Skip to content

Normalize cloned grad_absmax metrics for structured progress payloads #2913

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
bghira merged 3 commits into from
Jul 27, 2026
Merged

Normalize cloned grad_absmax metrics for structured progress payloads #2913

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
5e31303
Initial plan
Copilot Jul 27, 2026
874ee7a
fix: normalize cloned grad absmax metrics
Copilot Jul 27, 2026
6be65fd
Merge branch 'main' of ssh://github.com/bghira/SimpleTuner into copil…
Jul 27, 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
5 changes: 4 additions & 1 deletion simpletuner/helpers/training/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -598,7 +598,10 @@ def _update_grad_metrics(
elif (
not require_value_method or self.config.grad_clip_method == "value"
) and not self.config.use_deepspeed_optimizer:
target_logs[f"{prefix}grad_absmax"] = self.grad_norm
grad_value = self.grad_norm
if clone_norm_value:
grad_value = float(self.grad_norm.clone().detach())
target_logs[f"{prefix}grad_absmax"] = grad_value

def _config_uses_bitsandbytes(self) -> bool:
if not getattr(self, "config", None):
Expand Down
17 changes: 16 additions & 1 deletion tests/test_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1249,6 +1249,19 @@ def test_update_grad_metrics_logs_absmax_without_deepspeed(self):
self.assertIn("grad_absmax", logs)
self.assertIs(logs["grad_absmax"], trainer.grad_norm)

def test_update_grad_metrics_clones_absmax_value_when_requested(self):
trainer = self._build_trainer_for_grad_logging(
grad_clip_method="value",
use_deepspeed=False,
grad_value=torch.tensor(1.2),
)
logs = {}
trainer._update_grad_metrics(logs, clone_norm_value=True)
self.assertIn("grad_absmax", logs)
self.assertEqual(
logs["grad_absmax"], float(trainer.grad_norm.clone().detach())
)

def test_update_grad_metrics_clones_norm_value_when_requested(self):
trainer = self._build_trainer_for_grad_logging(
grad_clip_method="norm",
Expand Down Expand Up @@ -1309,7 +1322,9 @@ def test_compose_training_progress_metrics_includes_grad_absmax(self):
)
metrics = trainer._compose_training_progress_metrics(epoch=1)
self.assertIn("grad_absmax", metrics)
self.assertIs(metrics["grad_absmax"], trainer.grad_norm)
self.assertEqual(
metrics["grad_absmax"], float(trainer.grad_norm.clone().detach())
)
self.assertNotIn("grad_norm", metrics)

def test_compose_training_progress_metrics_excludes_grad_with_deepspeed(self):
Expand Down
Loading