Spacing
Browse files- README.md +9 -1
- src/accelerator.py +1 -4
- src/commands/config/config_utils.py +0 -1
- src/commands/launch.py +0 -1
- src/data_loader.py +0 -8
- src/hooks.py +0 -7
- src/local_sgd.py +0 -2
- src/logging.py +0 -1
- src/optimizer.py +0 -1
- src/scheduler.py +0 -1
- src/state.py +0 -4
- src/tracking.py +0 -9
- src/utils/bnb.py +0 -2
- src/utils/dataclasses.py +0 -20
- src/utils/deepspeed.py +0 -6
- src/utils/launch.py +0 -1
- src/utils/megatron_lm.py +0 -11
- src/utils/modeling.py +0 -3
- src/utils/offload.py +0 -2
- src/utils/operations.py +0 -12
README.md
CHANGED
@@ -37,4 +37,12 @@ Using `regex` in VSCODE, use the following replacement:
|
|
37 |
<!--Copyright(.*\n)+-->
|
38 |
```
|
39 |
|
40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
<!--Copyright(.*\n)+-->
|
38 |
```
|
39 |
|
40 |
+
In the source:
|
41 |
+
```regex
|
42 |
+
"""
|
43 |
+
|
44 |
+
|
45 |
+
```
|
46 |
+
Then remove all import statements (as we only care about the content).
|
47 |
+
|
48 |
+
**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
@@ -1,3 +1,4 @@
|
|
|
|
1 |
logger = get_logger(__name__)
|
2 |
|
3 |
|
@@ -95,7 +96,6 @@ class Accelerator:
|
|
95 |
- **sync_gradients** (`bool`) -- Whether the gradients are currently being synced across all processes.
|
96 |
- **use_distributed** (`bool`) -- Whether the current configuration is for distributed training.
|
97 |
"""
|
98 |
-
|
99 |
def __init__(
|
100 |
self,
|
101 |
device_placement: bool = True,
|
@@ -2010,7 +2010,6 @@ class Accelerator:
|
|
2010 |
9
|
2011 |
```
|
2012 |
"""
|
2013 |
-
|
2014 |
try:
|
2015 |
recursively_apply(lambda x: x, input_data, error_on_other_type=True)
|
2016 |
all_tensors = True
|
@@ -2373,7 +2372,6 @@ class Accelerator:
|
|
2373 |
>>> accelerator.save_model(model, save_directory)
|
2374 |
```
|
2375 |
"""
|
2376 |
-
|
2377 |
if os.path.isfile(save_directory):
|
2378 |
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
|
2379 |
return
|
@@ -2865,7 +2863,6 @@ class Accelerator:
|
|
2865 |
>>> state_dict = accelerator.get_state_dict(net)
|
2866 |
```
|
2867 |
"""
|
2868 |
-
|
2869 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
2870 |
if self.deepspeed_config["zero_optimization"]["stage"] == 3:
|
2871 |
if model.zero_gather_16bit_weights_on_model_save():
|
|
|
1 |
+
|
2 |
logger = get_logger(__name__)
|
3 |
|
4 |
|
|
|
96 |
- **sync_gradients** (`bool`) -- Whether the gradients are currently being synced across all processes.
|
97 |
- **use_distributed** (`bool`) -- Whether the current configuration is for distributed training.
|
98 |
"""
|
|
|
99 |
def __init__(
|
100 |
self,
|
101 |
device_placement: bool = True,
|
|
|
2010 |
9
|
2011 |
```
|
2012 |
"""
|
|
|
2013 |
try:
|
2014 |
recursively_apply(lambda x: x, input_data, error_on_other_type=True)
|
2015 |
all_tensors = True
|
|
|
2372 |
>>> accelerator.save_model(model, save_directory)
|
2373 |
```
|
2374 |
"""
|
|
|
2375 |
if os.path.isfile(save_directory):
|
2376 |
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
|
2377 |
return
|
|
|
2863 |
>>> state_dict = accelerator.get_state_dict(net)
|
2864 |
```
|
2865 |
"""
|
|
|
2866 |
if self.distributed_type == DistributedType.DEEPSPEED:
|
2867 |
if self.deepspeed_config["zero_optimization"]["stage"] == 3:
|
2868 |
if model.zero_gather_16bit_weights_on_model_save():
|
src/commands/config/config_utils.py
CHANGED
@@ -67,7 +67,6 @@ class SubcommandHelpFormatter(argparse.RawDescriptionHelpFormatter):
|
|
67 |
"""
|
68 |
A custom formatter that will remove the usage line from the help message for subcommands.
|
69 |
"""
|
70 |
-
|
71 |
def _format_usage(self, usage, actions, groups, prefix):
|
72 |
usage = super()._format_usage(usage, actions, groups, prefix)
|
73 |
usage = usage.replace("<command> [<args>] ", "")
|
|
|
67 |
"""
|
68 |
A custom formatter that will remove the usage line from the help message for subcommands.
|
69 |
"""
|
|
|
70 |
def _format_usage(self, usage, actions, groups, prefix):
|
71 |
usage = super()._format_usage(usage, actions, groups, prefix)
|
72 |
usage = usage.replace("<command> [<args>] ", "")
|
src/commands/launch.py
CHANGED
@@ -22,7 +22,6 @@ class _CustomHelpAction(argparse._HelpAction):
|
|
22 |
called. This is useful for the case where the user is using a specific platform and only wants to see the arguments
|
23 |
for that platform.
|
24 |
"""
|
25 |
-
|
26 |
def __call__(self, parser, namespace, values, option_string=None):
|
27 |
if "accelerate" in sys.argv[0] and "launch" in sys.argv[1:]:
|
28 |
args = sys.argv[2:]
|
|
|
22 |
called. This is useful for the case where the user is using a specific platform and only wants to see the arguments
|
23 |
for that platform.
|
24 |
"""
|
|
|
25 |
def __call__(self, parser, namespace, values, option_string=None):
|
26 |
if "accelerate" in sys.argv[0] and "launch" in sys.argv[1:]:
|
27 |
args = sys.argv[2:]
|
src/data_loader.py
CHANGED
@@ -36,7 +36,6 @@ class SeedableRandomSampler(RandomSampler):
|
|
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 |
"""
|
39 |
-
|
40 |
def __init__(self, *args, **kwargs):
|
41 |
super().__init__(*args, **kwargs)
|
42 |
self.epoch = 0
|
@@ -246,7 +245,6 @@ class IterableDatasetShard(IterableDataset):
|
|
246 |
- the shard on process 0 to yield `[0, 1, 4, 5]` and the sampler on process 1 to yield `[2, 3, 6, 7]` if
|
247 |
this argument is set to `True`.
