【AutoParallelism】fix dataloader bug and add ci for static (#8014)

* fix dataloader bug and add ci for static

* change dy2st ci

* disable fused_rms_norm

* disable pir-executor

* add loss on V100

* polish

* polish

* delete dy2st ci

* delete dy2st ci

paddlenlp/trainer/auto_trainer.py

@@ -92,9 +92,12 @@ def _get_meshes_for_loader(self):
         def _get_mesh(pp_idx=0):
             return self.global_mesh.get_mesh_with_dim("pp")[pp_idx]
 
+        # Note(lizhiyu): If the values returned by `DataLoader` don't have the format `[images, labels]`,
+        # error may occurs here.
         meshes = []
-        for pp_idx in range(self.args.pipeline_parallel_degree):
-            meshes.append(_get_mesh(pp_idx))
+        meshes.append(_get_mesh(0))
+        if self.args.pipeline_parallel_degree > 1:
+            meshes.append(_get_mesh(self.args.pipeline_parallel_degree - 1))
         return meshes
 
     def _wrap_for_dist_loader(self, train_dataloader):