fix: fire on_step_end before incrementing global_step

on_step_begin(step=N) was followed by on_step_end(step=N+1) because
global_step was incremented before on_step_end fired. Move on_step_end
before the increment so begin/end see the same step number.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>

Author: 3 people

src/instructlab/training/main_ds.py

@@ -420,15 +420,14 @@ def _save_and_exit(checkpoint_location: str) -> None:
                     ckpt_path = os.path.join(args.output_dir, "hf_format", subdir)
                     callback_manager.fire("on_save", checkpoint_path=ckpt_path)
 
+            if callback_manager:
+                callback_manager.fire("on_step_end")
+
             global_step += 1
             if local_rank == 0:
                 inner_pb.update(1)
             torch.cuda.empty_cache()
 
-            if callback_manager:
-                callback_manager.context.step = global_step
-                callback_manager.fire("on_step_end")
-
         if args.checkpoint_at_epoch:
             base_logger.debug(f"Saving checkpoint at epoch {epoch}")
             save_checkpoint(