fix bug of mp (#52789)

python/paddle/distributed/fleet/layers/mpu/mp_ops.py

@@ -46,7 +46,15 @@ def _c_identity(tensor, group=None):
         class c_identity_eager(PyLayer):
             @staticmethod
             def forward(ctx, tensor):
-                return tensor
+                return _legacy_C_ops.c_identity(
+                    tensor,
+                    'use_calc_stream',
+                    True,
+                    'ring_id',
+                    group.id,
+                    'use_model_parallel',
+                    True,
+                )
 
             @staticmethod
             def backward(ctx, dy):
@@ -249,7 +257,15 @@ def _mp_allreduce(
 
             @staticmethod
             def backward(ctx, dy):
-                return dy
+                return _legacy_C_ops.c_identity(
+                    dy,
+                    'use_calc_stream',
+                    True,
+                    'ring_id',
+                    ctx.ring_id,
+                    'use_model_parallel',
+                    True,
+                )
 
         return mp_allreduce_eager.apply(
             tensor, group, use_calc_stream, use_model_parallel