-for-pytorch).
|
262 |
- **TVM** -- Uses Apach TVM for inference optimizations. [Read more](https://tvm.apache.org/)
|
|
|
263 |
"""
|
264 |
# Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
|
265 |
NO = "NO"
|
|
|
275 |
TENSORRT = "TENSORRT"
|
276 |
IPEX = "IPEX"
|
277 |
TVM = "TVM"
|
|
|
|
|
278 |
class LoggerType(BaseEnum):
|
279 |
"""Represents a type of supported experiment tracker
|
|
|
280 |
Values:
|
|
|
281 |
- **ALL** -- all available trackers in the environment that are supported
|
282 |
- **TENSORBOARD** -- TensorBoard as an experiment tracker
|
283 |
- **WANDB** -- wandb as an experiment tracker
|
|
|
292 |
MLFLOW = "mlflow"
|
293 |
CLEARML = "clearml"
|
294 |
DVCLIVE = "dvclive"
|
|
|
|
|
295 |
class PrecisionType(BaseEnum):
|
296 |
"""Represents a type of precision used on floating point values
|
|
|
297 |
Values:
|
|
|
298 |
- **NO** -- using full precision (FP32)
|
299 |
- **FP16** -- using half precision
|
300 |
- **BF16** -- using brain floating point precision
|
|
|
303 |
FP8 = "fp8"
|
304 |
FP16 = "fp16"
|
305 |
BF16 = "bf16"
|
|
|
|
|
306 |
class RNGType(BaseEnum):
|
307 |
TORCH = "torch"
|
308 |
CUDA = "cuda"
|
|
|
310 |
XLA = "xla"
|
311 |
XPU = "xpu"
|
312 |
GENERATOR = "generator"
|
|
|
|
|
313 |
class CustomDtype(enum.Enum):
|
314 |
r"""
|
315 |
An enum that contains multiple custom dtypes that can be used for `infer_auto_device_map`.
|
316 |
"""
|
317 |
FP8 = "fp8"
|
318 |
INT4 = "int4"
|
|
|
|
|
319 |
# data classes
|
|
|
|
|
320 |
@dataclass
|
321 |
class TensorInformation:
|
322 |
shape: torch.Size
|
323 |
dtype: torch.dtype
|
|
|
|
|
324 |
@dataclass
|
325 |
class ProjectConfiguration:
|
326 |
"""
|
|
|
337 |
default=False,
|
338 |
metadata={"help": "Whether saved states should be automatically iteratively named."},
|
339 |
)
|
|
|
340 |
total_limit: int = field(
|
341 |
default=None,
|
342 |
metadata={"help": "The maximum number of total saved states to keep."},
|
343 |
)
|
|
|
344 |
iteration: int = field(
|
345 |
default=0,
|
346 |
metadata={"help": "The current save iteration."},
|
347 |
)
|
|
|
348 |
save_on_each_node: bool = field(
|
349 |
default=False,
|
350 |
metadata={
|
|
|
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
|
365 |
class GradientAccumulationPlugin(KwargsHandler):
|
366 |
"""
|
|
|
379 |
"help": "Whether to synchronize setting the gradients when at the end of the dataloader. Should only be set to `False` if you know what you're doing."
|
380 |
},
|
381 |
)
|
|
|
|
|
382 |
@dataclass
|
383 |
class TorchDynamoPlugin(KwargsHandler):
|
384 |
"""
|
|
|
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 |
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()
|
412 |
return dynamo_config
|
|
|
|
|
413 |
@dataclass
|
414 |
class DeepSpeedPlugin:
|
415 |
"""
|
|
|
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:
|
469 |
gas = os.environ.get("ACCELERATE_GRADIENT_ACCUMULATION_STEPS", "auto")
|
470 |
self.gradient_accumulation_steps = int(gas) if gas.isdigit() else gas
|
|
|
471 |
if self.gradient_clipping is None:
|
472 |
gradient_clipping = os.environ.get("ACCELERATE_GRADIENT_CLIPPING", "none")
|
473 |
if gradient_clipping != "none":
|
474 |
self.gradient_clipping = float(gradient_clipping)
|
|
|
475 |
if self.zero_stage is None:
|
476 |
self.zero_stage = int(os.environ.get("ACCELERATE_DEEPSPEED_ZERO_STAGE", 2))
|
|
|
477 |
if self.offload_optimizer_device is None:
|
478 |
self.offload_optimizer_device = os.environ.get("ACCELERATE_DEEPSPEED_OFFLOAD_OPTIMIZER_DEVICE", "none")
|
|
|
479 |
if self.offload_param_device is None:
|
480 |
self.offload_param_device = os.environ.get("ACCELERATE_DEEPSPEED_OFFLOAD_PARAM_DEVICE", "none")
|
|
|
481 |
if self.offload_optimizer_nvme_path is None:
|
482 |
self.offload_optimizer_nvme_path = os.environ.get(
|
483 |
"ACCELERATE_DEEPSPEED_OFFLOAD_OPTIMIZER_NVME_PATH", "none"
|
484 |
)
|
|
|
485 |
if self.offload_param_nvme_path is None:
|
486 |
self.offload_param_nvme_path = os.environ.get("ACCELERATE_DEEPSPEED_OFFLOAD_PARAM_NVME_PATH", "none")
|
|
|
487 |
if self.zero3_save_16bit_model is None:
|
488 |
self.zero3_save_16bit_model = (
|
489 |
os.environ.get("ACCELERATE_DEEPSPEED_ZERO3_SAVE_16BIT_MODEL", "false") == "true"
|
490 |
)
|
|
|
491 |
if self.hf_ds_config is None:
|
492 |
self.hf_ds_config = os.environ.get("ACCELERATE_DEEPSPEED_CONFIG_FILE", "none")
|
493 |
if (
|
|
|
501 |
self.hf_ds_config.config["gradient_accumulation_steps"] = 1
|
502 |
if "zero_optimization" not in self.hf_ds_config.config:
|
503 |
raise ValueError("Please specify the ZeRO optimization config in the DeepSpeed config.")
|
|
|
504 |
self._deepspeed_config_checks()
|
505 |
plugin_to_config_mapping = {
|
506 |
"gradient_accumulation_steps": "gradient_accumulation_steps",
|
|
|
516 |
for key in kwargs.keys():
|
517 |
self.fill_match(key, **kwargs, must_match=False)
|
518 |
self.hf_ds_config.set_stage_and_offload()
|
|
|
519 |
# filling the missing values in the class attributes from the DeepSpeed config
|
520 |
# when using the DeepSpeed config file.
|
521 |
for key, value in plugin_to_config_mapping.items():
|
|
|
545 |
if self.gradient_clipping:
|
546 |
config["gradient_clipping"] = self.gradient_clipping
|
547 |
self.hf_ds_config = HfDeepSpeedConfig(config)
|
|
|
548 |
self.deepspeed_config = self.hf_ds_config.config
|
549 |
self.deepspeed_config["steps_per_print"] = float("inf") # this will stop deepspeed from logging @ stdout
|
550 |
if self.zero3_init_flag is None:
|
|
|
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)
|
560 |
if config is None:
|
561 |
return
|
|
|
562 |
if config.get(ds_key) == "auto":
|
563 |
if ds_key_long in kwargs:
|
564 |
config[ds_key] = kwargs[ds_key_long]
|
|
|
569 |
f"Please specify `{ds_key_long}` without `auto`(set to correct value) in the DeepSpeed config file or "
|
570 |
"pass it in kwargs."
|
571 |
)
|
|
|
572 |
if not must_match:
|
573 |
return
|
|
|
574 |
ds_val = config.get(ds_key)
|
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 |
"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 = {
|
|
|
605 |
elif mixed_precision == "bf16":
|
606 |
if "bf16" not in ds_config:
|
607 |
ds_config["bf16"] = {"enabled": True}
|
|
|
608 |
if mixed_precision != "no":
|
609 |
diff_dtype = "bf16" if mixed_precision == "fp16" else "fp16"
|
610 |
if str(ds_config.get(diff_dtype, {}).get("enabled", "False")).lower() == "true":
|
|
|
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:
|
622 |
if not is_transformers_available():
|
623 |
raise Exception(
|
|
|
634 |
ds_config["train_micro_batch_size_per_gpu"] = 1
|
635 |
if ds_config.get("train_batch_size", None) == "auto":
|
636 |
del ds_config["train_batch_size"]
|
|
|
637 |
if compare_versions("transformers", "<", "4.33"):
|
638 |
from transformers.deepspeed import HfDeepSpeedConfig
|
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
|
|
|
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",
|
|
|
669 |
env_variable_names_to_ignore = [
|
670 |
name.replace("ACCELERATE_", "").replace("DEEPSPEED_", "").lower() for name in env_variable_names_to_ignore
|
671 |
]
|
|
|
672 |
deepspeed_fields_from_accelerate_config = os.environ.get("ACCELERATE_CONFIG_DS_FIELDS", "").split(",")
|
|
|
673 |
if any(name in env_variable_names_to_ignore for name in deepspeed_fields_from_accelerate_config):
|
674 |
raise ValueError(
|
675 |
f"When using `deepspeed_config_file`, the following accelerate config variables will be ignored: {env_variable_names_to_ignore}.\n"
|
|
|
678 |
"The easiest method is to create a new config following the questionnaire via `accelerate config`.\n"
|
679 |
"It will only ask for the necessary config variables when using `deepspeed_config_file`."
|
680 |
)
|
|
|
|
|
681 |
@dataclass
|
682 |
class FullyShardedDataParallelPlugin:
|
683 |
"""
|
|
|
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_"
|
790 |
if self.sharding_strategy is None:
|
791 |
sharding_strategy = os.environ.get(prefix + "SHARDING_STRATEGY", "FULL_SHARD")
|
|
|
795 |
else int(sharding_strategy)
|
796 |
)
|
797 |
self.sharding_strategy = ShardingStrategy(sharding_strategy)
|
|
|
798 |
if self.cpu_offload is None:
|
799 |
if str_to_bool(os.environ.get(prefix + "OFFLOAD_PARAMS", "False")) == 1:
|
800 |
self.cpu_offload = CPUOffload(offload_params=True)
|
801 |
else:
|
802 |
self.cpu_offload = CPUOffload(offload_params=False)
|
|
|
803 |
if self.backward_prefetch is None:
|
804 |
prefetch_policy = os.environ.get(prefix + "BACKWARD_PREFETCH", "NO_PREFETCH")
|
805 |
if prefetch_policy != FSDP_BACKWARD_PREFETCH[-1]:
|
806 |
self.backward_prefetch = BackwardPrefetch(FSDP_BACKWARD_PREFETCH.index(prefetch_policy) + 1)
|
|
|
807 |
if self.state_dict_type is None:
|
808 |
state_dict_type_policy = os.environ.get(prefix + "STATE_DICT_TYPE", "FULL_STATE_DICT")
|
809 |
self.set_state_dict_type(state_dict_type_policy)
|
|
|
811 |
self.sync_module_states = str_to_bool(os.environ.get(prefix + "SYNC_MODULE_STATES", "True")) == 1
|
812 |
self.forward_prefetch = str_to_bool(os.environ.get(prefix + "FORWARD_PREFETCH", "False")) == 1
|
813 |
self.activation_checkpointing = str_to_bool(os.environ.get(prefix + "ACTIVATION_CHECKPOINTING", "False")) == 1
|
|
|
814 |
if self.sync_module_states:
|
815 |
if is_npu_available():
|
816 |
device = torch.npu.current_device()
|
|
|
823 |
"There are currently no available devices found, must be one of 'XPU', 'CUDA', or 'NPU'."
|
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.
|
|
|
830 |
Args:
|
831 |
module (`torch.nn.Module`): The module to get the class from.
|
832 |
name (`str`): The name of the class.
|
|
|
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 = (
|
847 |
",".join(model._no_split_modules) if getattr(model, "_no_split_modules", None) is not None else ""
|
848 |
)
|
|
|
859 |
raise Exception("Could not find the transformer layer class to wrap in the model.")
|
860 |
else:
|
861 |
transformer_cls_to_wrap.add(transformer_cls)
|
|
|
862 |
self.auto_wrap_policy = functools.partial(
|
863 |
transformer_auto_wrap_policy,
|
864 |
# Transformer layer class to wrap
|
|
|
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
|
|
|
878 |
else:
|
879 |
raise ValueError(f"Unknown mixed precision value: {mixed_precision}")
|
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,
|
886 |
FullStateDictConfig,
|
887 |
StateDictType,
|
888 |
)
|
|
|
889 |
self.state_dict_type = StateDictType(FSDP_STATE_DICT_TYPE.index(state_dict_type_policy) + 1)
|
|
|
890 |
if self.state_dict_type == StateDictType.FULL_STATE_DICT:
|
891 |
if self.state_dict_config is None:
|
892 |
self.state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
|
893 |
if self.optim_state_dict_config is None:
|
894 |
self.optim_state_dict_config = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
|
|
|
|
|
895 |
@dataclass
|
896 |
class MegatronLMPlugin:
|
897 |
"""
|
|
|
1023 |
default=False,
|
1024 |
metadata={"help": "Whether to return logits from the model."},
|
1025 |
)
|
|
|
1026 |
# custom train step args
|
1027 |
custom_train_step_class: Optional[Any] = field(
|
1028 |
default=None,
|
|
|
1032 |
default=None,
|
1033 |
metadata={"help": "Custom train step kwargs."},
|
1034 |
)
|
|
|
1035 |
# custom model args
|
1036 |
custom_model_provider_function: Optional[Callable] = field(
|
1037 |
default=None,
|
|
|
1041 |
default=None,
|
1042 |
metadata={"help": "Custom prepare model function."},
|
1043 |
)
|
|
|
1044 |
# remaining args such as enabling Alibi/ROPE positional embeddings,
|
1045 |
# wandb logging, Multi-Query Attention, etc.
|
1046 |
other_megatron_args: Optional[Dict[str, Any]] = field(
|
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:
|
|
|
1065 |
)
|
1066 |
if self.sequence_parallelism is None:
|
1067 |
self.sequence_parallelism = str_to_bool(os.environ.get(prefix + "SEQUENCE_PARALLELISM", "False")) == 1
|
|
|
1068 |
if self.pp_degree > 1 or self.use_distributed_optimizer:
|
1069 |
self.DDP_impl = "local"
|
1070 |
else:
|
1071 |
self.DDP_impl = "torch"
|
|
|
1072 |
if self.consumed_samples is not None:
|
1073 |
if len(self.consumed_samples) == 1:
|
1074 |
self.consumed_samples.extend([0, 0])
|
1075 |
elif len(self.consumed_samples) == 2:
|
1076 |
self.consumed_samples.append(0)
|
|
|
1077 |
self.megatron_lm_default_args = {
|
1078 |
"tensor_model_parallel_size": self.tp_degree,
|
1079 |
"pipeline_model_parallel_size": self.pp_degree,
|
|
|
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'
|
|
|
1163 |
self.decoder_seq_length = batch_data["labels"].shape[1]
|
1164 |
else:
|
1165 |
self.decoder_seq_length = max_position_embeddings
|
|
|
1166 |
self.megatron_lm_default_args["encoder_seq_length"] = self.encoder_seq_length
|
1167 |
self.megatron_lm_default_args["decoder_seq_length"] = self.decoder_seq_length
|
1168 |
else:
|
|
|
1170 |
"🤗 Accelerate Megatron-LM integration supports only BERT, GPT and T5 model. "
|
1171 |
"Please check the model you are using is one of those."
|
1172 |
)
|
|
|
1173 |
self.megatron_lm_default_args["model_type_name"] = model_type_name
|
1174 |
self.megatron_lm_default_args["num_layers"] = num_layers
|
1175 |
self.megatron_lm_default_args["hidden_size"] = hidden_size
|
|
|
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 |
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 |
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:
|
|
|
1206 |
self.megatron_lm_default_args["sgd_momentum"] = optimizer.defaults["momentum"]
|
1207 |
else:
|
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"]
|
|
|
1223 |
"Ignoring `lr_warmup_samples` as `lr_warmup_iters` based on scheduler is being used for training."
|
1224 |
)
|
1225 |
self.lr_warmup_samples = 0
|
|
|
1226 |
self.megatron_lm_default_args["train_iters"] = self.train_iters
|
1227 |
self.megatron_lm_default_args["lr_warmup_iters"] = self.lr_warmup_iters
|
1228 |
self.megatron_lm_default_args["train_samples"] = self.train_samples
|
|
|
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()
|
1241 |
parser = _add_logging_args(parser)
|
1242 |
logging_args = parser.parse_known_args()
|
|
|
1246 |
self.megatron_lm_default_args[key] = True
|
1247 |
elif key.startswith("no_log_"):
|
1248 |
self.megatron_lm_default_args[key.replace("no_", "")] = True
|
|
|
|
|
1249 |
@dataclass
|
1250 |
class BnbQuantizationConfig:
|
1251 |
"""
|
1252 |
A plugin to enable BitsAndBytes 4bit and 8bit quantization
|
1253 |
"""
|
1254 |
load_in_8bit: bool = field(default=False, metadata={"help": "enable 8bit quantization."})
|
|
|
1255 |
llm_int8_threshold: float = field(
|
1256 |
default=6.0, metadata={"help": "value of the outliner threshold. only relevant when load_in_8bit=True"}
|
1257 |
)
|
|
|
1258 |
load_in_4bit: bool = field(default=False, metadata={"help": "enable 4bit quantization."})
|
|
|
1259 |
bnb_4bit_quant_type: str = field(
|
1260 |
default="fp4",
|
1261 |
metadata={
|
1262 |
"help": "set the quantization data type in the `bnb.nn.Linear4Bit` layers. Options are {'fp4','np4'}."
|
1263 |
},
|
1264 |
)
|
|
|
1265 |
bnb_4bit_use_double_quant: bool = field(
|
1266 |
default=False,
|
1267 |
metadata={
|
1268 |
"help": "enable nested quantization where the quantization constants from the first quantization are quantized again."
|
1269 |
},
|
1270 |
)
|
|
|
1271 |
bnb_4bit_compute_dtype: bool = field(
|
1272 |
default="fp16",
|
1273 |
metadata={
|
|
|
1275 |
"fp32, but computation can be set to bf16 for speedups. Options are {'fp32','fp16','bf16'}."
|
1276 |
},
|
1277 |
)
|
|
|
1278 |
torch_dtype: torch.dtype = field(
|
1279 |
default=None,
|
1280 |
metadata={
|
|
|
1282 |
"to `torch.float16` for 8 bit model and use the same dtype as the compute dtype for 4 bit model "
|
1283 |
},
|
1284 |
)
|
|
|
1285 |
skip_modules: List[str] = field(
|
1286 |
default=None,
|
1287 |
metadata={
|
1288 |
"help": "an explicit list of the modules that we don't quantize. The dtype of these modules will be `torch_dtype`."
|
1289 |
},
|
1290 |
)
|
|
|
1291 |
keep_in_fp32_modules: List[str] = field(
|
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.
|
1298 |
"""
|
1299 |
if not isinstance(self.load_in_8bit, bool):
|
1300 |
raise ValueError("load_in_8bit must be a boolean")
|
|
|
1301 |
if not isinstance(self.load_in_4bit, bool):
|
1302 |
raise ValueError("load_in_4bit must be a boolean")
|
|
|
1303 |
if self.load_in_4bit and self.load_in_8bit:
|
1304 |
raise ValueError("load_in_4bit and load_in_8 can't be both True")
|
|
|
1305 |
if not self.load_in_4bit and not self.load_in_8bit:
|
1306 |
raise ValueError("load_in_4bit and load_in_8 can't be both False")
|
|
|
1307 |
if not isinstance(self.llm_int8_threshold, (int, float)):
|
1308 |
raise ValueError("llm_int8_threshold must be a float or an int")
|
|
|
1309 |
if not isinstance(self.bnb_4bit_quant_type, str):
|
1310 |
raise ValueError("bnb_4bit_quant_type must be a string")
|
1311 |
elif self.bnb_4bit_quant_type not in ["fp4", "nf4"]:
|
1312 |
raise ValueError(f"bnb_4bit_quant_type must be in ['fp4','nf4'] but found {self.bnb_4bit_quant_type}")
|
|
|
1313 |
if not isinstance(self.bnb_4bit_use_double_quant, bool):
|
1314 |
raise ValueError("bnb_4bit_use_double_quant must be a boolean")
|
|
|
1315 |
if isinstance(self.bnb_4bit_compute_dtype, str):
|
1316 |
if self.bnb_4bit_compute_dtype == "fp32":
|
1317 |
self.bnb_4bit_compute_dtype = torch.float32
|
|
|
1325 |
)
|
1326 |
elif not isinstance(self.bnb_4bit_compute_dtype, torch.dtype):
|
1327 |
raise ValueError("bnb_4bit_compute_dtype must be a string or a torch.dtype")
|
|
|
1328 |
if self.skip_modules is not None and not isinstance(self.skip_modules, list):
|
1329 |
raise ValueError("skip_modules must be a list of strings")
|
|
|
1330 |
if self.keep_in_fp32_modules is not None and not isinstance(self.keep_in_fp32_modules, list):
|
1331 |
raise ValueError("keep_in_fp_32_modules must be a list of strings")
|
|
|
1332 |
if self.load_in_4bit:
|
1333 |
self.target_dtype = CustomDtype.INT4
|
|
|
1334 |
if self.load_in_8bit:
|
1335 |
self.target_dtype = torch.int8
|
|
|
1336 |
if self.load_in_4bit and self.llm_int8_threshold != 6.0:
|
1337 |
warnings.warn("llm_int8_threshold can only be used for model loaded in 8bit")
|
|
|
1338 |
if isinstance(self.torch_dtype, str):
|
1339 |
if self.torch_dtype == "fp32":
|
1340 |
self.torch_dtype = torch.float32
|
|
|
1346 |
raise ValueError(f"torch_dtype must be in ['fp32','fp16','bf16'] but found {self.torch_dtype}")
|
1347 |
if self.load_in_8bit and self.torch_dtype is None:
|
1348 |
self.torch_dtype = torch.float16
|
|
|
1349 |
if self.load_in_4bit and self.torch_dtype is None:
|
1350 |
self.torch_dtype = self.bnb_4bit_compute_dtype
|
|
|
1351 |
if not isinstance(self.torch_dtype, torch.dtype):
|
1352 |
raise ValueError("torch_dtype must be a torch.dtype")
|
src/utils/deepspeed.py
CHANGED
@@ -1,18 +1,14 @@
|
|
1 |
class HfDeepSpeedConfig:
|
2 |
"""
|
3 |
This object contains a DeepSpeed configuration dictionary and can be quickly queried for things like zero stage.
|
4 |
-
|
5 |
A `weakref` of this object is stored in the module's globals to be able to access the config from areas where
|
6 |
things like the Trainer object is not available (e.g. `from_pretrained` and `_get_resized_embeddings`). Therefore
|
7 |
it's important that this object remains alive while the program is still running.
|
8 |
-
|
9 |
[`Trainer`] uses the `HfTrainerDeepSpeedConfig` subclass instead. That subclass has logic to sync the configuration
|
10 |
with values of [`TrainingArguments`] by replacing special placeholder values: `"auto"`. Without this special logic
|
11 |
the DeepSpeed configuration is not modified in any way.
|
12 |
-
|
13 |
Args:
|
14 |
config_file_or_dict (`Union[str, Dict]`): path to DeepSpeed config file or dict.
|
15 |
-
|
16 |
"""
|
17 |
def __init__(self, config_file_or_dict):
|
18 |
if isinstance(config_file_or_dict, dict):
|
@@ -30,17 +26,13 @@ class HfDeepSpeedConfig:
|
|
30 |
raise ValueError(
|
31 |
f"Expected a string path to an existing deepspeed config, or a dictionary, or a base64 encoded string. Received: {config_file_or_dict}"
|
32 |
)
|
33 |
-
|
34 |
self.config = config
|
35 |
-
|
36 |
self.set_stage_and_offload()
|
37 |
-
|
38 |
def set_stage_and_offload(self):
|
39 |
# zero stage - this is done as early as possible, before model is created, to allow
|
40 |
# ``is_deepspeed_zero3_enabled`` query and getting to the early deepspeed config object
|
41 |
# during ``zero.Init()`` which needs to know the dtype, and some other hparams.
|
42 |
self._stage = self.get_value("zero_optimization.stage", -1)
|
43 |
-
|
44 |
# offload
|
45 |
self._offload = False
|
46 |
if self.is_zero2() or self.is_zero3():
|
@@ -53,10 +45,8 @@ class HfDeepSpeedConfig:
|
|
53 |
)
|
54 |
if len(offload_devices & offload_devices_valid) > 0:
|
55 |
self._offload = True
|
56 |
-
|
57 |
def find_config_node(self, ds_key_long):
|
58 |
config = self.config
|
59 |
-
|
60 |
# find the config node of interest if it exists
|
61 |
nodes = ds_key_long.split(".")
|
62 |
ds_key = nodes.pop()
|
@@ -64,9 +54,7 @@ class HfDeepSpeedConfig:
|
|
64 |
config = config.get(node)
|
65 |
if config is None:
|
66 |
return None, ds_key
|
67 |
-
|
68 |
return config, ds_key
|
69 |
-
|
70 |
def get_value(self, ds_key_long, default=None):
|
71 |
"""
|
72 |
Returns the set value or `default` if no value is set
|
@@ -75,15 +63,12 @@ class HfDeepSpeedConfig:
|
|
75 |
if config is None:
|
76 |
return default
|
77 |
return config.get(ds_key, default)
|
78 |
-
|
79 |
def del_config_sub_tree(self, ds_key_long, must_exist=False):
|
80 |
"""
|
81 |
Deletes a sub-section of the config file if it's found.
|
82 |
-
|
83 |
Unless `must_exist` is `True` the section doesn't have to exist.
|
84 |
"""
|
85 |
config = self.config
|
86 |
-
|
87 |
# find the config node of interest if it exists
|
88 |
nodes = ds_key_long.split(".")
|
89 |
for node in nodes:
|
@@ -94,20 +79,16 @@ class HfDeepSpeedConfig:
|
|
94 |
raise ValueError(f"Can't find {ds_key_long} entry in the config: {self.config}")
|
95 |
else:
|
96 |
return
|
97 |
-
|
98 |
# if found remove it
|
99 |
if parent_config is not None:
|
100 |
parent_config.pop(node)
|
101 |
-
|
102 |
def is_true(self, ds_key_long):
|
103 |
"""
|
104 |
Returns `True`/``False` only if the value is set, always `False` otherwise. So use this method to ask the very
|
105 |
specific question of whether the value is set to `True` (and it's not set to `False`` or isn't set).
|
106 |
-
|
107 |
"""
|
108 |
value = self.get_value(ds_key_long)
|
109 |
return False if value is None else bool(value)
|
110 |
-
|
111 |
def is_false(self, ds_key_long):
|
112 |
"""
|
113 |
Returns `True`/``False` only if the value is set, always `False` otherwise. So use this method to ask the very
|
@@ -115,31 +96,23 @@ class HfDeepSpeedConfig:
|
|
115 |
"""
|
116 |
value = self.get_value(ds_key_long)
|
117 |
return False if value is None else not bool(value)
|
118 |
-
|
119 |
def is_zero2(self):
|
120 |
return self._stage == 2
|
121 |
-
|
122 |
def is_zero3(self):
|
123 |
return self._stage == 3
|
124 |
-
|
125 |
def is_offload(self):
|
126 |
return self._offload
|
127 |
-
|
128 |
-
|
129 |
class DeepSpeedEngineWrapper:
|
130 |
"""
|
131 |
Internal wrapper for deepspeed.runtime.engine.DeepSpeedEngine. This is used to follow conventional training loop.
|
132 |
-
|
133 |
Args:
|
134 |
engine (deepspeed.runtime.engine.DeepSpeedEngine): deepspeed engine to wrap
|
135 |
"""
|
136 |
def __init__(self, engine):
|
137 |
self.engine = engine
|
138 |
-
|
139 |
def backward(self, loss, **kwargs):
|
140 |
# runs backpropagation and handles mixed precision
|
141 |
self.engine.backward(loss, **kwargs)
|
142 |
-
|
143 |
# Deepspeed's `engine.step` performs the following operations:
|
144 |
# - gradient accumulation check
|
145 |
# - gradient clipping
|
@@ -151,12 +124,9 @@ class DeepSpeedEngineWrapper:
|
|
151 |
# and this plugin overrides the above calls with no-ops when Accelerate runs under
|
152 |
# Deepspeed, but allows normal functionality for non-Deepspeed cases thus enabling a simple
|
153 |
# training loop that works transparently under many training regimes.
|
154 |
-
|
155 |
-
|
156 |
class DeepSpeedOptimizerWrapper(AcceleratedOptimizer):
|
157 |
"""
|
158 |
Internal wrapper around a deepspeed optimizer.
|
159 |
-
|
160 |
Args:
|
161 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
162 |
The optimizer to wrap.
|
@@ -164,25 +134,19 @@ class DeepSpeedOptimizerWrapper(AcceleratedOptimizer):
|
|
164 |
def __init__(self, optimizer):
|
165 |
super().__init__(optimizer, device_placement=False, scaler=None)
|
166 |
self.__has_overflow__ = hasattr(self.optimizer, "overflow")
|
167 |
-
|
168 |
def zero_grad(self, set_to_none=None):
|
169 |
pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
|
170 |
-
|
171 |
def step(self):
|
172 |
pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
|
173 |
-
|
174 |
@property
|
175 |
def step_was_skipped(self):
|
176 |
"""Whether or not the optimizer step was done, or skipped because of gradient overflow."""
|
177 |
if self.__has_overflow__:
|
178 |
return self.optimizer.overflow
|
179 |
return False
|
180 |
-
|
181 |
-
|
182 |
class DeepSpeedSchedulerWrapper(AcceleratedScheduler):
|
183 |
"""
|
184 |
Internal wrapper around a deepspeed scheduler.
|
185 |
-
|
186 |
Args:
|
187 |
scheduler (`torch.optim.lr_scheduler.LambdaLR`):
|
188 |
The scheduler to wrap.
|
@@ -190,16 +154,12 @@ class DeepSpeedSchedulerWrapper(AcceleratedScheduler):
|
|
190 |
"""
|
191 |
def __init__(self, scheduler, optimizers):
|
192 |
super().__init__(scheduler, optimizers)
|
193 |
-
|
194 |
def step(self):
|
195 |
pass # `accelerator.backward(loss)` is doing that automatically. Therefore, its implementation is not needed
|
196 |
-
|
197 |
-
|
198 |
class DummyOptim:
|
199 |
"""
|
200 |
Dummy optimizer presents model parameters or param groups, this is primarily used to follow conventional training
|
201 |
loop when optimizer config is specified in the deepspeed config file.
|
202 |
-
|
203 |
Args:
|
204 |
lr (float):
|
205 |
Learning rate.
|
@@ -215,13 +175,10 @@ class DummyOptim:
|
|
215 |
self.lr = lr
|
216 |
self.weight_decay = weight_decay
|
217 |
self.kwargs = kwargs
|
218 |
-
|
219 |
-
|
220 |
class DummyScheduler:
|
221 |
"""
|
222 |
Dummy scheduler presents model parameters or param groups, this is primarily used to follow conventional training
|
223 |
loop when scheduler config is specified in the deepspeed config file.
|
224 |
-
|
225 |
Args:
|
226 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
227 |
The optimizer to wrap.
|
|
|
1 |
class HfDeepSpeedConfig:
|
2 |
"""
|
3 |
This object contains a DeepSpeed configuration dictionary and can be quickly queried for things like zero stage.
|
|
|
4 |
A `weakref` of this object is stored in the module's globals to be able to access the config from areas where
|
5 |
things like the Trainer object is not available (e.g. `from_pretrained` and `_get_resized_embeddings`). Therefore
|
6 |
it's important that this object remains alive while the program is still running.
|
|
|
7 |
[`Trainer`] uses the `HfTrainerDeepSpeedConfig` subclass instead. That subclass has logic to sync the configuration
|
8 |
with values of [`TrainingArguments`] by replacing special placeholder values: `"auto"`. Without this special logic
|
9 |
the DeepSpeed configuration is not modified in any way.
|
|
|
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):
|
|
|
26 |
raise ValueError(
|
27 |
f"Expected a string path to an existing deepspeed config, or a dictionary, or a base64 encoded string. Received: {config_file_or_dict}"
|
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
|
34 |
# during ``zero.Init()`` which needs to know the dtype, and some other hparams.
|
35 |
self._stage = self.get_value("zero_optimization.stage", -1)
|
|
|
36 |
# offload
|
37 |
self._offload = False
|
38 |
if self.is_zero2() or self.is_zero3():
|
|
|
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
|
51 |
nodes = ds_key_long.split(".")
|
52 |
ds_key = nodes.pop()
|
|
|
54 |
config = config.get(node)
|
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 |
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.
|
|
|
69 |
Unless `must_exist` is `True` the section doesn't have to exist.
|
70 |
"""
|
71 |
config = self.config
|
|
|
72 |
# find the config node of interest if it exists
|
73 |
nodes = ds_key_long.split(".")
|
74 |
for node in nodes:
|
|
|
79 |
raise ValueError(f"Can't find {ds_key_long} entry in the config: {self.config}")
|
80 |
else:
|
81 |
return
|
|
|
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
|
88 |
specific question of whether the value is set to `True` (and it's not set to `False`` or isn't set).
|
|
|
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 |
"""
|
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:
|
106 |
"""
|
107 |
Internal wrapper for deepspeed.runtime.engine.DeepSpeedEngine. This is used to follow conventional training loop.
|
|
|
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)
|
|
|
116 |
# Deepspeed's `engine.step` performs the following operations:
|
117 |
# - gradient accumulation check
|
118 |
# - gradient clipping
|
|
|
124 |
# and this plugin overrides the above calls with no-ops when Accelerate runs under
|
125 |
# Deepspeed, but allows normal functionality for non-Deepspeed cases thus enabling a simple
|
126 |
# training loop that works transparently under many training regimes.
|
|
|
|
|
127 |
class DeepSpeedOptimizerWrapper(AcceleratedOptimizer):
|
128 |
"""
|
129 |
Internal wrapper around a deepspeed optimizer.
|
|
|
130 |
Args:
|
131 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
132 |
The optimizer to wrap.
|
|
|
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__:
|
145 |
return self.optimizer.overflow
|
146 |
return False
|
|
|
|
|
147 |
class DeepSpeedSchedulerWrapper(AcceleratedScheduler):
|
148 |
"""
|
149 |
Internal wrapper around a deepspeed scheduler.
|
|
|
150 |
Args:
|
151 |
scheduler (`torch.optim.lr_scheduler.LambdaLR`):
|
152 |
The scheduler to wrap.
|
|
|
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:
|
160 |
"""
|
161 |
Dummy optimizer presents model parameters or param groups, this is primarily used to follow conventional training
|
162 |
loop when optimizer config is specified in the deepspeed config file.
|
|
|
163 |
Args:
|
164 |
lr (float):
|
165 |
Learning rate.
|
|
|
175 |
self.lr = lr
|
176 |
self.weight_decay = weight_decay
|
177 |
self.kwargs = kwargs
|
|
|
|
|
178 |
class DummyScheduler:
|
179 |
"""
|
180 |
Dummy scheduler presents model parameters or param groups, this is primarily used to follow conventional training
|
181 |
loop when scheduler config is specified in the deepspeed config file.
|
|
|
182 |
Args:
|
183 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
184 |
The optimizer to wrap.
|
src/utils/environment.py
CHANGED
@@ -1,7 +1,6 @@
|
|
1 |
def str_to_bool(value) -> int:
|
2 |
"""
|
3 |
Converts a string representation of truth to `True` (1) or `False` (0).
|
4 |
-
|
5 |
True values are `y`, `yes`, `t`, `true`, `on`, and `1`; False value are `n`, `no`, `f`, `false`, `off`, and `0`;
|
6 |
"""
|
7 |
value = value.lower()
|
@@ -11,8 +10,6 @@ def str_to_bool(value) -> int:
|
|
11 |
return 0
|
12 |
else:
|
13 |
raise ValueError(f"invalid truth value {value}")
|
14 |
-
|
15 |
-
|
16 |
def get_int_from_env(env_keys, default):
|
17 |
"""Returns the first positive env value found in the `env_keys` list or the default."""
|
18 |
for e in env_keys:
|
@@ -20,30 +17,21 @@ def get_int_from_env(env_keys, default):
|
|
20 |
if val >= 0:
|
21 |
return val
|
22 |
return default
|
23 |
-
|
24 |
-
|
25 |
def parse_flag_from_env(key, default=False):
|
26 |
"""Returns truthy value for `key` from the env if available else the default."""
|
27 |
value = os.environ.get(key, str(default))
|
28 |
return str_to_bool(value) == 1 # As its name indicates `str_to_bool` actually returns an int...
|
29 |
-
|
30 |
-
|
31 |
def parse_choice_from_env(key, default="no"):
|
32 |
value = os.environ.get(key, str(default))
|
33 |
return value
|
34 |
-
|
35 |
-
|
36 |
def are_libraries_initialized(*library_names: str) -> Dict[str, bool]:
|
37 |
"""
|
38 |
Checks if any of `library_names` are imported in the environment. Will return results as a `key:bool` pair.
|
39 |
"""
|
40 |
return [lib_name for lib_name in library_names if lib_name in sys.modules.keys()]
|
41 |
-
|
42 |
-
|
43 |
def get_gpu_info():
|
44 |
"""
|
45 |
Gets GPU count and names using `nvidia-smi` instead of torch to not initialize CUDA.
|
46 |
-
|
47 |
Largely based on the `gputil` library.
|
48 |
"""
|
49 |
if platform.system() == "Windows":
|
@@ -64,13 +52,10 @@ def get_gpu_info():
|
|
64 |
gpu_count = len(gpus)
|
65 |
gpu_names = [gpu.split(",")[1].strip() for gpu in gpus]
|
66 |
return gpu_names, gpu_count
|
67 |
-
|
68 |
-
|
69 |
def check_cuda_p2p_ib_support():
|
70 |
"""
|
71 |
Checks if the devices being used have issues with P2P and IB communications, namely any consumer GPU hardware after
|
72 |
the 3090.
|
73 |
-
|
74 |
Noteably uses `nvidia-smi` instead of torch to not initialize CUDA.
|
75 |
"""
|
76 |
try:
|
@@ -86,12 +71,9 @@ def check_cuda_p2p_ib_support():
|
|
86 |
except Exception:
|
87 |
pass
|
88 |
return True
|
89 |
-
|
90 |
-
|
91 |
def check_fp8_capability():
|
92 |
"""
|
93 |
Checks if all the current GPUs available support FP8.
|
94 |
-
|
95 |
Notably must initialize `torch.cuda` to check.
|
96 |
"""
|
97 |
cuda_device_capacity = torch.cuda.get_device_capability()
|
|
|
1 |
def str_to_bool(value) -> int:
|
2 |
"""
|
3 |
Converts a string representation of truth to `True` (1) or `False` (0).
|
|
|
4 |
True values are `y`, `yes`, `t`, `true`, `on`, and `1`; False value are `n`, `no`, `f`, `false`, `off`, and `0`;
|
5 |
"""
|
6 |
value = value.lower()
|
|
|
10 |
return 0
|
11 |
else:
|
12 |
raise ValueError(f"invalid truth value {value}")
|
|
|
|
|
13 |
def get_int_from_env(env_keys, default):
|
14 |
"""Returns the first positive env value found in the `env_keys` list or the default."""
|
15 |
for e in env_keys:
|
|
|
17 |
if val >= 0:
|
18 |
return val
|
19 |
return default
|
|
|
|
|
20 |
def parse_flag_from_env(key, default=False):
|
21 |
"""Returns truthy value for `key` from the env if available else the default."""
|
22 |
value = os.environ.get(key, str(default))
|
23 |
return str_to_bool(value) == 1 # As its name indicates `str_to_bool` actually returns an int...
|
|
|
|
|
24 |
def parse_choice_from_env(key, default="no"):
|
25 |
value = os.environ.get(key, str(default))
|
26 |
return value
|
|
|
|
|
27 |
def are_libraries_initialized(*library_names: str) -> Dict[str, bool]:
|
28 |
"""
|
29 |
Checks if any of `library_names` are imported in the environment. Will return results as a `key:bool` pair.
|
30 |
"""
|
31 |
return [lib_name for lib_name in library_names if lib_name in sys.modules.keys()]
|
|
|
|
|
32 |
def get_gpu_info():
|
33 |
"""
|
34 |
Gets GPU count and names using `nvidia-smi` instead of torch to not initialize CUDA.
|
|
|
35 |
Largely based on the `gputil` library.
|
36 |
"""
|
37 |
if platform.system() == "Windows":
|
|
|
52 |
gpu_count = len(gpus)
|
53 |
gpu_names = [gpu.split(",")[1].strip() for gpu in gpus]
|
54 |
return gpu_names, gpu_count
|
|
|
|
|
55 |
def check_cuda_p2p_ib_support():
|
56 |
"""
|
57 |
Checks if the devices being used have issues with P2P and IB communications, namely any consumer GPU hardware after
|
58 |
the 3090.
|
|
|
59 |
Noteably uses `nvidia-smi` instead of torch to not initialize CUDA.
|
60 |
"""
|
61 |
try:
|
|
|
71 |
except Exception:
|
72 |
pass
|
73 |
return True
|
|
|
|
|
74 |
def check_fp8_capability():
|
75 |
"""
|
76 |
Checks if all the current GPUs available support FP8.
|
|
|
77 |
Notably must initialize `torch.cuda` to check.