|
248 |
"""
|
249 |
-
|
250 |
def __init__(
|
251 |
self,
|
252 |
dataset: IterableDataset,
|
@@ -326,7 +324,6 @@ class DataLoaderStateMixin:
|
|
326 |
batch size
|
327 |
|
328 |
"""
|
329 |
-
|
330 |
def __init_subclass__(cls, **kwargs):
|
331 |
cls.end_of_dataloader = False
|
332 |
cls.remainder = -1
|
@@ -381,7 +378,6 @@ class DataLoaderShard(DataLoader, DataLoaderStateMixin):
|
|
381 |
|
382 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
383 |
"""
|
384 |
-
|
385 |
def __init__(
|
386 |
self,
|
387 |
dataset,
|
@@ -481,7 +477,6 @@ if is_tpu_available(check_device=False):
|
|
481 |
|
482 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
483 |
"""
|
484 |
-
|
485 |
def __init__(self, dataloader: DataLoaderShard, device: torch.device):
|
486 |
super().__init__(dataloader, device)
|
487 |
self._rng_types = self._loader.rng_types
|
@@ -530,7 +525,6 @@ class DataLoaderDispatcher(DataLoader, DataLoaderStateMixin):
|
|
530 |
|
531 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
532 |
"""
|
533 |
-
|
534 |
def __init__(
|
535 |
self, dataset, split_batches: bool = False, skip_batches=0, _drop_last: bool = False, slice_fn=None, **kwargs
|
536 |
):
|
@@ -907,7 +901,6 @@ class SkipBatchSampler(BatchSampler):
|
|
907 |
"""
|
908 |
A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
|
909 |
"""
|
910 |
-
|
911 |
def __init__(self, batch_sampler, skip_batches=0):
|
912 |
self.batch_sampler = batch_sampler
|
913 |
self.skip_batches = skip_batches
|
@@ -937,7 +930,6 @@ class SkipDataLoader(DataLoader):
|
|
937 |
kwargs:
|
938 |
All other keyword arguments to pass to the regular `DataLoader` initialization.
|
939 |
"""
|
940 |
-
|
941 |
def __init__(self, dataset, skip_batches=0, **kwargs):
|
942 |
super().__init__(dataset, **kwargs)
|
943 |
self.skip_batches = skip_batches
|
|
|
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 |
"""
|
|
|
39 |
def __init__(self, *args, **kwargs):
|
40 |
super().__init__(*args, **kwargs)
|
41 |
self.epoch = 0
|
|
|
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
|
246 |
this argument is set to `True`.
|
247 |
"""
|
|
|
248 |
def __init__(
|
249 |
self,
|
250 |
dataset: IterableDataset,
|
|
|
324 |
batch size
|
325 |
|
326 |
"""
|
|
|
327 |
def __init_subclass__(cls, **kwargs):
|
328 |
cls.end_of_dataloader = False
|
329 |
cls.remainder = -1
|
|
|
378 |
|
379 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
380 |
"""
|
|
|
381 |
def __init__(
|
382 |
self,
|
383 |
dataset,
|
|
|
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
|
|
|
525 |
|
526 |
- **total_dataset_length** (`int`) -- Total length of the inner dataset across all processes.
|
527 |
"""
|
|
|
528 |
def __init__(
|
529 |
self, dataset, split_batches: bool = False, skip_batches=0, _drop_last: bool = False, slice_fn=None, **kwargs
|
530 |
):
|
|
|
901 |
"""
|
902 |
A `torch.utils.data.BatchSampler` that skips the first `n` batches of another `torch.utils.data.BatchSampler`.
|
903 |
"""
|
|
|
904 |
def __init__(self, batch_sampler, skip_batches=0):
|
905 |
self.batch_sampler = batch_sampler
|
906 |
self.skip_batches = skip_batches
|
|
|
930 |
kwargs:
|
931 |
All other keyword arguments to pass to the regular `DataLoader` initialization.
|
932 |
"""
|
|
|
933 |
def __init__(self, dataset, skip_batches=0, **kwargs):
|
934 |
super().__init__(dataset, **kwargs)
|
935 |
self.skip_batches = skip_batches
|
src/hooks.py
CHANGED
@@ -7,7 +7,6 @@ class ModelHook:
|
|
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 |
-
|
11 |
no_grad = False
|
12 |
|
13 |
def init_hook(self, module):
|
@@ -60,7 +59,6 @@ class SequentialHook(ModelHook):
|
|
60 |
"""
|
61 |
A hook that can contain several hooks and iterates through them at each event.
|
62 |
"""
|
63 |
-
|
64 |
def __init__(self, *hooks):
|
65 |
self.hooks = hooks
|
66 |
|
@@ -109,7 +107,6 @@ def add_hook_to_module(module: nn.Module, hook: ModelHook, append: bool = False)
|
|
109 |
`torch.nn.Module`: The same module, with the hook attached (the module is modified in place, so the result can
|
110 |
be discarded).
|
111 |
"""
|
112 |
-
|
113 |
if append and (getattr(module, "_hf_hook", None) is not None):
|
114 |
old_hook = module._hf_hook
|
115 |
remove_hook_from_module(module)
|
@@ -151,7 +148,6 @@ def remove_hook_from_module(module: nn.Module, recurse=False):
|
|
151 |
`torch.nn.Module`: The same module, with the hook detached (the module is modified in place, so the result can
|
152 |
be discarded).
|
153 |
"""
|
154 |
-
|
155 |
if hasattr(module, "_hf_hook"):
|
156 |
module._hf_hook.detach_hook(module)
|
157 |
delattr(module, "_hf_hook")
|
@@ -186,7 +182,6 @@ class AlignDevicesHook(ModelHook):
|
|
186 |
place_submodules (`bool`, *optional*, defaults to `False`):
|
187 |
Whether to place the submodules on `execution_device` during the `init_hook` event.
|
188 |
"""
|
189 |
-
|
190 |
def __init__(
|
191 |
self,
|
192 |
execution_device: Optional[Union[int, str, torch.device]] = None,
|
@@ -539,7 +534,6 @@ class CpuOffload(ModelHook):
|
|
539 |
passed, its offload method will be called just before the forward of the model to which this hook is
|
540 |
attached.
|
541 |
"""
|
542 |
-
|
543 |
def __init__(
|
544 |
self,
|
545 |
execution_device: Optional[Union[str, int, torch.device]] = None,
|
@@ -564,7 +558,6 @@ class UserCpuOffloadHook:
|
|
564 |
A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
|
565 |
or remove it entirely.
|
566 |
"""
|
567 |
-
|
568 |
def __init__(self, model, hook):
|
569 |
self.model = model
|
570 |
self.hook = hook
|
|
|
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):
|
|
|
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 |
|
|
|
107 |
`torch.nn.Module`: The same module, with the hook attached (the module is modified in place, so the result can
|
108 |
be discarded).
|
109 |
"""
|
|
|
110 |
if append and (getattr(module, "_hf_hook", None) is not None):
|
111 |
old_hook = module._hf_hook
|
112 |
remove_hook_from_module(module)
|
|
|
148 |
`torch.nn.Module`: The same module, with the hook detached (the module is modified in place, so the result can
|
149 |
be discarded).
|
150 |
"""
|
|
|
151 |
if hasattr(module, "_hf_hook"):
|
152 |
module._hf_hook.detach_hook(module)
|
153 |
delattr(module, "_hf_hook")
|
|
|
182 |
place_submodules (`bool`, *optional*, defaults to `False`):
|
183 |
Whether to place the submodules on `execution_device` during the `init_hook` event.
|
184 |
"""
|
|
|
185 |
def __init__(
|
186 |
self,
|
187 |
execution_device: Optional[Union[int, str, torch.device]] = None,
|
|
|
534 |
passed, its offload method will be called just before the forward of the model to which this hook is
|
535 |
attached.
|
536 |
"""
|
|
|
537 |
def __init__(
|
538 |
self,
|
539 |
execution_device: Optional[Union[str, int, torch.device]] = None,
|
|
|
558 |
A simple hook grouping a model and a `ModelHook`, which provides easy APIs for to call the init method of the hook
|
559 |
or remove it entirely.
