[OPT]Set order for hybridparallel setting (#51781)

* set order for hybridparallel

* fix bug

* fix ->

* fix ->

* fix ->

* add topology

* fix utest

paddle/fluid/framework/distributed_strategy.proto

@@ -55,6 +55,7 @@ message HybridConfig {
   optional int32 mp_degree = 2 [ default = 1 ];
   optional int32 pp_degree = 3 [ default = 1 ];
   optional int32 sharding_degree = 4 [ default = 1 ];
+  repeated string order = 5 ;
 }
 
 message AMPConfig {

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

@@ -1673,6 +1673,8 @@ class DistributedStrategy:
 
             **pp_degree(int)**: set number of GPUs in a pipeline parallel group. Default 1
 
+            **order(list(string))**: set hybrid parallel dimensions, the order is from outside to inside. Default ['dp','pp','sharding','mp']
+
         Examples:
             .. code-block:: python
 
@@ -1681,7 +1683,8 @@ class DistributedStrategy:
                 strategy.hybrid_configs = {
                     "dp_degree": 1,
                     "mp_degree": 2,
-                    "pp_degree": 1}
+                    "pp_degree": 1,
+                    "order":['dp','pp','sharding','mp']}
 
         """
         return get_msg_dict(self.strategy.hybrid_configs)

python/paddle/distributed/fleet/fleet.py

@@ -405,14 +405,28 @@ class Fleet:
 
         self.dp_degree = max(self.dp_degree, 1)
 
+        d_hybrid_degree = {
+            "dp": ["data", self.dp_degree],
+            "pp": ['pipe', self.pp_degree],
+            "sharding": ['sharding', self.sharding_degree],
+            "mp": ['model', self.mp_degree],
+        }
+
+        order = self.hybrid_configs["order"]
+        if not order:
+            order = ['dp', 'pp', 'sharding', 'mp']
+        if order[:].sort() != list(d_hybrid_degree.keys())[:].sort():
+            assert False, "The order of hybrid_config setting is incorrect."
+
+        hybrid_group_names = []
+        dims = []
+        for h_name in order:
+            name, degree = d_hybrid_degree[h_name]
+            hybrid_group_names.append(name)
+            dims.append(degree)
+
         self._topology = tp.CommunicateTopology(
-            hybrid_group_names=["data", "pipe", "sharding", "model"],
-            dims=[
-                self.dp_degree,
-                self.pp_degree,
-                self.sharding_degree,
-                self.mp_degree,
-            ],
+            hybrid_group_names=hybrid_group_names, dims=dims
         )
 
         self._hcg = tp.HybridCommunicateGroup(self._topology)