cherry-pick from (PaddlePaddle#54937

Make FLAGS_force_align_vpp_grad_sum_order default to false (PaddlePaddle#54937)

* make FLAGS_force_align_vpp_grad_sum_order default to false

* polish code

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/hybrid_parallel_optimizer.py

@@ -50,7 +50,7 @@ def __init__(self, clip, hcg):
         self.not_sharding_stage1 = True
         self._vpp_chunk_num = None
         self._force_align_vpp_grad_sum_order = distutils.util.strtobool(
-            os.getenv('FLAGS_force_align_vpp_grad_sum_order', '1')
+            os.getenv('FLAGS_force_align_vpp_grad_sum_order', '0')
         )
 
     def _get_vpp_chunk_num(self, params_grads):
@@ -219,9 +219,10 @@ def _global_norm(self, global_norm_var_dist, global_norm_var_not_dist):
 
     @no_grad()
     def _dygraph_clip(self, params_grads):
-        chunk_num = self._get_vpp_chunk_num(params_grads)
-        if chunk_num > 0 and self._force_align_vpp_grad_sum_order:
-            return self._vpp_dygraph_clip(params_grads, chunk_num)
+        if self._force_align_vpp_grad_sum_order:
+            chunk_num = self._get_vpp_chunk_num(params_grads)
+            if chunk_num > 0:
+                return self._vpp_dygraph_clip(params_grads, chunk_num)
 
         sum_square_dist_fp16 = []
         sum_square_dist_bf16 = []