|
560 |
"""
|
|
|
561 |
def __init__(self, model, hook):
|
562 |
self.model = model
|
563 |
self.hook = hook
|
src/local_sgd.py
CHANGED
@@ -19,7 +19,6 @@ class LocalSGD:
|
|
19 |
Learning Representations. No. CONF. 2019.](https://arxiv.org/abs/1805.09767)
|
20 |
|
21 |
"""
|
22 |
-
|
23 |
def __enter__(self):
|
24 |
if self.enabled:
|
25 |
self.model_sync_obj = self.model.no_sync()
|
@@ -75,7 +74,6 @@ class LocalSGD:
|
|
75 |
"""
|
76 |
Synchronize + Average model parameters across all GPUs
|
77 |
"""
|
78 |
-
|
79 |
self.accelerator.wait_for_everyone()
|
80 |
with self.accelerator.autocast():
|
81 |
for param in self.model.parameters():
|
|
|
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()
|
|
|
74 |
"""
|
75 |
Synchronize + Average model parameters across all GPUs
|
76 |
"""
|
|
|
77 |
self.accelerator.wait_for_everyone()
|
78 |
with self.accelerator.autocast():
|
79 |
for param in self.model.parameters():
|
src/logging.py
CHANGED
@@ -7,7 +7,6 @@ class MultiProcessAdapter(logging.LoggerAdapter):
|
|
7 |
|
8 |
Does not require an `Accelerator` object to be created first.
|
9 |
"""
|
10 |
-
|
11 |
@staticmethod
|
12 |
def _should_log(main_process_only):
|
13 |
"Check if log should be performed"
|
|
|
7 |
|
8 |
Does not require an `Accelerator` object to be created first.
|
9 |
"""
|
|
|
10 |
@staticmethod
|
11 |
def _should_log(main_process_only):
|
12 |
"Check if log should be performed"
|
src/optimizer.py
CHANGED
@@ -24,7 +24,6 @@ class AcceleratedOptimizer(torch.optim.Optimizer):
|
|
24 |
scaler (`torch.cuda.amp.grad_scaler.GradScaler`, *optional*):
|
25 |
The scaler to use in the step function if training with mixed precision.
|
26 |
"""
|
27 |
-
|
28 |
def __init__(self, optimizer, device_placement=True, scaler=None):
|
29 |
self.optimizer = optimizer
|
30 |
self.scaler = scaler
|
|
|
24 |
scaler (`torch.cuda.amp.grad_scaler.GradScaler`, *optional*):
|
25 |
The scaler to use in the step function if training with mixed precision.
|
26 |
"""
|
|
|
27 |
def __init__(self, optimizer, device_placement=True, scaler=None):
|
28 |
self.optimizer = optimizer
|
29 |
self.scaler = scaler
|
src/scheduler.py
CHANGED
@@ -23,7 +23,6 @@ class AcceleratedScheduler:
|
|
23 |
regardless of the number of processes) or create batches on each process (so batch size is the original
|
24 |
batch size multiplied by the number of processes).
|
25 |
"""
|
26 |
-
|
27 |
def __init__(self, scheduler, optimizers, step_with_optimizer: bool = True, split_batches: bool = False):
|
28 |
self.scheduler = scheduler
|
29 |
self.optimizers = optimizers if isinstance(optimizers, (list, tuple)) else [optimizers]
|
|
|
23 |
regardless of the number of processes) or create batches on each process (so batch size is the original
|
24 |
batch size multiplied by the number of processes).
|
25 |
"""
|
|
|
26 |
def __init__(self, scheduler, optimizers, step_with_optimizer: bool = True, split_batches: bool = False):
|
27 |
self.scheduler = scheduler
|
28 |
self.optimizers = optimizers if isinstance(optimizers, (list, tuple)) else [optimizers]
|
src/state.py
CHANGED
@@ -30,7 +30,6 @@ class ThreadLocalSharedDict(threading.local):
|
|
30 |
|
31 |
See https://github.com/pytorch/xla/blob/r2.0/docs/pjrt.md#multithreading-on-tpu-v2v3
|
32 |
"""
|
33 |
-
|
34 |
def __init__(self, thread_local: bool = False):
|
35 |
self._storage = {}
|
36 |
|
@@ -67,7 +66,6 @@ class PartialState:
|
|
67 |
- **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
|
68 |
- **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
|
69 |
"""
|
70 |
-
|
71 |
_shared_state = SharedDict()
|
72 |
|
73 |
def __init__(self, cpu: bool = False, **kwargs):
|
@@ -684,7 +682,6 @@ class AcceleratorState:
|
|
684 |
- **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
|
685 |
- **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
|
686 |
"""
|
687 |
-
|
688 |
_shared_state = SharedDict()
|
689 |
|
690 |
def __init__(
|
@@ -946,7 +943,6 @@ class GradientState:
|
|
946 |
- **sync_with_dataloader** (`bool`) -- Whether the gradients should be synced at the end of the dataloader
|
947 |
iteration and the number of total steps reset
|
948 |
"""
|
949 |
-
|
950 |
_shared_state = SharedDict()
|
951 |
|
952 |
def __init__(self, gradient_accumulation_plugin: Optional[GradientAccumulationPlugin] = None):
|
|
|
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 |
|
|
|
66 |
- **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
|
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):
|
|
|
682 |
- **is_local_main_process** (`bool`) -- Whether or not the current process is the main one on the local node.
|
683 |
- **debug** (`bool`) -- Whether or not the current script is being run in debug mode.
|
684 |
"""
|
|
|
685 |
_shared_state = SharedDict()
|
686 |
|
687 |
def __init__(
|
|
|
943 |
- **sync_with_dataloader** (`bool`) -- Whether the gradients should be synced at the end of the dataloader
|
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):
|
src/tracking.py
CHANGED
@@ -37,7 +37,6 @@ def on_main_process(function):
|
|
37 |
Checks at function execution rather than initialization time, not triggering the initialization of the
|
38 |
`PartialState`.
|
39 |
"""
|
40 |
-
|
41 |
@wraps(function)
|
42 |
def execute_on_main_process(self, *args, **kwargs):
|
43 |
if getattr(self, "main_process_only", False):
|
@@ -69,7 +68,6 @@ class GeneralTracker:
|
|
69 |
Implementations can also include a `main_process_only` (`bool`) attribute to toggle if relevent logging, init, and
|
70 |
other functions should occur on the main process or across all processes (by default will use `True`)
|
71 |
"""
|
72 |
-
|
73 |
main_process_only = True
|
74 |
|
75 |
def __init__(self, _blank=False):
|
@@ -139,7 +137,6 @@ class TensorBoardTracker(GeneralTracker):
|
|
139 |
kwargs:
|
140 |
Additional key word arguments passed along to the `tensorboard.SummaryWriter.__init__` method.
|
141 |
"""
|
142 |
-
|
143 |
name = "tensorboard"
|
144 |
requires_logging_directory = True
|
145 |
|
@@ -248,7 +245,6 @@ class WandBTracker(GeneralTracker):
|
|
248 |
kwargs:
|
249 |
Additional key word arguments passed along to the `wandb.init` method.
|
250 |
"""
|
251 |
-
|
252 |
name = "wandb"
|
253 |
requires_logging_directory = False
|
254 |
main_process_only = False
|
@@ -373,7 +369,6 @@ class CometMLTracker(GeneralTracker):
|
|
373 |
kwargs:
|
374 |
Additional key word arguments passed along to the `Experiment.__init__` method.
|
375 |
"""
|
376 |
-
|
377 |
name = "comet_ml"
|
378 |
requires_logging_directory = False
|
379 |
|
@@ -452,7 +447,6 @@ class AimTracker(GeneralTracker):
|
|
452 |
kwargs:
|
453 |
Additional key word arguments passed along to the `Run.__init__` method.
|
454 |
"""
|
455 |
-
|
456 |
name = "aim"
|
457 |
requires_logging_directory = True
|
458 |
|
@@ -568,7 +562,6 @@ class MLflowTracker(GeneralTracker):
|
|
568 |
description is set on the resumed run. If a new run is being created, the description is set on the new
|
569 |
run.
