fix split_tensor of dp_pp_comm_overlap (PaddlePaddle#54310)

Author: haohongxiang

python/paddle/distributed/fleet/meta_parallel/pp_utils/utils.py

@@ -20,6 +20,7 @@
 from paddle import _legacy_C_ops
 from paddle.distributed.parallel import _split_tensors
 from paddle.fluid import core
+from paddle.framework import base as imperative_base
 
 __all__ = []
 
@@ -165,6 +166,7 @@ def add_grad(self, param):
         if self._all_params_checked_in:
             self._fused_allreduce_grads()
 
+    @imperative_base.no_grad
     def _fused_allreduce_grads(self):
         assert self._all_params_checked_in
         flattened_vars = []
@@ -188,6 +190,7 @@ def _fused_allreduce_grads(self):
                 )
             )
 
+    @imperative_base.no_grad
     def scale_and_split_grads(self):
         for task in self._tasks:
             task.wait()