|
78 |
"""
|
79 |
cuda_device_capacity = torch.cuda.get_device_capability()
|
src/utils/fsdp_utils.py
CHANGED
@@ -1,16 +1,12 @@
|
|
1 |
logger = get_logger(__name__)
|
2 |
-
|
3 |
-
|
4 |
def save_fsdp_model(fsdp_plugin, accelerator, model, output_dir, model_index=0):
|
5 |
os.makedirs(output_dir, exist_ok=True)
|
6 |
-
|
7 |
if fsdp_plugin.state_dict_type == StateDictType.FULL_STATE_DICT:
|
8 |
# FSDP raises error when single GPU is used with `offload_to_cpu=True` for FULL_STATE_DICT
|
9 |
# so, only enable it when num_processes>1
|
10 |
is_multi_process = accelerator.num_processes > 1
|
11 |
fsdp_plugin.state_dict_config.offload_to_cpu = is_multi_process
|
12 |
fsdp_plugin.state_dict_config.rank0_only = is_multi_process
|
13 |
-
|
14 |
with FSDP.state_dict_type(
|
15 |
model, fsdp_plugin.state_dict_type, fsdp_plugin.state_dict_config, fsdp_plugin.optim_state_dict_config
|
16 |
):
|
@@ -37,15 +33,12 @@ def save_fsdp_model(fsdp_plugin, accelerator, model, output_dir, model_index=0):
|
|
37 |
os.makedirs(ckpt_dir, exist_ok=True)
|
38 |
logger.info(f"Saving model to {ckpt_dir}")
|
39 |
state_dict = {"model": state_dict}
|
40 |
-
|
41 |
dist_cp.save_state_dict(
|
42 |
state_dict=state_dict,
|
43 |
storage_writer=dist_cp.FileSystemWriter(ckpt_dir),
|
44 |
planner=DefaultSavePlanner(),
|
45 |
)
|
46 |
logger.info(f"Model saved to {ckpt_dir}")
|
47 |
-
|
48 |
-
|
49 |
def load_fsdp_model(fsdp_plugin, accelerator, model, input_dir, model_index=0):
|
50 |
accelerator.wait_for_everyone()
|
51 |
if fsdp_plugin.state_dict_type == StateDictType.FULL_STATE_DICT:
|
@@ -97,8 +90,6 @@ def load_fsdp_model(fsdp_plugin, accelerator, model, input_dir, model_index=0):
|
|
97 |
logger.info(f"Model loaded from {ckpt_dir}")
|
98 |
load_result = model.load_state_dict(state_dict)
|
99 |
return load_result
|
100 |
-
|
101 |
-
|
102 |
def save_fsdp_optimizer(fsdp_plugin, accelerator, optimizer, model, output_dir, optimizer_index=0):
|
103 |
os.makedirs(output_dir, exist_ok=True)
|
104 |
with FSDP.state_dict_type(
|
@@ -124,8 +115,6 @@ def save_fsdp_optimizer(fsdp_plugin, accelerator, optimizer, model, output_dir,
|
|
124 |
planner=DefaultSavePlanner(),
|
125 |
)
|
126 |
logger.info(f"Optimizer state saved in {ckpt_dir}")
|
127 |
-
|
128 |
-
|
129 |
def load_fsdp_optimizer(fsdp_plugin, accelerator, optimizer, model, input_dir, optimizer_index=0):
|
130 |
accelerator.wait_for_everyone()
|
131 |
with FSDP.state_dict_type(
|
|
|
1 |
logger = get_logger(__name__)
|
|
|
|
|
2 |
def save_fsdp_model(fsdp_plugin, accelerator, model, output_dir, model_index=0):
|
3 |
os.makedirs(output_dir, exist_ok=True)
|
|
|
4 |
if fsdp_plugin.state_dict_type == StateDictType.FULL_STATE_DICT:
|
5 |
# FSDP raises error when single GPU is used with `offload_to_cpu=True` for FULL_STATE_DICT
|
6 |
# so, only enable it when num_processes>1
|
7 |
is_multi_process = accelerator.num_processes > 1
|
8 |
fsdp_plugin.state_dict_config.offload_to_cpu = is_multi_process
|
9 |
fsdp_plugin.state_dict_config.rank0_only = is_multi_process
|
|
|
10 |
with FSDP.state_dict_type(
|
11 |
model, fsdp_plugin.state_dict_type, fsdp_plugin.state_dict_config, fsdp_plugin.optim_state_dict_config
|
12 |
):
|
|
|
33 |
os.makedirs(ckpt_dir, exist_ok=True)
|
34 |
logger.info(f"Saving model to {ckpt_dir}")
|
35 |
state_dict = {"model": state_dict}
|
|
|
36 |
dist_cp.save_state_dict(
|
37 |
state_dict=state_dict,
|
38 |
storage_writer=dist_cp.FileSystemWriter(ckpt_dir),
|
39 |
planner=DefaultSavePlanner(),
|
40 |
)
|
41 |
logger.info(f"Model saved to {ckpt_dir}")
|
|
|
|
|
42 |
def load_fsdp_model(fsdp_plugin, accelerator, model, input_dir, model_index=0):
|
43 |
accelerator.wait_for_everyone()
|
44 |
if fsdp_plugin.state_dict_type == StateDictType.FULL_STATE_DICT:
|
|
|
90 |
logger.info(f"Model loaded from {ckpt_dir}")
|
91 |
load_result = model.load_state_dict(state_dict)
|
92 |
return load_result
|
|
|
|
|
93 |
def save_fsdp_optimizer(fsdp_plugin, accelerator, optimizer, model, output_dir, optimizer_index=0):
|
94 |
os.makedirs(output_dir, exist_ok=True)
|
95 |
with FSDP.state_dict_type(
|
|
|
115 |
planner=DefaultSavePlanner(),
|
116 |
)
|
117 |
logger.info(f"Optimizer state saved in {ckpt_dir}")
|
|
|
|
|
118 |
def load_fsdp_optimizer(fsdp_plugin, accelerator, optimizer, model, input_dir, optimizer_index=0):
|
119 |
accelerator.wait_for_everyone()
|
120 |
with FSDP.state_dict_type(
|
src/utils/imports.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
# Cache this result has it's a C FFI call which can be pretty time-consuming
|
2 |
_torch_distributed_available = torch.distributed.is_available()
|
3 |
-
|
4 |
-
|
5 |
def _is_package_available(pkg_name):
|
6 |
# Check we're not importing a "pkg_name" directory somewhere but the actual library by trying to grab the version
|
7 |
package_exists = importlib.util.find_spec(pkg_name) is not None
|
@@ -11,12 +9,8 @@ def _is_package_available(pkg_name):
|
|
11 |
return True
|
12 |
except importlib.metadata.PackageNotFoundError:
|
13 |
return False
|
14 |
-
|
15 |
-
|
16 |
def is_torch_distributed_available() -> bool:
|
17 |
return _torch_distributed_available
|
18 |
-
|
19 |
-
|
20 |
def is_ccl_available():
|
21 |
try:
|
22 |
pass
|
@@ -30,12 +24,8 @@ def is_ccl_available():
|
|
30 |
importlib.util.find_spec("torch_ccl") is not None
|
31 |
or importlib.util.find_spec("oneccl_bindings_for_pytorch") is not None
|
32 |
)
|
33 |
-
|
34 |
-
|
35 |
def get_ccl_version():
|
36 |
return importlib.metadata.version("oneccl_bind_pt")
|
37 |
-
|
38 |
-
|
39 |
def is_msamp_available():
|
40 |
package_exists = importlib.util.find_spec("msamp") is not None
|
41 |
if package_exists:
|
@@ -46,16 +36,10 @@ def is_msamp_available():
|
|
46 |
except importlib.metadata.PackageNotFoundError:
|
47 |
return False
|
48 |
return False
|
49 |
-
|
50 |
-
|
51 |
def is_transformer_engine_available():
|
52 |
return _is_package_available("transformer_engine")
|
53 |
-
|
54 |
-
|
55 |
def is_fp8_available():
|
56 |
return is_msamp_available() or is_transformer_engine_available()
|
57 |
-
|
58 |
-
|
59 |
def is_cuda_available():
|
60 |
"""
|
61 |
Checks if `cuda` is available via an `nvml-based` check which won't trigger the drivers and leave cuda
|
@@ -67,8 +51,6 @@ def is_cuda_available():
|
|
67 |
finally:
|
68 |
os.environ.pop("PYTORCH_NVML_BASED_CUDA_CHECK", None)
|
69 |
return available
|
70 |
-
|
71 |
-
|
72 |
@lru_cache
|
73 |
def is_tpu_available(check_device=True):
|
74 |
"Checks if `torch_xla` is installed and potentially if a TPU is in the environment"
|
@@ -84,12 +66,8 @@ def is_tpu_available(check_device=True):
|
|
84 |
except RuntimeError:
|
85 |
return False
|
86 |
return _tpu_available
|
87 |
-
|
88 |
-
|
89 |
def is_deepspeed_available():
|
90 |
return _is_package_available("deepspeed")
|
91 |
-
|
92 |
-
|
93 |
def is_bf16_available(ignore_tpu=False):
|
94 |
"Checks if bf16 is supported, optionally ignoring the TPU"
|
95 |
if is_tpu_available():
|
@@ -97,28 +75,20 @@ def is_bf16_available(ignore_tpu=False):
|
|
97 |
if is_cuda_available():
|
98 |
return torch.cuda.is_bf16_supported()
|
99 |
return True
|
100 |
-
|
101 |
-
|
102 |
def is_4bit_bnb_available():
|
103 |
package_exists = _is_package_available("bitsandbytes")
|
104 |
if package_exists:
|
105 |
bnb_version = version.parse(importlib.metadata.version("bitsandbytes"))
|
106 |
return compare_versions(bnb_version, ">=", "0.39.0")
|
107 |
return False
|
108 |
-
|
109 |
-
|
110 |
def is_8bit_bnb_available():
|
111 |
package_exists = _is_package_available("bitsandbytes")
|
112 |
if package_exists:
|
113 |
bnb_version = version.parse(importlib.metadata.version("bitsandbytes"))
|
114 |
return compare_versions(bnb_version, ">=", "0.37.2")
|
115 |
return False
|
116 |
-
|
117 |
-
|
118 |
def is_bnb_available():
|
119 |
return _is_package_available("bitsandbytes")
|
120 |
-
|
121 |
-
|
122 |
def is_megatron_lm_available():
|
123 |
if str_to_bool(os.environ.get("ACCELERATE_USE_MEGATRON_LM", "False")) == 1:
|
124 |
package_exists = importlib.util.find_spec("megatron") is not None
|
@@ -129,44 +99,26 @@ def is_megatron_lm_available():
|
|
129 |
except Exception as e:
|
130 |
warnings.warn(f"Parse Megatron version failed. Exception:{e}")
|
131 |
return False
|
132 |
-
|
133 |
-
|
134 |
def is_transformers_available():
|
135 |
return _is_package_available("transformers")
|
136 |
-
|
137 |
-
|
138 |
def is_datasets_available():
|
139 |
return _is_package_available("datasets")
|
140 |
-
|
141 |
-
|
142 |
def is_timm_available():
|
143 |
return _is_package_available("timm")
|
144 |
-
|
145 |
-
|
146 |
def is_aim_available():
|
147 |
package_exists = _is_package_available("aim")
|
148 |
if package_exists:
|
149 |
aim_version = version.parse(importlib.metadata.version("aim"))
|
150 |
return compare_versions(aim_version, "<", "4.0.0")
|
151 |
return False
|
152 |
-
|
153 |
-
|
154 |
def is_tensorboard_available():
|
155 |
return _is_package_available("tensorboard") or _is_package_available("tensorboardX")
|
156 |
-
|
157 |
-
|
158 |
def is_wandb_available():
|
159 |
return _is_package_available("wandb")
|
160 |
-
|
161 |
-
|
162 |
def is_comet_ml_available():
|
163 |
return _is_package_available("comet_ml")
|
164 |
-
|
165 |
-
|
166 |
def is_boto3_available():
|
167 |
return _is_package_available("boto3")
|
168 |
-
|
169 |
-
|
170 |
def is_rich_available():
|
171 |
if _is_package_available("rich"):
|
172 |
if "ACCELERATE_DISABLE_RICH" in os.environ:
|
@@ -176,28 +128,17 @@ def is_rich_available():
|
|
176 |
return not parse_flag_from_env("ACCELERATE_DISABLE_RICH", False)
|
177 |
return parse_flag_from_env("ACCELERATE_ENABLE_RICH", False)
|
178 |
return False
|
179 |
-
|
180 |
-
|
181 |
def is_sagemaker_available():
|
182 |
return _is_package_available("sagemaker")
|
183 |
-
|
184 |
-
|
185 |
def is_tqdm_available():
|
186 |
return _is_package_available("tqdm")
|
187 |
-
|
188 |
-
|
189 |
def is_clearml_available():
|
190 |
return _is_package_available("clearml")
|
191 |
-
|
192 |
-
|
193 |
def is_pandas_available():
|
194 |
return _is_package_available("pandas")
|
195 |
-
|
196 |
-
|
197 |
def is_mlflow_available():
|
198 |
if _is_package_available("mlflow"):
|
199 |
return True
|
200 |
-
|
201 |
if importlib.util.find_spec("mlflow") is not None:
|
202 |
try:
|
203 |
_ = importlib.metadata.metadata("mlflow-skinny")
|
@@ -205,16 +146,11 @@ def is_mlflow_available():
|
|
205 |
except importlib.metadata.PackageNotFoundError:
|
206 |
return False
|
207 |
return False
|
208 |
-
|
209 |
-
|
210 |
def is_mps_available():
|
211 |
return is_torch_version(">=", "1.12") and torch.backends.mps.is_available() and torch.backends.mps.is_built()
|
212 |
-
|
213 |
-
|
214 |
def is_ipex_available():
|
215 |
def get_major_and_minor_from_version(full_version):
|
216 |
return str(version.parse(full_version).major) + "." + str(version.parse(full_version).minor)
|
217 |
-
|
218 |
_torch_version = importlib.metadata.version("torch")
|
219 |
if importlib.util.find_spec("intel_extension_for_pytorch") is None:
|
220 |
return False
|
@@ -232,17 +168,13 @@ def is_ipex_available():
|
|
232 |
)
|
233 |
return False
|
234 |
return True
|
235 |
-
|
236 |
-
|
237 |
@lru_cache
|
238 |
def is_npu_available(check_device=False):
|
239 |
"Checks if `torch_npu` is installed and potentially if a NPU is in the environment"
|
240 |
if importlib.util.find_spec("torch") is None or importlib.util.find_spec("torch_npu") is None:
|
241 |
return False
|
242 |
-
|
243 |
import torch
|
244 |
import torch_npu # noqa: F401
|
245 |
-
|
246 |
if check_device:
|
247 |
try:
|
248 |
# Will raise a RuntimeError if no NPU is found
|
@@ -251,8 +183,6 @@ def is_npu_available(check_device=False):
|
|
251 |
except RuntimeError:
|
252 |
return False
|
253 |
return hasattr(torch, "npu") and torch.npu.is_available()
|
254 |
-
|
255 |
-
|
256 |
@lru_cache
|
257 |
def is_xpu_available(check_device=False):
|
258 |
"check if user disables it explicitly"
|
@@ -261,14 +191,11 @@ def is_xpu_available(check_device=False):
|
|
261 |
"Checks if `intel_extension_for_pytorch` is installed and potentially if a XPU is in the environment"
|
262 |
if is_ipex_available():
|
263 |
import torch
|
264 |
-
|
265 |
if is_torch_version("<=", "1.12"):
|
266 |
return False
|
267 |
else:
|
268 |
return False
|
269 |
-
|
270 |
import intel_extension_for_pytorch # noqa: F401
|
271 |
-
|
272 |
if check_device:
|
273 |
try:
|
274 |
# Will raise a RuntimeError if no XPU is found
|
@@ -277,7 +204,5 @@ def is_xpu_available(check_device=False):
|
|
277 |
except RuntimeError:
|
278 |
return False
|
279 |
return hasattr(torch, "xpu") and torch.xpu.is_available()
|
280 |
-
|
281 |
-
|
282 |
def is_dvclive_available():
|
283 |
return _is_package_available("dvclive")
|
|
|
1 |
# Cache this result has it's a C FFI call which can be pretty time-consuming
|
2 |
_torch_distributed_available = torch.distributed.is_available()
|
|
|
|
|
3 |
def _is_package_available(pkg_name):
|
4 |
# Check we're not importing a "pkg_name" directory somewhere but the actual library by trying to grab the version
|
5 |
package_exists = importlib.util.find_spec(pkg_name) is not None
|
|
|
9 |
return True
|
10 |
except importlib.metadata.PackageNotFoundError:
|
11 |
return False
|
|
|
|
|
12 |
def is_torch_distributed_available() -> bool:
|
13 |
return _torch_distributed_available
|
|
|
|
|
14 |
def is_ccl_available():
|
15 |
try:
|
16 |
pass
|
|
|
24 |
importlib.util.find_spec("torch_ccl") is not None
|
25 |
or importlib.util.find_spec("oneccl_bindings_for_pytorch") is not None
|
26 |
)
|
|
|
|
|
27 |
def get_ccl_version():
|
28 |
return importlib.metadata.version("oneccl_bind_pt")
|
|
|
|
|
29 |
def is_msamp_available():
|
30 |
package_exists = importlib.util.find_spec("msamp") is not None
|
31 |
if package_exists:
|
|
|
36 |
except importlib.metadata.PackageNotFoundError:
|
37 |
return False
|
38 |
return False
|
|
|
|
|
39 |
def is_transformer_engine_available():
|
40 |
return _is_package_available("transformer_engine")
|
|
|
|
|
41 |
def is_fp8_available():
|
42 |
return is_msamp_available() or is_transformer_engine_available()
|
|
|
|
|
43 |
def is_cuda_available():
|
44 |
"""
|
45 |
Checks if `cuda` is available via an `nvml-based` check which won't trigger the drivers and leave cuda
|
|
|
51 |
finally:
|
52 |
os.environ.pop("PYTORCH_NVML_BASED_CUDA_CHECK", None)
|
53 |
return available
|
|
|
|
|
54 |
@lru_cache
|
55 |
def is_tpu_available(check_device=True):
|
56 |
"Checks if `torch_xla` is installed and potentially if a TPU is in the environment"
|
|
|
66 |
except RuntimeError:
|
67 |
return False
|
68 |
return _tpu_available
|
|
|
|
|
69 |
def is_deepspeed_available():
|
70 |
return _is_package_available("deepspeed")
|
|
|
|
|
71 |
def is_bf16_available(ignore_tpu=False):
|
72 |
"Checks if bf16 is supported, optionally ignoring the TPU"
|
73 |
if is_tpu_available():
|
|
|
75 |
if is_cuda_available():
|
76 |
return torch.cuda.is_bf16_supported()
|
77 |
return True
|
|
|
|
|
78 |
def is_4bit_bnb_available():
|
79 |
package_exists = _is_package_available("bitsandbytes")
|
80 |
if package_exists:
|
81 |
bnb_version = version.parse(importlib.metadata.version("bitsandbytes"))
|
82 |
return compare_versions(bnb_version, ">=", "0.39.0")
|
83 |
return False
|
|
|
|
|
84 |
def is_8bit_bnb_available():
|
85 |
package_exists = _is_package_available("bitsandbytes")
|
86 |
if package_exists:
|
87 |
bnb_version = version.parse(importlib.metadata.version("bitsandbytes"))
|
88 |
return compare_versions(bnb_version, ">=", "0.37.2")
|
89 |
return False
|
|
|
|
|
90 |
def is_bnb_available():
|
91 |
return _is_package_available("bitsandbytes")
|
|
|
|
|
92 |
def is_megatron_lm_available():
|
93 |
if str_to_bool(os.environ.get("ACCELERATE_USE_MEGATRON_LM", "False")) == 1:
|
94 |
package_exists = importlib.util.find_spec("megatron") is not None
|
|
|
99 |
except Exception as e:
|
100 |
warnings.warn(f"Parse Megatron version failed. Exception:{e}")
|
101 |
return False
|
|
|
|
|
102 |
def is_transformers_available():
|
103 |
return _is_package_available("transformers")
|
|
|
|
|
104 |
def is_datasets_available():
|
105 |
return _is_package_available("datasets")
|
|
|
|
|
106 |
def is_timm_available():
|
107 |
return _is_package_available("timm")
|
|
|
|
|
108 |
def is_aim_available():
|
109 |
package_exists = _is_package_available("aim")
|
110 |
if package_exists:
|
111 |
aim_version = version.parse(importlib.metadata.version("aim"))
|
112 |
return compare_versions(aim_version, "<", "4.0.0")
|
113 |
return False
|
|
|
|
|
114 |
def is_tensorboard_available():
|
115 |
return _is_package_available("tensorboard") or _is_package_available("tensorboardX")
|
|
|
|
|
116 |
def is_wandb_available():
|
117 |
return _is_package_available("wandb")
|
|
|
|
|
118 |
def is_comet_ml_available():
|
119 |
return _is_package_available("comet_ml")
|
|
|
|
|
120 |
def is_boto3_available():
|
121 |
return _is_package_available("boto3")
|
|
|
|
|
122 |
def is_rich_available():
|
123 |
if _is_package_available("rich"):
|
124 |
if "ACCELERATE_DISABLE_RICH" in os.environ:
|
|
|
128 |
return not parse_flag_from_env("ACCELERATE_DISABLE_RICH", False)
|
129 |
return parse_flag_from_env("ACCELERATE_ENABLE_RICH", False)
|
130 |
return False
|
|
|
|
|
131 |
def is_sagemaker_available():
|
132 |
return _is_package_available("sagemaker")
|
|
|
|
|
133 |
def is_tqdm_available():
|
134 |
return _is_package_available("tqdm")
|
|
|
|
|
135 |
def is_clearml_available():
|
136 |
return _is_package_available("clearml")
|
|
|
|
|
137 |
def is_pandas_available():
|
138 |
return _is_package_available("pandas")
|
|
|
|
|
139 |
def is_mlflow_available():
|
140 |
if _is_package_available("mlflow"):
|
141 |
return True
|
|
|
142 |
if importlib.util.find_spec("mlflow") is not None:
|
143 |
try:
|
144 |
_ = importlib.metadata.metadata("mlflow-skinny")
|
|
|
146 |
except importlib.metadata.PackageNotFoundError:
|
147 |
return False
|
148 |
return False
|
|
|
|
|
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")
|
155 |
if importlib.util.find_spec("intel_extension_for_pytorch") is None:
|
156 |
return False
|
|
|
168 |
)
|
169 |
return False
|
170 |
return True
|
|
|
|
|
171 |
@lru_cache
|
172 |
def is_npu_available(check_device=False):
|
173 |
"Checks if `torch_npu` is installed and potentially if a NPU is in the environment"
|
174 |
if importlib.util.find_spec("torch") is None or importlib.util.find_spec("torch_npu") is None:
|
175 |
return False
|
|
|
176 |
import torch
|
177 |
import torch_npu # noqa: F401
|
|
|
178 |
if check_device:
|
179 |
try:
|
180 |
# Will raise a RuntimeError if no NPU is found
|
|
|
183 |
except RuntimeError:
|
184 |
return False
|
185 |
return hasattr(torch, "npu") and torch.npu.is_available()
|
|
|
|
|
186 |
@lru_cache
|
187 |
def is_xpu_available(check_device=False):
|
188 |
"check if user disables it explicitly"
|
|
|
191 |
"Checks if `intel_extension_for_pytorch` is installed and potentially if a XPU is in the environment"
|
192 |
if is_ipex_available():
|
193 |
import torch
|
|
|
194 |
if is_torch_version("<=", "1.12"):
|
195 |
return False
|
196 |
else:
|
197 |
return False
|
|
|
198 |
import intel_extension_for_pytorch # noqa: F401
|
|
|
199 |
if check_device:
|
200 |
try:
|
201 |
# Will raise a RuntimeError if no XPU is found
|
|
|
204 |
except RuntimeError:
|
205 |
return False
|
206 |
return hasattr(torch, "xpu") and torch.xpu.is_available()
|
|
|
|
|
207 |
def is_dvclive_available():
|
208 |
return _is_package_available("dvclive")
|
src/utils/launch.py
CHANGED
@@ -7,8 +7,6 @@ def _filter_args(args, parser, default_args=[]):
|
|
7 |
if key in vars(new_args).keys():
|
8 |
setattr(new_args, key, value)
|
9 |
return new_args
|
10 |
-
|
11 |
-
|
12 |
def prepare_simple_launcher_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict[str, str]]:
|
13 |
"""
|
14 |
Prepares and returns the command list and an environment with the correct simple launcher environment variables.
|
@@ -22,7 +20,6 @@ def prepare_simple_launcher_cmd_env(args: argparse.Namespace) -> Tuple[List[str]
|
|
22 |
cmd.append("-m")
|
23 |
cmd.append(args.training_script)
|
24 |
cmd.extend(args.training_script_args)
|
25 |
-
|
26 |
current_env = os.environ.copy()
|
27 |
current_env["ACCELERATE_USE_CPU"] = str(args.cpu or args.use_cpu)
|
28 |
if args.debug:
|
@@ -40,16 +37,13 @@ def prepare_simple_launcher_cmd_env(args: argparse.Namespace) -> Tuple[List[str]
|
|
40 |
elif args.num_processes > 1:
|
41 |
current_env["MASTER_ADDR"] = args.main_process_ip if args.main_process_ip is not None else "127.0.0.1"
|
42 |
current_env["MASTER_PORT"] = str(args.main_process_port) if args.main_process_port is not None else "29500"
|
43 |
-
|
44 |
try:
|
45 |
mixed_precision = PrecisionType(args.mixed_precision.lower())
|
46 |
except ValueError:
|
47 |
raise ValueError(
|
48 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
49 |
)
|
50 |
-
|
51 |
current_env["ACCELERATE_MIXED_PRECISION"] = str(mixed_precision)
|
52 |
-
|
53 |
try:
|
54 |
dynamo_backend = DynamoBackend(args.dynamo_backend.upper())
|
55 |
except ValueError:
|
@@ -60,14 +54,11 @@ def prepare_simple_launcher_cmd_env(args: argparse.Namespace) -> Tuple[List[str]
|
|
60 |
current_env["ACCELERATE_DYNAMO_MODE"] = args.dynamo_mode
|
61 |
current_env["ACCELERATE_DYNAMO_USE_FULLGRAPH"] = str(args.dynamo_use_fullgraph)
|
62 |
current_env["ACCELERATE_DYNAMO_USE_DYNAMIC"] = str(args.dynamo_use_dynamic)
|
63 |
-
|
64 |
current_env["OMP_NUM_THREADS"] = str(args.num_cpu_threads_per_process)
|
65 |
if is_ipex_available():
|
66 |
current_env["ACCELERATE_USE_IPEX"] = str(args.ipex).lower()
|
67 |
current_env["ACCELERATE_USE_XPU"] = str(args.use_xpu).lower()
|
68 |
return cmd, current_env
|
69 |
-
|
70 |
-
|
71 |
def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
72 |
"""
|
73 |
Prepares and returns an environment with the correct multi-GPU environment variables.
|
@@ -89,10 +80,8 @@ def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
|
89 |
setattr(args, "nproc_per_node", str(num_processes))
|
90 |
if main_process_port is not None:
|
91 |
setattr(args, "master_port", str(main_process_port))
|
92 |
-
|
93 |
if main_process_port is None:
|
94 |
main_process_port = 29500
|
95 |
-
|
96 |
# only need to check port availability in main process, in case we have to start multiple launchers on the same machine
|
97 |
# for some reasons like splitting log files.
|
98 |
need_port_check = num_machines <= 1 or int(args.machine_rank) == 0
|
@@ -102,14 +91,12 @@ def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
|
102 |
"Please specify a different port (such as using the `----main_process_port` flag or specifying a different `main_process_port` in your config file)"
|
103 |
" and rerun your script. To automatically use the next open port (on a single node), you can set this to `0`."
|
104 |
)
|
105 |
-
|
106 |
if args.module and args.no_python:
|
107 |
raise ValueError("--module and --no_python cannot be used together")
|
108 |
elif args.module:
|
109 |
setattr(args, "module", True)
|
110 |
elif args.no_python:
|
111 |
setattr(args, "no_python", True)
|
112 |
-
|
113 |
current_env = os.environ.copy()
|
114 |
if args.debug:
|
115 |
current_env["ACCELERATE_DEBUG_MODE"] = "true"
|
@@ -126,9 +113,7 @@ def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
|
126 |
mixed_precision = PrecisionType(mixed_precision)
|
127 |
except ValueError:
|
128 |
raise ValueError(f"Unknown mixed_precision mode: {mixed_precision}. Choose between {PrecisionType.list()}.")
|
129 |
-
|
130 |
current_env["ACCELERATE_MIXED_PRECISION"] = str(mixed_precision)
|
131 |
-
|
132 |
try:
|
133 |
dynamo_backend = DynamoBackend(args.dynamo_backend.upper())
|
134 |
except ValueError:
|
@@ -139,12 +124,10 @@ def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
|
139 |
current_env["ACCELERATE_DYNAMO_MODE"] = args.dynamo_mode
|
140 |
current_env["ACCELERATE_DYNAMO_USE_FULLGRAPH"] = str(args.dynamo_use_fullgraph)
|
141 |
current_env["ACCELERATE_DYNAMO_USE_DYNAMIC"] = str(args.dynamo_use_dynamic)
|
142 |
-
|
143 |
if args.use_fsdp:
|
144 |
current_env["ACCELERATE_USE_FSDP"] = "true"
|
145 |
if args.fsdp_cpu_ram_efficient_loading and not args.fsdp_sync_module_states:
|
146 |
raise ValueError("When using `--fsdp_cpu_ram_efficient_loading` set `--fsdp_sync_module_states` to `True`")
|
147 |
-
|
148 |
current_env["FSDP_SHARDING_STRATEGY"] = str(args.fsdp_sharding_strategy)
|
149 |
current_env["FSDP_OFFLOAD_PARAMS"] = str(args.fsdp_offload_params).lower()
|
150 |
current_env["FSDP_MIN_NUM_PARAMS"] = str(args.fsdp_min_num_params)
|
@@ -167,7 +150,6 @@ def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
|
167 |
current_env["FSDP_USE_ORIG_PARAMS"] = str(args.fsdp_use_orig_params).lower()
|
168 |
current_env["FSDP_CPU_RAM_EFFICIENT_LOADING"] = str(args.fsdp_cpu_ram_efficient_loading).lower()
|
169 |
current_env["FSDP_SYNC_MODULE_STATES"] = str(args.fsdp_sync_module_states).lower()
|
170 |
-
|
171 |
if args.use_megatron_lm:
|
172 |
prefix = "MEGATRON_LM_"
|
173 |
current_env["ACCELERATE_USE_MEGATRON_LM"] = "true"
|
@@ -182,11 +164,8 @@ def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
|
182 |
current_env[prefix + "RECOMPUTE_ACTIVATIONS"] = str(args.megatron_lm_recompute_activations)
|
183 |
if args.megatron_lm_use_distributed_optimizer is not None:
|
184 |
current_env[prefix + "USE_DISTRIBUTED_OPTIMIZER"] = str(args.megatron_lm_use_distributed_optimizer)
|
185 |
-
|
186 |
current_env["OMP_NUM_THREADS"] = str(args.num_cpu_threads_per_process)
|
187 |
return current_env
|
188 |
-
|
189 |
-
|
190 |
def prepare_deepspeed_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict[str, str]]:
|
191 |
"""
|
192 |
Prepares and returns the command list and an environment with the correct DeepSpeed environment variables.
|
@@ -196,12 +175,10 @@ def prepare_deepspeed_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict
|
|
196 |
main_process_ip = getattr(args, "main_process_ip")
|
197 |
main_process_port = getattr(args, "main_process_port")
|
198 |
cmd = None
|
199 |
-
|
200 |
# make sure launcher is not None
|
201 |
if args.deepspeed_multinode_launcher is None:
|
202 |
# set to default pdsh
|
203 |
setattr(args, "deepspeed_multinode_launcher", DEEPSPEED_MULTINODE_LAUNCHERS[0])
|
204 |
-
|
205 |
if num_machines > 1 and args.deepspeed_multinode_launcher != DEEPSPEED_MULTINODE_LAUNCHERS[1]:
|
206 |
cmd = ["deepspeed", "--no_local_rank"]
|
207 |
cmd.extend(["--hostfile", str(args.deepspeed_hostfile), "--launcher", str(args.deepspeed_multinode_launcher)])
|
@@ -243,10 +220,8 @@ def prepare_deepspeed_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict
|
|
243 |
setattr(args, "nproc_per_node", str(num_processes))
|
244 |
if main_process_port is not None:
|
245 |
setattr(args, "master_port", str(main_process_port))
|
246 |
-
|
247 |
if main_process_port is None:
|
248 |
main_process_port = 29500
|
249 |
-
|
250 |
# only need to check port availability in main process, in case we have to start multiple launchers on the same machine
|
251 |
# for some reasons like splitting log files.
|
252 |
need_port_check = num_machines <= 1 or int(args.machine_rank) == 0
|
@@ -256,14 +231,12 @@ def prepare_deepspeed_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict
|
|
256 |
"Please specify a different port (such as using the `----main_process_port` flag or specifying a different `main_process_port` in your config file)"
|
257 |
" and rerun your script. To automatically use the next open port (on a single node), you can set this to `0`."
|
258 |
)
|
259 |
-
|
260 |
if args.module and args.no_python:
|
261 |
raise ValueError("--module and --no_python cannot be used together")
|
262 |
elif args.module:
|
263 |
setattr(args, "module", True)
|
264 |
elif args.no_python:
|
265 |
setattr(args, "no_python", True)
|
266 |
-
|
267 |
current_env = os.environ.copy()
|
268 |
if args.debug:
|
269 |
current_env["ACCELERATE_DEBUG_MODE"] = "true"
|
@@ -281,7 +254,6 @@ def prepare_deepspeed_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict
|
|
281 |
raise ValueError(
|
282 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
283 |
)
|
284 |
-
|
285 |
current_env["PYTHONPATH"] = env_var_path_add("PYTHONPATH", os.path.abspath("."))
|
286 |
current_env["ACCELERATE_MIXED_PRECISION"] = str(mixed_precision)
|
287 |
current_env["ACCELERATE_CONFIG_DS_FIELDS"] = str(args.deepspeed_fields_from_accelerate_config).lower()
|
@@ -303,8 +275,6 @@ def prepare_deepspeed_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict
|
|
303 |
if args.deepspeed_config_file is not None:
|
304 |
current_env["ACCELERATE_DEEPSPEED_CONFIG_FILE"] = str(args.deepspeed_config_file)
|
305 |
return cmd, current_env
|
306 |
-
|
307 |
-
|
308 |
def prepare_tpu(
|
309 |
args: argparse.Namespace, current_env: Dict[str, str], pod: bool = False
|
310 |
) -> Tuple[argparse.Namespace, Dict[str, str]]:
|
@@ -323,23 +293,18 @@ def prepare_tpu(
|
|
323 |
args.vm = args.tpu_vm
|
324 |
args.tpu = args.tpu_name
|
325 |
return args, current_env
|
326 |
-
|
327 |
-
|
328 |
def _convert_nargs_to_dict(nargs: List[str]) -> Dict[str, str]:
|
329 |
if len(nargs) < 0:
|
330 |
return {}
|
331 |
# helper function to infer type for argsparser
|
332 |
-
|
333 |
def _infer_type(s):
|
334 |
try:
|
335 |
s = float(s)
|
336 |
-
|
337 |
if s // 1 == s:
|
338 |
return int(s)
|
339 |
return s
|
340 |
except ValueError:
|
341 |
return s
|
342 |
-
|
343 |
parser = argparse.ArgumentParser()
|
344 |
_, unknown = parser.parse_known_args(nargs)
|
345 |
for index, argument in enumerate(unknown):
|
@@ -360,20 +325,16 @@ def _convert_nargs_to_dict(nargs: List[str]) -> Dict[str, str]:
|
|
360 |
parser.add_argument(argument, type=_infer_type)
|
361 |
else:
|
362 |
parser.add_argument(argument, action=action)
|
363 |
-
|
364 |
return {
|
365 |
key: (literal_eval(value) if value in ("True", "False") else value)
|
366 |
for key, value in parser.parse_args(nargs).__dict__.items()
|
367 |
}
|
368 |
-
|
369 |
-
|
370 |
def prepare_sagemager_args_inputs(
|
371 |
sagemaker_config: SageMakerConfig, args: argparse.Namespace
|
372 |
) -> Tuple[argparse.Namespace, Dict[str, Any]]:
|
373 |
# configure environment
|
374 |
print("Configuring Amazon SageMaker environment")
|
375 |
os.environ["AWS_DEFAULT_REGION"] = sagemaker_config.region
|
376 |
-
|
377 |
# configure credentials
|
378 |
if sagemaker_config.profile is not None:
|
379 |
os.environ["AWS_PROFILE"] = sagemaker_config.profile
|
@@ -384,7 +345,6 @@ def prepare_sagemager_args_inputs(
|
|
384 |
raise EnvironmentError(
|
385 |
"You need to provide an aws_access_key_id and aws_secret_access_key when not using aws_profile"
|
386 |
)
|
387 |
-
|
388 |
# extract needed arguments
|
389 |
source_dir = os.path.dirname(args.training_script)
|
390 |
if not source_dir: # checks if string is empty
|
@@ -392,24 +352,20 @@ def prepare_sagemager_args_inputs(
|
|
392 |
entry_point = os.path.basename(args.training_script)
|
393 |
if not entry_point.endswith(".py"):
|
394 |
raise ValueError(f'Your training script should be a python script and not "{entry_point}"')
|
395 |
-
|
396 |
print("Converting Arguments to Hyperparameters")
|
397 |
hyperparameters = _convert_nargs_to_dict(args.training_script_args)
|
398 |
-
|
399 |
try:
|
400 |
mixed_precision = PrecisionType(args.mixed_precision.lower())
|
401 |
except ValueError:
|
402 |
raise ValueError(
|
403 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
404 |
)
|
405 |
-
|
406 |
try:
|
407 |
dynamo_backend = DynamoBackend(args.dynamo_backend.upper())
|
408 |
except ValueError:
|
409 |
raise ValueError(
|
410 |
f"Unknown dynamo backend: {args.dynamo_backend.upper()}. Choose between {DynamoBackend.list()}."
|
411 |
)
|
412 |
-
|
413 |
# Environment variables to be set for use during training job
|
414 |
environment = {
|
415 |
"ACCELERATE_USE_SAGEMAKER": "true",
|
@@ -424,7 +380,6 @@ def prepare_sagemager_args_inputs(
|
|
424 |
distribution = None
|
425 |
if sagemaker_config.distributed_type == SageMakerDistributedType.DATA_PARALLEL:
|
426 |
distribution = {"smdistributed": {"dataparallel": {"enabled": True}}}
|
427 |
-
|
428 |
# configure sagemaker inputs
|
429 |
sagemaker_inputs = None
|
430 |
if sagemaker_config.sagemaker_inputs_file is not None:
|
@@ -437,7 +392,6 @@ def prepare_sagemager_args_inputs(
|
|
437 |
l = line.split("\t")
|
438 |
sagemaker_inputs[l[0]] = l[1].strip()
|
439 |
print(f"Loaded SageMaker Inputs: {sagemaker_inputs}")
|
440 |
-
|
441 |
# configure sagemaker metrics
|
442 |
sagemaker_metrics = None
|
443 |
if sagemaker_config.sagemaker_metrics_file is not None:
|
@@ -454,7 +408,6 @@ def prepare_sagemager_args_inputs(
|
|
454 |
}
|
455 |
sagemaker_metrics.append(metric_dict)
|
456 |
print(f"Loaded SageMaker Metrics: {sagemaker_metrics}")
|
457 |
-
|
458 |
# configure session
|
459 |
print("Creating Estimator")
|
460 |
args = {
|
@@ -474,12 +427,9 @@ def prepare_sagemager_args_inputs(
|
|
474 |
"environment": environment,
|
475 |
"metric_definitions": sagemaker_metrics,
|
476 |
}
|
477 |
-
|
478 |
if sagemaker_config.additional_args is not None:
|
479 |
args = merge_dicts(sagemaker_config.additional_args, args)
|
480 |
return args, sagemaker_inputs
|
481 |
-
|
482 |
-
|
483 |
def env_var_path_add(env_var_name, path_to_add):
|
484 |
"""
|
485 |
Extends a path-based environment variable's value with a new path and returns the updated value. It's up to the
|
@@ -488,12 +438,9 @@ def env_var_path_add(env_var_name, path_to_add):
|
|
488 |
paths = [p for p in os.environ.get(env_var_name, "").split(":") if len(p) > 0]
|
489 |
paths.append(str(path_to_add))
|
490 |
return ":".join(paths)
|
491 |
-
|
492 |
-
|
493 |
class PrepareForLaunch:
|
494 |
"""
|
495 |
Prepare a function that will launched in a distributed setup.
|
496 |
-
|
497 |
Args:
|
498 |
launcher (`Callable`):
|
499 |
The function to launch.
|
@@ -506,7 +453,6 @@ class PrepareForLaunch:
|
|
506 |
self.launcher = launcher
|
507 |
self.distributed_type = DistributedType(distributed_type)
|
508 |
self.debug = debug
|
509 |
-
|
510 |
def __call__(self, index, *args):
|
511 |
if self.debug:
|
512 |
world_size = int(os.environ.get("WORLD_SIZE"))
|
@@ -528,6 +474,5 @@ class PrepareForLaunch:
|
|
528 |
nproc = int(os.environ.get("NPROC", 1))
|
529 |
node_rank = int(os.environ.get("NODE_RANK", 0))
|
530 |
os.environ["RANK"] = str(nproc * node_rank + index)
|
531 |
-
|
532 |
os.environ["FORK_LAUNCHED"] = str(1)
|
533 |
self.launcher(*args)
|
|
|
7 |
if key in vars(new_args).keys():
|
8 |
setattr(new_args, key, value)
|
9 |
return new_args
|
|
|
|
|
10 |
def prepare_simple_launcher_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict[str, str]]:
|
11 |
"""
|
12 |
Prepares and returns the command list and an environment with the correct simple launcher environment variables.
|
|
|
20 |
cmd.append("-m")
|
21 |
cmd.append(args.training_script)
|
22 |
cmd.extend(args.training_script_args)
|
|
|
23 |
current_env = os.environ.copy()
|
24 |
current_env["ACCELERATE_USE_CPU"] = str(args.cpu or args.use_cpu)
|
25 |
if args.debug:
|
|
|
37 |
elif args.num_processes > 1:
|
38 |
current_env["MASTER_ADDR"] = args.main_process_ip if args.main_process_ip is not None else "127.0.0.1"
|
39 |
current_env["MASTER_PORT"] = str(args.main_process_port) if args.main_process_port is not None else "29500"
|
|
|
40 |
try:
|
41 |
mixed_precision = PrecisionType(args.mixed_precision.lower())
|
42 |
except ValueError:
|
43 |
raise ValueError(
|
44 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
45 |
)
|
|
|
46 |
current_env["ACCELERATE_MIXED_PRECISION"] = str(mixed_precision)
|
|
|
47 |
try:
|
48 |
dynamo_backend = DynamoBackend(args.dynamo_backend.upper())
|
49 |
except ValueError:
|
|
|
54 |
current_env["ACCELERATE_DYNAMO_MODE"] = args.dynamo_mode
|
55 |
current_env["ACCELERATE_DYNAMO_USE_FULLGRAPH"] = str(args.dynamo_use_fullgraph)
|
56 |
current_env["ACCELERATE_DYNAMO_USE_DYNAMIC"] = str(args.dynamo_use_dynamic)
|
|
|
57 |
current_env["OMP_NUM_THREADS"] = str(args.num_cpu_threads_per_process)
|
58 |
if is_ipex_available():
|
59 |
current_env["ACCELERATE_USE_IPEX"] = str(args.ipex).lower()
|
60 |
current_env["ACCELERATE_USE_XPU"] = str(args.use_xpu).lower()
|
61 |
return cmd, current_env
|
|
|
|
|
62 |
def prepare_multi_gpu_env(args: argparse.Namespace) -> Dict[str, str]:
|
63 |
"""
|
64 |
Prepares and returns an environment with the correct multi-GPU environment variables.
|
|
|
80 |
setattr(args, "nproc_per_node", str(num_processes))
|
81 |
if main_process_port is not None:
|
82 |
setattr(args, "master_port", str(main_process_port))
|
|
|
83 |
if main_process_port is None:
|
84 |
main_process_port = 29500
|
|
|
85 |
# only need to check port availability in main process, in case we have to start multiple launchers on the same machine
|
86 |
# for some reasons like splitting log files.
|
87 |
need_port_check = num_machines <= 1 or int(args.machine_rank) == 0
|
|
|
91 |
"Please specify a different port (such as using the `----main_process_port` flag or specifying a different `main_process_port` in your config file)"
|
92 |
" and rerun your script. To automatically use the next open port (on a single node), you can set this to `0`."
|
93 |
)
|
|
|
94 |
if args.module and args.no_python:
|
95 |
raise ValueError("--module and --no_python cannot be used together")
|
96 |
elif args.module:
|
97 |
setattr(args, "module", True)
|
98 |
elif args.no_python:
|
99 |
setattr(args, "no_python", True)
|
|
|
100 |
current_env = os.environ.copy()
|
101 |
if args.debug:
|
102 |
current_env["ACCELERATE_DEBUG_MODE"] = "true"
|
|
|
113 |
mixed_precision = PrecisionType(mixed_precision)
|
114 |
except ValueError:
|
115 |
raise ValueError(f"Unknown mixed_precision mode: {mixed_precision}. Choose between {PrecisionType.list()}.")