|
570 |
"""
|
571 |
-
|
572 |
name = "mlflow"
|
573 |
requires_logging_directory = False
|
574 |
|
@@ -697,7 +690,6 @@ class ClearMLTracker(GeneralTracker):
|
|
697 |
kwargs:
|
698 |
Kwargs passed along to the `Task.__init__` method.
|
699 |
"""
|
700 |
-
|
701 |
name = "clearml"
|
702 |
requires_logging_directory = False
|
703 |
|
@@ -857,7 +849,6 @@ class DVCLiveTracker(GeneralTracker):
|
|
857 |
accelerator.init_trackers(project_name="my_project", init_kwargs={"dvclive": {"dir": "my_directory"}})
|
858 |
```
|
859 |
"""
|
860 |
-
|
861 |
name = "dvclive"
|
862 |
requires_logging_directory = False
|
863 |
|
|
|
37 |
Checks at function execution rather than initialization time, not triggering the initialization of the
|
38 |
`PartialState`.
|
39 |
"""
|
|
|
40 |
@wraps(function)
|
41 |
def execute_on_main_process(self, *args, **kwargs):
|
42 |
if getattr(self, "main_process_only", False):
|
|
|
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):
|
|
|
137 |
kwargs:
|
138 |
Additional key word arguments passed along to the `tensorboard.SummaryWriter.__init__` method.
|
139 |
"""
|
|
|
140 |
name = "tensorboard"
|
141 |
requires_logging_directory = True
|
142 |
|
|
|
245 |
kwargs:
|
246 |
Additional key word arguments passed along to the `wandb.init` method.
|
247 |
"""
|
|
|
248 |
name = "wandb"
|
249 |
requires_logging_directory = False
|
250 |
main_process_only = False
|
|
|
369 |
kwargs:
|
370 |
Additional key word arguments passed along to the `Experiment.__init__` method.
|
371 |
"""
|
|
|
372 |
name = "comet_ml"
|
373 |
requires_logging_directory = False
|
374 |
|
|
|
447 |
kwargs:
|
448 |
Additional key word arguments passed along to the `Run.__init__` method.
|
449 |
"""
|
|
|
450 |
name = "aim"
|
451 |
requires_logging_directory = True
|
452 |
|
|
|
562 |
description is set on the resumed run. If a new run is being created, the description is set on the new
|
563 |
run.
|
564 |
"""
|
|
|
565 |
name = "mlflow"
|
566 |
requires_logging_directory = False
|
567 |
|
|
|
690 |
kwargs:
|
691 |
Kwargs passed along to the `Task.__init__` method.
|
692 |
"""
|
|
|
693 |
name = "clearml"
|
694 |
requires_logging_directory = False
|
695 |
|
|
|
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 |
|
src/utils/bnb.py
CHANGED
@@ -44,7 +44,6 @@ def load_and_quantize_model(
|
|
44 |
Returns:
|
45 |
`torch.nn.Module`: The quantized model
|
46 |
"""
|
47 |
-
|
48 |
load_in_4bit = bnb_quantization_config.load_in_4bit
|
49 |
load_in_8bit = bnb_quantization_config.load_in_8bit
|
50 |
|
@@ -246,7 +245,6 @@ def replace_with_bnb_layers(model, bnb_quantization_config, modules_to_not_conve
|
|
246 |
An array to track the current key of the recursion. This is used to check whether the current key (part of
|
247 |
it) is not in the list of modules to not convert.
|
248 |
"""
|
249 |
-
|
250 |
if modules_to_not_convert is None:
|
251 |
modules_to_not_convert = []
|
252 |
|
|
|
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 |
|
|
|
245 |
An array to track the current key of the recursion. This is used to check whether the current key (part of
|
246 |
it) is not in the list of modules to not convert.
|
247 |
"""
|
|
|
248 |
if modules_to_not_convert is None:
|
249 |
modules_to_not_convert = []
|
250 |
|
src/utils/dataclasses.py
CHANGED
@@ -5,7 +5,6 @@ class KwargsHandler:
|
|
5 |
"""
|
6 |
Internal mixin that implements a `to_kwargs()` method for a dataclass.
|
7 |
"""
|
8 |
-
|
9 |
def to_dict(self):
|
10 |
return copy.deepcopy(self.__dict__)
|
11 |
|
@@ -39,7 +38,6 @@ class AutocastKwargs(KwargsHandler):
|
|
39 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
40 |
```
|
41 |
"""
|
42 |
-
|
43 |
enabled: bool = True
|
44 |
cache_enabled: bool = None
|
45 |
|
@@ -70,7 +68,6 @@ class DistributedDataParallelKwargs(KwargsHandler):
|
|
70 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
71 |
```
|
72 |
"""
|
73 |
-
|
74 |
dim: int = 0
|
75 |
broadcast_buffers: bool = True
|
76 |
bucket_cap_mb: int = 25
|
@@ -103,7 +100,6 @@ class GradScalerKwargs(KwargsHandler):
|
|
103 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
104 |
```
|
105 |
"""
|
106 |
-
|
107 |
init_scale: float = 65536.0
|
108 |
growth_factor: float = 2.0
|
109 |
backoff_factor: float = 0.5
|
@@ -128,7 +124,6 @@ class InitProcessGroupKwargs(KwargsHandler):
|
|
128 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
129 |
```
|
130 |
"""
|
131 |
-
|
132 |
backend: Optional[str] = "nccl"
|
133 |
init_method: Optional[str] = None
|
134 |
timeout: timedelta = timedelta(seconds=1800)
|
@@ -197,7 +192,6 @@ class FP8RecipeKwargs(KwargsHandler):
|
|
197 |
are stored in FP8. If `fp8` is selected and deepspeed is enabled, will be used by default. (Not
|
198 |
available currently).
|
199 |
"""
|
200 |
-
|
201 |
backend: Backend = "msamp"
|
202 |
opt_level: OptLevel = "O2"
|
203 |
margin: int = 0
|
@@ -260,7 +254,6 @@ class DistributedType(str, enum.Enum):
|
|
260 |
- **DEEPSPEED** -- Using DeepSpeed.
|
261 |
- **TPU** -- Distributed on TPUs.
|
262 |
"""
|
263 |
-
|
264 |
# Subclassing str as well as Enum allows the `DistributedType` to be JSON-serializable out of the box.
|
265 |
NO = "NO"
|
266 |
MULTI_CPU = "MULTI_CPU"
|
@@ -283,7 +276,6 @@ class SageMakerDistributedType(str, enum.Enum):
|
|
283 |
- **DATA_PARALLEL** -- using sagemaker distributed data parallelism.
|
284 |
- **MODEL_PARALLEL** -- using sagemaker distributed model parallelism.
|
285 |
"""
|
286 |
-
|
287 |
# Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
|
288 |
NO = "NO"
|
289 |
DATA_PARALLEL = "DATA_PARALLEL"
|
@@ -299,7 +291,6 @@ class ComputeEnvironment(str, enum.Enum):
|
|
299 |
- **LOCAL_MACHINE** -- private/custom cluster hardware.
|
300 |
- **AMAZON_SAGEMAKER** -- Amazon SageMaker as compute environment.
