diff --git a/paddle/fluid/inference/tensorrt/plugin/gather_nd_op_plugin.h b/paddle/fluid/inference/tensorrt/plugin/gather_nd_op_plugin.h
index 841fb2f6fe399fd44ac91d29a15825ec53c93b69..a293cf69be3618993f29700df67f24f8b1353d1d 100644
--- a/paddle/fluid/inference/tensorrt/plugin/gather_nd_op_plugin.h
+++ b/paddle/fluid/inference/tensorrt/plugin/gather_nd_op_plugin.h
@@ -14,7 +14,6 @@
 
 #pragma once
 
-#include <thrust/device_vector.h>
 #include <string>
 #include <utility>
 #include <vector>
diff --git a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu
index ec4fcca6d74d0c5157e993a0be6b30640eb0b324..349716d1e2c30b12e70670068ae9f177080d870f 100644
--- a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu
+++ b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu
@@ -13,7 +13,9 @@
 // limitations under the License.
 
 #include <cuda_fp16.h>
+#include <thrust/device_vector.h>
 #include <algorithm>
+
 #include "paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h"
 
 namespace paddle {
@@ -61,9 +63,7 @@ void SplitPlugin::shareData(const SplitPlugin* another) {
   inner_cols_ = another->inner_cols_;
   same_shape_ = another->same_shape_;
   axis_shape_ = another->axis_shape_;
-  d_segment_offsets_ = another->d_segment_offsets_;
   segment_offsets_ = another->segment_offsets_;
-  d_output_ptrs_.resize(another->d_output_ptrs_.size(), nullptr);
 }
 
 int SplitPlugin::initialize() TRT_NOEXCEPT {
@@ -91,9 +91,7 @@ int SplitPlugin::initialize() TRT_NOEXCEPT {
     segment_offsets.push_back(segment_offsets.back() + output_length_[i]);
   }
   axis_shape_ = dims.d[axis_];
-  d_segment_offsets_ = segment_offsets;
   segment_offsets_ = std::move(segment_offsets);
-  d_output_ptrs_.resize(this->getNbOutputs(), nullptr);
   return 0;
 }
 
@@ -131,13 +129,18 @@ int SplitPlugin::enqueue(int batchSize, const void* const* inputs,
                          void* const* outputs, void* workspace,
                          cudaStream_t stream) TRT_NOEXCEPT {
 #endif
+  // this two thrust variables decalared here , not with in .h
+  // to avoid compiling error in cuda 11.6
+  thrust::device_vector<int> d_segment_offsets = segment_offsets_;
+  thrust::device_vector<float*> d_output_ptrs;
+  d_output_ptrs.resize(segment_offsets_.size(), nullptr);
   const int* d_segment_offsets_ptr =
-      thrust::raw_pointer_cast(&d_segment_offsets_[0]);
+      thrust::raw_pointer_cast(&d_segment_offsets[0]);
   float const* input_ptr = reinterpret_cast<float const*>(inputs[0]);
   float* const* h_odatas = reinterpret_cast<float* const*>(outputs);
-  float** output_ptrs = thrust::raw_pointer_cast(&d_output_ptrs_[0]);
+  float** output_ptrs = thrust::raw_pointer_cast(&d_output_ptrs[0]);
   PADDLE_ENFORCE_GPU_SUCCESS(cudaMemcpyAsync(
-      output_ptrs, h_odatas, d_output_ptrs_.size() * sizeof(float*),
+      output_ptrs, h_odatas, d_output_ptrs.size() * sizeof(float*),
       cudaMemcpyHostToDevice, stream));
 
   int outer_rows = outer_rows_ * batchSize;
@@ -148,7 +151,7 @@ int SplitPlugin::enqueue(int batchSize, const void* const* inputs,
             std::min((outer_rows_ - 1) / block.z + 1, 65535u));
 
   split_kernel<<<grid, block, 0, stream>>>(
-      d_segment_offsets_.size(), d_segment_offsets_ptr, input_ptr, output_ptrs,
+      segment_offsets_.size(), d_segment_offsets_ptr, input_ptr, output_ptrs,
       inner_cols_, axis_shape_, outer_rows);
   return cudaGetLastError() != cudaSuccess;
 }
diff --git a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h
index 7a41fe1d1eef2367aede87c578746ffc970f0e31..502c903dbfff639ae588561a285ffbfa62cf346d 100644
--- a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h
+++ b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h
@@ -14,7 +14,6 @@
 
 #pragma once
 
-#include <thrust/device_vector.h>
 #include <string>
 #include <utility>
 #include <vector>
@@ -92,8 +91,6 @@ class SplitPlugin : public PluginTensorRTV2Ext {
   bool same_shape_;
   std::vector<int> output_length_;
   std::vector<int> segment_offsets_;
-  thrust::device_vector<int> d_segment_offsets_;
-  thrust::device_vector<float*> d_output_ptrs_;
 
  private:
   void shareData(const SplitPlugin* another);