|
|
|
116 |
current_env["ACCELERATE_MIXED_PRECISION"] = str(mixed_precision)
|
|
|
117 |
try:
|
118 |
dynamo_backend = DynamoBackend(args.dynamo_backend.upper())
|
119 |
except ValueError:
|
|
|
124 |
current_env["ACCELERATE_DYNAMO_MODE"] = args.dynamo_mode
|
125 |
current_env["ACCELERATE_DYNAMO_USE_FULLGRAPH"] = str(args.dynamo_use_fullgraph)
|
126 |
current_env["ACCELERATE_DYNAMO_USE_DYNAMIC"] = str(args.dynamo_use_dynamic)
|
|
|
127 |
if args.use_fsdp:
|
128 |
current_env["ACCELERATE_USE_FSDP"] = "true"
|
129 |
if args.fsdp_cpu_ram_efficient_loading and not args.fsdp_sync_module_states:
|
130 |
raise ValueError("When using `--fsdp_cpu_ram_efficient_loading` set `--fsdp_sync_module_states` to `True`")
|
|
|
131 |
current_env["FSDP_SHARDING_STRATEGY"] = str(args.fsdp_sharding_strategy)
|
132 |
current_env["FSDP_OFFLOAD_PARAMS"] = str(args.fsdp_offload_params).lower()
|
133 |
current_env["FSDP_MIN_NUM_PARAMS"] = str(args.fsdp_min_num_params)
|
|
|
150 |
current_env["FSDP_USE_ORIG_PARAMS"] = str(args.fsdp_use_orig_params).lower()
|
151 |
current_env["FSDP_CPU_RAM_EFFICIENT_LOADING"] = str(args.fsdp_cpu_ram_efficient_loading).lower()
|
152 |
current_env["FSDP_SYNC_MODULE_STATES"] = str(args.fsdp_sync_module_states).lower()
|
|
|
153 |
if args.use_megatron_lm:
|
154 |
prefix = "MEGATRON_LM_"
|
155 |
current_env["ACCELERATE_USE_MEGATRON_LM"] = "true"
|
|
|
164 |
current_env[prefix + "RECOMPUTE_ACTIVATIONS"] = str(args.megatron_lm_recompute_activations)
|
165 |
if args.megatron_lm_use_distributed_optimizer is not None:
|
166 |
current_env[prefix + "USE_DISTRIBUTED_OPTIMIZER"] = str(args.megatron_lm_use_distributed_optimizer)
|
|
|
167 |
current_env["OMP_NUM_THREADS"] = str(args.num_cpu_threads_per_process)
|
168 |
return current_env
|
|
|
|
|
169 |
def prepare_deepspeed_cmd_env(args: argparse.Namespace) -> Tuple[List[str], Dict[str, str]]:
|
170 |
"""
|
171 |
Prepares and returns the command list and an environment with the correct DeepSpeed environment variables.
|
|
|
175 |
main_process_ip = getattr(args, "main_process_ip")
|
176 |
main_process_port = getattr(args, "main_process_port")
|
177 |
cmd = None
|
|
|
178 |
# make sure launcher is not None
|
179 |
if args.deepspeed_multinode_launcher is None:
|
180 |
# set to default pdsh
|
181 |
setattr(args, "deepspeed_multinode_launcher", DEEPSPEED_MULTINODE_LAUNCHERS[0])
|
|
|
182 |
if num_machines > 1 and args.deepspeed_multinode_launcher != DEEPSPEED_MULTINODE_LAUNCHERS[1]:
|
183 |
cmd = ["deepspeed", "--no_local_rank"]
|
184 |
cmd.extend(["--hostfile", str(args.deepspeed_hostfile), "--launcher", str(args.deepspeed_multinode_launcher)])
|
|
|
220 |
setattr(args, "nproc_per_node", str(num_processes))
|
221 |
if main_process_port is not None:
|
222 |
setattr(args, "master_port", str(main_process_port))
|
|
|
223 |
if main_process_port is None:
|
224 |
main_process_port = 29500
|
|
|
225 |
# only need to check port availability in main process, in case we have to start multiple launchers on the same machine
|
226 |
# for some reasons like splitting log files.
|
227 |
need_port_check = num_machines <= 1 or int(args.machine_rank) == 0
|
|
|
231 |
"Please specify a different port (such as using the `----main_process_port` flag or specifying a different `main_process_port` in your config file)"
|
232 |
" and rerun your script. To automatically use the next open port (on a single node), you can set this to `0`."
|
233 |
)
|
|
|
234 |
if args.module and args.no_python:
|
235 |
raise ValueError("--module and --no_python cannot be used together")
|
236 |
elif args.module:
|
237 |
setattr(args, "module", True)
|
238 |
elif args.no_python:
|
239 |
setattr(args, "no_python", True)
|
|
|
240 |
current_env = os.environ.copy()
|
241 |
if args.debug:
|
242 |
current_env["ACCELERATE_DEBUG_MODE"] = "true"
|
|
|
254 |
raise ValueError(
|
255 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
256 |
)
|
|
|
257 |
current_env["PYTHONPATH"] = env_var_path_add("PYTHONPATH", os.path.abspath("."))
|
258 |
current_env["ACCELERATE_MIXED_PRECISION"] = str(mixed_precision)
|
259 |
current_env["ACCELERATE_CONFIG_DS_FIELDS"] = str(args.deepspeed_fields_from_accelerate_config).lower()
|
|
|
275 |
if args.deepspeed_config_file is not None:
|
276 |
current_env["ACCELERATE_DEEPSPEED_CONFIG_FILE"] = str(args.deepspeed_config_file)
|
277 |
return cmd, current_env
|
|
|
|
|
278 |
def prepare_tpu(
|
279 |
args: argparse.Namespace, current_env: Dict[str, str], pod: bool = False
|
280 |
) -> Tuple[argparse.Namespace, Dict[str, str]]:
|
|
|
293 |
args.vm = args.tpu_vm
|
294 |
args.tpu = args.tpu_name
|
295 |
return args, current_env
|
|
|
|
|
296 |
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)
|
|
|
303 |
if s // 1 == s:
|
304 |
return int(s)
|
305 |
return s
|
306 |
except ValueError:
|
307 |
return s
|
|
|
308 |
parser = argparse.ArgumentParser()
|
309 |
_, unknown = parser.parse_known_args(nargs)
|
310 |
for index, argument in enumerate(unknown):
|
|
|
325 |
parser.add_argument(argument, type=_infer_type)
|
326 |
else:
|
327 |
parser.add_argument(argument, action=action)
|
|
|
328 |
return {
|
329 |
key: (literal_eval(value) if value in ("True", "False") else value)
|
330 |
for key, value in parser.parse_args(nargs).__dict__.items()
|
331 |
}
|
|
|
|
|
332 |
def prepare_sagemager_args_inputs(
|
333 |
sagemaker_config: SageMakerConfig, args: argparse.Namespace
|
334 |
) -> Tuple[argparse.Namespace, Dict[str, Any]]:
|
335 |
# configure environment
|
336 |
print("Configuring Amazon SageMaker environment")
|
337 |
os.environ["AWS_DEFAULT_REGION"] = sagemaker_config.region
|
|
|
338 |
# configure credentials
|
339 |
if sagemaker_config.profile is not None:
|
340 |
os.environ["AWS_PROFILE"] = sagemaker_config.profile
|
|
|
345 |
raise EnvironmentError(
|
346 |
"You need to provide an aws_access_key_id and aws_secret_access_key when not using aws_profile"
|
347 |
)
|
|
|
348 |
# extract needed arguments
|
349 |
source_dir = os.path.dirname(args.training_script)
|
350 |
if not source_dir: # checks if string is empty
|
|
|
352 |
entry_point = os.path.basename(args.training_script)
|
353 |
if not entry_point.endswith(".py"):
|
354 |
raise ValueError(f'Your training script should be a python script and not "{entry_point}"')
|
|
|
355 |
print("Converting Arguments to Hyperparameters")
|
356 |
hyperparameters = _convert_nargs_to_dict(args.training_script_args)
|
|
|
357 |
try:
|
358 |
mixed_precision = PrecisionType(args.mixed_precision.lower())
|
359 |
except ValueError:
|
360 |
raise ValueError(
|
361 |
f"Unknown mixed_precision mode: {args.mixed_precision.lower()}. Choose between {PrecisionType.list()}."
|
362 |
)
|
|
|
363 |
try:
|
364 |
dynamo_backend = DynamoBackend(args.dynamo_backend.upper())
|
365 |
except ValueError:
|
366 |
raise ValueError(
|
367 |
f"Unknown dynamo backend: {args.dynamo_backend.upper()}. Choose between {DynamoBackend.list()}."
|
368 |
)
|
|
|
369 |
# Environment variables to be set for use during training job
|
370 |
environment = {
|
371 |
"ACCELERATE_USE_SAGEMAKER": "true",
|
|
|
380 |
distribution = None
|
381 |
if sagemaker_config.distributed_type == SageMakerDistributedType.DATA_PARALLEL:
|
382 |
distribution = {"smdistributed": {"dataparallel": {"enabled": True}}}
|
|
|
383 |
# configure sagemaker inputs
|
384 |
sagemaker_inputs = None
|
385 |
if sagemaker_config.sagemaker_inputs_file is not None:
|
|
|
392 |
l = line.split("\t")
|
393 |
sagemaker_inputs[l[0]] = l[1].strip()
|
394 |
print(f"Loaded SageMaker Inputs: {sagemaker_inputs}")
|
|
|
395 |
# configure sagemaker metrics
|
396 |
sagemaker_metrics = None
|
397 |
if sagemaker_config.sagemaker_metrics_file is not None:
|
|
|
408 |
}
|
409 |
sagemaker_metrics.append(metric_dict)
|
410 |
print(f"Loaded SageMaker Metrics: {sagemaker_metrics}")
|
|
|
411 |
# configure session
|
412 |
print("Creating Estimator")
|
413 |
args = {
|
|
|
427 |
"environment": environment,
|
428 |
"metric_definitions": sagemaker_metrics,
|
429 |
}
|
|
|
430 |
if sagemaker_config.additional_args is not None:
|
431 |
args = merge_dicts(sagemaker_config.additional_args, args)
|
432 |
return args, sagemaker_inputs
|
|
|
|
|
433 |
def env_var_path_add(env_var_name, path_to_add):
|
434 |
"""
|
435 |
Extends a path-based environment variable's value with a new path and returns the updated value. It's up to the
|
|
|
438 |
paths = [p for p in os.environ.get(env_var_name, "").split(":") if len(p) > 0]
|
439 |
paths.append(str(path_to_add))
|
440 |
return ":".join(paths)
|
|
|
|
|
441 |
class PrepareForLaunch:
|
442 |
"""
|
443 |
Prepare a function that will launched in a distributed setup.
|
|
|
444 |
Args:
|
445 |
launcher (`Callable`):
|
446 |
The function to launch.
|
|
|
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"))
|
|
|
474 |
nproc = int(os.environ.get("NPROC", 1))
|
475 |
node_rank = int(os.environ.get("NODE_RANK", 0))
|
476 |
os.environ["RANK"] = str(nproc * node_rank + index)
|
|
|
477 |
os.environ["FORK_LAUNCHED"] = str(1)
|
478 |
self.launcher(*args)
|
src/utils/megatron_lm.py
CHANGED
@@ -37,8 +37,6 @@ def model_provider_func(pre_process=True, post_process=True, add_encoder=True, a
|
|
37 |
else:
|
38 |
raise ValueError(f"Unsupported model type: {args.model_type_name}")
|
39 |
return model
|
40 |
-
|
41 |
-
|
42 |
def prepare_model(accelerator):
|
43 |
accelerator.print("Preparing model")
|
44 |
args = get_args()
|
@@ -58,13 +56,10 @@ def prepare_model(accelerator):
|
|
58 |
args.pipeline_model_parallel_split_rank = args.pipeline_model_parallel_size // 2
|
59 |
model = get_model(model_provider_func, model_type)
|
60 |
return model
|
61 |
-
|
62 |
-
|
63 |
# dataloader utilities
|
64 |
class MegatronLMDummyDataLoader:
|
65 |
"""
|
66 |
Dummy dataloader presents model parameters or param groups, this is primarily used to follow conventional training
|
67 |
-
|
68 |
Args:
|
69 |
**dataset_kwargs: Megatron data arguments.
|
70 |
"""
|
@@ -76,12 +71,10 @@ class MegatronLMDummyDataLoader:
|
|
76 |
self.dataset_args = vars(data_args[0])
|
77 |
self.dataset_args.update(dataset_kwargs)
|
78 |
self.dataset_args["megatron_dataset_flag"] = True
|
79 |
-
|
80 |
def set_megatron_data_args(self):
|
81 |
args = get_args()
|
82 |
for key, value in self.dataset_args.items():
|
83 |
setattr(args, key, value)
|
84 |
-
|
85 |
def get_train_valid_test_datasets_provider(self):
|
86 |
def train_valid_test_datasets_provider(train_val_test_num_samples):
|
87 |
"""Build train, valid, and test datasets."""
|
@@ -127,15 +120,12 @@ class MegatronLMDummyDataLoader:
|
|
127 |
from megatron.data.dataset_utils import build_train_valid_test_datasets
|
128 |
train_ds, valid_ds, test_ds = build_train_valid_test_datasets(**dataset_args)
|
129 |
return train_ds, valid_ds, test_ds
|
130 |
-
|
131 |
return train_valid_test_datasets_provider
|
132 |
-
|
133 |
def build_pretraining_data_loader(self, dataset, consumed_samples):
|
134 |
if dataset is None:
|
135 |
return None
|
136 |
args = get_args()
|
137 |
micro_batch_size = args.micro_batch_size * args.num_micro_batches
|
138 |
-
|
139 |
# Megatron sampler
|
140 |
if args.dataloader_type == "single":
|
141 |
batch_sampler = MegatronPretrainingSampler(
|
@@ -157,24 +147,18 @@ class MegatronLMDummyDataLoader:
|
|
157 |
)
|
158 |
else:
|
159 |
raise Exception("{} dataloader type is not supported.".format(args.dataloader_type))
|
160 |
-
|
161 |
# Torch dataloader.
|
162 |
return torch.utils.data.DataLoader(
|
163 |
dataset, batch_sampler=batch_sampler, num_workers=args.num_workers, pin_memory=True
|
164 |
)
|
165 |
-
|
166 |
def build_train_valid_test_data_iterators(self):
|
167 |
def cyclic_iter(iter):
|
168 |
while True:
|
169 |
for x in iter:
|
170 |
yield x
|
171 |
-
|
172 |
args = get_args()
|
173 |
-
|
174 |
(train_dataloader, valid_dataloader, test_dataloader) = (None, None, None)
|
175 |
-
|
176 |
print_rank_0("> building train, validation, and test datasets ...")
|
177 |
-
|
178 |
# Backward compatibility, assume fixed batch size.
|
179 |
if args.iteration > 0 and args.consumed_train_samples == 0:
|
180 |
assert args.train_samples is None, "only backward compatiblity support for iteration-based training"
|
@@ -184,7 +168,6 @@ class MegatronLMDummyDataLoader:
|
|
184 |
args.consumed_valid_samples = (
|
185 |
(args.iteration // args.eval_interval) * args.eval_iters * args.global_batch_size
|
186 |
)
|
187 |
-
|
188 |
# Data loader only on rank 0 of each model parallel group.
|
189 |
if mpu.get_tensor_model_parallel_rank() == 0:
|
190 |
# Number of train/valid/test samples.
|
@@ -203,16 +186,13 @@ class MegatronLMDummyDataLoader:
|
|
203 |
print_rank_0(" train: {}".format(train_val_test_num_samples[0]))
|
204 |
print_rank_0(" validation: {}".format(train_val_test_num_samples[1]))
|
205 |
print_rank_0(" test: {}".format(train_val_test_num_samples[2]))
|
206 |
-
|
207 |
# Build the datasets.
|
208 |
train_valid_test_datasets_provider = self.get_train_valid_test_datasets_provider()
|
209 |
train_ds, valid_ds, test_ds = train_valid_test_datasets_provider(train_val_test_num_samples)
|
210 |
-
|
211 |
# Build dataloders.
|
212 |
train_dataloader = self.build_pretraining_data_loader(train_ds, args.consumed_train_samples)
|
213 |
valid_dataloader = self.build_pretraining_data_loader(valid_ds, args.consumed_valid_samples)
|
214 |
test_dataloader = self.build_pretraining_data_loader(test_ds, 0)
|
215 |
-
|
216 |
# Flags to know if we need to do training/validation/testing.
|
217 |
do_train = train_dataloader is not None and args.train_iters > 0
|
218 |
do_valid = valid_dataloader is not None and args.eval_iters > 0
|
@@ -221,7 +201,6 @@ class MegatronLMDummyDataLoader:
|
|
221 |
flags = torch.cuda.LongTensor([int(do_train), int(do_valid), int(do_test)])
|
222 |
else:
|
223 |
flags = torch.cuda.LongTensor([0, 0, 0])
|
224 |
-
|
225 |
# Broadcast num tokens.
|
226 |
torch.distributed.broadcast(
|
227 |
flags, mpu.get_tensor_model_parallel_src_rank(), group=mpu.get_tensor_model_parallel_group()
|
@@ -229,39 +208,31 @@ class MegatronLMDummyDataLoader:
|
|
229 |
args.do_train = flags[0].item()
|
230 |
args.do_valid = flags[1].item()
|
231 |
args.do_test = flags[2].item()
|
232 |
-
|
233 |
# Build iterators.
|
234 |
dl_type = args.dataloader_type
|
235 |
assert dl_type in ["single", "cyclic"]
|
236 |
-
|
237 |
if train_dataloader is not None:
|
238 |
train_data_iterator = (
|
239 |
iter(train_dataloader) if dl_type == "single" else iter(cyclic_iter(train_dataloader))
|
240 |
)
|
241 |
else:
|
242 |
train_data_iterator = None
|
243 |
-
|
244 |
if valid_dataloader is not None:
|
245 |
valid_data_iterator = (
|
246 |
iter(valid_dataloader) if dl_type == "single" else iter(cyclic_iter(valid_dataloader))
|
247 |
)
|
248 |
else:
|
249 |
valid_data_iterator = None
|
250 |
-
|
251 |
if test_dataloader is not None:
|
252 |
test_data_iterator = iter(test_dataloader) if dl_type == "single" else iter(cyclic_iter(test_dataloader))
|
253 |
else:
|
254 |
test_data_iterator = None
|
255 |
-
|
256 |
return train_data_iterator, valid_data_iterator, test_data_iterator
|
257 |
-
|
258 |
-
|
259 |
def prepare_data_loader(accelerator, dataloader):
|
260 |
accelerator.print("Preparing dataloader")
|
261 |
args = get_args()
|
262 |
if not args.megatron_dataset_flag:
|
263 |
from ..data_loader import _PYTORCH_DATALOADER_KWARGS, prepare_data_loader
|
264 |
-
|
265 |
args = get_args()
|
266 |
micro_batch_size = args.micro_batch_size * args.num_micro_batches
|
267 |
kwargs = {k: getattr(dataloader, k, _PYTORCH_DATALOADER_KWARGS[k]) for k in _PYTORCH_DATALOADER_KWARGS}
|
@@ -276,7 +247,6 @@ def prepare_data_loader(accelerator, dataloader):
|
|
276 |
else:
|
277 |
del kwargs["batch_sampler"]
|
278 |
kwargs["batch_size"] = micro_batch_size
|
279 |
-
|
280 |
dataloader = torch.utils.data.DataLoader(dataloader.dataset, **kwargs)
|
281 |
return prepare_data_loader(
|
282 |
dataloader,
|
@@ -303,38 +273,28 @@ def prepare_data_loader(accelerator, dataloader):
|
|
303 |
test_data_iterator,
|
304 |
) = dataloader.build_train_valid_test_data_iterators()
|
305 |
return train_data_iterator, valid_data_iterator, test_data_iterator
|
306 |
-
|
307 |
-
|
308 |
# optimizer utilities
|
309 |
class MegatronLMOptimizerWrapper(AcceleratedOptimizer):
|
310 |
def __init__(self, optimizer):
|
311 |
super().__init__(optimizer, device_placement=False, scaler=None)
|
312 |
-
|
313 |
def zero_grad(self, set_to_none=None):
|
314 |
pass # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
|
315 |
-
|
316 |
def step(self):
|
317 |
pass # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
|
318 |
-
|
319 |
@property
|
320 |
def step_was_skipped(self):
|
321 |
"""Whether or not the optimizer step was done, or skipped because of gradient overflow."""
|
322 |
return self.optimizer.skipped_iter
|
323 |
-
|
324 |
-
|
325 |
def prepare_optimizer(accelerator, model):
|
326 |
accelerator.print("Preparing optimizer")
|
327 |
args = get_args()
|
328 |
optimizer = get_megatron_optimizer(model, args.no_wd_decay_cond, args.scale_lr_cond, args.lr_mult)
|
329 |
return optimizer
|
330 |
-
|
331 |
-
|
332 |
# scheduler utilities
|
333 |
class MegatronLMDummyScheduler:
|
334 |
"""
|
335 |
Dummy scheduler presents model parameters or param groups, this is primarily used to follow conventional training
|
336 |
loop when scheduler config is specified in the deepspeed config file.
|
337 |
-
|
338 |
Args:
|
339 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
340 |
The optimizer to wrap.
|
@@ -350,43 +310,29 @@ class MegatronLMDummyScheduler:
|
|
350 |
self.total_num_steps = total_num_steps
|
351 |
self.warmup_num_steps = warmup_num_steps
|
352 |
self.kwargs = kwargs
|
353 |
-
|
354 |
-
|
355 |
class MegatronLMSchedulerWrapper(AcceleratedScheduler):
|
356 |
def __init__(self, scheduler, optimizers):
|
357 |
super().__init__(scheduler, optimizers)
|
358 |
-
|
359 |
def step(self, *args, **kwargs):
|
360 |
return # `model(**batch)` is doing that automatically. Therefore, it's implementation is not needed
|
361 |
-
|
362 |
-
|
363 |
def prepare_scheduler(accelerator, optimizer, scheduler):
|
364 |
accelerator.print("Preparing scheduler")
|
365 |
scheduler = get_optimizer_param_scheduler(optimizer)
|
366 |
return scheduler
|
367 |
-
|
368 |
-
|
369 |
class AbstractTrainStep(ABC):
|
370 |
"""Abstract class for batching, forward pass and loss handler."""
|
371 |
-
|
372 |
def __init__(self, name):
|
373 |
super().__init__()
|
374 |
self.name = name
|
375 |
-
|
376 |
def get_batch_func(self):
|
377 |
pass
|
378 |
-
|
379 |
def get_forward_step_func(self):
|
380 |
pass
|
381 |
-
|
382 |
def get_loss_func(self):
|
383 |
pass
|
384 |
-
|
385 |
-
|
386 |
class BertTrainStep(AbstractTrainStep):
|
387 |
"""
|
388 |
Bert train step class.
|
389 |
-
|
390 |
Args:
|
391 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
392 |
"""
|
@@ -399,22 +345,18 @@ class BertTrainStep(AbstractTrainStep):
|
|
399 |
self.model_output_class = None
|
400 |
else:
|
401 |
self.model_output_class = SequenceClassifierOutput
|
402 |
-
|
403 |
def get_batch_func(self, megatron_dataset_flag):
|
404 |
def get_batch_megatron(data_iterator):
|
405 |
"""Build the batch."""
|
406 |
-
|
407 |
# Items and their type.
|
408 |
keys = ["text", "types", "labels", "is_random", "loss_mask", "padding_mask"]
|
409 |
datatype = torch.int64
|
410 |
-
|
411 |
# Broadcast data.
|
412 |
if data_iterator is not None:
|
413 |
data = next(data_iterator)
|
414 |
else:
|
415 |
data = None
|
416 |
data_b = mpu.broadcast_data(keys, data, datatype)
|
417 |
-
|
418 |
# Unpack.
|
419 |
tokens = data_b["text"].long()
|
420 |
types = data_b["types"].long()
|
@@ -422,14 +364,11 @@ class BertTrainStep(AbstractTrainStep):
|
|
422 |
loss_mask = data_b["loss_mask"].float()
|
423 |
lm_labels = data_b["labels"].long()
|
424 |
padding_mask = data_b["padding_mask"].long()
|
425 |
-
|
426 |
return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
|
427 |
-
|
428 |
def get_batch_transformer(data_iterator):
|
429 |
"""Build the batch."""
|
430 |
data = next(data_iterator)
|
431 |
data = send_to_device(data, torch.cuda.current_device())
|
432 |
-
|
433 |
# Unpack.
|
434 |
tokens = data["input_ids"].long()
|
435 |
padding_mask = data["attention_mask"].long()
|
@@ -447,34 +386,27 @@ class BertTrainStep(AbstractTrainStep):
|
|
447 |
sentence_order = data["next_sentence_label"].long()
|
448 |
else:
|
449 |
sentence_order = None
|
450 |
-
|
451 |
return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
|
452 |
-
|
453 |
if megatron_dataset_flag:
|
454 |
return get_batch_megatron
|
455 |
else:
|
456 |
return get_batch_transformer
|
457 |
-
|
458 |
def get_loss_func(self, pretraining_flag, num_labels):
|
459 |
def loss_func_pretrain(loss_mask, sentence_order, output_tensor):
|
460 |
lm_loss_, sop_logits = output_tensor
|
461 |
-
|
462 |
lm_loss_ = lm_loss_.float()
|
463 |
loss_mask = loss_mask.float()
|
464 |
lm_loss = torch.sum(lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
|
465 |
-
|
466 |
if sop_logits is not None:
|
467 |
sop_loss = F.cross_entropy(sop_logits.view(-1, 2).float(), sentence_order.view(-1), ignore_index=-1)
|
468 |
sop_loss = sop_loss.float()
|
469 |
loss = lm_loss + sop_loss
|
470 |
averaged_losses = average_losses_across_data_parallel_group([lm_loss, sop_loss])
|
471 |
return loss, {"lm loss": averaged_losses[0], "sop loss": averaged_losses[1]}
|
472 |
-
|
473 |
else:
|
474 |
loss = lm_loss
|
475 |
averaged_losses = average_losses_across_data_parallel_group([lm_loss])
|
476 |
return loss, {"lm loss": averaged_losses[0]}
|
477 |
-
|
478 |
def loss_func_finetune(labels, logits):
|
479 |
if num_labels == 1:
|
480 |
# We are doing regression
|
@@ -488,12 +420,10 @@ class BertTrainStep(AbstractTrainStep):
|
|
488 |
loss = loss_fct(logits, labels)
|
489 |
averaged_losses = average_losses_across_data_parallel_group([loss])
|
490 |
return loss, {"loss": averaged_losses[0]}
|
491 |
-
|
492 |
if pretraining_flag:
|
493 |
return loss_func_pretrain
|
494 |
else:
|
495 |
return loss_func_finetune
|
496 |
-
|
497 |
def get_forward_step_func(self, pretraining_flag, bert_binary_head):
|
498 |
def forward_step(data_iterator, model):
|
499 |
"""Forward step."""
|
@@ -507,14 +437,10 @@ class BertTrainStep(AbstractTrainStep):
|
|
507 |
else:
|
508 |
logits = model(tokens, padding_mask, tokentype_ids=types)
|
509 |
return logits, partial(self.loss_func, labels)
|
510 |
-
|
511 |
return forward_step
|
512 |
-
|
513 |
-
|
514 |
class GPTTrainStep(AbstractTrainStep):
|
515 |
"""
|
516 |
GPT train step class.
|
517 |
-
|
518 |
Args:
|
519 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
520 |
"""
|
@@ -534,38 +460,31 @@ class GPTTrainStep(AbstractTrainStep):
|
|
534 |
self.model_output_class = None
|
535 |
else:
|
536 |
self.model_output_class = CausalLMOutputWithCrossAttentions
|
537 |
-
|
538 |
def get_batch_func(self, megatron_dataset_flag):
|
539 |
def get_batch_megatron(data_iterator):
|
540 |
"""Generate a batch"""
|
541 |
# Items and their type.
|
542 |
keys = ["text"]
|
543 |
datatype = torch.int64
|
544 |
-
|
545 |
# Broadcast data.
|
546 |
if data_iterator is not None:
|
547 |
data = next(data_iterator)
|
548 |
else:
|
549 |
data = None
|
550 |
data_b = mpu.broadcast_data(keys, data, datatype)
|
551 |
-
|
552 |
# Unpack.
|
553 |
tokens_ = data_b["text"].long()
|
554 |
labels = tokens_[:, 1:].contiguous()
|
555 |
tokens = tokens_[:, :-1].contiguous()
|
556 |
-
|
557 |
# Get the masks and postition ids.
|
558 |
attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids(
|
559 |
tokens, self.eod_token, self.reset_position_ids, self.reset_attention_mask, self.eod_mask_loss
|
560 |
)
|
561 |
-
|
562 |
return tokens, labels, loss_mask, attention_mask, position_ids
|
563 |
-
|
564 |
def get_batch_transformer(data_iterator):
|
565 |
data = next(data_iterator)
|
566 |
data = {"input_ids": data["input_ids"]}
|
567 |
data = send_to_device(data, torch.cuda.current_device())
|
568 |
-
|
569 |
tokens_ = data["input_ids"].long()
|
570 |
padding = torch.zeros((tokens_.shape[0], 1), dtype=tokens_.dtype, device=tokens_.device) + self.eod_token
|
571 |
tokens_ = torch.concat([tokens_, padding], dim=1)
|
@@ -576,15 +495,12 @@ class GPTTrainStep(AbstractTrainStep):
|
|
576 |
tokens, self.eod_token, self.reset_position_ids, self.reset_attention_mask, True
|
577 |
)
|
578 |
return tokens, labels, loss_mask, attention_mask, position_ids
|
579 |
-
|
580 |
if megatron_dataset_flag:
|
581 |
return get_batch_megatron
|
582 |
else:
|
583 |
return get_batch_transformer
|
584 |
-
|
585 |
def get_loss_func(self):
|
586 |
args = get_args()
|
587 |
-
|
588 |
def loss_func(loss_mask, output_tensor):
|
589 |
if args.return_logits:
|
590 |
losses, logits = output_tensor
|
@@ -593,33 +509,24 @@ class GPTTrainStep(AbstractTrainStep):
|
|
593 |
losses = losses.float()
|
594 |
loss_mask = loss_mask.view(-1).float()
|
595 |
loss = torch.sum(losses.view(-1) * loss_mask) / loss_mask.sum()
|
596 |
-
|
597 |
# Reduce loss for logging.
|
598 |
averaged_loss = average_losses_across_data_parallel_group([loss])
|
599 |
-
|
600 |
output_dict = {"lm loss": averaged_loss[0]}
|
601 |
if args.return_logits:
|
602 |
output_dict.update({"logits": logits})
|
603 |
return loss, output_dict
|
604 |
-
|
605 |
return loss_func
|
606 |
-
|
607 |
def get_forward_step_func(self):
|
608 |
def forward_step(data_iterator, model):
|
609 |
"""Forward step."""
|
610 |
# Get the batch.
|
611 |
tokens, labels, loss_mask, attention_mask, position_ids = self.get_batch(data_iterator)
|
612 |
output_tensor = model(tokens, position_ids, attention_mask, labels=labels)
|
613 |
-
|
614 |
return output_tensor, partial(self.loss_func, loss_mask)
|
615 |
-
|
616 |
return forward_step
|
617 |
-
|
618 |
-
|
619 |
class T5TrainStep(AbstractTrainStep):
|
620 |
"""
|
621 |
T5 train step class.
|
622 |
-
|
623 |
Args:
|
624 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
625 |
"""
|
@@ -632,7 +539,6 @@ class T5TrainStep(AbstractTrainStep):
|
|
632 |
self.model_output_class = None
|
633 |
else:
|
634 |
self.model_output_class = Seq2SeqLMOutput
|
635 |
-
|
636 |
@staticmethod
|
637 |
def attn_mask_postprocess(attention_mask):
|
638 |
# We create a 3D attention mask from a 2D tensor mask.
|
@@ -645,13 +551,11 @@ class T5TrainStep(AbstractTrainStep):
|
|
645 |
# Convert attention mask to binary:
|
646 |
extended_attention_mask = attention_mask_bss < 0.5
|
647 |
return extended_attention_mask
|
648 |
-
|
649 |
@staticmethod
|
650 |
def get_decoder_mask(seq_length, device):
|
651 |
attention_mask = torch.tril(torch.ones((1, seq_length, seq_length), device=device))
|
652 |
attention_mask = attention_mask < 0.5
|
653 |
return attention_mask
|
654 |
-
|
655 |
@staticmethod
|
656 |
def get_enc_dec_mask(attention_mask, dec_seq_length, device):
|
657 |
batch_size, _ = attention_mask.shape
|
@@ -663,38 +567,30 @@ class T5TrainStep(AbstractTrainStep):
|
|
663 |
attention_mask_bss = attention_mask_bs1 * attention_mask_b1s
|
664 |
extended_attention_mask = attention_mask_bss < 0.5
|
665 |
return extended_attention_mask
|
666 |
-
|
667 |
def get_batch_func(self, megatron_dataset_flag):
|
668 |
def get_batch_megatron(data_iterator):
|
669 |
"""Build the batch."""
|
670 |
-
|
671 |
keys = ["text_enc", "text_dec", "labels", "loss_mask", "enc_mask", "dec_mask", "enc_dec_mask"]
|
672 |
datatype = torch.int64
|
673 |
-
|
674 |
# Broadcast data.
|
675 |
if data_iterator is not None:
|
676 |
data = next(data_iterator)
|
677 |
else:
|
678 |
data = None
|
679 |
data_b = mpu.broadcast_data(keys, data, datatype)
|
680 |
-
|
681 |
# Unpack.
|
682 |
tokens_enc = data_b["text_enc"].long()
|
683 |
tokens_dec = data_b["text_dec"].long()
|
684 |
labels = data_b["labels"].long()
|
685 |
loss_mask = data_b["loss_mask"].float()
|
686 |
-
|
687 |
enc_mask = data_b["enc_mask"] < 0.5
|
688 |
dec_mask = data_b["dec_mask"] < 0.5
|
689 |
enc_dec_mask = data_b["enc_dec_mask"] < 0.5
|
690 |
-
|
691 |
return tokens_enc, tokens_dec, loss_mask, labels, enc_mask, dec_mask, enc_dec_mask
|
692 |
-
|
693 |
def get_batch_transformer(data_iterator):
|
694 |
"""Build the batch."""
|
695 |
data = next(data_iterator)
|
696 |
data = send_to_device(data, torch.cuda.current_device())
|
697 |
-
|
698 |
tokens_enc = data["input_ids"].long()
|
699 |
labels = data["labels"].long()
|
700 |
loss_mask = (labels != -100).to(torch.float)
|
@@ -710,26 +606,19 @@ class T5TrainStep(AbstractTrainStep):
|
|
710 |
enc_dec_mask = T5TrainStep.get_enc_dec_mask(
|
711 |
data["attention_mask"].long(), tokens_dec.shape[1], tokens_dec.device
|
712 |
)
|
713 |
-
|
714 |
return tokens_enc, tokens_dec, loss_mask, labels, enc_mask, dec_mask, enc_dec_mask
|
715 |
-
|
716 |
if megatron_dataset_flag:
|
717 |
return get_batch_megatron
|
718 |
else:
|
719 |
return get_batch_transformer
|
720 |
-
|
721 |
def get_loss_func(self):
|
722 |
def loss_func(loss_mask, output_tensor):
|
723 |
lm_loss_ = output_tensor.float()
|
724 |
lm_loss = torch.sum(lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
|
725 |
-
|
726 |
loss = lm_loss
|
727 |
averaged_losses = average_losses_across_data_parallel_group([lm_loss])
|
728 |
-
|
729 |
return loss, {"lm loss": averaged_losses[0]}
|
730 |
-
|
731 |
return loss_func
|
732 |
-
|
733 |
def get_forward_step_func(self):
|
734 |
def forward_step(data_iterator, model):
|
735 |
"""Forward step."""
|
@@ -741,20 +630,14 @@ class T5TrainStep(AbstractTrainStep):
|
|
741 |
output_tensor = model(
|
742 |
tokens_enc, tokens_dec, enc_mask, dec_mask, enc_dec_mask, tokentype_ids=None, lm_labels=lm_labels
|
743 |
)
|
744 |
-
|
745 |
return output_tensor, partial(self.loss_func, loss_mask)
|
746 |
-
|
747 |
return forward_step
|
748 |
-
|
749 |
-
|
750 |
# intialize megatron setup
|
751 |
def initialize(accelerator, extra_args_provider=None, args_defaults={}):
|
752 |
accelerator.print("Initializing Megatron-LM")
|
753 |
assert torch.cuda.is_available(), "Megatron requires CUDA."
|
754 |
-
|
755 |
# Parse arguments
|
756 |
args = parse_args(extra_args_provider, ignore_unknown_args=True)
|
757 |
-
|
758 |
# Set defaults
|
759 |
for key, value in args_defaults.items():
|
760 |
if getattr(args, key, None) is not None:
|
@@ -767,17 +650,13 @@ def initialize(accelerator, extra_args_provider=None, args_defaults={}):
|
|
767 |
flush=True,
|
768 |
)
|
769 |
setattr(args, key, value)
|
770 |
-
|
771 |
if args.use_checkpoint_args or args_defaults.get("use_checkpoint_args", False):
|
772 |
assert args.load is not None, "--use-checkpoints-args requires --load argument"
|
773 |
load_args_from_checkpoint(args)
|
774 |
-
|
775 |
validate_args(args)
|
776 |
-
|
777 |
# set global args, build tokenizer, and set adlr-autoresume,
|
778 |
# tensorboard-writer, and timers.
|
779 |
set_global_variables(args)
|
780 |
-
|
781 |
# torch.distributed initialization
|
782 |
def finish_mpu_init():
|
783 |
args = get_args()
|
@@ -791,7 +670,6 @@ def initialize(accelerator, extra_args_provider=None, args_defaults={}):
|
|
791 |
assert args.local_rank == device, "expected local-rank to be the same as rank % device-count."
|
792 |
else:
|
793 |
args.local_rank = device
|
794 |
-
|
795 |
# Set the tensor model-parallel, pipeline model-parallel, and
|
796 |
# data-parallel communicators.
|
797 |
if mpu.model_parallel_is_initialized():
|
@@ -803,23 +681,17 @@ def initialize(accelerator, extra_args_provider=None, args_defaults={}):
|
|
803 |
args.virtual_pipeline_model_parallel_size,
|
804 |
args.pipeline_model_parallel_split_rank,
|
805 |
)
|
806 |
-
|
807 |
# Random seeds for reproducibility.
|
808 |
if args.rank == 0:
|
809 |
print("> setting random seeds to {} ...".format(args.seed))
|
810 |
_set_random_seed(args.seed, args.data_parallel_random_init)
|
811 |
-
|
812 |
args = get_args()
|
813 |
-
|
814 |
# Megatron's MPU is the master. Complete initialization right away.
|
815 |
finish_mpu_init()
|
816 |
-
|
817 |
# Autoresume.
|
818 |
_init_autoresume()
|
819 |
-
|
820 |
# Compile dependencies.
|
821 |
_compile_dependencies()
|
822 |
-
|
823 |
# Set pytorch JIT layer fusion options and warmup JIT functions.
|
824 |
set_jit_fusion_options()
|
825 |
args = get_args()
|
@@ -829,12 +701,9 @@ def initialize(accelerator, extra_args_provider=None, args_defaults={}):
|
|
829 |
else:
|
830 |
args.bert_binary_head = False
|
831 |
args.iteration = 0
|
832 |
-
|
833 |
-
|
834 |
class MegatronEngine(torch.nn.Module):
|
835 |
"""
|
836 |
Megatron-LM model wrapper
|
837 |
-
|
838 |
Args:
|
839 |
accelerator (:class:`~accelerate.Accelerator`): The accelerator object to use.
|
840 |
model: Megatron-LM model
|
@@ -861,7 +730,6 @@ class MegatronEngine(torch.nn.Module):
|
|
861 |
else:
|
862 |
raise ValueError(f"Unsupported model type: {args.model_type_name}")
|
863 |
self.optimizer.skipped_iter = False
|
864 |
-
|
865 |
# Tracking loss.
|
866 |
self.total_loss_dict = {}
|
867 |
self.eval_total_loss_dict = {}
|
@@ -869,26 +737,21 @@ class MegatronEngine(torch.nn.Module):
|
|
869 |
self.report_memory_flag = True
|
870 |
if args.tensorboard_dir is not None:
|
871 |
write_args_to_tensorboard()
|
872 |
-
|
873 |
def train(self):
|
874 |
for model_module in self.module:
|
875 |
model_module.train()
|
876 |
self.log_eval_results()
|
877 |
-
|
878 |
def eval(self):
|
879 |
for model_module in self.module:
|
880 |
model_module.eval()
|
881 |
-
|
882 |
def train_step(self, **batch_data):
|
883 |
"""
|
884 |
Training step for Megatron-LM
|
885 |
-
|
886 |
Args:
|
887 |
batch_data (:obj:`dict`): The batch data to train on.
|
888 |
"""
|
889 |
args = get_args()
|
890 |
timers = get_timers()
|
891 |
-
|
892 |
if len(batch_data) > 0:
|
893 |
data_chunks = []
|
894 |
if args.num_micro_batches > 1:
|
@@ -901,7 +764,6 @@ class MegatronEngine(torch.nn.Module):
|
|
901 |
)
|
902 |
else:
|
903 |
data_chunks = [batch_data]
|
904 |
-
|
905 |
if len(self.module) > 1:
|
906 |
batch_data_iterator = (
|
907 |
[iter(data_chunks) for _ in range(len(self.module))]
|
@@ -910,13 +772,11 @@ class MegatronEngine(torch.nn.Module):
|
|
910 |
)
|
911 |
else:
|
912 |
batch_data_iterator = iter(data_chunks) if len(batch_data) > 0 else None
|
913 |
-
|
914 |
# Set grad to zero.
|
915 |
if args.DDP_impl == "local" and args.use_contiguous_buffers_in_local_ddp:
|
916 |
for partition in self.module:
|
917 |
partition.zero_grad_buffer()
|
918 |
self.optimizer.zero_grad()
|
919 |
-
|
920 |
# Forward pass.
|
921 |
forward_backward_func = get_forward_backward_func()
|
922 |
losses_reduced = forward_backward_func(
|
@@ -927,27 +787,22 @@ class MegatronEngine(torch.nn.Module):
|
|
927 |
None,
|
928 |
forward_only=False,
|
929 |
)
|
930 |
-
|
931 |
# Empty unused memory.
|
932 |
if args.empty_unused_memory_level >= 1:
|
933 |
torch.cuda.empty_cache()
|
934 |
-
|
935 |
# Reduce gradients.
|
936 |
timers("backward-reduce-model-grads").start()
|
937 |
self.optimizer.reduce_model_grads(args, timers)
|
938 |
timers("backward-reduce-model-grads").stop()
|
939 |
-
|
940 |
# Update parameters.
|
941 |
timers("optimizer").start()
|
942 |
update_successful, grad_norm, num_zeros_in_grad = self.optimizer.step(args, timers)
|
943 |
timers("optimizer").stop()
|
944 |
-
|
945 |
# Gather params.
|
946 |
if update_successful:
|
947 |
timers("backward-gather-model-params").start()
|
948 |
self.optimizer.gather_model_params(args, timers)
|
949 |
timers("backward-gather-model-params").stop()
|
950 |
-
|
951 |
# Update learning rate.
|
952 |
if update_successful:
|
953 |
if self.scheduler is not None:
|
@@ -956,17 +811,13 @@ class MegatronEngine(torch.nn.Module):
|
|
956 |
skipped_iter = 0
|
957 |
else:
|
958 |
skipped_iter = 1
|
959 |
-
|
960 |
self.optimizer.skipped_iter = not update_successful
|
961 |
-
|
962 |
# Empty unused memory.
|
963 |
if args.empty_unused_memory_level >= 2:
|
964 |
torch.cuda.empty_cache()
|
965 |
-
|
966 |
args.consumed_train_samples += (
|
967 |
mpu.get_data_parallel_world_size() * args.micro_batch_size * get_num_microbatches()
|
968 |
)
|
969 |
-
|
970 |
if mpu.is_pipeline_last_stage(ignore_virtual=True):
|
971 |
# Average loss across microbatches.
|
972 |
loss_reduced = {}
|
@@ -978,11 +829,9 @@ class MegatronEngine(torch.nn.Module):
|
|
978 |
loss_reduced[key] = torch.concat(losses_reduced_for_key)
|
979 |
return loss_reduced, skipped_iter, grad_norm, num_zeros_in_grad
|
980 |
return {}, skipped_iter, grad_norm, num_zeros_in_grad
|
981 |
-
|
982 |
def eval_step(self, **batch_data):
|
983 |
"""
|
984 |
Evaluation step for Megatron-LM
|
985 |
-
|
986 |
Args:
|
987 |
batch_data (:obj:`dict`): The batch data to evaluate on.
