fix

Author: HydrogenSulfate

paddle/fluid/primitive/decomp_rule/decomp_rule/composite.h

@@ -608,10 +608,10 @@ Tensor masked_fill_decomp(const Tensor& x,
     // NOTE: use add operator to get broadcast shape implicitly,
     // which is not efficient yet, should be improved in the future.
     Tensor dummy_x =
-        backend::full_with_tensor<T>(shape64<T>(x), 0.0, x.dtype());
-    Tensor dummy_y =
-        backend::full_with_tensor<T>(shape64<T>(mask), 0.0, x.dtype());
-    dummy = dummy_x + dummy_y;
+        backend::full_with_tensor<T>(shape64<T>(x), 0.0, x.dtype(), x.place());
+    Tensor dummy_y = backend::full_with_tensor<T>(
+        shape64<T>(mask), 0.0, x.dtype(), x.place());
+    Tensor dummy = dummy_x + dummy_y;
     Tensor mask_expanded = expand<T>(mask, shape64<T>(dummy));
     Tensor v_expanded = expand<T>(v, shape64<T>(dummy));
     return where<T>(mask_expanded, v_expanded, x);