|
301 |
"""
|
302 |
-
|
303 |
# Subclassing str as well as Enum allows the `ComputeEnvironment` to be JSON-serializable out of the box.
|
304 |
LOCAL_MACHINE = "LOCAL_MACHINE"
|
305 |
AMAZON_SAGEMAKER = "AMAZON_SAGEMAKER"
|
@@ -336,7 +327,6 @@ class DynamoBackend(str, BaseEnum):
|
|
336 |
- **TVM** -- Uses Apach TVM for inference optimizations. [Read more](https://tvm.apache.org/)
|
337 |
|
338 |
"""
|
339 |
-
|
340 |
# Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
|
341 |
NO = "NO"
|
342 |
EAGER = "EAGER"
|
@@ -364,7 +354,6 @@ class LoggerType(BaseEnum):
|
|
364 |
- **COMETML** -- comet_ml as an experiment tracker
|
365 |
- **DVCLIVE** -- dvclive as an experiment tracker
|
366 |
"""
|
367 |
-
|
368 |
ALL = "all"
|
369 |
AIM = "aim"
|
370 |
TENSORBOARD = "tensorboard"
|
@@ -384,7 +373,6 @@ class PrecisionType(BaseEnum):
|
|
384 |
- **FP16** -- using half precision
|
385 |
- **BF16** -- using brain floating point precision
|
386 |
"""
|
387 |
-
|
388 |
NO = "no"
|
389 |
FP8 = "fp8"
|
390 |
FP16 = "fp16"
|
@@ -404,7 +392,6 @@ class CustomDtype(enum.Enum):
|
|
404 |
r"""
|
405 |
An enum that contains multiple custom dtypes that can be used for `infer_auto_device_map`.
|
406 |
"""
|
407 |
-
|
408 |
FP8 = "fp8"
|
409 |
INT4 = "int4"
|
410 |
|
@@ -423,7 +410,6 @@ class ProjectConfiguration:
|
|
423 |
"""
|
424 |
Configuration for the Accelerator object based on inner-project needs.
|
425 |
"""
|
426 |
-
|
427 |
project_dir: str = field(default=None, metadata={"help": "A path to a directory for storing data."})
|
428 |
logging_dir: str = field(
|
429 |
default=None,
|
@@ -471,7 +457,6 @@ class GradientAccumulationPlugin(KwargsHandler):
|
|
471 |
"""
|
472 |
A plugin to configure gradient accumulation behavior.
|
473 |
"""
|
474 |
-
|
475 |
num_steps: int = field(default=None, metadata={"help": "The number of steps to accumulate gradients for."})
|
476 |
adjust_scheduler: bool = field(
|
477 |
default=True,
|
@@ -492,7 +477,6 @@ class TorchDynamoPlugin(KwargsHandler):
|
|
492 |
"""
|
493 |
This plugin is used to compile a model with PyTorch 2.0
|
494 |
"""
|
495 |
-
|
496 |
backend: DynamoBackend = field(
|
497 |
default=None,
|
498 |
metadata={"help": f"Possible options are {[b.value.lower() for b in DynamoBackend]}"},
|
@@ -528,7 +512,6 @@ class DeepSpeedPlugin:
|
|
528 |
"""
|
529 |
This plugin is used to integrate DeepSpeed.
|
530 |
"""
|
531 |
-
|
532 |
hf_ds_config: Any = field(
|
533 |
default=None,
|
534 |
metadata={
|
@@ -828,7 +811,6 @@ class FullyShardedDataParallelPlugin:
|
|
828 |
"""
|
829 |
This plugin is used to enable fully sharded data parallelism.
|
830 |
"""
|
831 |
-
|
832 |
sharding_strategy: "typing.Any" = field(
|
833 |
default=None,
|
834 |
metadata={
|
@@ -1062,7 +1044,6 @@ class MegatronLMPlugin:
|
|
1062 |
Plugin for Megatron-LM to enable tensor, pipeline, sequence and data parallelism. Also to enable selective
|
1063 |
activation recomputation and optimized fused kernels.
|
1064 |
"""
|
1065 |
-
|
1066 |
tp_degree: int = field(default=None, metadata={"help": "tensor parallelism degree."})
|
1067 |
pp_degree: int = field(default=None, metadata={"help": "pipeline parallelism degree."})
|
1068 |
num_micro_batches: int = field(default=None, metadata={"help": "number of micro-batches."})
|
@@ -1436,7 +1417,6 @@ class BnbQuantizationConfig:
|
|
1436 |
"""
|
1437 |
A plugin to enable BitsAndBytes 4bit and 8bit quantization
|
1438 |
"""
|
1439 |
-
|
1440 |
load_in_8bit: bool = field(default=False, metadata={"help": "enable 8bit quantization."})
|
1441 |
|
1442 |
llm_int8_threshold: float = field(
|
|
|
5 |
"""
|
6 |
Internal mixin that implements a `to_kwargs()` method for a dataclass.
|
7 |
"""
|
|
|
8 |
def to_dict(self):
|
9 |
return copy.deepcopy(self.__dict__)
|
10 |
|
|
|
38 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
39 |
```
|
40 |
"""
|
|
|
41 |
enabled: bool = True
|
42 |
cache_enabled: bool = None
|
43 |
|
|
|
68 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
69 |
```
|
70 |
"""
|
|
|
71 |
dim: int = 0
|
72 |
broadcast_buffers: bool = True
|
73 |
bucket_cap_mb: int = 25
|
|
|
100 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
101 |
```
|
102 |
"""
|
|
|
103 |
init_scale: float = 65536.0
|
104 |
growth_factor: float = 2.0
|
105 |
backoff_factor: float = 0.5
|
|
|
124 |
accelerator = Accelerator(kwargs_handlers=[kwargs])
|
125 |
```
|
126 |
"""
|
|
|
127 |
backend: Optional[str] = "nccl"
|
128 |
init_method: Optional[str] = None
|
129 |
timeout: timedelta = timedelta(seconds=1800)
|
|
|
192 |
are stored in FP8. If `fp8` is selected and deepspeed is enabled, will be used by default. (Not
|
193 |
available currently).
|
194 |
"""
|
|
|
195 |
backend: Backend = "msamp"
|
196 |
opt_level: OptLevel = "O2"
|
197 |
margin: int = 0
|
|
|
254 |
- **DEEPSPEED** -- Using DeepSpeed.
|
255 |
- **TPU** -- Distributed on TPUs.
|
256 |
"""
|
|
|
257 |
# Subclassing str as well as Enum allows the `DistributedType` to be JSON-serializable out of the box.
|
258 |
NO = "NO"
|
259 |
MULTI_CPU = "MULTI_CPU"
|
|
|
276 |
- **DATA_PARALLEL** -- using sagemaker distributed data parallelism.
|
277 |
- **MODEL_PARALLEL** -- using sagemaker distributed model parallelism.
|
278 |
"""
|
|
|
279 |
# Subclassing str as well as Enum allows the `SageMakerDistributedType` to be JSON-serializable out of the box.
|
280 |
NO = "NO"
|
281 |
DATA_PARALLEL = "DATA_PARALLEL"
|
|
|
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"
|
|
|
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"
|
332 |
EAGER = "EAGER"
|
|
|
354 |
- **COMETML** -- comet_ml as an experiment tracker
|
355 |
- **DVCLIVE** -- dvclive as an experiment tracker
|
356 |
"""
|
|
|
357 |
ALL = "all"
|
358 |
AIM = "aim"
|
359 |
TENSORBOARD = "tensorboard"
|
|
|
373 |
- **FP16** -- using half precision
|
374 |
- **BF16** -- using brain floating point precision
|
375 |
"""
|
|
|
376 |
NO = "no"
|
377 |
FP8 = "fp8"
|
378 |
FP16 = "fp16"
|
|
|
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 |
|
|
|
410 |
"""
|
411 |
Configuration for the Accelerator object based on inner-project needs.
|
412 |
"""
|
|
|
413 |
project_dir: str = field(default=None, metadata={"help": "A path to a directory for storing data."})
|
414 |
logging_dir: str = field(
|
415 |
default=None,
|
|
|
457 |
"""
|
458 |
A plugin to configure gradient accumulation behavior.
