Fix logging inconsistency: unify metrics logging to use global_steps

- Add global_steps property to PeftTrainer (returns train_steps for compatibility)
- Update DPO trainer to use global_steps for aux metrics logging
- Update progress bar to use global_steps for initial_steps
- Ensures consistent step counting across all RL algorithms
- Resolves confusion about max_steps parameter interpretation

Author: joaobenedetmachado

tunix/sft/dpo/dpo_trainer.py

@@ -262,7 +262,7 @@ def _prepare_inputs(
 
   @override
   def _post_process_train_step(self, aux: Any) -> None:
-    m, s = self._mode, self._train_steps
+    m, s = self._mode, self.global_steps
     self.metrics_logger.log("rewards/chosen", aux["rewards/chosen"], m, s)
     self.metrics_logger.log("rewards/rejected", aux["rewards/rejected"], m, s)
     self.metrics_logger.log("rewards/margin", aux["rewards/margin"], m, s)
@@ -274,7 +274,7 @@ def _post_process_train_step(self, aux: Any) -> None:
 
   @override
   def _post_process_eval_step(self, aux: Any) -> None:
-    m, s = self._mode, self._train_steps
+    m, s = self._mode, self.global_steps
     self.metrics_logger.log("rewards/chosen", aux["rewards/chosen"], m, s)
     self.metrics_logger.log("rewards/rejected", aux["rewards/rejected"], m, s)
     self.metrics_logger.log("rewards/margin", aux["rewards/margin"], m, s)

tunix/sft/peft_trainer.py

@@ -586,7 +586,7 @@ def train(
     if self.config.max_steps is not None and self._pbar is None:
       self._pbar = progress_bar.ProgressBar(
           metrics_logger=self.metrics_logger,
-          initial_steps=self._train_steps,
+          initial_steps=self.global_steps,
           max_steps=self.config.max_steps,
           description=self.config.pbar_description,
       )
@@ -713,6 +713,11 @@ def train_steps(self) -> int:
     """Returns the number of train steps taken."""
     return self._train_steps
 
+  @property
+  def global_steps(self) -> int:
+    """Returns the number of global steps taken (same as train_steps for compatibility)."""
+    return self._train_steps
+
   @property
   def iter_steps(self) -> int:
     """Returns the number of iterator steps taken."""