|
988 |
"""
|
@@ -995,7 +844,6 @@ class MegatronEngine(torch.nn.Module):
|
|
995 |
)
|
996 |
else:
|
997 |
data_chunks = [batch_data]
|
998 |
-
|
999 |
if len(self.module) > 1:
|
1000 |
batch_data_iterator = [iter(data_chunks) for _ in range(len(self.module))]
|
1001 |
else:
|
@@ -1012,11 +860,9 @@ class MegatronEngine(torch.nn.Module):
|
|
1012 |
# Empty unused memory
|
1013 |
if args.empty_unused_memory_level >= 1:
|
1014 |
torch.cuda.empty_cache()
|
1015 |
-
|
1016 |
args.consumed_valid_samples += (
|
1017 |
mpu.get_data_parallel_world_size() * args.micro_batch_size * get_num_microbatches()
|
1018 |
)
|
1019 |
-
|
1020 |
if mpu.is_pipeline_last_stage(ignore_virtual=True):
|
1021 |
# Average loss across microbatches.
|
1022 |
loss_reduced = {}
|
@@ -1029,7 +875,6 @@ class MegatronEngine(torch.nn.Module):
|
|
1029 |
return loss_reduced
|
1030 |
else:
|
1031 |
return {}
|
1032 |
-
|
1033 |
def forward(self, **batch_data):
|
1034 |
# During training, we use train_step()
|
1035 |
# model(**batch_data) performs following operations by delegating it to `self.train_step`:
|
@@ -1077,12 +922,10 @@ class MegatronEngine(torch.nn.Module):
|
|
1077 |
self.eval_total_loss_dict[key + "_num_iters"] = self.eval_total_loss_dict.get(
|
1078 |
key + "_num_iters", torch.cuda.FloatTensor([0.0])
|
1079 |
) + torch.cuda.FloatTensor([1.0])
|
1080 |
-
|
1081 |
loss = torch.tensor(0.0, device=args.local_rank)
|
1082 |
for key in loss_dict:
|
1083 |
if len(loss_dict[key].shape) == 0:
|
1084 |
loss += loss_dict[key]
|
1085 |
-
|
1086 |
logits = None
|
1087 |
if "logits" in loss_dict:
|
1088 |
logits = loss_dict["logits"]
|
@@ -1090,7 +933,6 @@ class MegatronEngine(torch.nn.Module):
|
|
1090 |
if self.train_step_handler.model_output_class is not None:
|
1091 |
return self.train_step_handler.model_output_class(loss=loss, logits=logits)
|
1092 |
return loss
|
1093 |
-
|
1094 |
def log_eval_results(self):
|
1095 |
args = get_args()
|
1096 |
if args.tensorboard_dir is None or self.iteration == 0:
|
@@ -1110,13 +952,11 @@ class MegatronEngine(torch.nn.Module):
|
|
1110 |
writer.add_scalar(f"{key} validation", value.item(), self.iteration)
|
1111 |
if args.pretraining_flag:
|
1112 |
writer.add_scalar(f"{key} validation ppl", ppl, self.iteration)
|
1113 |
-
|
1114 |
length = len(string) + 1
|
1115 |
print_rank_last("-" * length)
|
1116 |
print_rank_last(string)
|
1117 |
print_rank_last("-" * length)
|
1118 |
self.eval_total_loss_dict = {}
|
1119 |
-
|
1120 |
def save_checkpoint(self, output_dir):
|
1121 |
self.log_eval_results()
|
1122 |
args = get_args()
|
@@ -1124,7 +964,6 @@ class MegatronEngine(torch.nn.Module):
|
|
1124 |
torch.distributed.barrier()
|
1125 |
save_checkpoint(self.iteration, self.module, self.optimizer, self.scheduler)
|
1126 |
torch.distributed.barrier()
|
1127 |
-
|
1128 |
def load_checkpoint(self, input_dir):
|
1129 |
args = get_args()
|
1130 |
args.load = input_dir
|
@@ -1136,7 +975,6 @@ class MegatronEngine(torch.nn.Module):
|
|
1136 |
self.iteration = iteration
|
1137 |
if args.fp16 and self.iteration == 0:
|
1138 |
self.optimizer.reload_model_params()
|
1139 |
-
|
1140 |
def megatron_generate(
|
1141 |
self,
|
1142 |
inputs,
|
@@ -1153,7 +991,6 @@ class MegatronEngine(torch.nn.Module):
|
|
1153 |
"""
|
1154 |
Generate method for GPT2 model. This method is used for inference. Supports both greedy and beam search along
|
1155 |
with sampling. Refer the Megatron-LM repo for more details
|
1156 |
-
|
1157 |
Args:
|
1158 |
inputs (torch.Tensor): input ids
|
1159 |
attention_mask (torch.Tensor, optional): attention mask. Defaults to None.
|
@@ -1172,33 +1009,25 @@ class MegatronEngine(torch.nn.Module):
|
|
1172 |
args = get_args()
|
1173 |
if args.model_type_name != "gpt":
|
1174 |
raise NotImplementedError("Generate method is not implemented for this model")
|
1175 |
-
|
1176 |
if args.data_parallel_size > 1:
|
1177 |
raise ValueError("Generate method requires data parallelism to be 1")
|
1178 |
-
|
1179 |
if args.sequence_parallel:
|
1180 |
raise ValueError("Generate method requires sequence parallelism to be False")
|
1181 |
-
|
1182 |
if args.recompute_granularity is not None:
|
1183 |
raise ValueError("Checkpoint activations cannot be set for inference")
|
1184 |
-
|
1185 |
if args.vocab_file is None:
|
1186 |
raise ValueError("Vocab file is required for inference")
|
1187 |
-
|
1188 |
# Prepare inputs
|
1189 |
if max_length is None and max_new_tokens is None:
|
1190 |
raise ValueError("`max_length` or `max_new_tokens` are required for inference")
|
1191 |
-
|
1192 |
if temperature is None:
|
1193 |
temperature = 1.0
|
1194 |
elif not (0.0 < temperature <= 100.0):
|
1195 |
raise ValueError("temperature must be a positive number less than or equal to 100.0")
|
1196 |
-
|
1197 |
if top_k is None:
|
1198 |
top_k = 0
|
1199 |
elif not (0 <= top_k <= 1000):
|
1200 |
raise ValueError("top_k must be a positive number less than or equal to 1000")
|
1201 |
-
|
1202 |
if top_p is None:
|
1203 |
top_p = 0.0
|
1204 |
elif top_p > 0.0 and top_k > 0.0:
|
@@ -1206,19 +1035,15 @@ class MegatronEngine(torch.nn.Module):
|
|
1206 |
else:
|
1207 |
if not (0.0 <= top_p <= 1.0):
|
1208 |
raise ValueError("top_p must be less than or equal to 1.0")
|
1209 |
-
|
1210 |
top_p_decay = kwargs.get("top_p_decay", 0.0)
|
1211 |
if not (0.0 <= top_p_decay <= 1.0):
|
1212 |
raise ValueError("top_p_decay must be less than or equal to 1.0")
|
1213 |
-
|
1214 |
top_p_bound = kwargs.get("top_p_bound", 0.0)
|
1215 |
if not (0.0 <= top_p_bound <= 1.0):
|
1216 |
raise ValueError("top_p_bound must be less than or equal to 1.0")
|
1217 |
-
|
1218 |
add_BOS = kwargs.get("add_BOS", False)
|
1219 |
if not (isinstance(add_BOS, bool)):
|
1220 |
raise ValueError("add_BOS must be a boolean")
|
1221 |
-
|
1222 |
beam_width = num_beams
|
1223 |
if beam_width is not None:
|
1224 |
if not isinstance(beam_width, int):
|
@@ -1227,17 +1052,13 @@ class MegatronEngine(torch.nn.Module):
|
|
1227 |
raise ValueError("beam_width must be greater than 0")
|
1228 |
if inputs.shape[0] > 1:
|
1229 |
return "When doing beam_search, batch size must be 1"
|
1230 |
-
|
1231 |
tokenizer = get_tokenizer()
|
1232 |
-
|
1233 |
stop_token = kwargs.get("stop_token", tokenizer.eod)
|
1234 |
if stop_token is not None:
|
1235 |
if not isinstance(stop_token, int):
|
1236 |
raise ValueError("stop_token must be an integer")
|
1237 |
-
|
1238 |
if length_penalty is None:
|
1239 |
length_penalty = 1.0
|
1240 |
-
|
1241 |
sizes_list = None
|
1242 |
prompts_tokens_tensor = None
|
1243 |
prompts_length_tensor = None
|
@@ -1247,12 +1068,10 @@ class MegatronEngine(torch.nn.Module):
|
|
1247 |
prompts_length_tensor = torch.cuda.LongTensor([inputs.shape[1]] * inputs.shape[0])
|
1248 |
else:
|
1249 |
prompts_length_tensor = attention_mask.sum(axis=-1).cuda()
|
1250 |
-
|
1251 |
if max_new_tokens is None:
|
1252 |
max_new_tokens = max_length - inputs.shape[1]
|
1253 |
if max_new_tokens <= 0:
|
1254 |
raise ValueError("max_new_tokens must be greater than 0")
|
1255 |
-
|
1256 |
if add_BOS:
|
1257 |
max_length = max_new_tokens + inputs.shape[1] + 1
|
1258 |
# making sure that `max_length` is a multiple of 4 to leverage fused kernels
|
@@ -1269,22 +1088,18 @@ class MegatronEngine(torch.nn.Module):
|
|
1269 |
max_new_tokens = max_length - inputs.shape[1]
|
1270 |
padding = torch.cuda.LongTensor([[tokenizer.eod] * max_new_tokens] * inputs.shape[0])
|
1271 |
prompts_tokens_tensor = torch.concat([inputs.cuda(), padding], axis=-1)
|
1272 |
-
|
1273 |
# We need the sizes of these tensors for the boradcast
|
1274 |
sizes_list = [
|
1275 |
prompts_tokens_tensor.size(0), # Batch size
|
1276 |
prompts_tokens_tensor.size(1),
|
1277 |
] # Sequence lenght
|
1278 |
-
|
1279 |
# First, broadcast the sizes.
|
1280 |
sizes_tensor = broadcast_int_list(2, int_list=sizes_list, rank=0)
|
1281 |
-
|
1282 |
# Now that we have the sizes, we can boradcast the tokens
|
1283 |
# and length tensors.
|
1284 |
sizes = sizes_tensor.tolist()
|
1285 |
context_tokens_tensor = broadcast_tensor(sizes, torch.int64, tensor=prompts_tokens_tensor, rank=0)
|
1286 |
context_length_tensor = broadcast_tensor(sizes[0], torch.int64, tensor=prompts_length_tensor, rank=0)
|
1287 |
-
|
1288 |
# Run the inference
|
1289 |
random_seed = kwargs.get("random_seed", 0)
|
1290 |
torch.random.manual_seed(random_seed)
|
@@ -1313,27 +1128,20 @@ class MegatronEngine(torch.nn.Module):
|
|
1313 |
use_eod_token_for_early_termination=True,
|
1314 |
)
|
1315 |
return tokens
|
1316 |
-
|
1317 |
-
|
1318 |
# other utilities
|
1319 |
def avg_losses_across_data_parallel_group(losses):
|
1320 |
"""
|
1321 |
Average losses across data parallel group.
|
1322 |
-
|
1323 |
Args:
|
1324 |
losses (List[Tensor]): List of losses to average across data parallel group.
|
1325 |
"""
|
1326 |
return average_losses_across_data_parallel_group(losses)
|
1327 |
-
|
1328 |
-
|
1329 |
def gather_across_data_parallel_groups(tensor):
|
1330 |
"""
|
1331 |
Recursively gather tensor in a nested list/tuple/dictionary of tensors from data parallel ranks.
|
1332 |
-
|
1333 |
Args:
|
1334 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
1335 |
The data to gather across data parallel ranks.
|
1336 |
-
|
1337 |
"""
|
1338 |
def _gpu_gather_one(tensor):
|
1339 |
if tensor.ndim == 0:
|
@@ -1344,5 +1152,4 @@ def gather_across_data_parallel_groups(tensor):
|
|
1344 |
]
|
1345 |
torch.distributed.all_gather(output_tensors, tensor, group=mpu.get_data_parallel_group())
|
1346 |
return torch.cat(output_tensors, dim=0)
|
1347 |
-
|
1348 |
return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)
|
|
|
37 |
else:
|
38 |
raise ValueError(f"Unsupported model type: {args.model_type_name}")
|
39 |
return model
|
|
|
|
|
40 |
def prepare_model(accelerator):
|
41 |
accelerator.print("Preparing model")
|
42 |
args = get_args()
|
|
|
56 |
args.pipeline_model_parallel_split_rank = args.pipeline_model_parallel_size // 2
|
57 |
model = get_model(model_provider_func, model_type)
|
58 |
return model
|
|
|
|
|
59 |
# dataloader utilities
|
60 |
class MegatronLMDummyDataLoader:
|
61 |
"""
|
62 |
Dummy dataloader presents model parameters or param groups, this is primarily used to follow conventional training
|
|
|
63 |
Args:
|
64 |
**dataset_kwargs: Megatron data arguments.
|
65 |
"""
|
|
|
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."""
|
|
|
120 |
from megatron.data.dataset_utils import build_train_valid_test_datasets
|
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
|
127 |
args = get_args()
|
128 |
micro_batch_size = args.micro_batch_size * args.num_micro_batches
|
|
|
129 |
# Megatron sampler
|
130 |
if args.dataloader_type == "single":
|
131 |
batch_sampler = MegatronPretrainingSampler(
|
|
|
147 |
)
|
148 |
else:
|
149 |
raise Exception("{} dataloader type is not supported.".format(args.dataloader_type))
|
|
|
150 |
# Torch dataloader.
|
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
|
|
|
159 |
args = get_args()
|
|
|
160 |
(train_dataloader, valid_dataloader, test_dataloader) = (None, None, None)
|
|
|
161 |
print_rank_0("> building train, validation, and test datasets ...")
|
|
|
162 |
# Backward compatibility, assume fixed batch size.
|
163 |
if args.iteration > 0 and args.consumed_train_samples == 0:
|
164 |
assert args.train_samples is None, "only backward compatiblity support for iteration-based training"
|
|
|
168 |
args.consumed_valid_samples = (
|
169 |
(args.iteration // args.eval_interval) * args.eval_iters * args.global_batch_size
|
170 |
)
|
|
|
171 |
# Data loader only on rank 0 of each model parallel group.
|
172 |
if mpu.get_tensor_model_parallel_rank() == 0:
|
173 |
# Number of train/valid/test samples.
|
|
|
186 |
print_rank_0(" train: {}".format(train_val_test_num_samples[0]))
|
187 |
print_rank_0(" validation: {}".format(train_val_test_num_samples[1]))
|
188 |
print_rank_0(" test: {}".format(train_val_test_num_samples[2]))
|
|
|
189 |
# Build the datasets.
|
190 |
train_valid_test_datasets_provider = self.get_train_valid_test_datasets_provider()
|
191 |
train_ds, valid_ds, test_ds = train_valid_test_datasets_provider(train_val_test_num_samples)
|
|
|
192 |
# Build dataloders.
|
193 |
train_dataloader = self.build_pretraining_data_loader(train_ds, args.consumed_train_samples)
|
194 |
valid_dataloader = self.build_pretraining_data_loader(valid_ds, args.consumed_valid_samples)
|
195 |
test_dataloader = self.build_pretraining_data_loader(test_ds, 0)
|
|
|
196 |
# Flags to know if we need to do training/validation/testing.
|
197 |
do_train = train_dataloader is not None and args.train_iters > 0
|
198 |
do_valid = valid_dataloader is not None and args.eval_iters > 0
|
|
|
201 |
flags = torch.cuda.LongTensor([int(do_train), int(do_valid), int(do_test)])
|
202 |
else:
|
203 |
flags = torch.cuda.LongTensor([0, 0, 0])
|
|
|
204 |
# Broadcast num tokens.
|
205 |
torch.distributed.broadcast(
|
206 |
flags, mpu.get_tensor_model_parallel_src_rank(), group=mpu.get_tensor_model_parallel_group()
|
|
|
208 |
args.do_train = flags[0].item()
|
209 |
args.do_valid = flags[1].item()
|
210 |
args.do_test = flags[2].item()
|
|
|
211 |
# Build iterators.
|
212 |
dl_type = args.dataloader_type
|
213 |
assert dl_type in ["single", "cyclic"]
|
|
|
214 |
if train_dataloader is not None:
|
215 |
train_data_iterator = (
|
216 |
iter(train_dataloader) if dl_type == "single" else iter(cyclic_iter(train_dataloader))
|
217 |
)
|
218 |
else:
|
219 |
train_data_iterator = None
|
|
|
220 |
if valid_dataloader is not None:
|
221 |
valid_data_iterator = (
|
222 |
iter(valid_dataloader) if dl_type == "single" else iter(cyclic_iter(valid_dataloader))
|
223 |
)
|
224 |
else:
|
225 |
valid_data_iterator = None
|
|
|
226 |
if test_dataloader is not None:
|
227 |
test_data_iterator = iter(test_dataloader) if dl_type == "single" else iter(cyclic_iter(test_dataloader))
|
228 |
else:
|
229 |
test_data_iterator = None
|
|
|
230 |
return train_data_iterator, valid_data_iterator, test_data_iterator
|
|
|
|
|
231 |
def prepare_data_loader(accelerator, dataloader):
|
232 |
accelerator.print("Preparing dataloader")
|
233 |
args = get_args()
|
234 |
if not args.megatron_dataset_flag:
|
235 |
from ..data_loader import _PYTORCH_DATALOADER_KWARGS, prepare_data_loader
|
|
|
236 |
args = get_args()
|
237 |
micro_batch_size = args.micro_batch_size * args.num_micro_batches
|
238 |
kwargs = {k: getattr(dataloader, k, _PYTORCH_DATALOADER_KWARGS[k]) for k in _PYTORCH_DATALOADER_KWARGS}
|
|
|
247 |
else:
|
248 |
del kwargs["batch_sampler"]
|
249 |
kwargs["batch_size"] = micro_batch_size
|
|
|
250 |
dataloader = torch.utils.data.DataLoader(dataloader.dataset, **kwargs)
|
251 |
return prepare_data_loader(
|
252 |
dataloader,
|
|
|
273 |
test_data_iterator,
|
274 |
) = dataloader.build_train_valid_test_data_iterators()
|
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
|
|
|
|
|
288 |
def prepare_optimizer(accelerator, model):
|
289 |
accelerator.print("Preparing optimizer")
|
290 |
args = get_args()
|
291 |
optimizer = get_megatron_optimizer(model, args.no_wd_decay_cond, args.scale_lr_cond, args.lr_mult)
|
292 |
return optimizer
|
|
|
|
|
293 |
# scheduler utilities
|
294 |
class MegatronLMDummyScheduler:
|
295 |
"""
|
296 |
Dummy scheduler presents model parameters or param groups, this is primarily used to follow conventional training
|
297 |
loop when scheduler config is specified in the deepspeed config file.
|
|
|
298 |
Args:
|
299 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
300 |
The optimizer to wrap.
|
|
|
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):
|
319 |
accelerator.print("Preparing scheduler")
|
320 |
scheduler = get_optimizer_param_scheduler(optimizer)
|
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):
|
334 |
"""
|
335 |
Bert train step class.
|
|
|
336 |
Args:
|
337 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
338 |
"""
|
|
|
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"]
|
353 |
datatype = torch.int64
|
|
|
354 |
# Broadcast data.
|
355 |
if data_iterator is not None:
|
356 |
data = next(data_iterator)
|
357 |
else:
|
358 |
data = None
|
359 |
data_b = mpu.broadcast_data(keys, data, datatype)
|
|
|
360 |
# Unpack.
|
361 |
tokens = data_b["text"].long()
|
362 |
types = data_b["types"].long()
|
|
|
364 |
loss_mask = data_b["loss_mask"].float()
|
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())
|
|
|
372 |
# Unpack.
|
373 |
tokens = data["input_ids"].long()
|
374 |
padding_mask = data["attention_mask"].long()
|
|
|
386 |
sentence_order = data["next_sentence_label"].long()
|
387 |
else:
|
388 |
sentence_order = None
|
|
|
389 |
return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
|
|
|
390 |
if megatron_dataset_flag:
|
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()
|
399 |
lm_loss = torch.sum(lm_loss_.view(-1) * loss_mask.reshape(-1)) / loss_mask.sum()
|
|
|
400 |
if sop_logits is not None:
|
401 |
sop_loss = F.cross_entropy(sop_logits.view(-1, 2).float(), sentence_order.view(-1), ignore_index=-1)
|
402 |
sop_loss = sop_loss.float()
|
403 |
loss = lm_loss + sop_loss
|
404 |
averaged_losses = average_losses_across_data_parallel_group([lm_loss, sop_loss])
|
405 |
return loss, {"lm loss": averaged_losses[0], "sop loss": averaged_losses[1]}
|
|
|
406 |
else:
|
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
|
|
|
420 |
loss = loss_fct(logits, labels)
|
421 |
averaged_losses = average_losses_across_data_parallel_group([loss])
|
422 |
return loss, {"loss": averaged_losses[0]}
|
|
|
423 |
if pretraining_flag:
|
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."""
|
|
|
437 |
else:
|
438 |
logits = model(tokens, padding_mask, tokentype_ids=types)
|
439 |
return logits, partial(self.loss_func, labels)
|
|
|
440 |
return forward_step
|
|
|
|
|
441 |
class GPTTrainStep(AbstractTrainStep):
|
442 |
"""
|
443 |
GPT train step class.
|
|
|
444 |
Args:
|
445 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
446 |
"""
|
|
|
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"]
|
468 |
datatype = torch.int64
|
|
|
469 |
# Broadcast data.
|
470 |
if data_iterator is not None:
|
471 |
data = next(data_iterator)
|
472 |
else:
|
473 |
data = None
|
474 |
data_b = mpu.broadcast_data(keys, data, datatype)
|
|
|
475 |
# Unpack.
|
476 |
tokens_ = data_b["text"].long()
|
477 |
labels = tokens_[:, 1:].contiguous()
|
478 |
tokens = tokens_[:, :-1].contiguous()
|
|
|
479 |
# Get the masks and postition ids.
|
480 |
attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids(
|
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())
|
|
|
488 |
tokens_ = data["input_ids"].long()
|
489 |
padding = torch.zeros((tokens_.shape[0], 1), dtype=tokens_.dtype, device=tokens_.device) + self.eod_token
|
490 |
tokens_ = torch.concat([tokens_, padding], dim=1)
|
|
|
495 |
tokens, self.eod_token, self.reset_position_ids, self.reset_attention_mask, True
|
496 |
)
|
497 |
return tokens, labels, loss_mask, attention_mask, position_ids
|
|
|
498 |
if megatron_dataset_flag:
|
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
|
|
|
509 |
losses = losses.float()
|
510 |
loss_mask = loss_mask.view(-1).float()
|
511 |
loss = torch.sum(losses.view(-1) * loss_mask) / loss_mask.sum()
|
|
|
512 |
# Reduce loss for logging.
|
513 |
averaged_loss = average_losses_across_data_parallel_group([loss])
|
|
|
514 |
output_dict = {"lm loss": averaged_loss[0]}
|
515 |
if args.return_logits:
|
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)
|
524 |
output_tensor = model(tokens, position_ids, attention_mask, labels=labels)
|
|
|
525 |
return output_tensor, partial(self.loss_func, loss_mask)
|
|
|
526 |
return forward_step
|
|
|
|
|
527 |
class T5TrainStep(AbstractTrainStep):
|
528 |
"""
|
529 |
T5 train step class.
|
|
|
530 |
Args:
|
531 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
532 |
"""
|
|
|
539 |
self.model_output_class = None
|
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.
|
|
|
551 |
# Convert attention mask to binary:
|
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
|
|
|
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
|
|
|
575 |
# Broadcast data.
|
576 |
if data_iterator is not None:
|
577 |
data = next(data_iterator)
|
578 |
else:
|
579 |
data = None
|
580 |
data_b = mpu.broadcast_data(keys, data, datatype)
|
|
|
581 |
# Unpack.
|
582 |
tokens_enc = data_b["text_enc"].long()
|
583 |
tokens_dec = data_b["text_dec"].long()
|
584 |
labels = data_b["labels"].long()
|
585 |
loss_mask = data_b["loss_mask"].float()
|
|
|
586 |
enc_mask = data_b["enc_mask"] < 0.5
|
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())
|
|
|
594 |
tokens_enc = data["input_ids"].long()
|
595 |
labels = data["labels"].long()
|
596 |
loss_mask = (labels != -100).to(torch.float)
|
|
|
606 |
enc_dec_mask = T5TrainStep.get_enc_dec_mask(
|
607 |
data["attention_mask"].long(), tokens_dec.shape[1], tokens_dec.device
|
608 |
)
|
|
|
609 |
return tokens_enc, tokens_dec, loss_mask, labels, enc_mask, dec_mask, enc_dec_mask
|
|
|
610 |
if megatron_dataset_flag:
|
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."""
|
|
|
630 |
output_tensor = model(
|
631 |
tokens_enc, tokens_dec, enc_mask, dec_mask, enc_dec_mask, tokentype_ids=None, lm_labels=lm_labels
|
632 |
)
|
|
|
633 |
return output_tensor, partial(self.loss_func, loss_mask)
|
|
|
634 |
return forward_step
|
|
|
|
|
635 |
# intialize megatron setup
|
636 |
def initialize(accelerator, extra_args_provider=None, args_defaults={}):
|
637 |
accelerator.print("Initializing Megatron-LM")
|
638 |
assert torch.cuda.is_available(), "Megatron requires CUDA."
|
|
|
639 |
# Parse arguments
|
640 |
args = parse_args(extra_args_provider, ignore_unknown_args=True)
|
|
|
641 |
# Set defaults
|
642 |
for key, value in args_defaults.items():
|
643 |
if getattr(args, key, None) is not None:
|
|
|
650 |
flush=True,
|
651 |
)
|
652 |
setattr(args, key, value)
|
|
|
653 |
if args.use_checkpoint_args or args_defaults.get("use_checkpoint_args", False):
|
654 |
assert args.load is not None, "--use-checkpoints-args requires --load argument"
|
655 |
load_args_from_checkpoint(args)
|
|
|
656 |
validate_args(args)
|
|
|
657 |
# set global args, build tokenizer, and set adlr-autoresume,
|
658 |
# tensorboard-writer, and timers.
|
659 |
set_global_variables(args)
|
|
|
660 |
# torch.distributed initialization
|
661 |
def finish_mpu_init():
|
662 |
args = get_args()
|
|
|
670 |
assert args.local_rank == device, "expected local-rank to be the same as rank % device-count."
|
671 |
else:
|
672 |
args.local_rank = device
|
|
|
673 |
# Set the tensor model-parallel, pipeline model-parallel, and
|
674 |
# data-parallel communicators.
|
675 |
if mpu.model_parallel_is_initialized():
|
|
|
681 |
args.virtual_pipeline_model_parallel_size,
|
682 |
args.pipeline_model_parallel_split_rank,
|
683 |
)
|
|
|
684 |
# Random seeds for reproducibility.
|
685 |
if args.rank == 0:
|
686 |
print("> setting random seeds to {} ...".format(args.seed))
|
687 |
_set_random_seed(args.seed, args.data_parallel_random_init)
|
|
|
688 |
args = get_args()
|
|
|
689 |
# Megatron's MPU is the master. Complete initialization right away.
|
690 |
finish_mpu_init()
|
|
|
691 |
# Autoresume.
|
692 |
_init_autoresume()
|
|
|
693 |
# Compile dependencies.
|
694 |
_compile_dependencies()
|
|
|
695 |
# Set pytorch JIT layer fusion options and warmup JIT functions.
|
696 |
set_jit_fusion_options()
|
697 |
args = get_args()
|
|
|
701 |
else:
|
702 |
args.bert_binary_head = False
|
703 |
args.iteration = 0
|
|
|
|
|
704 |
class MegatronEngine(torch.nn.Module):
|
705 |
"""
|
706 |
Megatron-LM model wrapper
|
|
|
707 |
Args:
|
708 |
accelerator (:class:`~accelerate.Accelerator`): The accelerator object to use.
|
709 |
model: Megatron-LM model
|
|
|
730 |
else:
|
731 |
raise ValueError(f"Unsupported model type: {args.model_type_name}")
|
732 |
self.optimizer.skipped_iter = False
|
|
|
733 |
# Tracking loss.
|
734 |
self.total_loss_dict = {}
|
735 |
self.eval_total_loss_dict = {}
|
|
|
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
|
|
|
750 |
Args:
|
751 |
batch_data (:obj:`dict`): The batch data to train on.
|
752 |
"""
|
753 |
args = get_args()
|
754 |
timers = get_timers()
|
|
|
755 |
if len(batch_data) > 0:
|
756 |
data_chunks = []
|
757 |
if args.num_micro_batches > 1:
|
|
|
764 |
)
|
765 |
else:
|
766 |
data_chunks = [batch_data]
|
|
|
767 |
if len(self.module) > 1:
|
768 |
batch_data_iterator = (
|
769 |
[iter(data_chunks) for _ in range(len(self.module))]
|
|
|
772 |
)
|
773 |
else:
|
774 |
batch_data_iterator = iter(data_chunks) if len(batch_data) > 0 else None
|
|
|
775 |
# Set grad to zero.
|
776 |
if args.DDP_impl == "local" and args.use_contiguous_buffers_in_local_ddp:
|
777 |
for partition in self.module:
|
778 |
partition.zero_grad_buffer()
|
779 |
self.optimizer.zero_grad()
|
|
|
780 |
# Forward pass.
|
781 |
forward_backward_func = get_forward_backward_func()
|
782 |
losses_reduced = forward_backward_func(
|
|
|
787 |
None,
|
788 |
forward_only=False,
|
789 |
)
|
|
|
790 |
# Empty unused memory.
|
791 |
if args.empty_unused_memory_level >= 1:
|
792 |
torch.cuda.empty_cache()
|
|
|
793 |
# Reduce gradients.
|
794 |
timers("backward-reduce-model-grads").start()
|
795 |
self.optimizer.reduce_model_grads(args, timers)
|
796 |
timers("backward-reduce-model-grads").stop()
|
|
|
797 |
# Update parameters.
|
798 |
timers("optimizer").start()
|
799 |
update_successful, grad_norm, num_zeros_in_grad = self.optimizer.step(args, timers)
|
800 |
timers("optimizer").stop()
|
|
|
801 |
# Gather params.
|
802 |
if update_successful:
|
803 |
timers("backward-gather-model-params").start()
|
804 |
self.optimizer.gather_model_params(args, timers)
|
805 |
timers("backward-gather-model-params").stop()
|
|
|
806 |
# Update learning rate.
|
807 |
if update_successful:
|
808 |
if self.scheduler is not None:
|
|
|
811 |
skipped_iter = 0
|
812 |
else:
|
813 |
skipped_iter = 1
|
|
|
814 |
self.optimizer.skipped_iter = not update_successful
|
|
|
815 |
# Empty unused memory.
|
816 |
if args.empty_unused_memory_level >= 2:
|
817 |
torch.cuda.empty_cache()
|
|
|
818 |
args.consumed_train_samples += (
|
819 |
mpu.get_data_parallel_world_size() * args.micro_batch_size * get_num_microbatches()
|
820 |
)
|
|
|
821 |
if mpu.is_pipeline_last_stage(ignore_virtual=True):
|
822 |
# Average loss across microbatches.
|
823 |
loss_reduced = {}
|
|
|
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
|
|
|
835 |
Args:
|
836 |
batch_data (:obj:`dict`): The batch data to evaluate on.
|
837 |
"""
|
|
|
844 |
)
|
845 |
else:
|
846 |
data_chunks = [batch_data]
|
|
|
847 |
if len(self.module) > 1:
|
848 |
batch_data_iterator = [iter(data_chunks) for _ in range(len(self.module))]
|
849 |
else:
|
|
|
860 |
# Empty unused memory
|
861 |
if args.empty_unused_memory_level >= 1:
|
862 |
torch.cuda.empty_cache()
|
|
|
863 |
args.consumed_valid_samples += (
|
864 |
mpu.get_data_parallel_world_size() * args.micro_batch_size * get_num_microbatches()
|
865 |
)
|
|
|
866 |
if mpu.is_pipeline_last_stage(ignore_virtual=True):
|
867 |
# Average loss across microbatches.
|
868 |
loss_reduced = {}
|
|
|
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`:
|
|
|
922 |
self.eval_total_loss_dict[key + "_num_iters"] = self.eval_total_loss_dict.get(
|
923 |
key + "_num_iters", torch.cuda.FloatTensor([0.0])
|
924 |
) + torch.cuda.FloatTensor([1.0])
|
|
|
925 |
loss = torch.tensor(0.0, device=args.local_rank)
|
926 |
for key in loss_dict:
|
927 |
if len(loss_dict[key].shape) == 0:
|
928 |
loss += loss_dict[key]
|
|
|
929 |
logits = None
|
930 |
if "logits" in loss_dict:
|
931 |
logits = loss_dict["logits"]
|
|
|
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:
|
|
|
952 |
writer.add_scalar(f"{key} validation", value.item(), self.iteration)
|
953 |
if args.pretraining_flag:
|
954 |
writer.add_scalar(f"{key} validation ppl", ppl, self.iteration)
|
|
|
955 |
length = len(string) + 1
|
956 |
print_rank_last("-" * length)
|
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 |
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 |
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,
|
|
|
991 |
"""
|
992 |
Generate method for GPT2 model. This method is used for inference. Supports both greedy and beam search along
|
993 |
with sampling. Refer the Megatron-LM repo for more details
|
|
|
994 |
Args:
|
995 |
inputs (torch.Tensor): input ids
|
996 |
attention_mask (torch.Tensor, optional): attention mask. Defaults to None.
|
|
|
1009 |
args = get_args()
|
1010 |
if args.model_type_name != "gpt":
|
1011 |
raise NotImplementedError("Generate method is not implemented for this model")
|
|
|
1012 |
if args.data_parallel_size > 1:
|
1013 |
raise ValueError("Generate method requires data parallelism to be 1")
|
|
|
1014 |
if args.sequence_parallel:
|
1015 |
raise ValueError("Generate method requires sequence parallelism to be False")
|
|
|
1016 |
if args.recompute_granularity is not None:
|
1017 |
raise ValueError("Checkpoint activations cannot be set for inference")
|
|
|
1018 |
if args.vocab_file is None:
|
1019 |
raise ValueError("Vocab file is required for inference")
|
|
|
1020 |
# Prepare inputs
|
1021 |
if max_length is None and max_new_tokens is None:
|
1022 |
raise ValueError("`max_length` or `max_new_tokens` are required for inference")
|
|
|
1023 |
if temperature is None:
|
1024 |
temperature = 1.0
|
1025 |
elif not (0.0 < temperature <= 100.0):
|
1026 |
raise ValueError("temperature must be a positive number less than or equal to 100.0")
|
|
|
1027 |
if top_k is None:
|
1028 |
top_k = 0
|
1029 |
elif not (0 <= top_k <= 1000):
|
1030 |
raise ValueError("top_k must be a positive number less than or equal to 1000")
|
|
|
1031 |
if top_p is None:
|
1032 |
top_p = 0.0
|
1033 |
elif top_p > 0.0 and top_k > 0.0:
|
|
|
1035 |
else:
|
1036 |
if not (0.0 <= top_p <= 1.0):
|
1037 |
raise ValueError("top_p must be less than or equal to 1.0")
|
|
|
1038 |
top_p_decay = kwargs.get("top_p_decay", 0.0)
|
1039 |
if not (0.0 <= top_p_decay <= 1.0):
|
1040 |
raise ValueError("top_p_decay must be less than or equal to 1.0")
|
|
|
1041 |
top_p_bound = kwargs.get("top_p_bound", 0.0)
|
1042 |
if not (0.0 <= top_p_bound <= 1.0):
|
1043 |
raise ValueError("top_p_bound must be less than or equal to 1.0")
|
|
|
1044 |
add_BOS = kwargs.get("add_BOS", False)
|
1045 |
if not (isinstance(add_BOS, bool)):
|
1046 |
raise ValueError("add_BOS must be a boolean")
|
|
|
1047 |
beam_width = num_beams
|
1048 |
if beam_width is not None:
|
1049 |
if not isinstance(beam_width, int):
|
|
|
1052 |
raise ValueError("beam_width must be greater than 0")
|
1053 |
if inputs.shape[0] > 1:
|
1054 |
return "When doing beam_search, batch size must be 1"
|
|
|
1055 |
tokenizer = get_tokenizer()
|
|
|
1056 |
stop_token = kwargs.get("stop_token", tokenizer.eod)
|
1057 |
if stop_token is not None:
|
1058 |
if not isinstance(stop_token, int):
|
1059 |
raise ValueError("stop_token must be an integer")
|
|
|
1060 |
if length_penalty is None:
|
1061 |
length_penalty = 1.0
|
|
|
1062 |
sizes_list = None
|
1063 |
prompts_tokens_tensor = None
|
1064 |
prompts_length_tensor = None
|
|
|
1068 |
prompts_length_tensor = torch.cuda.LongTensor([inputs.shape[1]] * inputs.shape[0])
|
1069 |
else:
|
1070 |
prompts_length_tensor = attention_mask.sum(axis=-1).cuda()
|
|
|
1071 |
if max_new_tokens is None:
|
1072 |
max_new_tokens = max_length - inputs.shape[1]
|
1073 |
if max_new_tokens <= 0:
|
1074 |
raise ValueError("max_new_tokens must be greater than 0")
|
|
|
1075 |
if add_BOS:
|
1076 |
max_length = max_new_tokens + inputs.shape[1] + 1
|
1077 |
# making sure that `max_length` is a multiple of 4 to leverage fused kernels
|
|
|
1088 |
max_new_tokens = max_length - inputs.shape[1]
|
1089 |
padding = torch.cuda.LongTensor([[tokenizer.eod] * max_new_tokens] * inputs.shape[0])
|
1090 |
prompts_tokens_tensor = torch.concat([inputs.cuda(), padding], axis=-1)
|
|
|
1091 |
# We need the sizes of these tensors for the boradcast
|
1092 |
sizes_list = [
|
1093 |
prompts_tokens_tensor.size(0), # Batch size
|
1094 |
prompts_tokens_tensor.size(1),
|
1095 |
] # Sequence lenght
|
|
|
1096 |
# First, broadcast the sizes.
|
1097 |
sizes_tensor = broadcast_int_list(2, int_list=sizes_list, rank=0)
|
|
|
1098 |
# Now that we have the sizes, we can boradcast the tokens
|
1099 |
# and length tensors.
|
1100 |
sizes = sizes_tensor.tolist()
|
1101 |
context_tokens_tensor = broadcast_tensor(sizes, torch.int64, tensor=prompts_tokens_tensor, rank=0)
|
1102 |
context_length_tensor = broadcast_tensor(sizes[0], torch.int64, tensor=prompts_length_tensor, rank=0)
|
|
|
1103 |
# Run the inference
|
1104 |
random_seed = kwargs.get("random_seed", 0)
|
1105 |
torch.random.manual_seed(random_seed)
|
|
|
1128 |
use_eod_token_for_early_termination=True,
|
1129 |
)
|
1130 |
return tokens
|
|
|
|
|
1131 |
# other utilities
|
1132 |
def avg_losses_across_data_parallel_group(losses):
|
1133 |
"""
|
1134 |
Average losses across data parallel group.
|
|
|
1135 |
Args:
|
1136 |
losses (List[Tensor]): List of losses to average across data parallel group.
|
1137 |
"""
|
1138 |
return average_losses_across_data_parallel_group(losses)
|
|
|
|
|
1139 |
def gather_across_data_parallel_groups(tensor):
|
1140 |
"""
|
1141 |
Recursively gather tensor in a nested list/tuple/dictionary of tensors from data parallel ranks.
|
|
|
1142 |
Args:
|
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:
|
|
|
1152 |
]
|
1153 |
torch.distributed.all_gather(output_tensors, tensor, group=mpu.get_data_parallel_group())
|
1154 |
return torch.cat(output_tensors, dim=0)
|
|
|
1155 |
return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)
|
src/utils/memory.py
CHANGED
@@ -6,19 +6,15 @@ def release_memory(*objects):
|
|
6 |
"""
|
7 |
Releases memory from `objects` by setting them to `None` and calls `gc.collect()` and `torch.cuda.empty_cache()`.
|
8 |
Returned objects should be reassigned to the same variables.
|
9 |
-
|
10 |
Args:
|
11 |
objects (`Iterable`):
|
12 |
An iterable of objects
|
13 |
Returns:
|
14 |
A list of `None` objects to replace `objects`
|
15 |
-
|
16 |
Example:
|
17 |
-
|
18 |
```python
|
19 |
>>> import torch
|
20 |
>>> from accelerate.utils import release_memory
|
21 |
-
|
22 |
>>> a = torch.ones(1000, 1000).cuda()
|
23 |
>>> b = torch.ones(1000, 1000).cuda()
|
24 |
>>> a, b = release_memory(a, b)
|
@@ -36,12 +32,9 @@ def release_memory(*objects):
|
|
36 |
else:
|
37 |
torch.cuda.empty_cache()
|
38 |
return objects
|
39 |
-
|
40 |
-
|
41 |
def should_reduce_batch_size(exception: Exception) -> bool:
|
42 |
"""
|
43 |
Checks if `exception` relates to CUDA out-of-memory, CUDNN not supported, or CPU out-of-memory
|
44 |
-
|
45 |
Args:
|
46 |
exception (`Exception`):
|
47 |
An exception
|
@@ -54,40 +47,28 @@ def should_reduce_batch_size(exception: Exception) -> bool:
|
|
54 |
if isinstance(exception, RuntimeError) and len(exception.args) == 1:
|
55 |
return any(err in exception.args[0] for err in _statements)
|
56 |
return False
|
57 |
-
|
58 |
-
|
59 |
def find_executable_batch_size(function: callable = None, starting_batch_size: int = 128):
|
60 |
"""
|
61 |
A basic decorator that will try to execute `function`. If it fails from exceptions related to out-of-memory or
|
62 |
CUDNN, the batch size is cut in half and passed to `function`
|
63 |
-
|
64 |
`function` must take in a `batch_size` parameter as its first argument.