|
459 |
"""
|
|
|
460 |
num_steps: int = field(default=None, metadata={"help": "The number of steps to accumulate gradients for."})
|
461 |
adjust_scheduler: bool = field(
|
462 |
default=True,
|
|
|
477 |
"""
|
478 |
This plugin is used to compile a model with PyTorch 2.0
|
479 |
"""
|
|
|
480 |
backend: DynamoBackend = field(
|
481 |
default=None,
|
482 |
metadata={"help": f"Possible options are {[b.value.lower() for b in DynamoBackend]}"},
|
|
|
512 |
"""
|
513 |
This plugin is used to integrate DeepSpeed.
|
514 |
"""
|
|
|
515 |
hf_ds_config: Any = field(
|
516 |
default=None,
|
517 |
metadata={
|
|
|
811 |
"""
|
812 |
This plugin is used to enable fully sharded data parallelism.
|
813 |
"""
|
|
|
814 |
sharding_strategy: "typing.Any" = field(
|
815 |
default=None,
|
816 |
metadata={
|
|
|
1044 |
Plugin for Megatron-LM to enable tensor, pipeline, sequence and data parallelism. Also to enable selective
|
1045 |
activation recomputation and optimized fused kernels.
|
1046 |
"""
|
|
|
1047 |
tp_degree: int = field(default=None, metadata={"help": "tensor parallelism degree."})
|
1048 |
pp_degree: int = field(default=None, metadata={"help": "pipeline parallelism degree."})
|
1049 |
num_micro_batches: int = field(default=None, metadata={"help": "number of micro-batches."})
|
|
|
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(
|
src/utils/deepspeed.py
CHANGED
@@ -14,7 +14,6 @@ class HfDeepSpeedConfig:
|
|
14 |
config_file_or_dict (`Union[str, Dict]`): path to DeepSpeed config file or dict.
|
15 |
|
16 |
"""
|
17 |
-
|
18 |
def __init__(self, config_file_or_dict):
|
19 |
if isinstance(config_file_or_dict, dict):
|
20 |
# Don't modify user's data should they want to reuse it (e.g. in tests), because once we
|
@@ -134,7 +133,6 @@ class DeepSpeedEngineWrapper:
|
|
134 |
Args:
|
135 |
engine (deepspeed.runtime.engine.DeepSpeedEngine): deepspeed engine to wrap
|
136 |
"""
|
137 |
-
|
138 |
def __init__(self, engine):
|
139 |
self.engine = engine
|
140 |
|
@@ -163,7 +161,6 @@ class DeepSpeedOptimizerWrapper(AcceleratedOptimizer):
|
|
163 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
164 |
The optimizer to wrap.
|
165 |
"""
|
166 |
-
|
167 |
def __init__(self, optimizer):
|
168 |
super().__init__(optimizer, device_placement=False, scaler=None)
|
169 |
self.__has_overflow__ = hasattr(self.optimizer, "overflow")
|
@@ -191,7 +188,6 @@ class DeepSpeedSchedulerWrapper(AcceleratedScheduler):
|
|
191 |
The scheduler to wrap.
|
192 |
optimizers (one or a list of `torch.optim.Optimizer`):
|
193 |
"""
|
194 |
-
|
195 |
def __init__(self, scheduler, optimizers):
|
196 |
super().__init__(scheduler, optimizers)
|
197 |
|
@@ -214,7 +210,6 @@ class DummyOptim:
|
|
214 |
**kwargs:
|
215 |
Other arguments.
|
216 |
"""
|
217 |
-
|
218 |
def __init__(self, params, lr=0.001, weight_decay=0, **kwargs):
|
219 |
self.params = params
|
220 |
self.lr = lr
|
@@ -239,7 +234,6 @@ class DummyScheduler:
|
|
239 |
**kwargs:
|
240 |
Other arguments.
|
241 |
"""
|
242 |
-
|
243 |
def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, lr_scheduler_callable=None, **kwargs):
|
244 |
self.optimizer = optimizer
|
245 |
self.total_num_steps = total_num_steps
|
|
|
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):
|
19 |
# Don't modify user's data should they want to reuse it (e.g. in tests), because once we
|
|
|
133 |
Args:
|
134 |
engine (deepspeed.runtime.engine.DeepSpeedEngine): deepspeed engine to wrap
|
135 |
"""
|
|
|
136 |
def __init__(self, engine):
|
137 |
self.engine = engine
|
138 |
|
|
|
161 |
optimizer (`torch.optim.optimizer.Optimizer`):
|
162 |
The optimizer to wrap.
|
163 |
"""
|
|
|
164 |
def __init__(self, optimizer):
|
165 |
super().__init__(optimizer, device_placement=False, scaler=None)
|
166 |
self.__has_overflow__ = hasattr(self.optimizer, "overflow")
|
|
|
188 |
The scheduler to wrap.
|
189 |
optimizers (one or a list of `torch.optim.Optimizer`):
|
190 |
"""
|
|
|
191 |
def __init__(self, scheduler, optimizers):
|
192 |
super().__init__(scheduler, optimizers)
|
193 |
|
|
|
210 |
**kwargs:
|
211 |
Other arguments.
|
212 |
"""
|
|
|
213 |
def __init__(self, params, lr=0.001, weight_decay=0, **kwargs):
|
214 |
self.params = params
|
215 |
self.lr = lr
|
|
|
234 |
**kwargs:
|
235 |
Other arguments.
|
236 |
"""
|
|
|
237 |
def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, lr_scheduler_callable=None, **kwargs):
|
238 |
self.optimizer = optimizer
|
239 |
self.total_num_steps = total_num_steps
|
src/utils/launch.py
CHANGED
@@ -502,7 +502,6 @@ class PrepareForLaunch:
|
|
502 |
debug (`bool`, *optional*, defaults to `False`):
|
503 |
Whether or not this is a debug launch.
|
504 |
"""
|
505 |
-
|
506 |
def __init__(self, launcher, distributed_type="NO", debug=False):
|
507 |
self.launcher = launcher
|
508 |
self.distributed_type = DistributedType(distributed_type)
|
|
|
502 |
debug (`bool`, *optional*, defaults to `False`):
|
503 |
Whether or not this is a debug launch.
|
504 |
"""
|
|
|
505 |
def __init__(self, launcher, distributed_type="NO", debug=False):
|
506 |
self.launcher = launcher
|
507 |
self.distributed_type = DistributedType(distributed_type)
|
src/utils/megatron_lm.py
CHANGED
@@ -68,7 +68,6 @@ class MegatronLMDummyDataLoader:
|
|
68 |
Args:
|
69 |
**dataset_kwargs: Megatron data arguments.
|
70 |
"""
|
71 |
-
|
72 |
def __init__(self, **dataset_kwargs):
|
73 |
parser = argparse.ArgumentParser()
|
74 |
parser = _add_data_args(parser)
|
@@ -346,7 +345,6 @@ class MegatronLMDummyScheduler:
|
|
346 |
**kwargs:
|
347 |
Other arguments.
