fix the checkpoint recovery

finetrainers/trainer/sft_trainer/trainer.py

@@ -325,21 +325,8 @@ class SFTTrainer:
         resume_from_checkpoint = self.args.resume_from_checkpoint
         if resume_from_checkpoint == "latest":
             resume_from_checkpoint = -1
-            
-        # Store the load result
-        load_successful = False
         if resume_from_checkpoint is not None:
-            load_successful = self.checkpointer.load(resume_from_checkpoint)
-            
-        # If loading succeeded and we have a specific checkpoint path
-        if load_successful and isinstance(resume_from_checkpoint, str) and resume_from_checkpoint != "latest":
-            try:
-                step = int(resume_from_checkpoint.split("_")[-1])
-                self.state.train_state.step = step
-                logger.info(f"Explicitly setting training step to {step} based on checkpoint path")
-            except (ValueError, IndexError):
-                logger.warning(f"Could not parse step number from checkpoint path: {resume_from_checkpoint}")
-                
+            self.checkpointer.load(resume_from_checkpoint)
 
     def _train(self) -> None:
         logger.info("Starting training")

vms/ui/project/services/training.py

@@ -1097,6 +1097,11 @@ class TrainingService:
                 latest_checkpoint = max(checkpoints, key=lambda x: int(x.name.split("_")[-1]))
                 checkpoint_step = int(latest_checkpoint.name.split("_")[-1])
                 logger.info(f"Found checkpoint at step {checkpoint_step}")
+
+                # both options are valid, but imho it is easier to just return "latest"
+                # under the hood Finetrainers will convert ("latest") to (-1)
+                #latest_checkpoint = int(checkpoint_step)
+                latest_checkpoint = "latest"
             else:
                 logger.warning("No checkpoints found for recovery")
                 # Set buttons for no active training