|
65 |
-
|
66 |
Args:
|
67 |
function (`callable`, *optional*):
|
68 |
A function to wrap
|
69 |
starting_batch_size (`int`, *optional*):
|
70 |
The batch size to try and fit into memory
|
71 |
-
|
72 |
Example:
|
73 |
-
|
74 |
```python
|
75 |
>>> from accelerate.utils import find_executable_batch_size
|
76 |
-
|
77 |
-
|
78 |
>>> @find_executable_batch_size(starting_batch_size=128)
|
79 |
... def train(batch_size, model, optimizer):
|
80 |
... ...
|
81 |
-
|
82 |
-
|
83 |
>>> train(model, optimizer)
|
84 |
```
|
85 |
"""
|
86 |
if function is None:
|
87 |
return functools.partial(find_executable_batch_size, starting_batch_size=starting_batch_size)
|
88 |
-
|
89 |
batch_size = starting_batch_size
|
90 |
-
|
91 |
def decorator(*args, **kwargs):
|
92 |
nonlocal batch_size
|
93 |
gc.collect()
|
@@ -122,5 +103,4 @@ def find_executable_batch_size(function: callable = None, starting_batch_size: i
|
|
122 |
batch_size //= 2
|
123 |
else:
|
124 |
raise
|
125 |
-
|
126 |
return decorator
|
|
|
6 |
"""
|
7 |
Releases memory from `objects` by setting them to `None` and calls `gc.collect()` and `torch.cuda.empty_cache()`.
|
8 |
Returned objects should be reassigned to the same variables.
|
|
|
9 |
Args:
|
10 |
objects (`Iterable`):
|
11 |
An iterable of objects
|
12 |
Returns:
|
13 |
A list of `None` objects to replace `objects`
|
|
|
14 |
Example:
|
|
|
15 |
```python
|
16 |
>>> import torch
|
17 |
>>> from accelerate.utils import release_memory
|
|
|
18 |
>>> a = torch.ones(1000, 1000).cuda()
|
19 |
>>> b = torch.ones(1000, 1000).cuda()
|
20 |
>>> a, b = release_memory(a, b)
|
|
|
32 |
else:
|
33 |
torch.cuda.empty_cache()
|
34 |
return objects
|
|
|
|
|
35 |
def should_reduce_batch_size(exception: Exception) -> bool:
|
36 |
"""
|
37 |
Checks if `exception` relates to CUDA out-of-memory, CUDNN not supported, or CPU out-of-memory
|
|
|
38 |
Args:
|
39 |
exception (`Exception`):
|
40 |
An exception
|
|
|
47 |
if isinstance(exception, RuntimeError) and len(exception.args) == 1:
|
48 |
return any(err in exception.args[0] for err in _statements)
|
49 |
return False
|
|
|
|
|
50 |
def find_executable_batch_size(function: callable = None, starting_batch_size: int = 128):
|
51 |
"""
|
52 |
A basic decorator that will try to execute `function`. If it fails from exceptions related to out-of-memory or
|
53 |
CUDNN, the batch size is cut in half and passed to `function`
|
|
|
54 |
`function` must take in a `batch_size` parameter as its first argument.
|
|
|
55 |
Args:
|
56 |
function (`callable`, *optional*):
|
57 |
A function to wrap
|
58 |
starting_batch_size (`int`, *optional*):
|
59 |
The batch size to try and fit into memory
|
|
|
60 |
Example:
|
|
|
61 |
```python
|
62 |
>>> from accelerate.utils import find_executable_batch_size
|
|
|
|
|
63 |
>>> @find_executable_batch_size(starting_batch_size=128)
|
64 |
... def train(batch_size, model, optimizer):
|
65 |
... ...
|
|
|
|
|
66 |
>>> train(model, optimizer)
|
67 |
```
|
68 |
"""
|
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()
|
|
|
103 |
batch_size //= 2
|
104 |
else:
|
105 |
raise
|
|
|
106 |
return decorator
|
src/utils/modeling.py
CHANGED
@@ -1,13 +1,9 @@
|
|
1 |
WEIGHTS_INDEX_NAME = "pytorch_model.bin.index.json"
|
2 |
-
|
3 |
logger = logging.getLogger(__name__)
|
4 |
-
|
5 |
-
|
6 |
def check_device_same(first_device, second_device):
|
7 |
"""
|
8 |
Utility method to check if two `torch` devices are similar. When dealing with CUDA devices, torch throws `False`
|
9 |
for `torch.device("cuda") == torch.device("cuda:0")` whereas they should be the same
|
10 |
-
|
11 |
Args:
|
12 |
first_device (`torch.device`):
|
13 |
First device to check
|
@@ -16,29 +12,21 @@ def check_device_same(first_device, second_device):
|
|
16 |
"""
|
17 |
if first_device.type != second_device.type:
|
18 |
return False
|
19 |
-
|
20 |
if first_device.type == "cuda" and first_device.index is None:
|
21 |
# In case the first_device is a cuda device and have
|
22 |
# the index attribute set to `None`, default it to `0`
|
23 |
first_device = torch.device("cuda", index=0)
|
24 |
-
|
25 |
if second_device.type == "cuda" and second_device.index is None:
|
26 |
# In case the second_device is a cuda device and have
|
27 |
# the index attribute set to `None`, default it to `0`
|
28 |
second_device = torch.device("cuda", index=0)
|
29 |
-
|
30 |
return first_device == second_device
|
31 |
-
|
32 |
-
|
33 |
def convert_file_size_to_int(size: Union[int, str]):
|
34 |
"""
|
35 |
Converts a size expressed as a string with digits an unit (like `"5MB"`) to an integer (in bytes).
|
36 |
-
|
37 |
Args:
|
38 |
size (`int` or `str`): The size to convert. Will be directly returned if an `int`.
|
39 |
-
|
40 |
Example:
|
41 |
-
|
42 |
```py
|
43 |
>>> convert_file_size_to_int("1MiB")
|
44 |
1048576
|
@@ -68,18 +56,13 @@ def convert_file_size_to_int(size: Union[int, str]):
|
|
68 |
mem_size = int_size // 8 if size.endswith("b") else int_size
|
69 |
except ValueError:
|
70 |
raise ValueError(err_msg)
|
71 |
-
|
72 |
if mem_size <= 0:
|
73 |
raise ValueError(err_msg)
|
74 |
return mem_size
|
75 |
-
|
76 |
-
|
77 |
def dtype_byte_size(dtype: torch.dtype):
|
78 |
"""
|
79 |
Returns the size (in bytes) occupied by one parameter of type `dtype`.
|
80 |
-
|
81 |
Example:
|
82 |
-
|
83 |
```py
|
84 |
>>> dtype_byte_size(torch.float32)
|
85 |
4
|
@@ -96,8 +79,6 @@ def dtype_byte_size(dtype: torch.dtype):
|
|
96 |
raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
|
97 |
bit_size = int(bit_search.groups()[0])
|
98 |
return bit_size // 8
|
99 |
-
|
100 |
-
|
101 |
def id_tensor_storage(tensor: torch.Tensor) -> Tuple[torch.device, int, int]:
|
102 |
"""
|
103 |
Unique identifier to a tensor storage. Multiple different tensors can share the same underlying storage. For
|
@@ -130,29 +111,21 @@ def id_tensor_storage(tensor: torch.Tensor) -> Tuple[torch.device, int, int]:
|
|
130 |
storage_ptr = 0
|
131 |
# On torch >=2.0 this is the tensor size
|
132 |
storage_size = tensor.nelement() * _SIZE[tensor.dtype]
|
133 |
-
|
134 |
return tensor.device, storage_ptr, storage_size
|
135 |
-
|
136 |
-
|
137 |
def shard_checkpoint(
|
138 |
state_dict: Dict[str, torch.Tensor], max_shard_size: Union[int, str] = "10GB", weights_name: str = WEIGHTS_NAME
|
139 |
):
|
140 |
"""
|
141 |
Splits a model state dictionary in sub-checkpoints so that the final size of each sub-checkpoint does not exceed a
|
142 |
given size.
|
143 |
-
|
144 |
The sub-checkpoints are determined by iterating through the `state_dict` in the order of its keys, so there is no
|
145 |
optimization made to make each sub-checkpoint as close as possible to the maximum size passed. For example, if the
|
146 |
limit is 10GB and we have weights of sizes [6GB, 6GB, 2GB, 6GB, 2GB, 2GB] they will get sharded as [6GB], [6+2GB],
|
147 |
[6+2+2GB] and not [6+2+2GB], [6+2GB], [6GB].
|
148 |
-
|
149 |
<Tip warning={true}>
|
150 |
-
|
151 |
If one of the model's weight is bigger that `max_sahrd_size`, it will end up in its own sub-checkpoint which will
|
152 |
have a size greater than `max_shard_size`.
|
153 |
-
|
154 |
</Tip>
|
155 |
-
|
156 |
Args:
|
157 |
state_dict (`Dict[str, torch.Tensor]`): The state dictionary of a model to save.
|
158 |
max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
|
@@ -162,12 +135,10 @@ def shard_checkpoint(
|
|
162 |
The name of the model save file.
|
163 |
"""
|
164 |
max_shard_size = convert_file_size_to_int(max_shard_size)
|
165 |
-
|
166 |
sharded_state_dicts = [{}]
|
167 |
last_block_size = 0
|
168 |
total_size = 0
|
169 |
storage_id_to_block = {}
|
170 |
-
|
171 |
for key, weight in state_dict.items():
|
172 |
# when bnb serialization is used the weights in the state dict can be strings
|
173 |
# check: https://github.com/huggingface/transformers/pull/24416 for more details
|
@@ -175,29 +146,23 @@ def shard_checkpoint(
|
|
175 |
continue
|
176 |
else:
|
177 |
storage_id = id_tensor_storage(weight)
|
178 |
-
|
179 |
# If a `weight` shares the same underlying storage as another tensor, we put `weight` in the same `block`
|
180 |
if storage_id in storage_id_to_block:
|
181 |
block_id = storage_id_to_block[storage_id]
|
182 |
sharded_state_dicts[block_id][key] = weight
|
183 |
continue
|
184 |
-
|
185 |
weight_size = weight.numel() * dtype_byte_size(weight.dtype)
|
186 |
-
|
187 |
# If this weight is going to tip up over the maximal size, we split.
|
188 |
if last_block_size + weight_size > max_shard_size:
|
189 |
sharded_state_dicts.append({})
|
190 |
last_block_size = 0
|
191 |
-
|
192 |
sharded_state_dicts[-1][key] = weight
|
193 |
last_block_size += weight_size
|
194 |
total_size += weight_size
|
195 |
storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
|
196 |
-
|
197 |
# If we only have one shard, we return it
|
198 |
if len(sharded_state_dicts) == 1:
|
199 |
return {weights_name: sharded_state_dicts[0]}, None
|
200 |
-
|
201 |
# Otherwise, let's build the index
|
202 |
weight_map = {}
|
203 |
shards = {}
|
@@ -209,13 +174,10 @@ def shard_checkpoint(
|
|
209 |
shards[shard_file] = shard
|
210 |
for key in shard.keys():
|
211 |
weight_map[key] = shard_file
|
212 |
-
|
213 |
# Add the metadata
|
214 |
metadata = {"total_size": total_size}
|
215 |
index = {"metadata": metadata, "weight_map": weight_map}
|
216 |
return shards, index
|
217 |
-
|
218 |
-
|
219 |
def set_module_tensor_to_device(
|
220 |
module: nn.Module,
|
221 |
tensor_name: str,
|
@@ -227,7 +189,6 @@ def set_module_tensor_to_device(
|
|
227 |
"""
|
228 |
A helper function to set a given tensor (parameter of buffer) of a module on a specific device (note that doing
|
229 |
`param.to(device)` creates a new tensor not linked to the parameter, which is why we need this function).
|
230 |
-
|
231 |
Args:
|
232 |
module (`torch.nn.Module`):
|
233 |
The module in which the tensor we want to move lives.
|
@@ -252,30 +213,24 @@ def set_module_tensor_to_device(
|
|
252 |
raise ValueError(f"{module} has no attribute {split}.")
|
253 |
module = new_module
|
254 |
tensor_name = splits[-1]
|
255 |
-
|
256 |
if tensor_name not in module._parameters and tensor_name not in module._buffers:
|
257 |
raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.")
|
258 |
is_buffer = tensor_name in module._buffers
|
259 |
old_value = getattr(module, tensor_name)
|
260 |
-
|
261 |
if old_value.device == torch.device("meta") and device not in ["meta", torch.device("meta")] and value is None:
|
262 |
raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {device}.")
|
263 |
-
|
264 |
if value is not None:
|
265 |
if old_value.shape != value.shape:
|
266 |
raise ValueError(
|
267 |
f'Trying to set a tensor of shape {value.shape} in "{tensor_name}" (which has shape {old_value.shape}), this look incorrect.'
|
268 |
)
|
269 |
-
|
270 |
if dtype is None:
|
271 |
# For compatibility with PyTorch load_state_dict which converts state dict dtype to existing dtype in model
|
272 |
value = value.to(old_value.dtype)
|
273 |
elif not str(value.dtype).startswith(("torch.uint", "torch.int", "torch.bool")):
|
274 |
value = value.to(dtype)
|
275 |
-
|
276 |
param = module._parameters[tensor_name] if tensor_name in module._parameters else None
|
277 |
param_cls = type(param)
|
278 |
-
|
279 |
device_quantization = None
|
280 |
with torch.no_grad():
|
281 |
# leave it on cpu first before moving them to cuda
|
@@ -296,7 +251,6 @@ def set_module_tensor_to_device(
|
|
296 |
if dtype is not None and device in ["meta", torch.device("meta")]:
|
297 |
if not str(old_value.dtype).startswith(("torch.uint", "torch.int", "torch.bool")):
|
298 |
new_value = new_value.to(dtype)
|
299 |
-
|
300 |
if not is_buffer:
|
301 |
module._parameters[tensor_name] = param_cls(new_value, requires_grad=old_value.requires_grad)
|
302 |
elif isinstance(value, torch.Tensor):
|
@@ -352,15 +306,12 @@ def set_module_tensor_to_device(
|
|
352 |
torch.npu.empty_cache()
|
353 |
else:
|
354 |
torch.cuda.empty_cache()
|
355 |
-
|
356 |
-
|
357 |
def named_module_tensors(
|
358 |
module: nn.Module, include_buffers: bool = True, recurse: bool = False, remove_non_persistent: bool = False
|
359 |
):
|
360 |
"""
|
361 |
A helper function that gathers all the tensors (parameters + buffers) of a given module. If `include_buffers=True`
|
362 |
it's the same as doing `module.named_parameters(recurse=recurse) + module.named_buffers(recurse=recurse)`.
|
363 |
-
|
364 |
Args:
|
365 |
module (`torch.nn.Module`):
|
366 |
The module we want the tensors on.
|
@@ -374,7 +325,6 @@ def named_module_tensors(
|
|
374 |
"""
|
375 |
for named_parameter in module.named_parameters(recurse=recurse):
|
376 |
yield named_parameter
|
377 |
-
|
378 |
if include_buffers:
|
379 |
non_persistent_buffers = set()
|
380 |
if remove_non_persistent:
|
@@ -383,12 +333,9 @@ def named_module_tensors(
|
|
383 |
name, _ = named_buffer
|
384 |
if name not in non_persistent_buffers:
|
385 |
yield named_buffer
|
386 |
-
|
387 |
-
|
388 |
def get_non_persistent_buffers(module: nn.Module, recurse: bool = False):
|
389 |
"""
|
390 |
Gather all non persistent buffers of a given modules into a set
|
391 |
-
|
392 |
Args:
|
393 |
module (`nn.Module`):
|
394 |
The module we want the non persistent buffers on.
|
@@ -399,10 +346,7 @@ def get_non_persistent_buffers(module: nn.Module, recurse: bool = False):
|
|
399 |
if recurse:
|
400 |
for _, m in module.named_modules():
|
401 |
non_persistent_buffers_set |= m._non_persistent_buffers_set
|
402 |
-
|
403 |
return non_persistent_buffers_set
|
404 |
-
|
405 |
-
|
406 |
class FindTiedParametersResult(list):
|
407 |
"""
|
408 |
This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
|
@@ -410,19 +354,14 @@ class FindTiedParametersResult(list):
|
|
410 |
"""
|
411 |
def __init__(self, *args, **kwargs):
|
412 |
super().__init__(*args, **kwargs)
|
413 |
-
|
414 |
def values(self):
|
415 |
# TODO: at the next Transformers release (4.28.0) issue a deprecation warning here.
|
416 |
return sum([x[1:] for x in self], [])
|
417 |
-
|
418 |
-
|
419 |
def check_tied_parameters_in_config(model: nn.Module):
|
420 |
"""
|
421 |
Check if there is any indication in the given model that some weights should be tied.
|
422 |
-
|
423 |
Args:
|
424 |
model (`torch.nn.Module`): The model to inspect
|
425 |
-
|
426 |
Returns:
|
427 |
bool: True if the model needs to have tied weights
|
428 |
"""
|
@@ -430,7 +369,6 @@ def check_tied_parameters_in_config(model: nn.Module):
|
|
430 |
has_tied_word_embedding = False
|
431 |
has_tied_encoder_decoder = False
|
432 |
has_tied_module = False
|
433 |
-
|
434 |
if "PreTrainedModel" in [c.__name__ for c in inspect.getmro(model.__class__)]:
|
435 |
has_tied_word_embedding = (
|
436 |
hasattr(model, "config")
|
@@ -443,10 +381,7 @@ def check_tied_parameters_in_config(model: nn.Module):
|
|
443 |
and getattr(model.config, "tie_encoder_decoder", False)
|
444 |
)
|
445 |
has_tied_module = any(hasattr(module, "_tie_weights") for module in model.modules())
|
446 |
-
|
447 |
return any([has_tied_word_embedding, has_tied_encoder_decoder, has_tied_module])
|
448 |
-
|
449 |
-
|
450 |
def _get_param_device(param, device_map):
|
451 |
if param in device_map:
|
452 |
return device_map[param]
|
@@ -455,19 +390,14 @@ def _get_param_device(param, device_map):
|
|
455 |
raise ValueError(f"The `device_map` does not contain the module {param}.")
|
456 |
else:
|
457 |
return _get_param_device(parent_param, device_map)
|
458 |
-
|
459 |
-
|
460 |
def check_tied_parameters_on_same_device(tied_params, device_map):
|
461 |
"""
|
462 |
Check if tied parameters are on the same device
|
463 |
-
|
464 |
Args:
|
465 |
tied_params (`List[List[str]]`):
|
466 |
A list of lists of parameter names being all tied together.
|
467 |
-
|
468 |
device_map (`Dict[str, Union[int, str, torch.device]]`):
|
469 |
A map that specifies where each submodule should go.
|
470 |
-
|
471 |
"""
|
472 |
for tie_param in tied_params:
|
473 |
tie_param_devices = {}
|
@@ -478,31 +408,21 @@ def check_tied_parameters_on_same_device(tied_params, device_map):
|
|
478 |
f"Tied parameters are on different devices: {tie_param_devices}. "
|
479 |
"Please modify your custom device map or set `device_map='auto'`. "
|
480 |
)
|
481 |
-
|
482 |
-
|
483 |
def find_tied_parameters(model: nn.Module, **kwargs):
|
484 |
"""
|
485 |
Find the tied parameters in a given model.
|
486 |
-
|
487 |
<Tip warning={true}>
|
488 |
-
|
489 |
The signature accepts keyword arguments, but they are for the recursive part of this function and you should ignore
|
490 |
them.
|
491 |
-
|
492 |
</Tip>
|
493 |
-
|
494 |
Args:
|
495 |
model (`torch.nn.Module`): The model to inspect.
|
496 |
-
|
497 |
Returns:
|
498 |
List[List[str]]: A list of lists of parameter names being all tied together.
|
499 |
-
|
500 |
Example:
|
501 |
-
|
502 |
```py
|
503 |
>>> from collections import OrderedDict
|
504 |
>>> import torch.nn as nn
|
505 |
-
|
506 |
>>> model = nn.Sequential(OrderedDict([("linear1", nn.Linear(4, 4)), ("linear2", nn.Linear(4, 4))]))
|
507 |
>>> model.linear2.weight = model.linear1.weight
|
508 |
>>> find_tied_parameters(model)
|
@@ -513,7 +433,6 @@ def find_tied_parameters(model: nn.Module, **kwargs):
|
|
513 |
named_parameters = kwargs.get("named_parameters", None)
|
514 |
prefix = kwargs.get("prefix", "")
|
515 |
result = kwargs.get("result", {})
|
516 |
-
|
517 |
if named_parameters is None:
|
518 |
named_parameters = {n: p for n, p in model.named_parameters()}
|
519 |
else:
|
@@ -529,19 +448,14 @@ def find_tied_parameters(model: nn.Module, **kwargs):
|
|
529 |
if new_name not in result:
|
530 |
result[new_name] = []
|
531 |
result[new_name].append(full_name)
|
532 |
-
|
533 |
# Once we have treated direct parameters, we move to the child modules.
|
534 |
for name, child in model.named_children():
|
535 |
child_name = name if prefix == "" else f"{prefix}.{name}"
|
536 |
find_tied_parameters(child, named_parameters=named_parameters, prefix=child_name, result=result)
|
537 |
-
|
538 |
return FindTiedParametersResult([sorted([weight] + list(set(tied))) for weight, tied in result.items()])
|
539 |
-
|
540 |
-
|
541 |
def retie_parameters(model, tied_params):
|
542 |
"""
|
543 |
Reties tied parameters in a given model if the link was broken (for instance when adding hooks).
|
544 |
-
|
545 |
Args:
|
546 |
model (`torch.nn.Module`):
|
547 |
The model in which to retie parameters.
|
@@ -567,8 +481,6 @@ def retie_parameters(model, tied_params):
|
|
567 |
for split in splits[:-1]:
|
568 |
module = getattr(module, split)
|
569 |
setattr(module, splits[-1], param_to_tie)
|
570 |
-
|
571 |
-
|
572 |
def _get_proper_dtype(dtype: Union[str, torch.device]) -> torch.dtype:
|
573 |
"""
|
574 |
Just does torch.dtype(dtype) if necessary.
|
@@ -578,8 +490,6 @@ def _get_proper_dtype(dtype: Union[str, torch.device]) -> torch.dtype:
|
|
578 |
dtype = dtype.replace("torch.", "")
|
579 |
dtype = getattr(torch, dtype)
|
580 |
return dtype
|
581 |
-
|
582 |
-
|
583 |
def compute_module_sizes(
|
584 |
model: nn.Module,
|
585 |
dtype: Optional[Union[str, torch.device]] = None,
|
@@ -605,10 +515,7 @@ def compute_module_sizes(
|
|
605 |
name_parts = name.split(".")
|
606 |
for idx in range(len(name_parts) + 1):
|
607 |
module_sizes[".".join(name_parts[:idx])] += size
|
608 |
-
|
609 |
return module_sizes
|
610 |
-
|
611 |
-
|
612 |
def get_max_layer_size(
|
613 |
modules: List[Tuple[str, torch.nn.Module]], module_sizes: Dict[str, int], no_split_module_classes: List[str]
|
614 |
):
|
@@ -617,7 +524,6 @@ def get_max_layer_size(
|
|
617 |
definition of a layer being:
|
618 |
- a module with no direct children (just parameters and buffers)
|
619 |
- a module whose class name is in the list `no_split_module_classes`
|
620 |
-
|
621 |
Args:
|
622 |
modules (`List[Tuple[str, torch.nn.Module]]`):
|
623 |
The list of named modules where we want to determine the maximum layer size.
|
@@ -625,7 +531,6 @@ def get_max_layer_size(
|
|
625 |
A dictionary mapping each layer name to its size (as generated by `compute_module_sizes`).
|
626 |
no_split_module_classes (`List[str]`):
|
627 |
A list of class names for layers we don't want to be split.
|
628 |
-
|
629 |
Returns:
|
630 |
`Tuple[int, List[str]]`: The maximum size of a layer with the list of layer names realizing that maximum size.
|
631 |
"""
|
@@ -646,18 +551,14 @@ def get_max_layer_size(
|
|
646 |
else:
|
647 |
modules_to_treat = [(f"{module_name}.{n}", v) for n, v in modules_children] + modules_to_treat
|
648 |
return max_size, layer_names
|
649 |
-
|
650 |
-
|
651 |
def get_max_memory(max_memory: Optional[Dict[Union[int, str], Union[int, str]]] = None):
|
652 |
"""
|
653 |
Get the maximum memory available if nothing is passed, converts string to int otherwise.
|
654 |
"""
|
655 |
import psutil
|
656 |
-
|
657 |
if max_memory is None:
|
658 |
if not (torch.cuda.is_available() or is_npu_available() or is_xpu_available()):
|
659 |
max_memory = {}
|
660 |
-
|
661 |
else:
|
662 |
# Make sure CUDA is initialized on each GPU to have the right memory info.
|
663 |
if is_npu_available():
|
@@ -678,11 +579,9 @@ def get_max_memory(max_memory: Optional[Dict[Union[int, str], Union[int, str]]]
|
|
678 |
else:
|
679 |
max_memory["cpu"] = psutil.virtual_memory().available
|
680 |
return max_memory
|
681 |
-
|
682 |
for key in max_memory:
|
683 |
if isinstance(max_memory[key], str):
|
684 |
max_memory[key] = convert_file_size_to_int(max_memory[key])
|
685 |
-
|
686 |
# Need to sort the device by type to make sure that we allocate the gpu first.
|
687 |
# As gpu/npu/xpu are represented by int, we need to sort them first.
|
688 |
gpu_devices = [k for k in max_memory.keys() if isinstance(k, int)]
|
@@ -706,10 +605,7 @@ def get_max_memory(max_memory: Optional[Dict[Union[int, str], Union[int, str]]]
|
|
706 |
f"Device {k} is not recognized, available devices are integers(for GPU/XPU), 'mps', 'cpu' and 'disk'"
|
707 |
)
|
708 |
max_memory = {k: max_memory[k] for k in all_devices}
|
709 |
-
|
710 |
return max_memory
|
711 |
-
|
712 |
-
|
713 |
def clean_device_map(device_map: Dict[str, Union[int, str, torch.device]], module_name: str = ""):
|
714 |
"""
|
715 |
Cleans a device_map by grouping all submodules that go on the same device together.
|
@@ -721,21 +617,16 @@ def clean_device_map(device_map: Dict[str, Union[int, str, torch.device]], modul
|
|
721 |
for k in [k for k in device_map if k.startswith(prefix)]:
|
722 |
del device_map[k]
|
723 |
device_map[module_name] = values[0]
|
724 |
-
|
725 |
# Recurse over the children
|
726 |
children_modules = [k for k in device_map.keys() if k.startswith(prefix) and len(k) > len(module_name)]
|
727 |
idx = len(module_name.split(".")) + 1 if len(module_name) > 0 else 1
|
728 |
children_modules = set(".".join(k.split(".")[:idx]) for k in children_modules)
|
729 |
for child in children_modules:
|
730 |
clean_device_map(device_map, module_name=child)
|
731 |
-
|
732 |
return device_map
|
733 |
-
|
734 |
-
|
735 |
def load_offloaded_weights(model, index, offload_folder):
|
736 |
"""
|
737 |
Loads the weights from the offload folder into the model.
|
738 |
-
|
739 |
Args:
|
740 |
model (`torch.nn.Module`):
|
741 |
The model to load the weights into.
|
@@ -760,8 +651,6 @@ def load_offloaded_weights(model, index, offload_folder):
|
|
760 |
tensor_file = os.path.join(offload_folder, f"{param_name}.dat")
|
761 |
weight = load_offloaded_weight(tensor_file, metadata)
|
762 |
set_module_tensor_to_device(model, param_name, "cpu", value=weight, fp16_statistics=fp16_statistics)
|
763 |
-
|
764 |
-
|
765 |
def get_balanced_memory(
|
766 |
model: nn.Module,
|
767 |
max_memory: Optional[Dict[Union[int, str], Union[int, str]]] = None,
|
@@ -772,14 +661,10 @@ def get_balanced_memory(
|
|
772 |
):
|
773 |
"""
|
774 |
Compute a `max_memory` dictionary for [`infer_auto_device_map`] that will balance the use of each available GPU.
|
775 |
-
|
776 |
<Tip>
|
777 |
-
|
778 |
All computation is done analyzing sizes and dtypes of the model parameters. As a result, the model can be on the
|
779 |
meta device (as it would if initialized within the `init_empty_weights` context manager).
|
780 |
-
|
781 |
</Tip>
|
782 |
-
|
783 |
Args:
|
784 |
model (`torch.nn.Module`):
|
785 |
The model to analyze.
|
@@ -800,7 +685,6 @@ def get_balanced_memory(
|
|
800 |
# Get default / clean up max_memory
|
801 |
user_not_set_max_memory = max_memory is None
|
802 |
max_memory = get_max_memory(max_memory)
|
803 |
-
|
804 |
if is_npu_available():
|
805 |
num_devices = len([d for d in max_memory if torch.device(d).type == "npu" and max_memory[d] > 0])
|
806 |
elif is_xpu_available():
|
@@ -817,10 +701,8 @@ def get_balanced_memory(
|
|
817 |
)
|
818 |
else:
|
819 |
num_devices = len([d for d in max_memory if torch.device(d).type == "cuda" and max_memory[d] > 0])
|
820 |
-
|
821 |
if num_devices == 0:
|
822 |
return max_memory
|
823 |
-
|
824 |
if num_devices == 1:
|
825 |
# We cannot do low_zero on just one GPU, but we will still reserve some memory for the buffer
|
826 |
low_zero = False
|
@@ -834,10 +716,8 @@ def get_balanced_memory(
|
|
834 |
"You can set `max_memory` in to a higher value to use more memory (at your own risk)."
|
835 |
)
|
836 |
break # only one device
|
837 |
-
|
838 |
module_sizes = compute_module_sizes(model, dtype=dtype, special_dtypes=special_dtypes)
|
839 |
per_gpu = module_sizes[""] // (num_devices - 1 if low_zero else num_devices)
|
840 |
-
|
841 |
# We can't just set the memory to model_size // num_devices as it will end being too small: each GPU will get
|
842 |
# slightly less layers and some layers will end up offload at the end. So this function computes a buffer size to
|
843 |
# add which is the biggest of:
|
@@ -847,7 +727,6 @@ def get_balanced_memory(
|
|
847 |
no_split_module_classes = []
|
848 |
elif not isinstance(no_split_module_classes, (list, tuple)):
|
849 |
no_split_module_classes = [no_split_module_classes]
|
850 |
-
|
851 |
# Identify the size of the no_split_block modules
|
852 |
if len(no_split_module_classes) > 0:
|
853 |
no_split_children = {}
|
@@ -860,13 +739,11 @@ def get_balanced_memory(
|
|
860 |
class_name = submodule.__class__.__name__
|
861 |
if class_name in no_split_module_classes and class_name not in no_split_children:
|
862 |
no_split_children[class_name] = size
|
863 |
-
|
864 |
if set(no_split_children.keys()) == set(no_split_module_classes):
|
865 |
break
|
866 |
buffer = max(no_split_children.values()) if len(no_split_children) > 0 else 0
|
867 |
else:
|
868 |
buffer = 0
|
869 |
-
|
870 |
# Compute mean of final modules. In the first dict of module sizes, leaves are the parameters
|
871 |
leaves = [n for n in module_sizes if len([p for p in module_sizes if n == "" or p.startswith(n + ".")]) == 0]
|
872 |
module_sizes = {n: v for n, v in module_sizes.items() if n not in leaves}
|
@@ -875,7 +752,6 @@ def get_balanced_memory(
|
|
875 |
mean_leaves = int(sum([module_sizes[n] for n in leaves]) / max(len(leaves), 1))
|
876 |
buffer = int(1.25 * max(buffer, mean_leaves))
|
877 |
per_gpu += buffer
|
878 |
-
|
879 |
# Sorted list of GPUs id (we may have some gpu ids not included in the our max_memory list - let's ignore them)
|
880 |
gpus_idx_list = list(
|
881 |
sorted(
|
@@ -885,14 +761,10 @@ def get_balanced_memory(
|
|
885 |
# The last device is left with max_memory just in case the buffer is not enough.
|
886 |
for idx in gpus_idx_list[:-1]:
|
887 |
max_memory[idx] = min(max_memory[0] if low_zero and idx == 0 else per_gpu, max_memory[idx])
|
888 |
-
|
889 |
if low_zero:
|
890 |
min_zero = max(0, module_sizes[""] - sum([max_memory[i] for i in range(1, num_devices)]))
|
891 |
max_memory[0] = min(min_zero, max_memory[0])
|
892 |
-
|
893 |
return max_memory
|
894 |
-
|
895 |
-
|
896 |
def calculate_maximum_sizes(model: torch.nn.Module):
|
897 |
"Computes the total size of the model and its largest layer"
|
898 |
sizes = compute_module_sizes(model)
|
@@ -900,7 +772,6 @@ def calculate_maximum_sizes(model: torch.nn.Module):
|
|
900 |
no_split_modules = getattr(model, "_no_split_modules", None)
|
901 |
if no_split_modules is None:
|
902 |
no_split_modules = []
|
903 |
-
|
904 |
modules_to_treat = (
|
905 |
list(model.named_parameters(recurse=False))
|
906 |
+ list(model.named_children())
|
@@ -909,8 +780,6 @@ def calculate_maximum_sizes(model: torch.nn.Module):
|
|
909 |
largest_layer = get_max_layer_size(modules_to_treat, sizes, no_split_modules)
|
910 |
total_size = sizes[""]
|
911 |
return total_size, largest_layer
|
912 |
-
|
913 |
-
|
914 |
def infer_auto_device_map(
|
915 |
model: nn.Module,
|
916 |
max_memory: Optional[Dict[Union[int, str], Union[int, str]]] = None,
|
@@ -929,14 +798,10 @@ def infer_auto_device_map(
|
|
929 |
- if offload to the CPU is needed,we don't exceed the RAM available on the CPU.
|
930 |
- if offload to the disk is needed, there is always room left on the CPU to put back the layer offloaded on disk
|
931 |
that has the largest size.
|
932 |
-
|
933 |
<Tip>
|
934 |
-
|
935 |
All computation is done analyzing sizes and dtypes of the model parameters. As a result, the model can be on the
|
936 |
meta device (as it would if initialized within the `init_empty_weights` context manager).
|
937 |
-
|
938 |
</Tip>
|
939 |
-
|
940 |
Args:
|
941 |
model (`torch.nn.Module`):
|
942 |
The model to analyze.
|
@@ -961,12 +826,10 @@ def infer_auto_device_map(
|
|
961 |
no_split_module_classes = []
|
962 |
elif not isinstance(no_split_module_classes, (list, tuple)):
|
963 |
no_split_module_classes = [no_split_module_classes]
|
964 |
-
|
965 |
devices = list(max_memory.keys())
|
966 |
if "disk" not in devices:
|
967 |
devices.append("disk")
|
968 |
gpus = [device for device in devices if device not in ["cpu", "disk"]]
|
969 |
-
|
970 |
# Devices that need to keep space for a potential offloaded layer.
|
971 |
if "mps" in gpus:
|
972 |
main_devices = ["mps"]
|
@@ -974,19 +837,15 @@ def infer_auto_device_map(
|
|
974 |
main_devices = [gpus[0], "cpu"]
|
975 |
else:
|
976 |
main_devices = ["cpu"]
|
977 |
-
|
978 |
module_sizes = compute_module_sizes(model, dtype=dtype, special_dtypes=special_dtypes)
|
979 |
tied_parameters = find_tied_parameters(model)
|
980 |
-
|
981 |
if check_tied_parameters_in_config(model) and len(tied_parameters) == 0:
|
982 |
logger.warn(
|
983 |
"The model weights are not tied. Please use the `tie_weights` method before using the `infer_auto_device` function."
|
984 |
)
|
985 |
-
|
986 |
device_map = OrderedDict()
|
987 |
current_device = 0
|
988 |
current_memory_used = 0
|
989 |
-
|
990 |
# Direct submodules and parameters
|
991 |
modules_to_treat = (
|
992 |
list(model.named_parameters(recurse=False))
|
@@ -995,7 +854,6 @@ def infer_auto_device_map(
|
|
995 |
)
|
996 |
# Initialize maximum largest layer, to know which space to keep in memory
|
997 |
max_layer_size, max_layer_names = get_max_layer_size(modules_to_treat, module_sizes, no_split_module_classes)
|
998 |
-
|
999 |
# Ready ? This is going to be a bit messy.
|
1000 |
while len(modules_to_treat) > 0:
|
1001 |
name, module = modules_to_treat.pop(0)
|
@@ -1011,7 +869,6 @@ def infer_auto_device_map(
|
|
1011 |
)
|
1012 |
# Assess size needed
|
1013 |
module_size = module_sizes[name]
|
1014 |
-
|
1015 |
# We keep relevant tied parameters only: one of the tied parameters in the group is inside the current module
|
1016 |
# and the other is not.
|
1017 |
tied_param_goups = [
|
@@ -1025,7 +882,6 @@ def infer_auto_device_map(
|
|
1025 |
tied_params = sum([[p for p in tied_group if name not in p] for tied_group in tied_param_goups], [])
|
1026 |
if verbose and len(tied_params) > 0:
|
1027 |
print(f" So those parameters need to be taken into account {tied_params}")
|
1028 |
-
|
1029 |
device = devices[current_device]
|
1030 |
current_max_size = max_memory[device] if device != "disk" else None
|
1031 |
# Reduce max size available by the largest layer.
|
@@ -1059,7 +915,6 @@ def infer_auto_device_map(
|
|
1059 |
module_sizes,
|
1060 |
no_split_module_classes,
|
1061 |
)
|
1062 |
-
|
1063 |
# Case 2, it fits! We're not entirely out of the wood though, because we may have some tied parameters.
|
1064 |
elif len(tied_params) > 0:
|
1065 |
# First locate all tied modules
|
@@ -1074,12 +929,10 @@ def infer_auto_device_map(
|
|
1074 |
f" It looks like {name} is going to fit on {devices[current_device]} but we have tied "
|
1075 |
f"parameters to account for.\n - Names {tied_params}\n - Module names {tied_module_names}"
|
1076 |
)
|
1077 |
-
|
1078 |
# Let's see if it all fits first
|
1079 |
module_size_with_ties = module_size
|
1080 |
for tied_param, tied_module_name in zip(tied_params, tied_module_names):
|
1081 |
module_size_with_ties += module_sizes[tied_module_name] - module_sizes[tied_param]
|
1082 |
-
|
1083 |
if current_max_size is None or current_memory_used + module_size_with_ties <= current_max_size:
|
1084 |
# We really really fit!
|
1085 |
if verbose:
|
@@ -1094,7 +947,6 @@ def infer_auto_device_map(
|
|
1094 |
]
|
1095 |
modules_to_treat.pop(tied_module_index)
|
1096 |
device_map[tied_module_name] = devices[current_device]
|
1097 |
-
|
1098 |
else:
|
1099 |
# We don't fit with the tied modules. Next question is: can we split one of the tied modules to make it
|
1100 |
# smaller or do we need to go on the next device?
|
@@ -1109,13 +961,11 @@ def infer_auto_device_map(
|
|
1109 |
if len(tied_module_children) == 0 or tied_module.__class__.__name__ in no_split_module_classes:
|
1110 |
# can't break this one.
|
1111 |
continue
|
1112 |
-
|
1113 |
if verbose:
|
1114 |
print(f"Splitting {tied_module_name}.")
|
1115 |
tied_module_children = list(tied_module.named_parameters(recurse=False)) + tied_module_children
|
1116 |
tied_module_children = [(f"{tied_module_name}.{n}", v) for n, v in tied_module_children]
|
1117 |
tied_module_index = [i for i, (n, _) in enumerate(modules_to_treat) if n == tied_module_name][0]
|
1118 |
-
|
1119 |
modules_to_treat = (
|
1120 |
[(name, module)]
|
1121 |
+ modules_to_treat[:tied_module_index]
|
@@ -1130,7 +980,6 @@ def infer_auto_device_map(
|
|
1130 |
)
|
1131 |
split_happened = True
|
1132 |
break
|
1133 |
-
|
1134 |
if not split_happened:
|
1135 |
# If the tied module is not split, we go to the next device
|
1136 |
if verbose:
|
@@ -1138,7 +987,6 @@ def infer_auto_device_map(
|
|
1138 |
current_device += 1
|
1139 |
modules_to_treat = [(name, module)] + modules_to_treat
|
1140 |
current_memory_used = 0
|
1141 |
-
|
1142 |
else:
|
1143 |
if verbose:
|
1144 |
if current_max_size is None:
|
@@ -1150,16 +998,12 @@ def infer_auto_device_map(
|
|
1150 |
)
|
1151 |
current_memory_used += module_size
|
1152 |
device_map[name] = devices[current_device]
|
1153 |
-
|
1154 |
if clean_result:
|
1155 |
device_map = clean_device_map(device_map)
|
1156 |
return device_map
|
1157 |
-
|
1158 |
-
|
1159 |
def check_device_map(model: nn.Module, device_map: Dict[str, Union[int, str, torch.device]]):
|
1160 |
"""
|
1161 |
Checks a device map covers everything in a given model.
|
1162 |
-
|
1163 |
Args:
|
1164 |
model (`torch.nn.Module`): The model to check the device map against.
|
1165 |
device_map (`Dict[str, Union[int, str, torch.device]]`): The device map to check.
|
@@ -1180,13 +1024,10 @@ def check_device_map(model: nn.Module, device_map: Dict[str, Union[int, str, tor
|
|
1180 |
raise ValueError(
|
1181 |
f"The device_map provided does not give any device for the following parameters: {non_covered_params}"
|
1182 |
)
|
1183 |
-
|
1184 |
-
|
1185 |
def load_state_dict(checkpoint_file, device_map=None):
|
1186 |
"""
|
1187 |
Load a checkpoint from a given file. If the checkpoint is in the safetensors format and a device map is passed, the
|
1188 |
weights can be fast-loaded directly on the GPU.
|
1189 |
-
|
1190 |
Args:
|
1191 |
checkpoint_file (`str`): The path to the checkpoint to load.
|
1192 |
device_map (`Dict[str, Union[int, str, torch.device]]`, *optional*):
|
@@ -1197,14 +1038,12 @@ def load_state_dict(checkpoint_file, device_map=None):
|
|
1197 |
with safe_open(checkpoint_file, framework="pt") as f:
|
1198 |
metadata = f.metadata()
|
1199 |
weight_names = f.keys()
|
1200 |
-
|
1201 |
if metadata is None:
|
1202 |
logger.warn(
|
1203 |
f"The safetensors archive passed at {checkpoint_file} does not contain metadata. "
|
1204 |
"Make sure to save your model with the `save_pretrained` method. Defaulting to 'pt' metadata."