|
348 |
"""
|
349 |
-
|
350 |
def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, **kwargs):
|
351 |
self.optimizer = optimizer
|
352 |
self.total_num_steps = total_num_steps
|
@@ -392,7 +390,6 @@ class BertTrainStep(AbstractTrainStep):
|
|
392 |
Args:
|
393 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
394 |
"""
|
395 |
-
|
396 |
def __init__(self, args):
|
397 |
super().__init__("BertTrainStep")
|
398 |
self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
|
@@ -521,7 +518,6 @@ class GPTTrainStep(AbstractTrainStep):
|
|
521 |
Args:
|
522 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
523 |
"""
|
524 |
-
|
525 |
def __init__(self, args):
|
526 |
super().__init__("GPTTrainStep")
|
527 |
self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
|
@@ -627,7 +623,6 @@ class T5TrainStep(AbstractTrainStep):
|
|
627 |
Args:
|
628 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
629 |
"""
|
630 |
-
|
631 |
def __init__(self, args):
|
632 |
super().__init__("T5TrainStep")
|
633 |
self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
|
@@ -846,7 +841,6 @@ class MegatronEngine(torch.nn.Module):
|
|
846 |
optimizer: Megatron-LM optimizer
|
847 |
lr_scheduler: Megatron-LM lr scheduler
|
848 |
"""
|
849 |
-
|
850 |
def __init__(self, accelerator, model, optimizer, scheduler):
|
851 |
super(MegatronEngine, self).__init__()
|
852 |
self.module = model
|
@@ -892,7 +886,6 @@ class MegatronEngine(torch.nn.Module):
|
|
892 |
Args:
|
893 |
batch_data (:obj:`dict`): The batch data to train on.
|
894 |
"""
|
895 |
-
|
896 |
args = get_args()
|
897 |
timers = get_timers()
|
898 |
|
@@ -993,7 +986,6 @@ class MegatronEngine(torch.nn.Module):
|
|
993 |
Args:
|
994 |
batch_data (:obj:`dict`): The batch data to evaluate on.
|
995 |
"""
|
996 |
-
|
997 |
args = get_args()
|
998 |
data_chunks = []
|
999 |
if args.num_micro_batches > 1:
|
@@ -1176,7 +1168,6 @@ class MegatronEngine(torch.nn.Module):
|
|
1176 |
length_penalty (float, optional): length penalty for beam search. Defaults to None.
|
1177 |
kwargs: additional key-value arguments
|
1178 |
"""
|
1179 |
-
|
1180 |
# checking if required arguments are passed
|
1181 |
args = get_args()
|
1182 |
if args.model_type_name != "gpt":
|
@@ -1332,7 +1323,6 @@ def avg_losses_across_data_parallel_group(losses):
|
|
1332 |
Args:
|
1333 |
losses (List[Tensor]): List of losses to average across data parallel group.
|
1334 |
"""
|
1335 |
-
|
1336 |
return average_losses_across_data_parallel_group(losses)
|
1337 |
|
1338 |
|
@@ -1345,7 +1335,6 @@ def gather_across_data_parallel_groups(tensor):
|
|
1345 |
The data to gather across data parallel ranks.
|
1346 |
|
1347 |
"""
|
1348 |
-
|
1349 |
def _gpu_gather_one(tensor):
|
1350 |
if tensor.ndim == 0:
|
1351 |
tensor = tensor.clone()[None]
|
|
|
68 |
Args:
|
69 |
**dataset_kwargs: Megatron data arguments.
|
70 |
"""
|
|
|
71 |
def __init__(self, **dataset_kwargs):
|
72 |
parser = argparse.ArgumentParser()
|
73 |
parser = _add_data_args(parser)
|
|
|
345 |
**kwargs:
|
346 |
Other arguments.
|
347 |
"""
|
|
|
348 |
def __init__(self, optimizer, total_num_steps=None, warmup_num_steps=0, **kwargs):
|
349 |
self.optimizer = optimizer
|
350 |
self.total_num_steps = total_num_steps
|
|
|
390 |
Args:
|
391 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
392 |
"""
|
|
|
393 |
def __init__(self, args):
|
394 |
super().__init__("BertTrainStep")
|
395 |
self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
|
|
|
518 |
Args:
|
519 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
520 |
"""
|
|
|
521 |
def __init__(self, args):
|
522 |
super().__init__("GPTTrainStep")
|
523 |
self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
|
|
|
623 |
Args:
|
624 |
args (`argparse.Namespace`): Megatron-LM arguments.
|
625 |
"""
|
|
|
626 |
def __init__(self, args):
|
627 |
super().__init__("T5TrainStep")
|
628 |
self.get_batch = self.get_batch_func(args.megatron_dataset_flag)
|
|
|
841 |
optimizer: Megatron-LM optimizer
|
842 |
lr_scheduler: Megatron-LM lr scheduler
|
843 |
"""
|
|
|
844 |
def __init__(self, accelerator, model, optimizer, scheduler):
|
845 |
super(MegatronEngine, self).__init__()
|
846 |
self.module = model
|
|
|
886 |
Args:
|
887 |
batch_data (:obj:`dict`): The batch data to train on.
|
888 |
"""
|
|
|
889 |
args = get_args()
|
890 |
timers = get_timers()
|
891 |
|
|
|
986 |
Args:
|
987 |
batch_data (:obj:`dict`): The batch data to evaluate on.
|
988 |
"""
|
|
|
989 |
args = get_args()
|
990 |
data_chunks = []
|
991 |
if args.num_micro_batches > 1:
|
|
|
1168 |
length_penalty (float, optional): length penalty for beam search. Defaults to None.
|
1169 |
kwargs: additional key-value arguments
|
1170 |
"""
|
|
|
1171 |
# checking if required arguments are passed
|
1172 |
args = get_args()
|
1173 |
if args.model_type_name != "gpt":
|
|
|
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 |
|
|
|
1335 |
The data to gather across data parallel ranks.
|
1336 |
|
1337 |
"""
|
|
|
1338 |
def _gpu_gather_one(tensor):
|
1339 |
if tensor.ndim == 0:
|
1340 |
tensor = tensor.clone()[None]
|
src/utils/modeling.py
CHANGED
@@ -395,7 +395,6 @@ def get_non_persistent_buffers(module: nn.Module, recurse: bool = False):
|
|
395 |
recurse (`bool`, *optional*, defaults to `False`):
|
396 |
Whether or not to go look in every submodule or just return the direct non persistent buffers.
|
397 |
"""
|
398 |
-
|
399 |
non_persistent_buffers_set = module._non_persistent_buffers_set
|
400 |
if recurse:
|
401 |
for _, m in module.named_modules():
|
@@ -409,7 +408,6 @@ class FindTiedParametersResult(list):
|
|
409 |
This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
|
410 |
a list or on the `values` method as in the future this will be removed.
|
411 |
"""
|
412 |
-
|
413 |
def __init__(self, *args, **kwargs):
|
414 |
super().__init__(*args, **kwargs)
|
415 |
|
@@ -428,7 +426,6 @@ def check_tied_parameters_in_config(model: nn.Module):
|
|
428 |
Returns:
|
429 |
bool: True if the model needs to have tied weights
|
430 |
"""
|
431 |
-
|
432 |
# based on model.tie_weights() method
|
433 |
has_tied_word_embedding = False
|
434 |
has_tied_encoder_decoder = False
|
|
|
395 |
recurse (`bool`, *optional*, defaults to `False`):
|
396 |
Whether or not to go look in every submodule or just return the direct non persistent buffers.
|
397 |
"""
|
|
|
398 |
non_persistent_buffers_set = module._non_persistent_buffers_set
|
399 |
if recurse:
|
400 |
for _, m in module.named_modules():
|
|
|
408 |
This is a subclass of a list to handle backward compatibility for Transformers. Do not rely on the fact this is not
|
409 |
a list or on the `values` method as in the future this will be removed.
|
410 |
"""
|
|
|
411 |
def __init__(self, *args, **kwargs):
|
412 |
super().__init__(*args, **kwargs)
|
413 |
|
|
|
426 |
Returns:
|
427 |
bool: True if the model needs to have tied weights
|
428 |
"""
|
|
|
429 |
# based on model.tie_weights() method
|
430 |
has_tied_word_embedding = False
|
431 |
has_tied_encoder_decoder = False
|
src/utils/offload.py
CHANGED
@@ -85,7 +85,6 @@ class PrefixedDataset(Mapping):
|
|
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.
