Skip to content

[SDL] StatefulDistributedSampler add epoch to state_dict #1515

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

Merged
meta-codesync merged 3 commits into from
Dec 3, 2025
Merged

[SDL] StatefulDistributedSampler add epoch to state_dict #1515

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8cd24d5
StatefulDistributedSampler add epoch to state_dict
aelavender Nov 18, 2025
23e0ddd
update tests with new state dict format
aelavender Nov 19, 2025
9981716
remove testing print
aelavender Dec 2, 2025
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
11 changes: 8 additions & 3 deletions test/stateful_dataloader/test_sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,10 +96,10 @@ def test_sampler_state_dict(self):
def test_sampler_load_state_dict(self):

sampler = StatefulDistributedSampler(self.dataset, num_replicas=10, rank=0)
sampler.load_state_dict({"yielded": 3})
sampler.load_state_dict({"epoch": 0, "yielded": 3})
self.assertEqual(sampler.next_yielded, 3)
with self.assertRaises(ValueError):
sampler.load_state_dict({"yielded": -1})
sampler.load_state_dict({"epoch": 0, "yielded": -1})

def test_sampler_next_yielded(self):

Expand All @@ -108,7 +108,12 @@ def test_sampler_next_yielded(self):
next(iterator) # advance the iterator
self.assertEqual(sampler.yielded, 1)
self.assertIsNone(sampler.next_yielded)
sampler.load_state_dict({StatefulDistributedSampler._YIELDED: 5})
sampler.load_state_dict(
{
StatefulDistributedSampler._EPOCH: 0,
StatefulDistributedSampler._YIELDED: 5,
}
)
self.assertEqual(sampler.next_yielded, 5)
iterator = iter(sampler)
next(iterator) # advance the iterator again
Expand Down
7 changes: 6 additions & 1 deletion torchdata/stateful_dataloader/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,6 +181,7 @@ def __iter__(self):

class StatefulDistributedSampler(torch.utils.data.distributed.DistributedSampler):
_YIELDED = "yielded"
_EPOCH = "epoch"

def __init__(
self,
Expand All @@ -196,6 +197,7 @@ def __init__(
self.next_yielded = None

def __iter__(self):
print(f"Calling __iter__... {self.yielded=} {self.next_yielded=}")
Copy link
Contributor

@divyanshk divyanshk Nov 19, 2025

Choose a reason for hiding this comment

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

We can skip the print ?

Copy link
Contributor Author

@aelavender aelavender Dec 2, 2025

Choose a reason for hiding this comment

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

Sorry, will remove

self.yielded = 0
if self.next_yielded is not None:
self.yielded = self.next_yielded
Expand All @@ -206,9 +208,12 @@ def __iter__(self):
yield idx

def state_dict(self) -> Dict[str, Any]:
return {self._YIELDED: self.yielded}
return {self._YIELDED: self.yielded, self._EPOCH: self.epoch, "NEXT_YIELDED": self.next_yielded}

def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
if self._EPOCH not in state_dict:
raise ValueError("Invalid state_dict")
self.set_epoch(state_dict[self._EPOCH])
Copy link
Contributor

@divyanshk divyanshk Nov 19, 2025

Choose a reason for hiding this comment

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

If users are used to calling set_epoch in a certain in the main loop, then doing self.set_epoch(..) on load_state_dict might interfere with that ?

Copy link
Contributor Author

@aelavender aelavender Dec 2, 2025

Choose a reason for hiding this comment

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

If the user is trying to set_epoch to something other than what was previously saved, then they are likely going to get repeated indices in the current epoch, which seems like broken behaviour. Perhaps we should warn in that case.

if self._YIELDED not in state_dict:
raise ValueError("Invalid state_dict")
if state_dict[self._YIELDED] < 0:
Expand Down