[bugfix] fix megatron kto pp + sp (#7882)

Author: Jintao-Huang

swift/megatron/trainers/kto_trainer.py

@@ -128,6 +128,8 @@ def forward_step(self, data_iterator, model):
         kl_data.pop('loss_scale', None)
 
         length = self._get_kto_length(data)
+        if self.args.sequence_parallel:
+            length //= mpu.get_tensor_model_parallel_world_size()
         with torch.no_grad(), self.null_ref_context() as ref_models:
             ref_model = ref_models[vp_stage or 0]
             if self.args.calculate_KL: