bug fix, test=develop (#32753)

Author: lilong12

python/paddle/distributed/fleet/base/topology.py

@@ -147,7 +147,7 @@ def __init__(self, topology):
         debug_str = "HybridParallelInfo: rank_id: %d, dp_degree: %d, " \
                     "mp_degree: %d, pp_degree: %d" % (self.global_rank, self._dp_degree,
                     self._mp_degree,self._pp_degree)
-        debug_str += "dp_group: %s, mp_group: %s, pp_group: %s, check/clip group: %s" % (
+        debug_str += ", dp_group: %s, mp_group: %s, pp_group: %s, check/clip group: %s" % (
             self._dp_group, self._mp_group, self._pp_group, self._check_group)
         logger.info(debug_str)
 

python/paddle/distributed/fleet/meta_parallel/pipeline_parallel.py

@@ -125,9 +125,9 @@ def _forward(self, cache_id):
             self._recv_activations(cache_id)
 
         if isinstance(self.caches['inputs'][cache_id], tuple):
-            inputs = tuple(t.clone() for t in self.caches['inputs'][cache_id])
+            inputs = tuple(t for t in self.caches['inputs'][cache_id])
         else:
-            inputs = self.caches['inputs'][cache_id].clone()
+            inputs = self.caches['inputs'][cache_id]
 
         self._clear_grads(inputs)
         outputs = self._layers.forward(inputs)