|
87 |
"""
|
88 |
-
|
89 |
def __init__(self, dataset: Mapping, prefix: str):
|
90 |
self.dataset = dataset
|
91 |
self.prefix = prefix
|
@@ -113,7 +112,6 @@ class OffloadedWeightsLoader(Mapping):
|
|
113 |
A dictionary from weight name to their information (`dtype`/ `shape` or safetensors filename). Will default
|
114 |
to the index saved in `save_folder`.
|
115 |
"""
|
116 |
-
|
117 |
def __init__(
|
118 |
self,
|
119 |
state_dict: Dict[str, torch.Tensor] = None,
|
|
|
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.
|
87 |
"""
|
|
|
88 |
def __init__(self, dataset: Mapping, prefix: str):
|
89 |
self.dataset = dataset
|
90 |
self.prefix = prefix
|
|
|
112 |
A dictionary from weight name to their information (`dtype`/ `shape` or safetensors filename). Will default
|
113 |
to the index saved in `save_folder`.
|
114 |
"""
|
|
|
115 |
def __init__(
|
116 |
self,
|
117 |
state_dict: Dict[str, torch.Tensor] = None,
|
src/utils/operations.py
CHANGED
@@ -150,7 +150,6 @@ def get_data_structure(data):
|
|
150 |
Returns:
|
151 |
The same data structure as `data` with [`~utils.TensorInformation`] instead of tensors.
|
152 |
"""
|
153 |
-
|
154 |
def _get_data_structure(tensor):
|
155 |
return TensorInformation(shape=tensor.shape, dtype=tensor.dtype)
|
156 |
|
@@ -168,7 +167,6 @@ def get_shape(data):
|
|
168 |
Returns:
|
169 |
The same data structure as `data` with lists of tensor shapes instead of tensors.
|
170 |
"""
|
171 |
-
|
172 |
def _get_shape(tensor):
|
173 |
return list(tensor.shape)
|
174 |
|
@@ -182,7 +180,6 @@ def initialize_tensors(data_structure):
|
|
182 |
Returns:
|
183 |
The same data structure as `data` with tensors instead of [`~utils.TensorInformation`].
|
184 |
"""
|
185 |
-
|
186 |
def _initialize_tensor(tensor_info):
|
187 |
return torch.empty(*tensor_info.shape, dtype=tensor_info.dtype)
|
188 |
|
@@ -222,7 +219,6 @@ def listify(data):
|
|
222 |
Returns:
|
223 |
The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
|
224 |
"""
|
225 |
-
|
226 |
def _convert_to_list(tensor):
|
227 |
tensor = tensor.detach().cpu()
|
228 |
if tensor.dtype == torch.bfloat16:
|
@@ -293,7 +289,6 @@ class DistributedOperationException(Exception):
|
|
293 |
An exception class for distributed operations. Raised if the operation cannot be performed due to the shape of the
|
294 |
tensors.
|
295 |
"""
|
296 |
-
|
297 |
pass
|
298 |
|
299 |
|
@@ -301,7 +296,6 @@ def verify_operation(function):
|
|
301 |
"""
|
302 |
Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
|
303 |
"""
|
304 |
-
|
305 |
@wraps(function)
|
306 |
def wrapper(*args, **kwargs):
|
307 |
if PartialState().distributed_type == DistributedType.NO or not PartialState().debug:
|
@@ -337,7 +331,6 @@ def chained_operation(function):
|
|
337 |
Checks that `verify_operation` failed and if so reports a more helpful error chaining the existing
|
338 |
`DistributedOperationException`.
|
339 |
"""
|
340 |
-
|
341 |
@wraps(function)
|
342 |
def wrapper(*args, **kwargs):
|
343 |
try:
|
@@ -469,7 +462,6 @@ def slice_tensors(data, tensor_slice, process_index=None, num_processes=None):
|
|
469 |
Returns:
|
470 |
The same data structure as `data` with all the tensors slices.
|
471 |
"""
|
472 |
-
|
473 |
def _slice_tensor(tensor, tensor_slice):
|
474 |
return tensor[tensor_slice]
|
475 |
|
@@ -518,7 +510,6 @@ def pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
|
518 |
pad_first (`bool`, *optional*, defaults to `False`):
|
519 |
Whether to pad at the beginning or the end.
|
520 |
"""
|
521 |
-
|
522 |
def _pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
523 |
if getattr(tensor, "is_nested", False):
|
524 |
warnings.warn(
|
@@ -572,7 +563,6 @@ def reduce(tensor, reduction="mean", scale=1.0):
|
|
572 |
Returns:
|
573 |
The same data structure as `data` with all the tensors reduced.
|
574 |
"""
|
575 |
-
|
576 |
def _reduce_across_processes(tensor, reduction="mean", scale=1.0):
|
577 |
state = PartialState()
|
578 |
cloned_tensor = tensor.clone()
|
@@ -602,7 +592,6 @@ def convert_to_fp32(tensor):
|
|
602 |
Returns:
|
603 |
The same data structure as `tensor` with all tensors that were in FP16/BF16 precision converted to FP32.
|
604 |
"""
|
605 |
-
|
606 |
def _convert_to_fp32(tensor):
|
607 |
return tensor.float()
|
608 |
|
@@ -624,7 +613,6 @@ class ConvertOutputsToFp32:
|
|
624 |
Returns:
|
625 |
The same function as `model_forward` but with converted outputs.
|
626 |
"""
|
627 |
-
|
628 |
def __init__(self, model_forward):
|
629 |
self.model_forward = model_forward
|
630 |
update_wrapper(self, model_forward)
|
|
|
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 |
|
|
|
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 |
|
|
|
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 |
|
|
|
219 |
Returns:
|
220 |
The same data structure as `data` with lists of numbers instead of `torch.Tensor`.
|
221 |
"""
|
|
|
222 |
def _convert_to_list(tensor):
|
223 |
tensor = tensor.detach().cpu()
|
224 |
if tensor.dtype == torch.bfloat16:
|
|
|
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 |
|
|
|
296 |
"""
|
297 |
Verifies that `tensor` is the same shape across all processes. Only ran if `PartialState().debug` is `True`.
|
298 |
"""
|
|
|
299 |
@wraps(function)
|
300 |
def wrapper(*args, **kwargs):
|
301 |
if PartialState().distributed_type == DistributedType.NO or not PartialState().debug:
|
|
|
331 |
Checks that `verify_operation` failed and if so reports a more helpful error chaining the existing
|
332 |
`DistributedOperationException`.
|
333 |
"""
|
|
|
334 |
@wraps(function)
|
335 |
def wrapper(*args, **kwargs):
|
336 |
try:
|
|
|
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 |
|
|
|
510 |
pad_first (`bool`, *optional*, defaults to `False`):
|
511 |
Whether to pad at the beginning or the end.
|
512 |
"""
|
|
|
513 |
def _pad_across_processes(tensor, dim=0, pad_index=0, pad_first=False):
|
514 |
if getattr(tensor, "is_nested", False):
|
515 |
warnings.warn(
|
|
|
563 |
Returns:
|
564 |
The same data structure as `data` with all the tensors reduced.
|
565 |
"""
|
|
|
566 |
def _reduce_across_processes(tensor, reduction="mean", scale=1.0):
|
567 |
state = PartialState()
|
568 |
cloned_tensor = tensor.clone()
|
|
|
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 |
|
|
|
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)
|