diff --git a/tensorflow/core/kernels/avgpooling_op.cc b/tensorflow/core/kernels/avgpooling_op.cc
index 1cc5a2d8a3e..ba38e1a188f 100644
--- a/tensorflow/core/kernels/avgpooling_op.cc
+++ b/tensorflow/core/kernels/avgpooling_op.cc
@@ -36,10 +36,10 @@ limitations under the License.
 #include "tensorflow/core/util/padding.h"
 #include "tensorflow/core/util/tensor_format.h"
 
-#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#if GOOGLE_CUDA
 #include "tensorflow/core/kernels/maxpooling_op_gpu.h"
 #include "tensorflow/core/kernels/pooling_ops_common_gpu.h"
-#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#endif  // GOOGLE_CUDA
 
 namespace tensorflow {
 
@@ -112,7 +112,7 @@ REGISTER_KERNEL_BUILDER(
     Name("AvgPool").Device(DEVICE_CPU).TypeConstraint<Eigen::half>("T"),
     AvgPoolingOp<CPUDevice, Eigen::half>);
 
-#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#if GOOGLE_CUDA
 template <typename T>
 class AvgPoolingOp<GPUDevice, T> : public UnaryOp<T> {
  public:
@@ -205,7 +205,7 @@ REGISTER_KERNEL_BUILDER(
 REGISTER_KERNEL_BUILDER(
     Name("AvgPool").Device(DEVICE_GPU).TypeConstraint<double>("T"),
     AvgPoolingOp<GPUDevice, double>);
-#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#endif  // GOOGLE_CUDA
 
 // The operation to compute AvgPool gradients.
 // It takes two inputs:
@@ -368,7 +368,7 @@ TF_CALL_float(REGISTER_CPU_KERNEL);
 TF_CALL_double(REGISTER_CPU_KERNEL);
 TF_CALL_half(REGISTER_CPU_KERNEL);
 
-#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#if GOOGLE_CUDA
 
 // A CUDNN based AvgPoolingGrad implementation. It includes the padding as the
 // candidates for the pooling operation.
@@ -577,6 +577,6 @@ REGISTER_KERNEL_BUILDER(Name("AvgPoolGrad")
                             .HostMemory("orig_input_shape"),
                         AvgPoolingGradOpCustomGPUKernel<Eigen::half>);
 
-#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#endif  // GOOGLE_CUDA
 
 }  // namespace tensorflow
diff --git a/tensorflow/stream_executor/gpu/gpu_driver.h b/tensorflow/stream_executor/gpu/gpu_driver.h
index 07b35192f03..57a87fb5bd2 100644
--- a/tensorflow/stream_executor/gpu/gpu_driver.h
+++ b/tensorflow/stream_executor/gpu/gpu_driver.h
@@ -19,16 +19,13 @@ limitations under the License.
 #define TENSORFLOW_STREAM_EXECUTOR_GPU_GPU_DRIVER_H_
 
 #include <stddef.h>
-#include "tensorflow/stream_executor/platform/port.h"
 
-#include "third_party/gpus/cuda/include/cuda.h"
 #include "tensorflow/stream_executor/device_options.h"
+#include "tensorflow/stream_executor/gpu/gpu_types.h"
 #include "tensorflow/stream_executor/lib/status.h"
 #include "tensorflow/stream_executor/lib/statusor.h"
 #include "tensorflow/stream_executor/platform/port.h"
 
-#include "tensorflow/stream_executor/gpu/gpu_types.h"
-
 namespace stream_executor {
 namespace gpu {