移除 global_step 自增逻辑并调整至循环末尾

src/trainer/model.py

@@ -570,7 +570,6 @@ class MoEModel(nn.Module):
                     scaler.step(optimizer)
                     scaler.update()
                     optimizer.zero_grad()
-                    global_step += 1
                     original_loss = loss.item() * grad_accum_steps
                     batch_loss_sum += original_loss
                     # 周期性评估（与原代码相同）
@@ -592,6 +591,7 @@ class MoEModel(nn.Module):
                         batch_loss_sum = 0.0
                     if processed_batches >= stop_batch:
                         break
+                    global_step += 1
 
     def load_from_state_dict(self, state_dict_path: Union[str, Path]):
         state_dict = torch.load(