[Auto Parallel]fix bugs in cluster to device meshes (#49892)

* fix bugs in cluster to device meshes

* add tests

* 1

python/paddle/distributed/auto_parallel/tuner/rule_based_tuner.py

@@ -516,6 +516,11 @@ class ClusterPartitionUtil:
 
     @staticmethod
     def complete_meshes(partitions: list, num: int):
+        if num == 2:
+            return [[1, 2], [2, 1]]
+        if num == 3:
+            return [[1, 2], [2, 1], [1]]
+        # special cases
         if len(partitions) == 1:
             partitions = ClusterPartitionUtil.factorization(num - 1)
             partitions.append([1])

python/paddle/fluid/tests/unittests/auto_parallel/test_cluster_partition.py

@@ -17,7 +17,7 @@ import unittest
 
 class TestClusterPartition(unittest.TestCase):
     def test_cluster_partition(self):
-        clusters = [(5, 8), (1, 8), (4, 8), (16, 8)]
+        clusters = [(5, 8), (1, 8), (4, 8), (16, 8), (2, 8), (3, 8)]
         from paddle.distributed.auto_parallel.tuner.rule_based_tuner import (
             ClusterPartitionUtil,
         )