Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Support saving and loading of step while saving and loading state #2765

Merged
merged 2 commits into from
Jul 1, 2024
+15 −1
Merged

Support saving and loading of step while saving and loading state #2765

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ade4ab1
Add feature to save step when saving state
bipinKrishnan May 11, 2024
ff08093
Update docstring for `load_accelerate_state`
bipinKrishnan Jun 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion src/accelerate/accelerator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2952,6 +2952,7 @@ def _inner(folder):
schedulers,
dataloaders,
self.state.process_index,
self.step,
self.scaler,
save_on_each_node=self.project_configuration.save_on_each_node,
safe_serialization=safe_serialization,
Expand Down Expand Up @@ -3099,7 +3100,7 @@ def _inner(folder):
else:
map_location = "cpu"

load_accelerator_state(
override_attributes = load_accelerator_state(
input_dir,
models,
optimizers,
Expand All @@ -3110,6 +3111,7 @@ def _inner(folder):
map_location,
**load_model_func_kwargs,
)
self.step = override_attributes["step"]
custom_checkpoints = [
f for f in os.listdir(input_dir) if re.search(r"^custom_checkpoint_\d+\.pkl$", f) is not None
]
Expand Down
12 changes: 12 additions & 0 deletions src/accelerate/checkpointing.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,7 @@ def save_accelerator_state(
schedulers: list,
dataloaders: list,
process_index: int,
step: int,
scaler: GradScaler = None,
save_on_each_node: bool = False,
safe_serialization: bool = True,
Expand Down Expand Up @@ -82,6 +83,8 @@ def save_accelerator_state(
A list of dataloader instances to save their sampler states
process_index (`int`):
The current process index in the Accelerator state
step (`int`):
The current step in the internal step tracker
scaler (`torch.cuda.amp.GradScaler`, *optional*):
An optional gradient scaler instance to save
save_on_each_node (`bool`, *optional*):
Expand Down Expand Up @@ -134,6 +137,7 @@ def save_accelerator_state(
# Random number generator states
states = {}
states_name = f"{RNG_STATE_NAME}_{process_index}.pkl"
states["step"] = step
states["random_state"] = random.getstate()
states["numpy_random_seed"] = np.random.get_state()
states["torch_manual_seed"] = torch.get_rng_state()
Expand Down Expand Up @@ -180,7 +184,12 @@ def load_accelerator_state(
What device to load the optimizer state onto. Should be one of either "cpu" or "on_device".
load_model_func_kwargs (`dict`, *optional*):
Additional arguments that can be passed to the model's `load_state_dict` method.

Returns:
`dict`: Contains the `Accelerator` attributes to override while loading the state.
"""
# stores the `Accelerator` attributes to override
override_attributes = dict()
if map_location not in [None, "cpu", "on_device"]:
raise TypeError(
"Unsupported optimizer map location passed, please choose one of `None`, `'cpu'`, or `'on_device'`"
Expand Down Expand Up @@ -240,6 +249,7 @@ def load_accelerator_state(
# Random states
try:
states = torch.load(input_dir.joinpath(f"{RNG_STATE_NAME}_{process_index}.pkl"))
override_attributes["step"] = states["step"]
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a breaking change for people upgrading accelerate. If previously saved pickled states objects don't contain a "step" key, then random states fail to initialize correctly when you resume a training that was checkpointed from a prior version of accelerate.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Correct, we've fixed this in a patch release

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah I see this now. Thanks!

random.setstate(states["random_state"])
np.random.set_state(states["numpy_random_seed"])
torch.set_rng_state(states["torch_manual_seed"])
Expand All @@ -253,6 +263,8 @@ def load_accelerator_state(
except Exception:
logger.info("Could not load random states")

return override_attributes


def save_custom_state(obj, path, index: int = 0, save_on_each_node: bool = False):
"""
Expand Down
Loading