Skip to content

[training] fix: use int64 for TrainState counters to prevent overflow #3312

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
yaoyu-33 merged 3 commits into from
Apr 16, 2026
+9 −9
Merged

[training] fix: use int64 for TrainState counters to prevent overflow #3312

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a74a3ed
[training] fix: use int64 for TrainState counters to prevent overflow
yaoyu-33 Apr 14, 2026
45f7814
[test] fix: update test_state assertions to expect int64 dtype
yaoyu-33 Apr 14, 2026
d4f7fe3
[training] fix: use uint64 instead of int64 for TrainState counters
yaoyu-33 Apr 16, 2026
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
8 changes: 4 additions & 4 deletions src/megatron/bridge/training/state.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,10 +70,10 @@ def state_dict(self) -> dict[str, torch.Tensor]:
their corresponding tensor representations.
"""
return {
"step": torch.tensor(self.step, dtype=torch.int32),
"consumed_train_samples": torch.tensor(self.consumed_train_samples, dtype=torch.int32),
"skipped_train_samples": torch.tensor(self.skipped_train_samples, dtype=torch.int32),
"consumed_valid_samples": torch.tensor(self.consumed_valid_samples, dtype=torch.int32),
"step": torch.tensor(self.step, dtype=torch.uint64),
"consumed_train_samples": torch.tensor(self.consumed_train_samples, dtype=torch.uint64),
"skipped_train_samples": torch.tensor(self.skipped_train_samples, dtype=torch.uint64),
"consumed_valid_samples": torch.tensor(self.consumed_valid_samples, dtype=torch.uint64),
"floating_point_operations_so_far": torch.tensor(
self.floating_point_operations_so_far, dtype=torch.float64
),
Expand Down
10 changes: 5 additions & 5 deletions tests/unit_tests/training/test_state.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,11 +86,11 @@ def test_state_dict_structure_and_types(self):
}
assert set(state_dict.keys()) == expected_keys

# Check tensor types
assert state_dict["step"].dtype == torch.int32
assert state_dict["consumed_train_samples"].dtype == torch.int32
assert state_dict["skipped_train_samples"].dtype == torch.int32
assert state_dict["consumed_valid_samples"].dtype == torch.int32
# Check tensor types (uint64 to avoid overflow on long training runs)
assert state_dict["step"].dtype == torch.uint64
assert state_dict["consumed_train_samples"].dtype == torch.uint64
assert state_dict["skipped_train_samples"].dtype == torch.uint64
assert state_dict["consumed_valid_samples"].dtype == torch.uint64
assert state_dict["floating_point_operations_so_far"].dtype == torch.float64
assert state_dict["do_train"].dtype == torch.bool
assert state_dict["do_valid"].dtype == torch.bool
Expand Down
Loading