|
1205 |
)
|
1206 |
metadata = {"format": "pt"}
|
1207 |
-
|
1208 |
if metadata.get("format") not in ["pt", "tf", "flax"]:
|
1209 |
raise OSError(
|
1210 |
f"The safetensors archive passed at {checkpoint_file} does not contain the valid metadata. Make sure "
|
@@ -1218,12 +1057,10 @@ def load_state_dict(checkpoint_file, device_map=None):
|
|
1218 |
# if we only have one device we can load everything directly
|
1219 |
if len(set(device_map.values())) == 1:
|
1220 |
return safe_load_file(checkpoint_file, device=list(device_map.values())[0])
|
1221 |
-
|
1222 |
devices = list(set(device_map.values()) - {"disk"})
|
1223 |
# cpu device should always exist as fallback option
|
1224 |
if "cpu" not in devices:
|
1225 |
devices.append("cpu")
|
1226 |
-
|
1227 |
# For each device, get the weights that go there
|
1228 |
device_weights = {device: [] for device in devices}
|
1229 |
for module_name, device in device_map.items():
|
@@ -1231,7 +1068,6 @@ def load_state_dict(checkpoint_file, device_map=None):
|
|
1231 |
device_weights[device].extend(
|
1232 |
[k for k in weight_names if k == module_name or k.startswith(module_name + ".")]
|
1233 |
)
|
1234 |
-
|
1235 |
# all weights that haven't defined a device should be loaded on CPU
|
1236 |
device_weights["cpu"].extend([k for k in weight_names if k not in sum(device_weights.values(), [])])
|
1237 |
tensors = {}
|
@@ -1256,22 +1092,17 @@ def load_state_dict(checkpoint_file, device_map=None):
|
|
1256 |
progress_bar.update()
|
1257 |
if progress_bar is not None:
|
1258 |
progress_bar.close()
|
1259 |
-
|
1260 |
return tensors
|
1261 |
else:
|
1262 |
return torch.load(checkpoint_file, map_location=torch.device("cpu"))
|
1263 |
-
|
1264 |
-
|
1265 |
def get_state_dict_offloaded_model(model: nn.Module):
|
1266 |
"""
|
1267 |
Returns the state dictionary for an offloaded model via iterative onloading
|
1268 |
-
|
1269 |
Args:
|
1270 |
model (`torch.nn.Module`):
|
1271 |
The offloaded model we want to save
|
1272 |
"""
|
1273 |
from ..hooks import AlignDevicesHook
|
1274 |
-
|
1275 |
state_dict = {}
|
1276 |
placeholders = set()
|
1277 |
for name, module in model.named_modules():
|
@@ -1293,7 +1124,6 @@ def get_state_dict_offloaded_model(model: nn.Module):
|
|
1293 |
module._hf_hook.execution_device = original_device
|
1294 |
else:
|
1295 |
module_state_dict = module.state_dict()
|
1296 |
-
|
1297 |
for key in module_state_dict:
|
1298 |
# ignore placeholder parameters that are still on the meta device
|
1299 |
if module_state_dict[key].device == torch.device("meta"):
|
@@ -1306,10 +1136,7 @@ def get_state_dict_offloaded_model(model: nn.Module):
|
|
1306 |
placeholders.remove(key)
|
1307 |
if placeholders:
|
1308 |
logger.warning(f"The following tensors were not saved because they were still on meta device: {placeholders}")
|
1309 |
-
|
1310 |
return state_dict
|
1311 |
-
|
1312 |
-
|
1313 |
def load_checkpoint_in_model(
|
1314 |
model: nn.Module,
|
1315 |
checkpoint: Union[str, os.PathLike],
|
@@ -1324,14 +1151,10 @@ def load_checkpoint_in_model(
|
|
1324 |
"""
|
1325 |
Loads a (potentially sharded) checkpoint inside a model, potentially sending weights to a given device as they are
|
1326 |
loaded.
|
1327 |
-
|
1328 |
<Tip warning={true}>
|
1329 |
-
|
1330 |
Once loaded across devices, you still need to call [`dispatch_model`] on your model to make it able to run. To
|
1331 |
group the checkpoint loading and dispatch in one single call, use [`load_checkpoint_and_dispatch`].
|
1332 |
-
|
1333 |
</Tip>
|
1334 |
-
|
1335 |
Args:
|
1336 |
model (`torch.nn.Module`):
|
1337 |
The model in which we want to load a checkpoint.
|
@@ -1357,32 +1180,26 @@ def load_checkpoint_in_model(
|
|
1357 |
A list of the modules that we keep in `torch.float32` dtype.
|
1358 |
offload_8bit_bnb (`bool`, *optional*):
|
1359 |
Whether or not to enable offload of 8-bit modules on cpu/disk.
|
1360 |
-
|
1361 |
"""
|
1362 |
if offload_8bit_bnb:
|
1363 |
from .bnb import quantize_and_offload_8bit
|
1364 |
-
|
1365 |
tied_params = find_tied_parameters(model)
|
1366 |
-
|
1367 |
if check_tied_parameters_in_config(model) and len(tied_params) == 0:
|
1368 |
logger.warn(
|
1369 |
"The model weights are not tied. Please use the `tie_weights` method before using the `infer_auto_device` function."
|
1370 |
)
|
1371 |
if device_map is not None:
|
1372 |
check_tied_parameters_on_same_device(tied_params, device_map)
|
1373 |
-
|
1374 |
if offload_folder is None and device_map is not None and "disk" in device_map.values():
|
1375 |
raise ValueError(
|
1376 |
"At least one of the model submodule will be offloaded to disk, please pass along an `offload_folder`."
|
1377 |
)
|
1378 |
elif offload_folder is not None and device_map is not None and "disk" in device_map.values():
|
1379 |
os.makedirs(offload_folder, exist_ok=True)
|
1380 |
-
|
1381 |
if isinstance(dtype, str):
|
1382 |
# We accept "torch.float16" or just "float16"
|
1383 |
dtype = dtype.replace("torch.", "")
|
1384 |
dtype = getattr(torch, dtype)
|
1385 |
-
|
1386 |
checkpoint_files = None
|
1387 |
index_filename = None
|
1388 |
if os.path.isfile(checkpoint):
|
@@ -1416,24 +1233,19 @@ def load_checkpoint_in_model(
|
|
1416 |
"`checkpoint` should be the path to a file containing a whole state dict, or the index of a sharded "
|
1417 |
f"checkpoint, or a folder containing a sharded checkpoint or the whole state dict, but got {checkpoint}."
|
1418 |
)
|
1419 |
-
|
1420 |
if index_filename is not None:
|
1421 |
checkpoint_folder = os.path.split(index_filename)[0]
|
1422 |
with open(index_filename, "r") as f:
|
1423 |
index = json.loads(f.read())
|
1424 |
-
|
1425 |
if "weight_map" in index:
|
1426 |
index = index["weight_map"]
|
1427 |
checkpoint_files = sorted(list(set(index.values())))
|
1428 |
checkpoint_files = [os.path.join(checkpoint_folder, f) for f in checkpoint_files]
|
1429 |
-
|
1430 |
# Logic for missing/unexepected keys goes here.
|
1431 |
-
|
1432 |
offload_index = {}
|
1433 |
if offload_state_dict:
|
1434 |
state_dict_folder = tempfile.mkdtemp()
|
1435 |
state_dict_index = {}
|
1436 |
-
|
1437 |
buffer_names = [name for name, _ in model.named_buffers()]
|
1438 |
for checkpoint_file in checkpoint_files:
|
1439 |
checkpoint = load_state_dict(checkpoint_file, device_map=device_map)
|
@@ -1444,9 +1256,7 @@ def load_checkpoint_in_model(
|
|
1444 |
# skip SCB parameter (for 8-bit serialization)
|
1445 |
if "SCB" in param_name:
|
1446 |
continue
|
1447 |
-
|
1448 |
module_name = param_name
|
1449 |
-
|
1450 |
while len(module_name) > 0 and module_name not in device_map:
|
1451 |
module_name = ".".join(module_name.split(".")[:-1])
|
1452 |
if module_name == "" and "" not in device_map:
|
@@ -1463,13 +1273,11 @@ def load_checkpoint_in_model(
|
|
1463 |
break
|
1464 |
if proceed:
|
1465 |
new_dtype = torch.float32
|
1466 |
-
|
1467 |
if "weight" in param_name and param_name.replace("weight", "SCB") in checkpoint.keys():
|
1468 |
if param.dtype == torch.int8:
|
1469 |
fp16_statistics = checkpoint[param_name.replace("weight", "SCB")]
|
1470 |
else:
|
1471 |
fp16_statistics = None
|
1472 |
-
|
1473 |
if param_device == "disk":
|
1474 |
if offload_buffers or param_name not in buffer_names:
|
1475 |
if new_dtype is None:
|
@@ -1501,25 +1309,18 @@ def load_checkpoint_in_model(
|
|
1501 |
dtype=new_dtype,
|
1502 |
fp16_statistics=fp16_statistics,
|
1503 |
)
|
1504 |
-
|
1505 |
# Force Python to clean up.
|
1506 |
del checkpoint
|
1507 |
gc.collect()
|
1508 |
-
|
1509 |
save_offload_index(offload_index, offload_folder)
|
1510 |
-
|
1511 |
# Load back offloaded state dict on CPU
|
1512 |
if offload_state_dict:
|
1513 |
load_offloaded_weights(model, state_dict_index, state_dict_folder)
|
1514 |
shutil.rmtree(state_dict_folder)
|
1515 |
-
|
1516 |
retie_parameters(model, tied_params)
|
1517 |
-
|
1518 |
-
|
1519 |
def get_mixed_precision_context_manager(native_amp: bool = False, autocast_kwargs: AutocastKwargs = None):
|
1520 |
"""
|
1521 |
Return a context manager for autocasting mixed precision
|
1522 |
-
|
1523 |
Args:
|
1524 |
native_amp (`bool`, *optional*, defaults to False):
|
1525 |
Whether mixed precision is actually enabled.
|
|
|
1 |
WEIGHTS_INDEX_NAME = "pytorch_model.bin.index.json"
|
|
|
2 |
logger = logging.getLogger(__name__)
|
|
|
|
|
3 |
def check_device_same(first_device, second_device):
|
4 |
"""
|
5 |
Utility method to check if two `torch` devices are similar. When dealing with CUDA devices, torch throws `False`
|
6 |
for `torch.device("cuda") == torch.device("cuda:0")` whereas they should be the same
|
|
|
7 |
Args:
|
8 |
first_device (`torch.device`):
|
9 |
First device to check
|
|
|
12 |
"""
|
13 |
if first_device.type != second_device.type:
|
14 |
return False
|
|
|
15 |
if first_device.type == "cuda" and first_device.index is None:
|
16 |
# In case the first_device is a cuda device and have
|
17 |
# the index attribute set to `None`, default it to `0`
|
18 |
first_device = torch.device("cuda", index=0)
|
|
|
19 |
if second_device.type == "cuda" and second_device.index is None:
|
20 |
# In case the second_device is a cuda device and have
|
21 |
# the index attribute set to `None`, default it to `0`
|
22 |
second_device = torch.device("cuda", index=0)
|
|
|
23 |
return first_device == second_device
|
|
|
|
|
24 |
def convert_file_size_to_int(size: Union[int, str]):
|
25 |
"""
|
26 |
Converts a size expressed as a string with digits an unit (like `"5MB"`) to an integer (in bytes).
|
|
|
27 |
Args:
|
28 |
size (`int` or `str`): The size to convert. Will be directly returned if an `int`.
|
|
|
29 |
Example:
|
|
|
30 |
```py
|
31 |
>>> convert_file_size_to_int("1MiB")
|
32 |
1048576
|
|
|
56 |
mem_size = int_size // 8 if size.endswith("b") else int_size
|
57 |
except ValueError:
|
58 |
raise ValueError(err_msg)
|
|
|
59 |
if mem_size <= 0:
|
60 |
raise ValueError(err_msg)
|
61 |
return mem_size
|
|
|
|
|
62 |
def dtype_byte_size(dtype: torch.dtype):
|
63 |
"""
|
64 |
Returns the size (in bytes) occupied by one parameter of type `dtype`.
|
|
|
65 |
Example:
|
|
|
66 |
```py
|
67 |
>>> dtype_byte_size(torch.float32)
|
68 |
4
|
|
|
79 |
raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
|
80 |
bit_size = int(bit_search.groups()[0])
|
81 |
return bit_size // 8
|
|
|
|
|
82 |
def id_tensor_storage(tensor: torch.Tensor) -> Tuple[torch.device, int, int]:
|
83 |
"""
|
84 |
Unique identifier to a tensor storage. Multiple different tensors can share the same underlying storage. For
|
|
|
111 |
storage_ptr = 0
|
112 |
# On torch >=2.0 this is the tensor size
|
113 |
storage_size = tensor.nelement() * _SIZE[tensor.dtype]
|
|
|
114 |
return tensor.device, storage_ptr, storage_size
|
|
|
|
|
115 |
def shard_checkpoint(
|
116 |
state_dict: Dict[str, torch.Tensor], max_shard_size: Union[int, str] = "10GB", weights_name: str = WEIGHTS_NAME
|
117 |
):
|
118 |
"""
|
119 |
Splits a model state dictionary in sub-checkpoints so that the final size of each sub-checkpoint does not exceed a
|
120 |
given size.
|
|
|
121 |
The sub-checkpoints are determined by iterating through the `state_dict` in the order of its keys, so there is no
|
122 |
optimization made to make each sub-checkpoint as close as possible to the maximum size passed. For example, if the
|
123 |
limit is 10GB and we have weights of sizes [6GB, 6GB, 2GB, 6GB, 2GB, 2GB] they will get sharded as [6GB], [6+2GB],
|
124 |
[6+2+2GB] and not [6+2+2GB], [6+2GB], [6GB].
|
|
|
125 |
<Tip warning={true}>
|
|
|
126 |
If one of the model's weight is bigger that `max_sahrd_size`, it will end up in its own sub-checkpoint which will
|
127 |
have a size greater than `max_shard_size`.
|
|
|
128 |
</Tip>
|
|
|
129 |
Args:
|
130 |
state_dict (`Dict[str, torch.Tensor]`): The state dictionary of a model to save.
|
131 |
max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
|
|
|
135 |
The name of the model save file.
|
136 |
"""
|
137 |
max_shard_size = convert_file_size_to_int(max_shard_size)
|
|
|
138 |
sharded_state_dicts = [{}]
|
139 |
last_block_size = 0
|
140 |
total_size = 0
|
141 |
storage_id_to_block = {}
|
|
|
142 |
for key, weight in state_dict.items():
|
143 |
# when bnb serialization is used the weights in the state dict can be strings
|
144 |
# check: https://github.com/huggingface/transformers/pull/24416 for more details
|
|
|
146 |
continue
|
147 |
else:
|
148 |
storage_id = id_tensor_storage(weight)
|
|
|
149 |
# If a `weight` shares the same underlying storage as another tensor, we put `weight` in the same `block`
|
150 |
if storage_id in storage_id_to_block:
|
151 |
block_id = storage_id_to_block[storage_id]
|
152 |
sharded_state_dicts[block_id][key] = weight
|
153 |
continue
|
|
|
154 |
weight_size = weight.numel() * dtype_byte_size(weight.dtype)
|
|
|
155 |
# If this weight is going to tip up over the maximal size, we split.
|
156 |
if last_block_size + weight_size > max_shard_size:
|
157 |
sharded_state_dicts.append({})
|
158 |
last_block_size = 0
|
|
|
159 |
sharded_state_dicts[-1][key] = weight
|
160 |
last_block_size += weight_size
|
161 |
total_size += weight_size
|
162 |
storage_id_to_block[storage_id] = len(sharded_state_dicts) - 1
|
|
|
163 |
# If we only have one shard, we return it
|
164 |
if len(sharded_state_dicts) == 1:
|
165 |
return {weights_name: sharded_state_dicts[0]}, None
|
|
|
166 |
# Otherwise, let's build the index
|
167 |
weight_map = {}
|
168 |
shards = {}
|
|
|
174 |
shards[shard_file] = shard
|
175 |
for key in shard.keys():
|
176 |
weight_map[key] = shard_file
|
|
|
177 |
# Add the metadata
|
178 |
metadata = {"total_size": total_size}
|
179 |
index = {"metadata": metadata, "weight_map": weight_map}
|
180 |
return shards, index
|
|
|
|
|
181 |
def set_module_tensor_to_device(
|
182 |
module: nn.Module,
|
183 |
tensor_name: str,
|
|
|
189 |
"""
|
190 |
A helper function to set a given tensor (parameter of buffer) of a module on a specific device (note that doing
|
191 |
`param.to(device)` creates a new tensor not linked to the parameter, which is why we need this function).
|
|
|
192 |
Args:
|
193 |
module (`torch.nn.Module`):
|
194 |
The module in which the tensor we want to move lives.
|
|
|
213 |
raise ValueError(f"{module} has no attribute {split}.")
|
214 |
module = new_module
|
215 |
tensor_name = splits[-1]
|
|
|
216 |
if tensor_name not in module._parameters and tensor_name not in module._buffers:
|
217 |
raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.")
|
218 |
is_buffer = tensor_name in module._buffers
|
219 |
old_value = getattr(module, tensor_name)
|
|
|
220 |
if old_value.device == torch.device("meta") and device not in ["meta", torch.device("meta")] and value is None:
|
221 |
raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {device}.")
|
|
|
222 |
if value is not None:
|
223 |
if old_value.shape != value.shape:
|
224 |
raise ValueError(
|
225 |
f'Trying to set a tensor of shape {value.shape} in "{tensor_name}" (which has shape {old_value.shape}), this look incorrect.'
|
226 |
)
|
|
|
227 |
if dtype is None:
|
228 |
# For compatibility with PyTorch load_state_dict which converts state dict dtype to existing dtype in model
|
229 |
value = value.to(old_value.dtype)
|
230 |
elif not str(value.dtype).startswith(("torch.uint", "torch.int", "torch.bool")):
|
231 |
value = value.to(dtype)
|
|
|
232 |
param = module._parameters[tensor_name] if tensor_name in module._parameters else None
|
233 |
param_cls = type(param)
|
|
|
234 |
device_quantization = None
|
235 |
with torch.no_grad():
|
236 |
# leave it on cpu first before moving them to cuda
|
|
|
251 |
if dtype is not None and device in ["meta", torch.device("meta")]:
|
252 |
if not str(old_value.dtype).startswith(("torch.uint", "torch.int", "torch.bool")):
|
253 |
new_value = new_value.to(dtype)
|
|
|
254 |
if not is_buffer:
|
255 |
module._parameters[tensor_name] = param_cls(new_value, requires_grad=old_value.requires_grad)
|
256 |
elif isinstance(value, torch.Tensor):
|
|
|
306 |
torch.npu.empty_cache()
|
307 |
else:
|
308 |
torch.cuda.empty_cache()
|
|
|
|
|
309 |
def named_module_tensors(
|
310 |
module: nn.Module, include_buffers: bool = True, recurse: bool = False, remove_non_persistent: bool = False
|
311 |
):
|
312 |
"""
|
313 |
A helper function that gathers all the tensors (parameters + buffers) of a given module. If `include_buffers=True`
|
314 |
it's the same as doing `module.named_parameters(recurse=recurse) + module.named_buffers(recurse=recurse)`.
|
|
|
315 |
Args:
|
316 |
module (`torch.nn.Module`):
|
317 |
The module we want the tensors on.
|
|
|
325 |
"""
|
326 |
for named_parameter in module.named_parameters(recurse=recurse):
|
327 |
yield named_parameter
|
|
|
328 |
if include_buffers:
|
329 |
non_persistent_buffers = set()
|
330 |
if remove_non_persistent:
|
|
|
333 |
name, _ = named_buffer
|
334 |
if name not in non_persistent_buffers:
|
335 |
yield named_buffer
|
|
|
|
|
336 |
def get_non_persistent_buffers(module: nn.Module, recurse: bool = False):
|
337 |
"""
|
338 |
Gather all non persistent buffers of a given modules into a set
|
|
|
339 |
Args:
|
340 |
module (`nn.Module`):
|
341 |
The module we want the non persistent buffers on.
|
|
|
346 |
if recurse:
|
347 |
for _, m in module.named_modules():
|
348 |
non_persistent_buffers_set |= m._non_persistent_buffers_set
|
|
|
349 |
return non_persistent_buffers_set
|
|
|
|
|
350 |
class FindTiedParametersResult(list):
|
351 |
"""
|
352 |
This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
|
|
|
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], [])
|
|
|
|
|
360 |
def check_tied_parameters_in_config(model: nn.Module):
|
361 |
"""
|
362 |
Check if there is any indication in the given model that some weights should be tied.
|
|
|
363 |
Args:
|
364 |
model (`torch.nn.Module`): The model to inspect
|
|
|
365 |
Returns:
|
366 |
bool: True if the model needs to have tied weights
|
367 |
"""
|
|
|
369 |
has_tied_word_embedding = False
|
370 |
has_tied_encoder_decoder = False
|
371 |
has_tied_module = False
|
|
|
372 |
if "PreTrainedModel" in [c.__name__ for c in inspect.getmro(model.__class__)]:
|
373 |
has_tied_word_embedding = (
|
374 |
hasattr(model, "config")
|
|
|
381 |
and getattr(model.config, "tie_encoder_decoder", False)
|
382 |
)
|
383 |
has_tied_module = any(hasattr(module, "_tie_weights") for module in model.modules())
|
|
|
384 |
return any([has_tied_word_embedding, has_tied_encoder_decoder, has_tied_module])
|
|
|
|
|
385 |
def _get_param_device(param, device_map):
|
386 |
if param in device_map:
|
387 |
return device_map[param]
|
|
|
390 |
raise ValueError(f"The `device_map` does not contain the module {param}.")
|
391 |
else:
|
392 |
return _get_param_device(parent_param, device_map)
|
|
|
|
|
393 |
def check_tied_parameters_on_same_device(tied_params, device_map):
|
394 |
"""
|
395 |
Check if tied parameters are on the same device
|
|
|
396 |
Args:
|
397 |
tied_params (`List[List[str]]`):
|
398 |
A list of lists of parameter names being all tied together.
|
|
|
399 |
device_map (`Dict[str, Union[int, str, torch.device]]`):
|
400 |
A map that specifies where each submodule should go.
|
|
|
401 |
"""
|
402 |
for tie_param in tied_params:
|
403 |
tie_param_devices = {}
|
|
|
408 |
f"Tied parameters are on different devices: {tie_param_devices}. "
|
409 |
"Please modify your custom device map or set `device_map='auto'`. "
|
410 |
)
|
|
|
|
|
411 |
def find_tied_parameters(model: nn.Module, **kwargs):
|
412 |
"""
|
413 |
Find the tied parameters in a given model.
|
|
|
414 |
<Tip warning={true}>
|
|
|
415 |
The signature accepts keyword arguments, but they are for the recursive part of this function and you should ignore
|
416 |
them.
|
|
|
417 |
</Tip>
|
|
|
418 |
Args:
|
419 |
model (`torch.nn.Module`): The model to inspect.
|
|
|
420 |
Returns:
|
421 |
List[List[str]]: A list of lists of parameter names being all tied together.
|
|
|
422 |
Example:
|
|
|
423 |
```py
|
424 |
>>> from collections import OrderedDict
|
425 |
>>> import torch.nn as nn
|
|
|
426 |
>>> model = nn.Sequential(OrderedDict([("linear1", nn.Linear(4, 4)), ("linear2", nn.Linear(4, 4))]))
|
427 |
>>> model.linear2.weight = model.linear1.weight
|
428 |
>>> find_tied_parameters(model)
|
|
|
433 |
named_parameters = kwargs.get("named_parameters", None)
|
434 |
prefix = kwargs.get("prefix", "")
|
435 |
result = kwargs.get("result", {})
|
|
|
436 |
if named_parameters is None:
|
437 |
named_parameters = {n: p for n, p in model.named_parameters()}
|
438 |
else:
|
|
|
448 |
if new_name not in result:
|
449 |
result[new_name] = []
|
450 |
result[new_name].append(full_name)
|
|
|
451 |
# Once we have treated direct parameters, we move to the child modules.
|
452 |
for name, child in model.named_children():
|
453 |
child_name = name if prefix == "" else f"{prefix}.{name}"
|
454 |
find_tied_parameters(child, named_parameters=named_parameters, prefix=child_name, result=result)
|
|
|
455 |
return FindTiedParametersResult([sorted([weight] + list(set(tied))) for weight, tied in result.items()])
|
|
|
|
|
456 |
def retie_parameters(model, tied_params):
|
457 |
"""
|
458 |
Reties tied parameters in a given model if the link was broken (for instance when adding hooks).
|
|
|
459 |
Args:
|
460 |
model (`torch.nn.Module`):
|
461 |
The model in which to retie parameters.
|
|
|
481 |
for split in splits[:-1]:
|
482 |
module = getattr(module, split)
|
483 |
setattr(module, splits[-1], param_to_tie)
|
|
|
|
|
484 |
def _get_proper_dtype(dtype: Union[str, torch.device]) -> torch.dtype:
|
485 |
"""
|
486 |
Just does torch.dtype(dtype) if necessary.
|
|
|
490 |
dtype = dtype.replace("torch.", "")
|
491 |
dtype = getattr(torch, dtype)
|
492 |
return dtype
|
|
|
|
|
493 |
def compute_module_sizes(
|
494 |
model: nn.Module,
|
495 |
dtype: Optional[Union[str, torch.device]] = None,
|
|
|
515 |
name_parts = name.split(".")
|
516 |
for idx in range(len(name_parts) + 1):
|
517 |
module_sizes[".".join(name_parts[:idx])] += size
|
|
|
518 |
return module_sizes
|
|
|
|
|
519 |
def get_max_layer_size(
|
520 |
modules: List[Tuple[str, torch.nn.Module]], module_sizes: Dict[str, int], no_split_module_classes: List[str]
|
521 |
):
|
|
|
524 |
definition of a layer being:
|
525 |
- a module with no direct children (just parameters and buffers)
|
526 |
- a module whose class name is in the list `no_split_module_classes`
|
|
|
527 |
Args:
|
528 |
modules (`List[Tuple[str, torch.nn.Module]]`):
|
529 |
The list of named modules where we want to determine the maximum layer size.
|
|
|
531 |
A dictionary mapping each layer name to its size (as generated by `compute_module_sizes`).
|
532 |
no_split_module_classes (`List[str]`):
|
533 |
A list of class names for layers we don't want to be split.
|
|
|
534 |
Returns:
|
535 |
`Tuple[int, List[str]]`: The maximum size of a layer with the list of layer names realizing that maximum size.
|
536 |
"""
|
|
|
551 |
else:
|
552 |
modules_to_treat = [(f"{module_name}.{n}", v) for n, v in modules_children] + modules_to_treat
|
553 |
return max_size, layer_names
|
|
|
|
|
554 |
def get_max_memory(max_memory: Optional[Dict[Union[int, str], Union[int, str]]] = None):
|
555 |
"""
|
556 |
Get the maximum memory available if nothing is passed, converts string to int otherwise.
|
557 |
"""
|
558 |
import psutil
|
|
|
559 |
if max_memory is None:
|
560 |
if not (torch.cuda.is_available() or is_npu_available() or is_xpu_available()):
|
561 |
max_memory = {}
|
|
|
562 |
else:
|
563 |
# Make sure CUDA is initialized on each GPU to have the right memory info.
|
564 |
if is_npu_available():
|
|
|
579 |
else:
|
580 |
max_memory["cpu"] = psutil.virtual_memory().available
|
581 |
return max_memory
|
|
|
582 |
for key in max_memory:
|
583 |
if isinstance(max_memory[key], str):
|
584 |
max_memory[key] = convert_file_size_to_int(max_memory[key])
|
|
|
585 |
# Need to sort the device by type to make sure that we allocate the gpu first.
|
586 |
# As gpu/npu/xpu are represented by int, we need to sort them first.
|
587 |
gpu_devices = [k for k in max_memory.keys() if isinstance(k, int)]
|
|
|
605 |
f"Device {k} is not recognized, available devices are integers(for GPU/XPU), 'mps', 'cpu' and 'disk'"
|
606 |
)
|
607 |
max_memory = {k: max_memory[k] for k in all_devices}
|
|
|
608 |
return max_memory
|
|
|
|
|
609 |
def clean_device_map(device_map: Dict[str, Union[int, str, torch.device]], module_name: str = ""):
|
610 |
"""
|
611 |
Cleans a device_map by grouping all submodules that go on the same device together.
|
|
|
617 |
for k in [k for k in device_map if k.startswith(prefix)]:
|
618 |
del device_map[k]
|
619 |
device_map[module_name] = values[0]
|
|
|
620 |
# Recurse over the children
|
621 |
children_modules = [k for k in device_map.keys() if k.startswith(prefix) and len(k) > len(module_name)]
|
622 |
idx = len(module_name.split(".")) + 1 if len(module_name) > 0 else 1
|
623 |
children_modules = set(".".join(k.split(".")[:idx]) for k in children_modules)
|
624 |
for child in children_modules:
|
625 |
clean_device_map(device_map, module_name=child)
|
|
|
626 |
return device_map
|
|
|
|
|
627 |
def load_offloaded_weights(model, index, offload_folder):
|
628 |
"""
|
629 |
Loads the weights from the offload folder into the model.
|
|
|
630 |
Args:
|
631 |
model (`torch.nn.Module`):
|
632 |
The model to load the weights into.
|
|
|
651 |
tensor_file = os.path.join(offload_folder, f"{param_name}.dat")
|
652 |
weight = load_offloaded_weight(tensor_file, metadata)
|
653 |
set_module_tensor_to_device(model, param_name, "cpu", value=weight, fp16_statistics=fp16_statistics)
|
|
|
|
|
654 |
def get_balanced_memory(
|
655 |
model: nn.Module,
|
656 |
max_memory: Optional[Dict[Union[int, str], Union[int, str]]] = None,
|
|
|
661 |
):
|
662 |
"""
|
663 |
Compute a `max_memory` dictionary for [`infer_auto_device_map`] that will balance the use of each available GPU.
|
|
|
664 |
<Tip>
|
|
|
665 |
All computation is done analyzing sizes and dtypes of the model parameters. As a result, the model can be on the
|
666 |
meta device (as it would if initialized within the `init_empty_weights` context manager).
|
|
|
667 |
</Tip>
|
|
|
668 |
Args:
|
669 |
model (`torch.nn.Module`):
|
670 |
The model to analyze.
|
|
|
685 |
# Get default / clean up max_memory
|
686 |
user_not_set_max_memory = max_memory is None
|
687 |
max_memory = get_max_memory(max_memory)
|
|
|
688 |
if is_npu_available():
|
689 |
num_devices = len([d for d in max_memory if torch.device(d).type == "npu" and max_memory[d] > 0])
|
690 |
elif is_xpu_available():
|
|
|
701 |
)
|
702 |
else:
|
703 |
num_devices = len([d for d in max_memory if torch.device(d).type == "cuda" and max_memory[d] > 0])
|
|
|
704 |
if num_devices == 0:
|
705 |
return max_memory
|
|
|
706 |
if num_devices == 1:
|
707 |
# We cannot do low_zero on just one GPU, but we will still reserve some memory for the buffer
|
708 |
low_zero = False
|
|
|
716 |
"You can set `max_memory` in to a higher value to use more memory (at your own risk)."
|
717 |
)
|
718 |
break # only one device
|
|
|
719 |
module_sizes = compute_module_sizes(model, dtype=dtype, special_dtypes=special_dtypes)
|
720 |
per_gpu = module_sizes[""] // (num_devices - 1 if low_zero else num_devices)
|
|
|
721 |
# We can't just set the memory to model_size // num_devices as it will end being too small: each GPU will get
|
722 |
# slightly less layers and some layers will end up offload at the end. So this function computes a buffer size to
|
723 |
# add which is the biggest of:
|
|
|
727 |
no_split_module_classes = []
|
728 |
elif not isinstance(no_split_module_classes, (list, tuple)):
|
729 |
no_split_module_classes = [no_split_module_classes]
|
|
|
730 |
# Identify the size of the no_split_block modules
|
731 |
if len(no_split_module_classes) > 0:
|
732 |
no_split_children = {}
|
|
|
739 |
class_name = submodule.__class__.__name__
|
740 |
if class_name in no_split_module_classes and class_name not in no_split_children:
|
741 |
no_split_children[class_name] = size
|
|
|
742 |
if set(no_split_children.keys()) == set(no_split_module_classes):
|
743 |
break
|
744 |
buffer = max(no_split_children.values()) if len(no_split_children) > 0 else 0
|
745 |
else:
|
746 |
buffer = 0
|
|
|
747 |
# Compute mean of final modules. In the first dict of module sizes, leaves are the parameters
|
748 |
leaves = [n for n in module_sizes if len([p for p in module_sizes if n == "" or p.startswith(n + ".")]) == 0]
|
749 |
module_sizes = {n: v for n, v in module_sizes.items() if n not in leaves}
|
|
|
752 |
mean_leaves = int(sum([module_sizes[n] for n in leaves]) / max(len(leaves), 1))
|
753 |
buffer = int(1.25 * max(buffer, mean_leaves))
|
754 |
per_gpu += buffer
|
|
|
755 |
# Sorted list of GPUs id (we may have some gpu ids not included in the our max_memory list - let's ignore them)
|
756 |
gpus_idx_list = list(
|
757 |
sorted(
|
|
|
761 |
# The last device is left with max_memory just in case the buffer is not enough.
|
762 |
for idx in gpus_idx_list[:-1]:
|
763 |
max_memory[idx] = min(max_memory[0] if low_zero and idx == 0 else per_gpu, max_memory[idx])
|
|
|
764 |
if low_zero:
|
765 |
min_zero = max(0, module_sizes[""] - sum([max_memory[i] for i in range(1, num_devices)]))
|
766 |
max_memory[0] = min(min_zero, max_memory[0])
|
|
|
767 |
return max_memory
|
|
|
|
|
768 |
def calculate_maximum_sizes(model: torch.nn.Module):
|
769 |
"Computes the total size of the model and its largest layer"
|
770 |
sizes = compute_module_sizes(model)
|
|
|
772 |
no_split_modules = getattr(model, "_no_split_modules", None)
|
773 |
if no_split_modules is None:
|
774 |
no_split_modules = []
|
|
|
775 |
modules_to_treat = (
|
776 |
list(model.named_parameters(recurse=False))
|
777 |
+ list(model.named_children())
|
|
|
780 |
largest_layer = get_max_layer_size(modules_to_treat, sizes, no_split_modules)
|
781 |
total_size = sizes[""]
|
782 |
return total_size, largest_layer
|
|
|
|
|
783 |
def infer_auto_device_map(
|
784 |
model: nn.Module,
|
785 |
max_memory: Optional[Dict[Union[int, str], Union[int, str]]] = None,
|
|
|
798 |
- if offload to the CPU is needed,we don't exceed the RAM available on the CPU.
|
799 |
- if offload to the disk is needed, there is always room left on the CPU to put back the layer offloaded on disk
|
800 |
that has the largest size.
|
|
|
801 |
<Tip>
|
|
|
802 |
All computation is done analyzing sizes and dtypes of the model parameters. As a result, the model can be on the
|
803 |
meta device (as it would if initialized within the `init_empty_weights` context manager).
|
|
|
804 |
</Tip>
|
|
|
805 |
Args:
|
806 |
model (`torch.nn.Module`):
|
807 |
The model to analyze.
|
|
|
826 |
no_split_module_classes = []
|
827 |
elif not isinstance(no_split_module_classes, (list, tuple)):
|
828 |
no_split_module_classes = [no_split_module_classes]
|
|
|
829 |
devices = list(max_memory.keys())
|
830 |
if "disk" not in devices:
|
831 |
devices.append("disk")
|
832 |
gpus = [device for device in devices if device not in ["cpu", "disk"]]
|
|
|
833 |
# Devices that need to keep space for a potential offloaded layer.
|
834 |
if "mps" in gpus:
|
835 |
main_devices = ["mps"]
|
|
|
837 |
main_devices = [gpus[0], "cpu"]
|
838 |
else:
|
839 |
main_devices = ["cpu"]
|
|
|
840 |
module_sizes = compute_module_sizes(model, dtype=dtype, special_dtypes=special_dtypes)
|
841 |
tied_parameters = find_tied_parameters(model)
|
|
|
842 |
if check_tied_parameters_in_config(model) and len(tied_parameters) == 0:
|
843 |
logger.warn(
|
844 |
"The model weights are not tied. Please use the `tie_weights` method before using the `infer_auto_device` function."
|
845 |
)
|
|
|
846 |
device_map = OrderedDict()
|
847 |
current_device = 0
|
848 |
current_memory_used = 0
|
|
|
849 |
# Direct submodules and parameters
|
850 |
modules_to_treat = (
|
851 |
list(model.named_parameters(recurse=False))
|
|
|
854 |
)
|
855 |
# Initialize maximum largest layer, to know which space to keep in memory
|
856 |
max_layer_size, max_layer_names = get_max_layer_size(modules_to_treat, module_sizes, no_split_module_classes)
|
|
|
857 |
# Ready ? This is going to be a bit messy.
|
858 |
while len(modules_to_treat) > 0:
|
859 |
name, module = modules_to_treat.pop(0)
|
|
|
869 |
)
|
870 |
# Assess size needed
|
871 |
module_size = module_sizes[name]
|
|
|
872 |
# We keep relevant tied parameters only: one of the tied parameters in the group is inside the current module
|
873 |
# and the other is not.
|
874 |
tied_param_goups = [
|
|
|
882 |
tied_params = sum([[p for p in tied_group if name not in p] for tied_group in tied_param_goups], [])
|
883 |
if verbose and len(tied_params) > 0:
|
884 |
print(f" So those parameters need to be taken into account {tied_params}")
|
|
|
885 |
device = devices[current_device]
|
886 |
current_max_size = max_memory[device] if device != "disk" else None
|
887 |
# Reduce max size available by the largest layer.
|
|
|
915 |
module_sizes,
|
916 |
no_split_module_classes,
|
917 |
)
|
|
|
918 |
# Case 2, it fits! We're not entirely out of the wood though, because we may have some tied parameters.
|
919 |
elif len(tied_params) > 0:
|
920 |
# First locate all tied modules
|
|
|
929 |
f" It looks like {name} is going to fit on {devices[current_device]} but we have tied "
|
930 |
f"parameters to account for.\n - Names {tied_params}\n - Module names {tied_module_names}"
|
931 |
)
|
|
|
932 |
# Let's see if it all fits first
|
933 |
module_size_with_ties = module_size
|
934 |
for tied_param, tied_module_name in zip(tied_params, tied_module_names):
|
935 |
module_size_with_ties += module_sizes[tied_module_name] - module_sizes[tied_param]
|
|
|
936 |
if current_max_size is None or current_memory_used + module_size_with_ties <= current_max_size:
|
937 |
# We really really fit!
|
938 |
if verbose:
|
|
|
947 |
]
|
948 |
modules_to_treat.pop(tied_module_index)
|
949 |
device_map[tied_module_name] = devices[current_device]
|
|
|
950 |
else:
|
951 |
# We don't fit with the tied modules. Next question is: can we split one of the tied modules to make it
|
952 |
# smaller or do we need to go on the next device?
|
|
|
961 |
if len(tied_module_children) == 0 or tied_module.__class__.__name__ in no_split_module_classes:
|
962 |
# can't break this one.
|
963 |
continue
|
|
|
964 |
if verbose:
|
965 |
print(f"Splitting {tied_module_name}.")
|
966 |
tied_module_children = list(tied_module.named_parameters(recurse=False)) + tied_module_children
|
967 |
tied_module_children = [(f"{tied_module_name}.{n}", v) for n, v in tied_module_children]
|
968 |
tied_module_index = [i for i, (n, _) in enumerate(modules_to_treat) if n == tied_module_name][0]
|
|
|
969 |
modules_to_treat = (
|
970 |
[(name, module)]
|
971 |
+ modules_to_treat[:tied_module_index]
|
|
|
980 |
)
|
981 |
split_happened = True
|
982 |
break
|
|
|
983 |
if not split_happened:
|
984 |
# If the tied module is not split, we go to the next device
|
985 |
if verbose:
|
|
|
987 |
current_device += 1
|
988 |
modules_to_treat = [(name, module)] + modules_to_treat
|
989 |
current_memory_used = 0
|
|
|
990 |
else:
|
991 |
if verbose:
|
992 |
if current_max_size is None:
|
|
|
998 |
)
|
999 |
current_memory_used += module_size
|
1000 |
device_map[name] = devices[current_device]
|
|
|
1001 |
if clean_result:
|
1002 |
device_map = clean_device_map(device_map)
|
1003 |
return device_map
|
|
|
|
|
1004 |
def check_device_map(model: nn.Module, device_map: Dict[str, Union[int, str, torch.device]]):
|
1005 |
"""
|
1006 |
Checks a device map covers everything in a given model.
|
|
|
1007 |
Args:
|
1008 |
model (`torch.nn.Module`): The model to check the device map against.
|
1009 |
device_map (`Dict[str, Union[int, str, torch.device]]`): The device map to check.
|
|
|
1024 |
raise ValueError(
|
1025 |
f"The device_map provided does not give any device for the following parameters: {non_covered_params}"
|
1026 |
)
|
|
|
|
|
1027 |
def load_state_dict(checkpoint_file, device_map=None):
|
1028 |
"""
|
1029 |
Load a checkpoint from a given file. If the checkpoint is in the safetensors format and a device map is passed, the
|
1030 |
weights can be fast-loaded directly on the GPU.
|
|
|
1031 |
Args:
|
1032 |
checkpoint_file (`str`): The path to the checkpoint to load.
|
1033 |
device_map (`Dict[str, Union[int, str, torch.device]]`, *optional*):
|
|
|
1038 |
with safe_open(checkpoint_file, framework="pt") as f:
|
1039 |
metadata = f.metadata()
|
1040 |
weight_names = f.keys()
|
|
|
1041 |
if metadata is None:
|
1042 |
logger.warn(
|
1043 |
f"The safetensors archive passed at {checkpoint_file} does not contain metadata. "
|
1044 |
"Make sure to save your model with the `save_pretrained` method. Defaulting to 'pt' metadata."
|
1045 |
)
|
1046 |
metadata = {"format": "pt"}
|
|
|
1047 |
if metadata.get("format") not in ["pt", "tf", "flax"]:
|
1048 |
raise OSError(
|
1049 |
f"The safetensors archive passed at {checkpoint_file} does not contain the valid metadata. Make sure "
|
|
|
1057 |
# if we only have one device we can load everything directly
|
1058 |
if len(set(device_map.values())) == 1:
|
1059 |
return safe_load_file(checkpoint_file, device=list(device_map.values())[0])
|
|
|
1060 |
devices = list(set(device_map.values()) - {"disk"})
|
1061 |
# cpu device should always exist as fallback option
|
1062 |
if "cpu" not in devices:
|
1063 |
devices.append("cpu")
|
|
|
1064 |
# For each device, get the weights that go there
|
1065 |
device_weights = {device: [] for device in devices}
|
1066 |
for module_name, device in device_map.items():
|
|
|
1068 |
device_weights[device].extend(
|
1069 |
[k for k in weight_names if k == module_name or k.startswith(module_name + ".")]
