use TileD

paddle/fluid/operators/expand_op_npu.cc

@@ -57,16 +57,13 @@ class ExpandNPUKernel : public framework::OpKernel<T> {
             "of dimensions (%d) of the input.",
             expand_times.size(), static_cast<size_t>(in_dims.size())));
     auto* out0 = context.Output<framework::LoDTensor>("Out");
+    out0->mutable_data<T>(context.device_context().GetPlace());
     framework::DDim out_dims(in_dims);
     for (size_t i = 0; i < expand_times.size(); ++i) {
       out_dims[i] *= expand_times[i];
     }
     out0->Resize(out_dims);
-
-    framework::LoDTensor expand_times_t;
-    TensorFromVector(expand_times, context.device_context(), &expand_times_t);
-
-    auto runner = NpuOpRunner("Tile", {*in0, expand_times_t}, {*out0});
+    auto runner = NpuOpRunner("TileD", {*in0}, {*out0}, {{"multiples", expand_times}});
     auto stream =
         context.template device_context<paddle::platform::NPUDeviceContext>()
             .stream();