|
1070 |
)
|
|
|
1071 |
# all weights that haven't defined a device should be loaded on CPU
|
1072 |
device_weights["cpu"].extend([k for k in weight_names if k not in sum(device_weights.values(), [])])
|
1073 |
tensors = {}
|
|
|
1092 |
progress_bar.update()
|
1093 |
if progress_bar is not None:
|
1094 |
progress_bar.close()
|
|
|
1095 |
return tensors
|
1096 |
else:
|
1097 |
return torch.load(checkpoint_file, map_location=torch.device("cpu"))
|
|
|
|
|
1098 |
def get_state_dict_offloaded_model(model: nn.Module):
|
1099 |
"""
|
1100 |
Returns the state dictionary for an offloaded model via iterative onloading
|
|
|
1101 |
Args:
|
1102 |
model (`torch.nn.Module`):
|
1103 |
The offloaded model we want to save
|
1104 |
"""
|
1105 |
from ..hooks import AlignDevicesHook
|
|
|
1106 |
state_dict = {}
|
1107 |
placeholders = set()
|
1108 |
for name, module in model.named_modules():
|
|
|
1124 |
module._hf_hook.execution_device = original_device
|
1125 |
else:
|
1126 |
module_state_dict = module.state_dict()
|
|
|
1127 |
for key in module_state_dict:
|
1128 |
# ignore placeholder parameters that are still on the meta device
|
1129 |
if module_state_dict[key].device == torch.device("meta"):
|
|
|
1136 |
placeholders.remove(key)
|
1137 |
if placeholders:
|
1138 |
logger.warning(f"The following tensors were not saved because they were still on meta device: {placeholders}")
|
|
|
1139 |
return state_dict
|
|
|
|
|
1140 |
def load_checkpoint_in_model(
|
1141 |
model: nn.Module,
|
1142 |
checkpoint: Union[str, os.PathLike],
|
|
|
1151 |
"""
|
1152 |
Loads a (potentially sharded) checkpoint inside a model, potentially sending weights to a given device as they are
|
1153 |
loaded.
|
|
|
1154 |
<Tip warning={true}>
|
|
|
1155 |
Once loaded across devices, you still need to call [`dispatch_model`] on your model to make it able to run. To
|
1156 |
group the checkpoint loading and dispatch in one single call, use [`load_checkpoint_and_dispatch`].
|
|
|
1157 |
</Tip>
|
|
|
1158 |
Args:
|
1159 |
model (`torch.nn.Module`):
|
1160 |
The model in which we want to load a checkpoint.
|
|
|
1180 |
A list of the modules that we keep in `torch.float32` dtype.
|
1181 |
offload_8bit_bnb (`bool`, *optional*):
|
1182 |
Whether or not to enable offload of 8-bit modules on cpu/disk.
|
|
|
1183 |
"""
|
1184 |
if offload_8bit_bnb:
|
1185 |
from .bnb import quantize_and_offload_8bit
|
|
|
1186 |
tied_params = find_tied_parameters(model)
|
|
|
1187 |
if check_tied_parameters_in_config(model) and len(tied_params) == 0:
|
1188 |
logger.warn(
|
1189 |
"The model weights are not tied. Please use the `tie_weights` method before using the `infer_auto_device` function."
|
1190 |
)
|
1191 |
if device_map is not None:
|
1192 |
check_tied_parameters_on_same_device(tied_params, device_map)
|
|
|
1193 |
if offload_folder is None and device_map is not None and "disk" in device_map.values():
|
1194 |
raise ValueError(
|
1195 |
"At least one of the model submodule will be offloaded to disk, please pass along an `offload_folder`."
|
1196 |
)
|
1197 |
elif offload_folder is not None and device_map is not None and "disk" in device_map.values():
|
1198 |
os.makedirs(offload_folder, exist_ok=True)
|
|
|
1199 |
if isinstance(dtype, str):
|
1200 |
# We accept "torch.float16" or just "float16"
|
1201 |
dtype = dtype.replace("torch.", "")
|
1202 |
dtype = getattr(torch, dtype)
|
|
|
1203 |
checkpoint_files = None
|
1204 |
index_filename = None
|
1205 |
if os.path.isfile(checkpoint):
|
|
|
1233 |
"`checkpoint` should be the path to a file containing a whole state dict, or the index of a sharded "
|
1234 |
f"checkpoint, or a folder containing a sharded checkpoint or the whole state dict, but got {checkpoint}."
|
1235 |
)
|
|
|
1236 |
if index_filename is not None:
|
1237 |
checkpoint_folder = os.path.split(index_filename)[0]
|
1238 |
with open(index_filename, "r") as f:
|
1239 |
index = json.loads(f.read())
|
|
|
1240 |
if "weight_map" in index:
|
1241 |
index = index["weight_map"]
|
1242 |
checkpoint_files = sorted(list(set(index.values())))
|
1243 |
checkpoint_files = [os.path.join(checkpoint_folder, f) for f in checkpoint_files]
|
|
|
1244 |
# Logic for missing/unexepected keys goes here.
|
|
|
1245 |
offload_index = {}
|
1246 |
if offload_state_dict:
|
1247 |
state_dict_folder = tempfile.mkdtemp()
|
1248 |
state_dict_index = {}
|
|
|
1249 |
buffer_names = [name for name, _ in model.named_buffers()]
|
1250 |
for checkpoint_file in checkpoint_files:
|
1251 |
checkpoint = load_state_dict(checkpoint_file, device_map=device_map)
|
|
|
1256 |
# skip SCB parameter (for 8-bit serialization)
|
1257 |
if "SCB" in param_name:
|
1258 |
continue
|
|
|
1259 |
module_name = param_name
|
|
|
1260 |
while len(module_name) > 0 and module_name not in device_map:
|
1261 |
module_name = ".".join(module_name.split(".")[:-1])
|
1262 |
if module_name == "" and "" not in device_map:
|
|
|
1273 |
break
|
1274 |
if proceed:
|
1275 |
new_dtype = torch.float32
|
|
|
1276 |
if "weight" in param_name and param_name.replace("weight", "SCB") in checkpoint.keys():
|
1277 |
if param.dtype == torch.int8:
|
1278 |
fp16_statistics = checkpoint[param_name.replace("weight", "SCB")]
|
1279 |
else:
|
1280 |
fp16_statistics = None
|
|
|
1281 |
if param_device == "disk":
|
1282 |
if offload_buffers or param_name not in buffer_names:
|
1283 |
if new_dtype is None:
|
|
|
1309 |
dtype=new_dtype,
|
1310 |
fp16_statistics=fp16_statistics,
|
1311 |
)
|
|
|
1312 |
# Force Python to clean up.
|
1313 |
del checkpoint
|
1314 |
gc.collect()
|
|
|
1315 |
save_offload_index(offload_index, offload_folder)
|
|
|
1316 |
# Load back offloaded state dict on CPU
|
1317 |
if offload_state_dict:
|
1318 |
load_offloaded_weights(model, state_dict_index, state_dict_folder)
|
1319 |
shutil.rmtree(state_dict_folder)
|
|
|
1320 |
retie_parameters(model, tied_params)
|
|
|
|
|
1321 |
def get_mixed_precision_context_manager(native_amp: bool = False, autocast_kwargs: AutocastKwargs = None):
|
1322 |
"""
|
1323 |
Return a context manager for autocasting mixed precision
|
|
|
1324 |
Args:
|
1325 |
native_amp (`bool`, *optional*, defaults to False):
|
1326 |
Whether mixed precision is actually enabled.
|
src/utils/offload.py
CHANGED
@@ -17,35 +17,26 @@ def offload_weight(weight, weight_name, offload_folder, index=None):
|
|
17 |
file_array[:] = array[:]
|
18 |
file_array.flush()
|
19 |
return index
|
20 |
-
|
21 |
-
|
22 |
def load_offloaded_weight(weight_file, weight_info):
|
23 |
shape = tuple(weight_info["shape"])
|
24 |
if shape == ():
|
25 |
# NumPy memory-mapped arrays can't have 0 dims so it was saved as 1d tensor
|
26 |
shape = (1,)
|
27 |
-
|
28 |
dtype = weight_info["dtype"]
|
29 |
if dtype == "bfloat16":
|
30 |
# NumPy does not support bfloat16 so this was saved as a int16
|
31 |
dtype = "int16"
|
32 |
-
|
33 |
weight = np.memmap(weight_file, dtype=dtype, shape=shape, mode="r")
|
34 |
-
|
35 |
if len(weight_info["shape"]) == 0:
|
36 |
weight = weight[0]
|
37 |
weight = torch.tensor(weight)
|
38 |
if weight_info["dtype"] == "bfloat16":
|
39 |
weight = weight.view(torch.bfloat16)
|
40 |
-
|
41 |
return weight
|
42 |
-
|
43 |
-
|
44 |
def save_offload_index(index, offload_folder):
|
45 |
if index is None or len(index) == 0:
|
46 |
# Nothing to save
|
47 |
return
|
48 |
-
|
49 |
offload_index_file = os.path.join(offload_folder, "index.json")
|
50 |
if os.path.isfile(offload_index_file):
|
51 |
with open(offload_index_file, "r", encoding="utf-8") as f:
|
@@ -53,15 +44,11 @@ def save_offload_index(index, offload_folder):
|
|
53 |
else:
|
54 |
current_index = {}
|
55 |
current_index.update(index)
|
56 |
-
|
57 |
with open(offload_index_file, "w", encoding="utf-8") as f:
|
58 |
json.dump(current_index, f, indent=2)
|
59 |
-
|
60 |
-
|
61 |
def offload_state_dict(save_dir: Union[str, os.PathLike], state_dict: Dict[str, torch.Tensor]):
|
62 |
"""
|
63 |
Offload a state dict in a given folder.
|
64 |
-
|
65 |
Args:
|
66 |
save_dir (`str` or `os.PathLike`):
|
67 |
The directory in which to offload the state dict.
|
@@ -72,15 +59,11 @@ def offload_state_dict(save_dir: Union[str, os.PathLike], state_dict: Dict[str,
|
|
72 |
index = {}
|
73 |
for name, parameter in state_dict.items():
|
74 |
index = offload_weight(parameter, name, save_dir, index=index)
|
75 |
-
|
76 |
# Update index
|
77 |
save_offload_index(index, save_dir)
|
78 |
-
|
79 |
-
|
80 |
class PrefixedDataset(Mapping):
|
81 |
"""
|
82 |
Will access keys in a given dataset by adding a prefix.
|
83 |
-
|
84 |
Args:
|
85 |
dataset (`Mapping`): Any map with string keys.
|
86 |
prefix (`str`): A prefix to add when trying to access any element in the underlying dataset.
|
@@ -88,21 +71,15 @@ class PrefixedDataset(Mapping):
|
|
88 |
def __init__(self, dataset: Mapping, prefix: str):
|
89 |
self.dataset = dataset
|
90 |
self.prefix = prefix
|
91 |
-
|
92 |
def __getitem__(self, key):
|
93 |
return self.dataset[f"{self.prefix}{key}"]
|
94 |
-
|
95 |
def __iter__(self):
|
96 |
return iter([key for key in self.dataset if key.startswith(self.prefix)])
|
97 |
-
|
98 |
def __len__(self):
|
99 |
return len(self.dataset)
|
100 |
-
|
101 |
-
|
102 |
class OffloadedWeightsLoader(Mapping):
|
103 |
"""
|
104 |
A collection that loads weights stored in a given state dict or memory-mapped on disk.
|
105 |
-
|
106 |
Args:
|
107 |
state_dict (`Dict[str, torch.Tensor]`, *optional*):
|
108 |
A dictionary parameter name to tensor.
|
@@ -121,7 +98,6 @@ class OffloadedWeightsLoader(Mapping):
|
|
121 |
):
|
122 |
if state_dict is None and save_folder is None and index is None:
|
123 |
raise ValueError("Need either a `state_dict`, a `save_folder` or an `index` containing offloaded weights.")
|
124 |
-
|
125 |
self.state_dict = {} if state_dict is None else state_dict
|
126 |
self.save_folder = save_folder
|
127 |
if index is None and save_folder is not None:
|
@@ -131,7 +107,6 @@ class OffloadedWeightsLoader(Mapping):
|
|
131 |
self.all_keys = list(self.state_dict.keys())
|
132 |
self.all_keys.extend([key for key in self.index if key not in self.all_keys])
|
133 |
self.device = device
|
134 |
-
|
135 |
def __getitem__(self, key: str):
|
136 |
# State dict gets priority
|
137 |
if key in self.state_dict:
|
@@ -147,28 +122,20 @@ class OffloadedWeightsLoader(Mapping):
|
|
147 |
# if failed to get_tensor on the device, such as bf16 on mps, try to load it on CPU first
|
148 |
with safe_open(weight_info["safetensors_file"], framework="pt", device="cpu") as f:
|
149 |
tensor = f.get_tensor(weight_info.get("weight_name", key))
|
150 |
-
|
151 |
if "dtype" in weight_info:
|
152 |
tensor = tensor.to(getattr(torch, weight_info["dtype"]))
|
153 |
-
|
154 |
if tensor.device != torch.device(device):
|
155 |
tensor = tensor.to(device)
|
156 |
return tensor
|
157 |
-
|
158 |
weight_file = os.path.join(self.save_folder, f"{key}.dat")
|
159 |
return load_offloaded_weight(weight_file, weight_info)
|
160 |
-
|
161 |
def __iter__(self):
|
162 |
return iter(self.all_keys)
|
163 |
-
|
164 |
def __len__(self):
|
165 |
return len(self.all_keys)
|
166 |
-
|
167 |
-
|
168 |
def extract_submodules_state_dict(state_dict: Dict[str, torch.Tensor], submodule_names: List[str]):
|
169 |
"""
|
170 |
Extract the sub state-dict corresponding to a list of given submodules.
|
171 |
-
|
172 |
Args:
|
173 |
state_dict (`Dict[str, torch.Tensor]`): The state dict to extract from.
|
174 |
submodule_names (`List[str]`): The list of submodule names we want to extract.
|
|
|
17 |
file_array[:] = array[:]
|
18 |
file_array.flush()
|
19 |
return index
|
|
|
|
|
20 |
def load_offloaded_weight(weight_file, weight_info):
|
21 |
shape = tuple(weight_info["shape"])
|
22 |
if shape == ():
|
23 |
# NumPy memory-mapped arrays can't have 0 dims so it was saved as 1d tensor
|
24 |
shape = (1,)
|
|
|
25 |
dtype = weight_info["dtype"]
|
26 |
if dtype == "bfloat16":
|
27 |
# NumPy does not support bfloat16 so this was saved as a int16
|
28 |
dtype = "int16"
|
|
|
29 |
weight = np.memmap(weight_file, dtype=dtype, shape=shape, mode="r")
|
|
|
30 |
if len(weight_info["shape"]) == 0:
|
31 |
weight = weight[0]
|
32 |
weight = torch.tensor(weight)
|
33 |
if weight_info["dtype"] == "bfloat16":
|
34 |
weight = weight.view(torch.bfloat16)
|
|
|
35 |
return weight
|
|
|
|
|
36 |
def save_offload_index(index, offload_folder):
|
37 |
if index is None or len(index) == 0:
|
38 |
# Nothing to save
|
39 |
return
|
|
|
40 |
offload_index_file = os.path.join(offload_folder, "index.json")
|
41 |
if os.path.isfile(offload_index_file):
|
42 |
with open(offload_index_file, "r", encoding="utf-8") as f:
|
|
|
44 |
else:
|
45 |
current_index = {}
|
46 |
current_index.update(index)
|
|
|
47 |
with open(offload_index_file, "w", encoding="utf-8") as f:
|
48 |
json.dump(current_index, f, indent=2)
|
|
|
|
|
49 |
def offload_state_dict(save_dir: Union[str, os.PathLike], state_dict: Dict[str, torch.Tensor]):
|
50 |
"""
|
51 |
Offload a state dict in a given folder.
|
|
|
52 |
Args:
|
53 |
save_dir (`str` or `os.PathLike`):
|
54 |
The directory in which to offload the state dict.
|
|
|
59 |
index = {}
|
60 |
for name, parameter in state_dict.items():
|
61 |
index = offload_weight(parameter, name, save_dir, index=index)
|
|
|
62 |
# Update index
|
63 |
save_offload_index(index, save_dir)
|
|
|
|
|
64 |
class PrefixedDataset(Mapping):
|
65 |
"""
|
66 |
Will access keys in a given dataset by adding a prefix.
|
|
|
67 |
Args:
|
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.
|
|
|
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):
|
81 |
"""
|
82 |
A collection that loads weights stored in a given state dict or memory-mapped on disk.
|
|
|
83 |
Args:
|
84 |
state_dict (`Dict[str, torch.Tensor]`, *optional*):
|
85 |
A dictionary parameter name to tensor.
|
|
|
98 |
):
|
99 |
if state_dict is None and save_folder is None and index is None:
|
100 |
raise ValueError("Need either a `state_dict`, a `save_folder` or an `index` containing offloaded weights.")
|
|
|
101 |
self.state_dict = {} if state_dict is None else state_dict
|
102 |
self.save_folder = save_folder
|
103 |
if index is None and save_folder is not None:
|
|
|
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:
|
|
|
122 |
# if failed to get_tensor on the device, such as bf16 on mps, try to load it on CPU first
|
123 |
with safe_open(weight_info["safetensors_file"], framework="pt", device="cpu") as f:
|
124 |
tensor = f.get_tensor(weight_info.get("weight_name", key))
|
|
|
125 |
if "dtype" in weight_info:
|
126 |
tensor = tensor.to(getattr(torch, weight_info["dtype"]))
|
|
|
127 |
if tensor.device != torch.device(device):
|
128 |
tensor = tensor.to(device)
|
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]):
|
137 |
"""
|
138 |
Extract the sub state-dict corresponding to a list of given submodules.
|
|
|
139 |
Args:
|
140 |
state_dict (`Dict[str, torch.Tensor]`): The state dict to extract from.
|
141 |
submodule_names (`List[str]`): The list of submodule names we want to extract.
|
src/utils/operations.py
CHANGED
@@ -3,8 +3,6 @@ A set of basic tensor ops compatible with tpu, gpu, and multigpu
|
|
3 |
"""
|
4 |
def is_torch_tensor(tensor):
|
5 |
return isinstance(tensor, torch.Tensor)
|
6 |
-
|
7 |
-
|
8 |
def is_torch_xpu_tensor(tensor):
|
9 |
return isinstance(
|
10 |
tensor,
|
@@ -16,12 +14,8 @@ def is_torch_xpu_tensor(tensor):
|
|
16 |
torch.xpu.DoubleTensor,
|
17 |
torch.xpu.BFloat16Tensor,
|
18 |
)
|
19 |
-
|
20 |
-
|
21 |
def is_tensor_information(tensor_info):
|
22 |
return isinstance(tensor_info, TensorInformation)
|
23 |
-
|
24 |
-
|
25 |
def is_namedtuple(data):
|
26 |
"""
|
27 |
Checks if `x` is a `namedtuple` or not. Can have false positives, but only if a user is trying to mimic a
|
@@ -35,8 +29,6 @@ def is_namedtuple(data):
|
|
35 |
if not isinstance(fields, tuple):
|
36 |
return False
|
37 |
return all(isinstance(member, str) for member in fields)
|
38 |
-
|
39 |
-
|
40 |
def honor_type(obj, generator):
|
41 |
"""
|
42 |
Cast a generator to the same type as obj (list, tuple, or namedtuple)
|
@@ -46,12 +38,9 @@ def honor_type(obj, generator):
|
|
46 |
return type(obj)(*list(generator))
|
47 |
else:
|
48 |
return type(obj)(generator)
|
49 |
-
|
50 |
-
|
51 |
def recursively_apply(func, data, *args, test_type=is_torch_tensor, error_on_other_type=False, **kwargs):
|
52 |
"""
|
53 |
Recursively apply a function on a data structure that is a nested list/tuple/dictionary of a given base type.
|
54 |
-
|
55 |
Args:
|
56 |
func (`callable`):
|
57 |
The function to recursively apply.
|
@@ -66,7 +55,6 @@ def recursively_apply(func, data, *args, test_type=is_torch_tensor, error_on_oth
|
|
66 |
`main_type`. If `False`, the function will leave objects of types different than `main_type` unchanged.
|
67 |
**kwargs:
|
68 |
Keyword arguments that will be passed to `func` when applied on the unpacked data.
|
69 |
-
|
70 |
Returns:
|
71 |
The same data structure as `data` with `func` applied to every object of type `main_type`.
|
72 |
"""
|
@@ -97,18 +85,14 @@ def recursively_apply(func, data, *args, test_type=is_torch_tensor, error_on_oth
|
|
97 |
f"objects that are valid for `{test_type.__name__}` should be passed."
|
98 |
)
|
99 |
return data
|
100 |
-
|
101 |
-
|
102 |
def send_to_device(tensor, device, non_blocking=False, skip_keys=None):
|
103 |
"""
|
104 |
Recursively sends the elements in a nested list/tuple/dictionary of tensors to a given device.
|
105 |
-
|
106 |
Args:
|
107 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
108 |
The data to send to a given device.
|
109 |
device (`torch.device`):
|
110 |
The device to send the data to.
|
111 |
-
|
112 |
Returns:
|
113 |
The same data structure as `tensor` with all tensors sent to the proper device.
|
114 |
"""
|
@@ -137,68 +121,49 @@ def send_to_device(tensor, device, non_blocking=False, skip_keys=None):
|
|
137 |
return tensor.to(device)
|
138 |
else:
|
139 |
return tensor
|
140 |
-
|
141 |
-
|
142 |
def get_data_structure(data):
|
143 |
"""
|
144 |
Recursively gathers the information needed to rebuild a nested list/tuple/dictionary of tensors.
|
145 |
-
|
146 |
Args:
|
147 |
data (nested list/tuple/dictionary of `torch.Tensor`):
|
148 |
The data to send to analyze.
|
149 |
-
|
150 |
Returns:
|
151 |
The same data structure as `data` with [`~utils.TensorInformation`] instead of tensors.
|
152 |
"""
|
153 |
def _get_data_structure(tensor):
|
154 |
return TensorInformation(shape=tensor.shape, dtype=tensor.dtype)
|
155 |
-
|
156 |
return recursively_apply(_get_data_structure, data)
|
157 |
-
|
158 |
-
|
159 |
def get_shape(data):
|
160 |
"""
|
161 |
Recursively gathers the shape of a nested list/tuple/dictionary of tensors as a list.
|
162 |
-
|
163 |
Args:
|
164 |
data (nested list/tuple/dictionary of `torch.Tensor`):
|
165 |
The data to send to analyze.
|
166 |
-
|
167 |
Returns:
|
168 |
The same data structure as `data` with lists of tensor shapes instead of tensors.
|
169 |
"""
|
170 |
def _get_shape(tensor):
|
171 |
return list(tensor.shape)
|
172 |
-
|
173 |
return recursively_apply(_get_shape, data)
|
174 |
-
|
175 |
-
|
176 |
def initialize_tensors(data_structure):
|
177 |
"""
|
178 |
Recursively initializes tensors from a nested list/tuple/dictionary of [`~utils.TensorInformation`].
|
179 |
-
|
180 |
Returns:
|
181 |
The same data structure as `data` with tensors instead of [`~utils.TensorInformation`].
|
182 |
"""
|
183 |
def _initialize_tensor(tensor_info):
|
184 |
return torch.empty(*tensor_info.shape, dtype=tensor_info.dtype)
|
185 |
-
|
186 |
return recursively_apply(_initialize_tensor, data_structure, test_type=is_tensor_information)
|
187 |
-
|
188 |
-
|
189 |
def find_batch_size(data):
|
190 |
"""
|
191 |
Recursively finds the batch size in a nested list/tuple/dictionary of lists of tensors.
|
192 |
-
|
193 |
Args:
|
194 |
data (nested list/tuple/dictionary of `torch.Tensor`): The data from which to find the batch size.
|
195 |
-
|
196 |
Returns:
|
197 |
`int`: The batch size.
|
198 |
"""
|
199 |
if isinstance(data, (tuple, list, Mapping)) and (len(data) == 0):
|
200 |
raise ValueError(f"Cannot find the batch size from empty {type(data)}.")
|
201 |
-
|
202 |
if isinstance(data, (tuple, list)):
|
203 |
return find_batch_size(data[0])
|
204 |
elif isinstance(data, Mapping):
|
@@ -207,15 +172,11 @@ def find_batch_size(data):
|
|
207 |
elif not isinstance(data, torch.Tensor):
|
208 |
raise TypeError(f"Can only find the batch size of tensors but got {type(data)}.")
|
209 |
return data.shape[0]
|
210 |
-
|
211 |
-
|
212 |
def listify(data):
|
213 |
"""
|
214 |
Recursively finds tensors in a nested list/tuple/dictionary and converts them to a list of numbers.
|
215 |
-
|
216 |
Args:
|
217 |
data (nested list/tuple/dictionary of `torch.Tensor`): The data from which to convert to regular numbers.
|
218 |
-
|
219 |
Returns:
|
220 |
The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
|
221 |
"""
|
@@ -227,40 +188,30 @@ def listify(data):
|
|
227 |
# Until Numpy adds bfloat16, we must convert float32.
|
228 |
tensor = tensor.to(torch.float32)
|
229 |
return tensor.tolist()
|
230 |
-
|
231 |
return recursively_apply(_convert_to_list, data)
|
232 |
-
|
233 |
-
|
234 |
def _tpu_gather(tensor):
|
235 |
def _tpu_gather_one(tensor):
|
236 |
if tensor.ndim == 0:
|
237 |
tensor = tensor.clone()[None]
|
238 |
-
|
239 |
# Can only gather contiguous tensors
|
240 |
if not tensor.is_contiguous():
|
241 |
tensor = tensor.contiguous()
|
242 |
return xm.all_gather(tensor)
|
243 |
-
|
244 |
res = recursively_apply(_tpu_gather_one, tensor, error_on_other_type=True)
|
245 |
xm.mark_step()
|
246 |
return res
|
247 |
-
|
248 |
-
|
249 |
def _gpu_gather(tensor):
|
250 |
state = PartialState()
|
251 |
if is_torch_version(">=", "1.13"):
|
252 |
gather_op = torch.distributed.all_gather_into_tensor
|
253 |
else:
|
254 |
gather_op = torch.distributed._all_gather_base
|
255 |
-
|
256 |
def _gpu_gather_one(tensor):
|
257 |
if tensor.ndim == 0:
|
258 |
tensor = tensor.clone()[None]
|
259 |
-
|
260 |
# Can only gather contiguous tensors
|
261 |
if not tensor.is_contiguous():
|
262 |
tensor = tensor.contiguous()
|
263 |
-
|
264 |
if state.backend is not None and state.backend != "gloo":
|
265 |
# We use `empty` as `all_gather_into_tensor` slightly
|
266 |
# differs from `all_gather` for better efficiency,
|
@@ -280,18 +231,13 @@ def _gpu_gather(tensor):
|
|
280 |
output_tensors = [torch.empty_like(tensor) for _ in range(state.num_processes)]
|
281 |
torch.distributed.all_gather(output_tensors, tensor)
|
282 |
return torch.cat(output_tensors, dim=0)
|
283 |
-
|
284 |
return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)
|
285 |
-
|
286 |
-
|
287 |
class DistributedOperationException(Exception):
|
288 |
"""
|
289 |
An exception class for distributed operations. Raised if the operation cannot be performed due to the shape of the
|
290 |
tensors.
|
291 |
"""
|
292 |
pass
|
293 |
-
|
294 |
-
|
295 |
def verify_operation(function):
|
296 |
"""
|
297 |
Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
|
@@ -322,10 +268,7 @@ def verify_operation(function):
|
|
322 |
f"\n\nOperation: `{operation}`\nInput shapes:\n - {process_shape_str}"
|
323 |
)
|
324 |
return function(*args, **kwargs)
|
325 |
-
|
326 |
return wrapper
|
327 |
-
|
328 |
-
|
329 |
def chained_operation(function):
|
330 |
"""
|
331 |
Checks that `verify_operation` failed and if so reports a more helpful error chaining the existing
|
@@ -340,19 +283,14 @@ def chained_operation(function):
|
|
340 |
raise DistributedOperationException(
|
341 |
f"Error found while calling `{operation}`. Please see the earlier error for more details."
|
342 |
) from e
|
343 |
-
|
344 |
return wrapper
|
345 |
-
|
346 |
-
|
347 |
@verify_operation
|
348 |
def gather(tensor):
|
349 |
"""
|
350 |
Recursively gather tensor in a nested list/tuple/dictionary of tensors from all devices.
|
351 |
-
|
352 |
Args:
|
353 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
354 |
The data to gather.
|
355 |
-
|
356 |
Returns:
|
357 |
The same data structure as `tensor` with all tensors sent to the proper device.
|
358 |
"""
|
@@ -362,23 +300,17 @@ def gather(tensor):
|
|
362 |
return _gpu_gather(tensor)
|
363 |
else:
|
364 |
return tensor
|
365 |
-
|
366 |
-
|
367 |
def _gpu_gather_object(object: Any):
|
368 |
output_objects = [None for _ in range(PartialState().num_processes)]
|
369 |
torch.distributed.all_gather_object(output_objects, object)
|
370 |
# all_gather_object returns a list of lists, so we need to flatten it
|
371 |
return [x for y in output_objects for x in y]
|
372 |
-
|
373 |
-
|
374 |
def gather_object(object: Any):
|
375 |
"""
|
376 |
Recursively gather object in a nested list/tuple/dictionary of objects from all devices.
|
377 |
-
|
378 |
Args:
|
379 |
object (nested list/tuple/dictionary of picklable object):
|
380 |
The data to gather.
|
381 |
-
|
382 |
Returns:
|
383 |
The same data structure as `object` with all the objects sent to every device.
|
384 |
"""
|
@@ -388,35 +320,26 @@ def gather_object(object: Any):
|
|
388 |
return _gpu_gather_object(object)
|
389 |
else:
|
390 |
return object
|
391 |
-
|
392 |
-
|
393 |
def _gpu_broadcast(data, src=0):
|
394 |
def _gpu_broadcast_one(tensor, src=0):
|
395 |
torch.distributed.broadcast(tensor, src=src)
|
396 |
return tensor
|
397 |
-
|
398 |
return recursively_apply(_gpu_broadcast_one, data, error_on_other_type=True, src=src)
|
399 |
-
|
400 |
-
|
401 |
def _tpu_broadcast(tensor, src=0, name="broadcast tensor"):
|
402 |
if isinstance(tensor, (list, tuple)):
|
403 |
return honor_type(tensor, (_tpu_broadcast(t, name=f"{name}_{i}") for i, t in enumerate(tensor)))
|
404 |
elif isinstance(tensor, Mapping):
|
405 |
return type(tensor)({k: _tpu_broadcast(v, name=f"{name}_{k}") for k, v in tensor.items()})
|
406 |
return xm.mesh_reduce(name, tensor, lambda x: x[src])
|
407 |
-
|
408 |
-
|
409 |
@verify_operation
|
410 |
def broadcast(tensor, from_process: int = 0):
|
411 |
"""
|
412 |
Recursively broadcast tensor in a nested list/tuple/dictionary of tensors to all devices.
|
413 |
-
|
414 |
Args:
|
415 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
416 |
The data to gather.
|
417 |
from_process (`int`, *optional*, defaults to 0):
|
418 |
The process from which to send the data
|
419 |
-
|
420 |
Returns:
|
421 |
The same data structure as `tensor` with all tensors broadcasted to the proper device.
|
422 |
"""
|
@@ -426,18 +349,14 @@ def broadcast(tensor, from_process: int = 0):
|
|
426 |
return _gpu_broadcast(tensor, src=from_process)
|
427 |
else:
|
428 |
return tensor
|
429 |
-
|
430 |
-
|
431 |
def broadcast_object_list(object_list, from_process: int = 0):
|
432 |
"""
|
433 |
Broadcast a list of picklable objects form one process to the others.
|
434 |
-
|
435 |
Args:
|
436 |
object_list (list of picklable objects):
|
437 |
The list of objects to broadcast. This list will be modified inplace.
|
438 |
from_process (`int`, *optional*, defaults to 0):
|
439 |
The process from which to send the data.
|
440 |
-
|
441 |
Returns:
|
442 |
The same list containing the objects from process 0.
|
443 |
"""
|
@@ -447,37 +366,28 @@ def broadcast_object_list(object_list, from_process: int = 0):
|
|
447 |
elif PartialState().distributed_type in TORCH_DISTRIBUTED_OPERATION_TYPES:
|
448 |
torch.distributed.broadcast_object_list(object_list, src=from_process)
|
449 |
return object_list
|
450 |
-
|
451 |
-
|
452 |
def slice_tensors(data, tensor_slice, process_index=None, num_processes=None):
|
453 |
"""
|
454 |
Recursively takes a slice in a nested list/tuple/dictionary of tensors.
|
455 |
-
|
456 |
Args:
|
457 |
data (nested list/tuple/dictionary of `torch.Tensor`):
|
458 |
The data to slice.
|
459 |
tensor_slice (`slice`):
|
460 |
The slice to take.
|
461 |
-
|
462 |
Returns:
|
463 |
The same data structure as `data` with all the tensors slices.
|
464 |
"""
|
465 |
def _slice_tensor(tensor, tensor_slice):
|
466 |
return tensor[tensor_slice]
|
467 |
-
|
468 |
return recursively_apply(_slice_tensor, data, tensor_slice)
|
469 |
-
|
470 |
-
|
471 |
def concatenate(data, dim=0):
|
472 |
"""
|
473 |
Recursively concatenate the tensors in a nested list/tuple/dictionary of lists of tensors with the same shape.
|
474 |
-
|
475 |
Args:
|
476 |
data (nested list/tuple/dictionary of lists of tensors `torch.Tensor`):
|
477 |
The data to concatenate.
|
478 |
dim (`int`, *optional*, defaults to 0):
|
479 |
The dimension on which to concatenate.
|
480 |
-
|
481 |
Returns:
|
482 |
The same data structure as `data` with all the tensors concatenated.
|
483 |
"""
|
@@ -488,18 +398,13 @@ def concatenate(data, dim=0):
|
|
488 |
elif not isinstance(data[0], torch.Tensor):
|
489 |
raise TypeError(f"Can only concatenate tensors but got {type(data[0])}")
|
490 |
return torch.cat(data, dim=dim)
|
491 |
-
|
492 |
-
|
493 |
class CannotPadNestedTensorWarning(UserWarning):
|
494 |
pass
|
495 |
-
|
496 |
-
|
497 |
@chained_operation
|
498 |
def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
499 |
"""
|
500 |
Recursively pad the tensors in a nested list/tuple/dictionary of tensors from all devices to the same size so they
|
501 |
can safely be gathered.
|
502 |
-
|
503 |
Args:
|
504 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
505 |
The data to gather.
|
@@ -519,7 +424,6 @@ def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
|
519 |
return tensor
|
520 |
if dim >= len(tensor.shape):
|
521 |
return tensor
|
522 |
-
|
523 |
# Gather all sizes
|
524 |
size = torch.tensor(tensor.shape, device=tensor.device)[None]
|
525 |
sizes = gather(size).cpu()
|
@@ -527,7 +431,6 @@ def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
|
527 |
max_size = max(s[dim] for s in sizes)
|
528 |
if max_size == tensor.shape[dim]:
|
529 |
return tensor
|
530 |
-
|
531 |
old_size = tensor.shape
|
532 |
new_size = list(old_size)
|
533 |
new_size[dim] = max_size
|
@@ -540,18 +443,14 @@ def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
|
540 |
indices = tuple(slice(0, old_size[dim]) if i == dim else slice(None) for i in range(len(new_size)))
|
541 |
new_tensor[indices] = tensor
|
542 |
return new_tensor
|
543 |
-
|
544 |
return recursively_apply(
|
545 |
_pad_across_processes, tensor, error_on_other_type=True, dim=dim, pad_index=pad_index, pad_first=pad_first
|
546 |
)
|
547 |
-
|
548 |
-
|
549 |
@verify_operation
|
550 |
def reduce(tensor, reduction="mean", scale=1.0):
|
551 |
"""
|
552 |
Recursively reduce the tensors in a nested list/tuple/dictionary of lists of tensors across all processes by the
|
553 |
mean of a given operation.
|
554 |
-
|
555 |
Args:
|
556 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
557 |
The data to reduce.
|
@@ -559,7 +458,6 @@ def reduce(tensor, reduction="mean", scale=1.0):
|
|
559 |
A reduction method. Can be of "mean", "sum", or "none"
|
560 |
scale (`float`, *optional*):
|
561 |
A default scaling value to be applied after the reduce, only valied on XLA.
|
562 |
-
|
563 |
Returns:
|
564 |
The same data structure as `data` with all the tensors reduced.
|
565 |
"""
|
@@ -575,73 +473,52 @@ def reduce(tensor, reduction="mean", scale=1.0):
|
|
575 |
if reduction == "mean":
|
576 |
cloned_tensor /= state.num_processes
|
577 |
return cloned_tensor
|
578 |
-
|
579 |
return recursively_apply(
|
580 |
_reduce_across_processes, tensor, error_on_other_type=True, reduction=reduction, scale=scale
|
581 |
)
|
582 |
-
|
583 |
-
|
584 |
def convert_to_fp32(tensor):
|
585 |
"""
|
586 |
Recursively converts the elements nested list/tuple/dictionary of tensors in FP16/BF16 precision to FP32.
|
587 |
-
|
588 |
Args:
|
589 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
590 |
The data to convert from FP16/BF16 to FP32.
|
591 |
-
|
592 |
Returns:
|
593 |
The same data structure as `tensor` with all tensors that were in FP16/BF16 precision converted to FP32.
|
594 |
"""
|
595 |
def _convert_to_fp32(tensor):
|
596 |
return tensor.float()
|
597 |
-
|
598 |
def _is_fp16_bf16_tensor(tensor):
|
599 |
return hasattr(tensor, "dtype") and tensor.dtype in (torch.float16, torch.bfloat16)
|
600 |
-
|
601 |
return recursively_apply(_convert_to_fp32, tensor, test_type=_is_fp16_bf16_tensor)
|
602 |
-
|
603 |
-
|
604 |
class ConvertOutputsToFp32:
|
605 |
"""
|
606 |
Decorator to apply to a function outputing tensors (like a model forward pass) that ensures the outputs in FP16
|
607 |
precision will be convert back to FP32.
|
608 |
-
|
609 |
Args:
|
610 |
model_forward (`Callable`):
|
611 |
The function which outputs we want to treat.
|
612 |
-
|
613 |
Returns:
|
614 |
The same function as `model_forward` but with converted outputs.
|
615 |
"""
|
616 |
def __init__(self, model_forward):
|
617 |
self.model_forward = model_forward
|
618 |
update_wrapper(self, model_forward)
|
619 |
-
|
620 |
def __call__(self, *args, **kwargs):
|
621 |
return convert_to_fp32(self.model_forward(*args, **kwargs))
|
622 |
-
|
623 |
def __getstate__(self):
|
624 |
raise pickle.PicklingError(
|
625 |
"Cannot pickle a prepared model with automatic mixed precision, please unwrap the model with `Accelerator.unwrap_model(model)` before pickling it."
|
626 |
)
|
627 |
-
|
628 |
-
|
629 |
def convert_outputs_to_fp32(model_forward):
|
630 |
model_forward = ConvertOutputsToFp32(model_forward)
|
631 |
-
|
632 |
def forward(*args, **kwargs):
|
633 |
return model_forward(*args, **kwargs)
|
634 |
-
|
635 |
# To act like a decorator so that it can be popped when doing `extract_model_from_parallel`
|
636 |
forward.__wrapped__ = model_forward
|
637 |
-
|
638 |
return forward
|
639 |
-
|
640 |
-
|
641 |
def find_device(data):
|
642 |
"""
|
643 |
Finds the device on which a nested dict/list/tuple of tensors lies (assuming they are all on the same device).
|
644 |
-
|
645 |
Args:
|
646 |
(nested list/tuple/dictionary of `torch.Tensor`): The data we want to know the device of.
|
647 |
"""
|
|
|
3 |
"""
|
4 |
def is_torch_tensor(tensor):
|
5 |
return isinstance(tensor, torch.Tensor)
|
|
|
|
|
6 |
def is_torch_xpu_tensor(tensor):
|
7 |
return isinstance(
|
8 |
tensor,
|
|
|
14 |
torch.xpu.DoubleTensor,
|
15 |
torch.xpu.BFloat16Tensor,
|
16 |
)
|
|
|
|
|
17 |
def is_tensor_information(tensor_info):
|
18 |
return isinstance(tensor_info, TensorInformation)
|
|
|
|
|
19 |
def is_namedtuple(data):
|
20 |
"""
|
21 |
Checks if `x` is a `namedtuple` or not. Can have false positives, but only if a user is trying to mimic a
|
|
|
29 |
if not isinstance(fields, tuple):
|
30 |
return False
|
31 |
return all(isinstance(member, str) for member in fields)
|
|
|
|
|
32 |
def honor_type(obj, generator):
|
33 |
"""
|
34 |
Cast a generator to the same type as obj (list, tuple, or namedtuple)
|
|
|
38 |
return type(obj)(*list(generator))
|
39 |
else:
|
40 |
return type(obj)(generator)
|
|
|
|
|
41 |
def recursively_apply(func, data, *args, test_type=is_torch_tensor, error_on_other_type=False, **kwargs):
|
42 |
"""
|
43 |
Recursively apply a function on a data structure that is a nested list/tuple/dictionary of a given base type.
|
|
|
44 |
Args:
|
45 |
func (`callable`):
|
46 |
The function to recursively apply.
|
|
|
55 |
`main_type`. If `False`, the function will leave objects of types different than `main_type` unchanged.
|
56 |
**kwargs:
|
57 |
Keyword arguments that will be passed to `func` when applied on the unpacked data.
|
|
|
58 |
Returns:
|
59 |
The same data structure as `data` with `func` applied to every object of type `main_type`.
|
60 |
"""
|
|
|
85 |
f"objects that are valid for `{test_type.__name__}` should be passed."
|
86 |
)
|
87 |
return data
|
|
|
|
|
88 |
def send_to_device(tensor, device, non_blocking=False, skip_keys=None):
|
89 |
"""
|
90 |
Recursively sends the elements in a nested list/tuple/dictionary of tensors to a given device.
|
|
|
91 |
Args:
|
92 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
93 |
The data to send to a given device.
|
94 |
device (`torch.device`):
|
95 |
The device to send the data to.
|
|
|
96 |
Returns:
|
97 |
The same data structure as `tensor` with all tensors sent to the proper device.
|
98 |
"""
|
|
|
121 |
return tensor.to(device)
|
122 |
else:
|
123 |
return tensor
|
|
|
|
|
124 |
def get_data_structure(data):
|
125 |
"""
|
126 |
Recursively gathers the information needed to rebuild a nested list/tuple/dictionary of tensors.
|
|
|
127 |
Args:
|
128 |
data (nested list/tuple/dictionary of `torch.Tensor`):
|
129 |
The data to send to analyze.
|
|
|
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)
|
|
|
|
|
136 |
def get_shape(data):
|
137 |
"""
|
138 |
Recursively gathers the shape of a nested list/tuple/dictionary of tensors as a list.
|
|
|
139 |
Args:
|
140 |
data (nested list/tuple/dictionary of `torch.Tensor`):
|
141 |
The data to send to analyze.
|
|
|
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)
|
|
|
|
|
148 |
def initialize_tensors(data_structure):
|
149 |
"""
|
150 |
Recursively initializes tensors from a nested list/tuple/dictionary of [`~utils.TensorInformation`].
|
|
|
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)
|
|
|
|
|
157 |
def find_batch_size(data):
|
158 |
"""
|
159 |
Recursively finds the batch size in a nested list/tuple/dictionary of lists of tensors.
|
|
|
160 |
Args:
|
161 |
data (nested list/tuple/dictionary of `torch.Tensor`): The data from which to find the batch size.
|
|
|
162 |
Returns:
|
163 |
`int`: The batch size.
|
164 |
"""
|
165 |
if isinstance(data, (tuple, list, Mapping)) and (len(data) == 0):
|
166 |
raise ValueError(f"Cannot find the batch size from empty {type(data)}.")
|
|
|
167 |
if isinstance(data, (tuple, list)):
|
168 |
return find_batch_size(data[0])
|
169 |
elif isinstance(data, Mapping):
|
|
|
172 |
elif not isinstance(data, torch.Tensor):
|
173 |
raise TypeError(f"Can only find the batch size of tensors but got {type(data)}.")
|
174 |
return data.shape[0]
|
|
|
|
|
175 |
def listify(data):
|
176 |
"""
|
177 |
Recursively finds tensors in a nested list/tuple/dictionary and converts them to a list of numbers.
|
|
|
178 |
Args:
|
179 |
data (nested list/tuple/dictionary of `torch.Tensor`): The data from which to convert to regular numbers.
|
|
|
180 |
Returns:
|
181 |
The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
|
182 |
"""
|
|
|
188 |
# Until Numpy adds bfloat16, we must convert float32.
|
189 |
tensor = tensor.to(torch.float32)
|
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]
|
|
|
196 |
# Can only gather contiguous tensors
|
197 |
if not tensor.is_contiguous():
|
198 |
tensor = tensor.contiguous()
|
199 |
return xm.all_gather(tensor)
|
|
|
200 |
res = recursively_apply(_tpu_gather_one, tensor, error_on_other_type=True)
|
201 |
xm.mark_step()
|
202 |
return res
|
|
|
|
|
203 |
def _gpu_gather(tensor):
|
204 |
state = PartialState()
|
205 |
if is_torch_version(">=", "1.13"):
|
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]
|
|
|
212 |
# Can only gather contiguous tensors
|
213 |
if not tensor.is_contiguous():
|
214 |
tensor = tensor.contiguous()
|
|
|
215 |
if state.backend is not None and state.backend != "gloo":
|
216 |
# We use `empty` as `all_gather_into_tensor` slightly
|
217 |
# differs from `all_gather` for better efficiency,
|
|
|
231 |
output_tensors = [torch.empty_like(tensor) for _ in range(state.num_processes)]
|
232 |
torch.distributed.all_gather(output_tensors, tensor)
|
233 |
return torch.cat(output_tensors, dim=0)
|
|
|
234 |
return recursively_apply(_gpu_gather_one, tensor, error_on_other_type=True)
|
|
|
|
|
235 |
class DistributedOperationException(Exception):
|
236 |
"""
|
237 |
An exception class for distributed operations. Raised if the operation cannot be performed due to the shape of the
|
238 |
tensors.
|
239 |
"""
|
240 |
pass
|
|
|
|
|
241 |
def verify_operation(function):
|
242 |
"""
|
243 |
Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
|
|
|
268 |
f"\n\nOperation: `{operation}`\nInput shapes:\n - {process_shape_str}"
|
269 |
)
|
270 |
return function(*args, **kwargs)
|
|
|
271 |
return wrapper
|
|
|
|
|
272 |
def chained_operation(function):
|
273 |
"""
|
274 |
Checks that `verify_operation` failed and if so reports a more helpful error chaining the existing
|
|
|
283 |
raise DistributedOperationException(
|
284 |
f"Error found while calling `{operation}`. Please see the earlier error for more details."
|
285 |
) from e
|
|
|
286 |
return wrapper
|
|
|
|
|
287 |
@verify_operation
|
288 |
def gather(tensor):
|
289 |
"""
|
290 |
Recursively gather tensor in a nested list/tuple/dictionary of tensors from all devices.
|
|
|
291 |
Args:
|
292 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
293 |
The data to gather.
|
|
|
294 |
Returns:
|
295 |
The same data structure as `tensor` with all tensors sent to the proper device.
|
296 |
"""
|
|
|
300 |
return _gpu_gather(tensor)
|
301 |
else:
|
302 |
return tensor
|
|
|
|
|
303 |
def _gpu_gather_object(object: Any):
|
304 |
output_objects = [None for _ in range(PartialState().num_processes)]
|
305 |
torch.distributed.all_gather_object(output_objects, object)
|
306 |
# all_gather_object returns a list of lists, so we need to flatten it
|
307 |
return [x for y in output_objects for x in y]
|
|
|
|
|
308 |
def gather_object(object: Any):
|
309 |
"""
|
310 |
Recursively gather object in a nested list/tuple/dictionary of objects from all devices.
|
|
|
311 |
Args:
|
312 |
object (nested list/tuple/dictionary of picklable object):
|
313 |
The data to gather.
|
|
|
314 |
Returns:
|
315 |
The same data structure as `object` with all the objects sent to every device.
|
316 |
"""
|
|
|
320 |
return _gpu_gather_object(object)
|
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
|
|
|
327 |
return recursively_apply(_gpu_broadcast_one, data, error_on_other_type=True, src=src)
|
|
|
|
|
328 |
def _tpu_broadcast(tensor, src=0, name="broadcast tensor"):
|
329 |
if isinstance(tensor, (list, tuple)):
|
330 |
return honor_type(tensor, (_tpu_broadcast(t, name=f"{name}_{i}") for i, t in enumerate(tensor)))
|
331 |
elif isinstance(tensor, Mapping):
|
332 |
return type(tensor)({k: _tpu_broadcast(v, name=f"{name}_{k}") for k, v in tensor.items()})
|
333 |
return xm.mesh_reduce(name, tensor, lambda x: x[src])
|
|
|
|
|
334 |
@verify_operation
|
335 |
def broadcast(tensor, from_process: int = 0):
|
336 |
"""
|
337 |
Recursively broadcast tensor in a nested list/tuple/dictionary of tensors to all devices.
|
|
|
338 |
Args:
|
339 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
340 |
The data to gather.
|
341 |
from_process (`int`, *optional*, defaults to 0):
|
342 |
The process from which to send the data
|
|
|
343 |
Returns:
|
344 |
The same data structure as `tensor` with all tensors broadcasted to the proper device.
|
345 |
"""
|
|
|
349 |
return _gpu_broadcast(tensor, src=from_process)
|
350 |
else:
|
351 |
return tensor
|
|
|
|
|
352 |
def broadcast_object_list(object_list, from_process: int = 0):
|
353 |
"""
|
354 |
Broadcast a list of picklable objects form one process to the others.
|
|
|
355 |
Args:
|
356 |
object_list (list of picklable objects):
|
357 |
The list of objects to broadcast. This list will be modified inplace.
|
358 |
from_process (`int`, *optional*, defaults to 0):
|
359 |
The process from which to send the data.
|
|
|
360 |
Returns:
|
361 |
The same list containing the objects from process 0.
|
362 |
"""
|
|
|
366 |
elif PartialState().distributed_type in TORCH_DISTRIBUTED_OPERATION_TYPES:
|
367 |
torch.distributed.broadcast_object_list(object_list, src=from_process)
|
368 |
return object_list
|
|
|
|
|
369 |
def slice_tensors(data, tensor_slice, process_index=None, num_processes=None):
|
370 |
"""
|
371 |
Recursively takes a slice in a nested list/tuple/dictionary of tensors.
|
|
|
372 |
Args:
|
373 |
data (nested list/tuple/dictionary of `torch.Tensor`):
|
374 |
The data to slice.
|
375 |
tensor_slice (`slice`):
|
376 |
The slice to take.
|
|
|
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)
|
|
|
|
|
383 |
def concatenate(data, dim=0):
|
384 |
"""
|
385 |
Recursively concatenate the tensors in a nested list/tuple/dictionary of lists of tensors with the same shape.
|
|
|
386 |
Args:
|
387 |
data (nested list/tuple/dictionary of lists of tensors `torch.Tensor`):
|
388 |
The data to concatenate.
|
389 |
dim (`int`, *optional*, defaults to 0):
|
390 |
The dimension on which to concatenate.
|
|
|
391 |
Returns:
|
392 |
The same data structure as `data` with all the tensors concatenated.
|
393 |
"""
|
|
|
398 |
elif not isinstance(data[0], torch.Tensor):
|
399 |
raise TypeError(f"Can only concatenate tensors but got {type(data[0])}")
|
400 |
return torch.cat(data, dim=dim)
|
|
|
|
|
401 |
class CannotPadNestedTensorWarning(UserWarning):
|
402 |
pass
|
|
|
|
|
403 |
@chained_operation
|
404 |
def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
405 |
"""
|
406 |
Recursively pad the tensors in a nested list/tuple/dictionary of tensors from all devices to the same size so they
|
407 |
can safely be gathered.
|
|
|
408 |
Args:
|
409 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
410 |
The data to gather.
|
|
|
424 |
return tensor
|
425 |
if dim >= len(tensor.shape):
|
426 |
return tensor
|
|
|
427 |
# Gather all sizes
|
428 |
size = torch.tensor(tensor.shape, device=tensor.device)[None]
|
429 |
sizes = gather(size).cpu()
|
|
|
431 |
max_size = max(s[dim] for s in sizes)
|
432 |
if max_size == tensor.shape[dim]:
|
433 |
return tensor
|
|
|
434 |
old_size = tensor.shape
|
435 |
new_size = list(old_size)
|
436 |
new_size[dim] = max_size
|
|
|
443 |
indices = tuple(slice(0, old_size[dim]) if i == dim else slice(None) for i in range(len(new_size)))
|
444 |
new_tensor[indices] = tensor
|
445 |
return new_tensor
|
|
|
446 |
return recursively_apply(
|
447 |
_pad_across_processes, tensor, error_on_other_type=True, dim=dim, pad_index=pad_index, pad_first=pad_first
|
448 |
)
|
|
|
|
|
449 |
@verify_operation
|
450 |
def reduce(tensor, reduction="mean", scale=1.0):
|
451 |
"""
|
452 |
Recursively reduce the tensors in a nested list/tuple/dictionary of lists of tensors across all processes by the
|
453 |
mean of a given operation.
|
|
|
454 |
Args:
|
455 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
456 |
The data to reduce.
|
|
|
458 |
A reduction method. Can be of "mean", "sum", or "none"
|
459 |
scale (`float`, *optional*):
|
460 |
A default scaling value to be applied after the reduce, only valied on XLA.
|
|
|
461 |
Returns:
|
462 |
The same data structure as `data` with all the tensors reduced.
|
463 |
"""
|
|
|
473 |
if reduction == "mean":
|
474 |
cloned_tensor /= state.num_processes
|
475 |
return cloned_tensor
|
|
|
476 |
return recursively_apply(
|
477 |
_reduce_across_processes, tensor, error_on_other_type=True, reduction=reduction, scale=scale
|
478 |
)
|
|
|
|
|
479 |
def convert_to_fp32(tensor):
|
480 |
"""
|
481 |
Recursively converts the elements nested list/tuple/dictionary of tensors in FP16/BF16 precision to FP32.
|
|
|
482 |
Args:
|
483 |
tensor (nested list/tuple/dictionary of `torch.Tensor`):
|
484 |
The data to convert from FP16/BF16 to FP32.
|
|
|
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)
|
|
|
|
|
493 |
class ConvertOutputsToFp32:
|
494 |
"""
|
495 |
Decorator to apply to a function outputing tensors (like a model forward pass) that ensures the outputs in FP16
|
496 |
precision will be convert back to FP32.
|
|
|
497 |
Args:
|
498 |
model_forward (`Callable`):
|
499 |
The function which outputs we want to treat.
|
|
|
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`
|
517 |
forward.__wrapped__ = model_forward
|
|
|
518 |
return forward
|
|
|
|
|
519 |
def find_device(data):
|
520 |
"""
|
521 |
Finds the device on which a nested dict/list/tuple of tensors lies (assuming they are all on the same device).
|
|
|
522 |
Args:
|
523 |
(nested list/tuple/dictionary of `torch.Tensor`): The data we want to know the device of.
|
524 |
"""
|
src/utils/other.py
CHANGED
@@ -1,10 +1,6 @@
|
|
1 |
logger = get_logger(__name__)
|
2 |
-
|
3 |
-
|
4 |
if is_tpu_available(check_device=False):
|
5 |
import torch_xla.core.xla_model as xm
|
6 |
-
|
7 |
-
|
8 |
def is_compiled_module(module):
|
9 |
"""
|
10 |
Check whether the module was compiled with torch.compile()
|
@@ -12,41 +8,30 @@ def is_compiled_module(module):
|
|
12 |
if is_torch_version("<", "2.0.0") or not hasattr(torch, "_dynamo"):
|
13 |
return False
|
14 |
return isinstance(module, torch._dynamo.eval_frame.OptimizedModule)
|
15 |
-
|
16 |
-
|
17 |
def extract_model_from_parallel(model, keep_fp32_wrapper: bool = True):
|
18 |
"""
|
19 |
Extract a model from its distributed containers.
|
20 |
-
|
21 |
Args:
|
22 |
model (`torch.nn.Module`):
|
23 |
The model to extract.
|
24 |
keep_fp32_wrapper (`bool`, *optional*):
|
25 |
Whether to remove mixed precision hooks from the model.
|
26 |
-
|
27 |
Returns:
|
28 |
`torch.nn.Module`: The extracted model.
|
29 |
"""
|
30 |
options = (torch.nn.parallel.DistributedDataParallel, torch.nn.DataParallel)
|
31 |
-
|
32 |
is_compiled = is_compiled_module(model)
|
33 |
if is_compiled:
|
34 |
compiled_model = model
|
35 |
model = model._orig_mod
|
36 |
-
|
37 |
if is_deepspeed_available():
|
38 |
from deepspeed import DeepSpeedEngine
|
39 |
-
|
40 |
options += (DeepSpeedEngine,)
|
41 |
-
|
42 |
if is_torch_version(">=", FSDP_PYTORCH_VERSION) and is_torch_distributed_available():
|
43 |
from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel as FSDP
|
44 |
-
|
45 |
options += (FSDP,)
|
46 |
-
|
47 |
while isinstance(model, options):
|
48 |
model = model.module
|
49 |
-
|
50 |
if not keep_fp32_wrapper:
|
51 |
forward = getattr(model, "forward")
|
52 |
original_forward = model.__dict__.pop("_original_forward", None)
|
@@ -58,31 +43,21 @@ def extract_model_from_parallel(model, keep_fp32_wrapper: bool = True):
|
|
58 |
model.forward = MethodType(forward, model)
|
59 |
if getattr(model, "_converted_to_transformer_engine", False):
|
60 |
convert_model(model, to_transformer_engine=False)
|
61 |
-
|
62 |
if is_compiled:
|
63 |
compiled_model._orig_mod = model
|
64 |
model = compiled_model
|
65 |
-
|
66 |
return model
|
67 |
-
|
68 |
-
|
69 |
def wait_for_everyone():
|
70 |
"""
|
71 |
Introduces a blocking point in the script, making sure all processes have reached this point before continuing.
|
72 |
-
|
73 |
<Tip warning={true}>
|
74 |
-
|
75 |
Make sure all processes will reach this instruction otherwise one of your processes will hang forever.
|
76 |
-
|
77 |
</Tip>
|
78 |
"""
|
79 |
PartialState().wait_for_everyone()
|
80 |
-
|
81 |
-
|
82 |
def clean_state_dict_for_safetensors(state_dict: dict):
|
83 |
"""
|
84 |
Cleans the state dictionary from a model and removes tensor aliasing if present.
|
85 |
-
|
86 |
Args:
|
87 |
state_dict (`dict`):
|
88 |
The state dictionary from a model
|
@@ -92,7 +67,6 @@ def clean_state_dict_for_safetensors(state_dict: dict):
|
|
92 |
for name, tensor in state_dict.items():
|
93 |
if not isinstance(tensor, str):
|
94 |
ptrs[id_tensor_storage(tensor)].append(name)
|
95 |
-
|
96 |
# These are all pointers of tensors with shared memory
|
97 |
shared_ptrs = {ptr: names for ptr, names in ptrs.items() if len(names) > 1}
|
98 |
warn_names = set()
|
@@ -112,12 +86,9 @@ def clean_state_dict_for_safetensors(state_dict: dict):
|
|
112 |
)
|
113 |
state_dict = {k: v.contiguous() if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()}
|
114 |
return state_dict
|
115 |
-
|
116 |
-
|
117 |
def save(obj, f, save_on_each_node: bool = False, safe_serialization: bool = False):
|
118 |
"""
|
119 |
Save the data to disk. Use in place of `torch.save()`.
|
120 |
-
|
121 |
Args:
|
122 |
obj:
|
123 |
The data to save
|
@@ -135,28 +106,21 @@ def save(obj, f, save_on_each_node: bool = False, safe_serialization: bool = Fal
|
|
135 |
obj = clean_state_dict_for_safetensors(obj)
|
136 |
else:
|
137 |
save_func = torch.save
|
138 |
-
|
139 |
if PartialState().distributed_type == DistributedType.TPU:
|
140 |
xm.save(obj, f)
|
141 |
elif PartialState().is_main_process and not save_on_each_node:
|
142 |
save_func(obj, f)
|
143 |
elif PartialState().is_local_main_process and save_on_each_node:
|
144 |
save_func(obj, f)
|
145 |
-
|
146 |
-
|
147 |
@contextmanager
|
148 |
def clear_environment():
|
149 |
"""
|
150 |
A context manager that will cache origin `os.environ` and replace it with a empty dictionary in this context.
|
151 |
-
|
152 |
When this context exits, the cached `os.environ` will be back.
|
153 |
-
|
154 |
Example:
|
155 |
-
|
156 |
```python
|
157 |
>>> import os
|
158 |
>>> from accelerate.utils import clear_environment
|
159 |
-
|
160 |
>>> os.environ["FOO"] = "bar"
|
161 |
>>> with clear_environment():
|
162 |
... print(os.environ)
|
@@ -164,32 +128,23 @@ def clear_environment():
|
|
164 |
... print(os.environ["FOO"])
|
165 |
{}
|
166 |
new_bar
|
167 |
-
|
168 |
>>> print(os.environ["FOO"])
|
169 |
bar
|
170 |
```
|
171 |
"""
|
172 |
_old_os_environ = os.environ
|
173 |
os.environ = dict()
|
174 |
-
|
175 |
yield
|
176 |
-
|
177 |
os.environ = _old_os_environ
|
178 |
-
|
179 |
-
|
180 |
@contextmanager
|
181 |
def patch_environment(**kwargs):
|
182 |
"""
|
183 |
A context manager that will add each keyword argument passed to `os.environ` and remove them when exiting.
|
184 |
-
|
185 |
Will convert the values in `kwargs` to strings and upper-case all the keys.
|
186 |
-
|
187 |
Example:
|
188 |
-
|
189 |
```python
|
190 |
>>> import os
|
191 |
>>> from accelerate.utils import patch_environment
|
192 |
-
|
193 |
>>> with patch_environment(FOO="bar"):
|
194 |
... print(os.environ["FOO"]) # prints "bar"
|
195 |
>>> print(os.environ["FOO"]) # raises KeyError
|
@@ -201,9 +156,7 @@ def patch_environment(**kwargs):
|
|
201 |
if key in os.environ:
|
202 |
existing_vars[key] = os.environ[key]
|
203 |
os.environ[key] = str(value)
|
204 |
-
|
205 |
yield
|
206 |
-
|
207 |
for key in kwargs:
|
208 |
key = key.upper()
|
209 |
if key in existing_vars:
|
@@ -211,8 +164,6 @@ def patch_environment(**kwargs):
|
|
211 |
os.environ[key] = existing_vars[key]
|
212 |
else:
|
213 |
os.environ.pop(key, None)
|
214 |
-
|
215 |
-
|
216 |
def get_pretty_name(obj):
|
217 |
"""
|
218 |
Gets a pretty name from `obj`.
|
@@ -224,12 +175,9 @@ def get_pretty_name(obj):
|
|
224 |
if hasattr(obj, "__name__"):
|
225 |
return obj.__name__
|
226 |
return str(obj)
|
227 |
-
|
228 |
-
|
229 |
def merge_dicts(source, destination):
|
230 |
"""
|
231 |
Recursively merges two dictionaries.
|
232 |
-
|
233 |
Args:
|
234 |
source (`dict`): The dictionary to merge into `destination`.
|
235 |
destination (`dict`): The dictionary to merge `source` into.
|
@@ -240,10 +188,7 @@ def merge_dicts(source, destination):
|
|
240 |
merge_dicts(value, node)
|
241 |
else:
|
242 |
destination[key] = value
|
243 |
-
|
244 |
return destination
|
245 |
-
|
246 |
-
|
247 |
def is_port_in_use(port: int = None) -> bool:
|
248 |
"""
|
249 |
Checks if a port is in use on `localhost`. Useful for checking if multiple `accelerate launch` commands have been
|
@@ -253,18 +198,13 @@ def is_port_in_use(port: int = None) -> bool:
|
|
253 |
port = 29500
|
254 |
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
|
255 |
return s.connect_ex(("localhost", port)) == 0
|
256 |
-
|
257 |
-
|
258 |
def convert_bytes(size):
|
259 |
"Converts `size` from bytes to the largest possible unit"
|
260 |
for x in ["bytes", "KB", "MB", "GB", "TB"]:
|
261 |
if size < 1024.0:
|
262 |
return f"{round(size, 2)} {x}"
|
263 |
size /= 1024.0
|
264 |
-
|
265 |
return f"{round(size, 2)} PB"
|
266 |
-
|
267 |
-
|
268 |
def check_os_kernel():
|
269 |
"""Warns if the kernel version is below the recommended minimum on Linux."""
|
270 |
# see issue #1929
|
@@ -272,7 +212,6 @@ def check_os_kernel():
|
|
272 |
system = info.system
|
273 |
if system != "Linux":
|
274 |
return
|
275 |
-
|
276 |
_, version, *_ = re.split(r"(\d+\.\d+\.\d+)", info.release)
|
277 |
min_version = "5.5.0"
|
278 |
if Version(version) < Version(min_version):
|
|
|
1 |
logger = get_logger(__name__)
|
|
|
|
|
2 |
if is_tpu_available(check_device=False):
|
3 |
import torch_xla.core.xla_model as xm
|
|
|
|
|
4 |
def is_compiled_module(module):
|
5 |
"""
|
6 |
Check whether the module was compiled with torch.compile()
|
|
|
8 |
if is_torch_version("<", "2.0.0") or not hasattr(torch, "_dynamo"):
|
9 |
return False
|
10 |
return isinstance(module, torch._dynamo.eval_frame.OptimizedModule)
|
|
|
|
|
11 |
def extract_model_from_parallel(model, keep_fp32_wrapper: bool = True):
|
12 |
"""
|
13 |
Extract a model from its distributed containers.
|
|
|
14 |
Args:
|
15 |
model (`torch.nn.Module`):
|
16 |
The model to extract.
|
17 |
keep_fp32_wrapper (`bool`, *optional*):
|
18 |
Whether to remove mixed precision hooks from the model.
|
|
|
19 |
Returns:
|
20 |
`torch.nn.Module`: The extracted model.
|
21 |
"""
|
22 |
options = (torch.nn.parallel.DistributedDataParallel, torch.nn.DataParallel)
|
|
|
23 |
is_compiled = is_compiled_module(model)
|
24 |
if is_compiled:
|
25 |
compiled_model = model
|
26 |
model = model._orig_mod
|
|
|
27 |
if is_deepspeed_available():
|
28 |
from deepspeed import DeepSpeedEngine
|
|
|
29 |
options += (DeepSpeedEngine,)
|
|
|
30 |
if is_torch_version(">=", FSDP_PYTORCH_VERSION) and is_torch_distributed_available():
|
31 |
from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel as FSDP
|
|
|
32 |
options += (FSDP,)
|
|
|
33 |
while isinstance(model, options):
|
34 |
model = model.module
|
|
|
35 |
if not keep_fp32_wrapper:
|
36 |
forward = getattr(model, "forward")
|
37 |
original_forward = model.__dict__.pop("_original_forward", None)
|
|
|
43 |
model.forward = MethodType(forward, model)
|
44 |
if getattr(model, "_converted_to_transformer_engine", False):
|
45 |
convert_model(model, to_transformer_engine=False)
|
|
|
46 |
if is_compiled:
|
47 |
compiled_model._orig_mod = model
|
48 |
model = compiled_model
|
|
|
49 |
return model
|
|
|
|
|
50 |
def wait_for_everyone():
|
51 |
"""
|
52 |
Introduces a blocking point in the script, making sure all processes have reached this point before continuing.
|
|
|
53 |
<Tip warning={true}>
|
|
|
54 |
Make sure all processes will reach this instruction otherwise one of your processes will hang forever.
|
|
|
55 |
</Tip>
|
56 |
"""
|
57 |
PartialState().wait_for_everyone()
|
|
|
|
|
58 |
def clean_state_dict_for_safetensors(state_dict: dict):
|
59 |
"""
|
60 |
Cleans the state dictionary from a model and removes tensor aliasing if present.
|
|
|
61 |
Args:
|
62 |
state_dict (`dict`):
|
63 |
The state dictionary from a model
|
|
|
67 |
for name, tensor in state_dict.items():
|
68 |
if not isinstance(tensor, str):
|
69 |
ptrs[id_tensor_storage(tensor)].append(name)
|
|
|
70 |
# These are all pointers of tensors with shared memory
|
71 |
shared_ptrs = {ptr: names for ptr, names in ptrs.items() if len(names) > 1}
|
72 |
warn_names = set()
|
|
|
86 |
)
|
87 |
state_dict = {k: v.contiguous() if isinstance(v, torch.Tensor) else v for k, v in state_dict.items()}
|
88 |
return state_dict
|
|
|
|
|
89 |
def save(obj, f, save_on_each_node: bool = False, safe_serialization: bool = False):
|
90 |
"""
|
91 |
Save the data to disk. Use in place of `torch.save()`.
|
|
|
92 |
Args:
|
93 |
obj:
|
94 |
The data to save
|
|
|
106 |
obj = clean_state_dict_for_safetensors(obj)
|
107 |
else:
|
108 |
save_func = torch.save
|
|
|
109 |
if PartialState().distributed_type == DistributedType.TPU:
|
110 |
xm.save(obj, f)
|
111 |
elif PartialState().is_main_process and not save_on_each_node:
|
112 |
save_func(obj, f)
|
113 |
elif PartialState().is_local_main_process and save_on_each_node:
|
114 |
save_func(obj, f)
|
|
|
|
|
115 |
@contextmanager
|
116 |
def clear_environment():
|
117 |
"""
|
118 |
A context manager that will cache origin `os.environ` and replace it with a empty dictionary in this context.
|
|
|
119 |
When this context exits, the cached `os.environ` will be back.
|
|
|
120 |
Example:
|
|
|
121 |
```python
|
122 |
>>> import os
|
123 |
>>> from accelerate.utils import clear_environment
|
|
|
124 |
>>> os.environ["FOO"] = "bar"
|
125 |
>>> with clear_environment():
|
126 |
... print(os.environ)
|
|
|
128 |
... print(os.environ["FOO"])
|
129 |
{}
|
130 |
new_bar
|
|
|
131 |
>>> print(os.environ["FOO"])
|
132 |
bar
|
133 |
```
|
134 |
"""
|
135 |
_old_os_environ = os.environ
|
136 |
os.environ = dict()
|
|
|
137 |
yield
|
|
|
138 |
os.environ = _old_os_environ
|
|
|
|
|
139 |
@contextmanager
|
140 |
def patch_environment(**kwargs):
|
141 |
"""
|
142 |
A context manager that will add each keyword argument passed to `os.environ` and remove them when exiting.
|
|
|
143 |
Will convert the values in `kwargs` to strings and upper-case all the keys.
|
|
|
144 |
Example:
|
|
|
145 |
```python
|
146 |
>>> import os
|
147 |
>>> from accelerate.utils import patch_environment
|
|
|
148 |
>>> with patch_environment(FOO="bar"):
|
149 |
... print(os.environ["FOO"]) # prints "bar"
|
150 |
>>> print(os.environ["FOO"]) # raises KeyError
|
|
|
156 |
if key in os.environ:
|
157 |
existing_vars[key] = os.environ[key]
|
158 |
os.environ[key] = str(value)
|
|
|
159 |
yield
|
|
|
160 |
for key in kwargs:
|
161 |
key = key.upper()
|
162 |
if key in existing_vars:
|
|
|
164 |
os.environ[key] = existing_vars[key]
|
165 |
else:
|
166 |
os.environ.pop(key, None)
|
|
|
|
|
167 |
def get_pretty_name(obj):
|
168 |
"""
|
169 |
Gets a pretty name from `obj`.
|
|
|
175 |
if hasattr(obj, "__name__"):
|
176 |
return obj.__name__
|
177 |
return str(obj)
|
|
|
|
|
178 |
def merge_dicts(source, destination):
|
179 |
"""
|
180 |
Recursively merges two dictionaries.
|
|
|
181 |
Args:
|
182 |
source (`dict`): The dictionary to merge into `destination`.
|
183 |
destination (`dict`): The dictionary to merge `source` into.
|
|
|
188 |
merge_dicts(value, node)
|
189 |
else:
|
190 |
destination[key] = value
|
|
|
191 |
return destination
|
|
|
|
|
192 |
def is_port_in_use(port: int = None) -> bool:
|
193 |
"""
|
194 |
Checks if a port is in use on `localhost`. Useful for checking if multiple `accelerate launch` commands have been
|
|
|
198 |
port = 29500
|
199 |
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
|
200 |
return s.connect_ex(("localhost", port)) == 0
|
|
|
|
|
201 |
def convert_bytes(size):
|
202 |
"Converts `size` from bytes to the largest possible unit"
|
203 |
for x in ["bytes", "KB", "MB", "GB", "TB"]:
|
204 |
if size < 1024.0:
|
205 |
return f"{round(size, 2)} {x}"
|
206 |
size /= 1024.0
|
|
|
207 |
return f"{round(size, 2)} PB"
|
|
|
|
|
208 |
def check_os_kernel():
|
209 |
"""Warns if the kernel version is below the recommended minimum on Linux."""
|
210 |
# see issue #1929
|
|
|
212 |
system = info.system
|
213 |
if system != "Linux":
|
214 |
return
|
|
|
215 |
_, version, *_ = re.split(r"(\d+\.\d+\.\d+)", info.release)
|
216 |
min_version = "5.5.0"
|
217 |
if Version(version) < Version(min_version):
|
src/utils/random.py
CHANGED
@@ -1,7 +1,6 @@
|
|
1 |
def set_seed(seed: int, device_specific: bool = False):
|
2 |
"""
|
3 |
Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`.
|
4 |
-
|
5 |
Args:
|
6 |
seed (`int`):
|
7 |
The seed to set.
|
@@ -22,8 +21,6 @@ def set_seed(seed: int, device_specific: bool = False):
|
|
22 |
# ^^ safe to call this function even if cuda is not available
|
23 |
if is_tpu_available():
|
24 |
xm.set_rng_state(seed)
|
25 |
-
|
26 |
-
|
27 |
def synchronize_rng_state(rng_type: Optional[RNGType] = None, generator: Optional[torch.Generator] = None):
|
28 |
# Get the proper rng state
|
29 |
if rng_type == RNGType.TORCH:
|
@@ -42,7 +39,6 @@ def synchronize_rng_state(rng_type: Optional[RNGType] = None, generator: Optiona
|
|
42 |
elif rng_type == RNGType.GENERATOR:
|
43 |
assert generator is not None, "Need a generator to synchronize its seed."
|
44 |
rng_state = generator.get_state()
|
45 |
-
|
46 |
# Broadcast the rng state from device 0 to other devices
|
47 |
state = AcceleratorState()
|
48 |
if state.distributed_type == DistributedType.TPU:
|
@@ -60,7 +56,6 @@ def synchronize_rng_state(rng_type: Optional[RNGType] = None, generator: Optiona
|
|
60 |
rng_state = rng_state.cpu()
|
61 |
elif state.distributed_type == DistributedType.MULTI_CPU:
|
62 |
torch.distributed.broadcast(rng_state, 0)
|
63 |
-
|
64 |
# Set the broadcast rng state
|
65 |
if rng_type == RNGType.TORCH:
|
66 |
torch.set_rng_state(rng_state)
|
@@ -74,8 +69,6 @@ def synchronize_rng_state(rng_type: Optional[RNGType] = None, generator: Optiona
|
|
74 |
xm.set_rng_state(rng_state.item())
|
75 |
elif rng_type == RNGType.GENERATOR:
|
76 |
generator.set_state(rng_state)
|
77 |
-
|
78 |
-
|
79 |
def synchronize_rng_states(rng_types: List[Union[str, RNGType]], generator: Optional[torch.Generator] = None):
|
80 |
for rng_type in rng_types:
|
81 |
synchronize_rng_state(RNGType(rng_type), generator=generator)
|
|
|
1 |
def set_seed(seed: int, device_specific: bool = False):
|
2 |
"""
|
3 |
Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`.
|
|
|
4 |
Args:
|
5 |
seed (`int`):
|
6 |
The seed to set.
|
|
|
21 |
# ^^ safe to call this function even if cuda is not available
|
22 |
if is_tpu_available():
|
23 |
xm.set_rng_state(seed)
|
|
|
|
|
24 |
def synchronize_rng_state(rng_type: Optional[RNGType] = None, generator: Optional[torch.Generator] = None):
|
25 |
# Get the proper rng state
|
26 |
if rng_type == RNGType.TORCH:
|
|
|
39 |
elif rng_type == RNGType.GENERATOR:
|
40 |
assert generator is not None, "Need a generator to synchronize its seed."
|
41 |
rng_state = generator.get_state()
|
|
|
42 |
# Broadcast the rng state from device 0 to other devices
|
43 |
state = AcceleratorState()
|
44 |
if state.distributed_type == DistributedType.TPU:
|
|
|
56 |
rng_state = rng_state.cpu()
|
57 |
elif state.distributed_type == DistributedType.MULTI_CPU:
|
58 |
torch.distributed.broadcast(rng_state, 0)
|
|
|
59 |
# Set the broadcast rng state
|
60 |
if rng_type == RNGType.TORCH:
|
61 |
torch.set_rng_state(rng_state)
|
|
|
69 |
xm.set_rng_state(rng_state.item())
|
70 |
elif rng_type == RNGType.GENERATOR:
|
71 |
generator.set_state(rng_state)
|
|
|
|
|
72 |
def synchronize_rng_states(rng_types: List[Union[str, RNGType]], generator: Optional[torch.Generator] = None):
|
73 |
for rng_type in rng_types:
|
74 |
synchronize_rng_state(RNGType(rng_type), generator=generator)
|
src/utils/rich.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
if is_rich_available():
|
2 |
from rich.traceback import install
|
3 |
-
|
4 |
install(show_locals=False)
|
5 |
-
|
6 |
else:
|
7 |
raise ModuleNotFoundError("To use the rich extension, install rich with `pip install rich`")
|
|
|
1 |
if is_rich_available():
|
2 |
from rich.traceback import install
|
|
|
3 |
install(show_locals=False)
|
|
|
4 |
else:
|
5 |
raise ModuleNotFoundError("To use the rich extension, install rich with `pip install rich`")
|
src/utils/torch_xla.py
CHANGED
@@ -1,23 +1,18 @@
|
|
1 |
def install_xla(upgrade: bool = False):
|
2 |
"""
|
3 |
Helper function to install appropriate xla wheels based on the `torch` version in Google Colaboratory.
|
4 |
-
|
5 |
Args:
|
6 |
upgrade (`bool`, *optional*, defaults to `False`):
|
7 |
Whether to upgrade `torch` and install the latest `torch_xla` wheels.
|
8 |
-
|
9 |
Example:
|
10 |
-
|
11 |
```python
|
12 |
>>> from accelerate.utils import install_xla
|
13 |
-
|
14 |
>>> install_xla(upgrade=True)
|
15 |
```
|
16 |
"""
|
17 |
in_colab = False
|
18 |
if "IPython" in sys.modules:
|
19 |
in_colab = "google.colab" in str(sys.modules["IPython"].get_ipython())
|
20 |
-
|
21 |
if in_colab:
|
22 |
if upgrade:
|
23 |
torch_install_cmd = ["pip", "install", "-U", "torch"]
|
|
|
1 |
def install_xla(upgrade: bool = False):
|
2 |
"""
|
3 |
Helper function to install appropriate xla wheels based on the `torch` version in Google Colaboratory.
|
|
|
4 |
Args:
|
5 |
upgrade (`bool`, *optional*, defaults to `False`):
|
6 |
Whether to upgrade `torch` and install the latest `torch_xla` wheels.
|
|
|
7 |
Example:
|
|
|
8 |
```python
|
9 |
>>> from accelerate.utils import install_xla
|
|
|
10 |
>>> install_xla(upgrade=True)
|
11 |
```
|
12 |
"""
|
13 |
in_colab = False
|
14 |
if "IPython" in sys.modules:
|
15 |
in_colab = "google.colab" in str(sys.modules["IPython"].get_ipython())
|
|
|
16 |
if in_colab:
|
17 |
if upgrade:
|
18 |
torch_install_cmd = ["pip", "install", "-U", "torch"]
|
src/utils/tqdm.py
CHANGED
@@ -1,7 +1,6 @@
|
|
1 |
def tqdm(main_process_only: bool = True, *args, **kwargs):
|
2 |
"""
|
3 |
Wrapper around `tqdm.tqdm` that optionally displays only on the main process.
|
4 |
-
|
5 |
Args:
|
6 |
main_process_only (`bool`, *optional*):
|
7 |
Whether to display the progress bar only on the main process
|
|
|
1 |
def tqdm(main_process_only: bool = True, *args, **kwargs):
|
2 |
"""
|
3 |
Wrapper around `tqdm.tqdm` that optionally displays only on the main process.
|
|
|
4 |
Args:
|
5 |
main_process_only (`bool`, *optional*):
|
6 |
Whether to display the progress bar only on the main process
|
src/utils/transformer_engine.py
CHANGED
@@ -16,13 +16,11 @@ def convert_model(model, to_transformer_engine=True, _convert_linear=True, _conv
|
|
16 |
module.weight.copy_(te_module.weight)
|
17 |
if has_bias:
|
18 |
module.bias.copy_(te_module.bias)
|
19 |
-
|
20 |
setattr(model, name, te_module)
|
21 |
elif isinstance(module, nn.LayerNorm) and to_transformer_engine and _convert_ln:
|
22 |
te_module = te.LayerNorm(module.normalized_shape[0], eps=module.eps, params_dtype=module.weight.dtype)
|
23 |
module.weight.copy_(te_module.weight)
|
24 |
module.bias.copy_(te_module.bias)
|
25 |
-
|
26 |
setattr(model, name, te_module)
|
27 |
elif isinstance(module, te.Linear) and not to_transformer_engine and _convert_linear:
|
28 |
has_bias = module.bias is not None
|
@@ -32,13 +30,11 @@ def convert_model(model, to_transformer_engine=True, _convert_linear=True, _conv
|
|
32 |
module.weight.copy_(new_module.weight)
|
33 |
if has_bias:
|
34 |
module.bias.copy_(new_module.bias)
|
35 |
-
|
36 |
setattr(model, name, new_module)
|
37 |
elif isinstance(module, te.LayerNorm) and not to_transformer_engine and _convert_ln:
|
38 |
new_module = nn.LayerNorm(module.normalized_shape[0], eps=module.eps, params_dtype=module.weight.dtype)
|
39 |
module.weight.copy_(new_module.weight)
|
40 |
module.bias.copy_(new_module.bias)
|
41 |
-
|
42 |
setattr(model, name, new_module)
|
43 |
else:
|
44 |
convert_model(
|
@@ -47,8 +43,6 @@ def convert_model(model, to_transformer_engine=True, _convert_linear=True, _conv
|
|
47 |
_convert_linear=_convert_linear,
|
48 |
_convert_ln=_convert_ln,
|
49 |
)
|
50 |
-
|
51 |
-
|
52 |
def has_transformer_engine_layers(model):
|
53 |
"""
|
54 |
Returns whether a given model has some `transformer_engine` layer or not.
|
|
|
16 |
module.weight.copy_(te_module.weight)
|
17 |
if has_bias:
|
18 |
module.bias.copy_(te_module.bias)
|
|
|
19 |
setattr(model, name, te_module)
|
20 |
elif isinstance(module, nn.LayerNorm) and to_transformer_engine and _convert_ln:
|
21 |
te_module = te.LayerNorm(module.normalized_shape[0], eps=module.eps, params_dtype=module.weight.dtype)
|
22 |
module.weight.copy_(te_module.weight)
|
23 |
module.bias.copy_(te_module.bias)
|
|
|
24 |
setattr(model, name, te_module)
|
25 |
elif isinstance(module, te.Linear) and not to_transformer_engine and _convert_linear:
|
26 |
has_bias = module.bias is not None
|
|
|
30 |
module.weight.copy_(new_module.weight)
|
31 |
if has_bias:
|
32 |
module.bias.copy_(new_module.bias)
|
|
|
33 |
setattr(model, name, new_module)
|
34 |
elif isinstance(module, te.LayerNorm) and not to_transformer_engine and _convert_ln:
|
35 |
new_module = nn.LayerNorm(module.normalized_shape[0], eps=module.eps, params_dtype=module.weight.dtype)
|
36 |
module.weight.copy_(new_module.weight)
|
37 |
module.bias.copy_(new_module.bias)
|
|
|
38 |
setattr(model, name, new_module)
|
39 |
else:
|
40 |
convert_model(
|
|
|
43 |
_convert_linear=_convert_linear,
|
44 |
_convert_ln=_convert_ln,
|
45 |
)
|
|
|
|
|
46 |
def has_transformer_engine_layers(model):
|
47 |
"""
|
48 |
Returns whether a given model has some `transformer_engine` layer or not.
|
src/utils/versions.py
CHANGED
@@ -1,10 +1,7 @@
|
|
1 |
torch_version = parse(importlib.metadata.version("torch"))
|
2 |
-
|
3 |
-
|
4 |
def compare_versions(library_or_version: Union[str, Version], operation: str, requirement_version: str):
|
5 |
"""
|
6 |
Compares a library version to some requirement using a given operation.
|
7 |
-
|
8 |
Args:
|
9 |
library_or_version (`str` or `packaging.version.Version`):
|
10 |
A library name or a version to check.
|
@@ -19,12 +16,9 @@ def compare_versions(library_or_version: Union[str, Version], operation: str, re
|
|
19 |
if isinstance(library_or_version, str):
|
20 |
library_or_version = parse(importlib.metadata.version(library_or_version))
|
21 |
return operation(library_or_version, parse(requirement_version))
|
22 |
-
|
23 |
-
|
24 |
def is_torch_version(operation: str, version: str):
|
25 |
"""
|
26 |
Compares the current PyTorch version to a given reference with an operation.
|
27 |
-
|
28 |
Args:
|
29 |
operation (`str`):
|
30 |
A string representation of an operator, such as `">"` or `"<="`
|
|
|
1 |
torch_version = parse(importlib.metadata.version("torch"))
|
|
|
|
|
2 |
def compare_versions(library_or_version: Union[str, Version], operation: str, requirement_version: str):
|
3 |
"""
|
4 |
Compares a library version to some requirement using a given operation.
|
|
|
5 |
Args:
|
6 |
library_or_version (`str` or `packaging.version.Version`):
|
7 |
A library name or a version to check.
|
|
|
16 |
if isinstance(library_or_version, str):
|
17 |
library_or_version = parse(importlib.metadata.version(library_or_version))
|
18 |
return operation(library_or_version, parse(requirement_version))
|
|
|
|
|
19 |
def is_torch_version(operation: str, version: str):
|
20 |
"""
|
21 |
Compares the current PyTorch version to a given reference with an operation.
|
|
|
22 |
Args:
|
23 |
operation (`str`):
|
24 |
A string representation of an operator, such as `">"` or `"<="`
|