Compiling from source, cuDNN version is not compatible? How can I change the cuDNN compile version?

I use mvn install java-0.5.0 in jetson with below environment. Finally it completed and generated the "tensorflow-core-api-0.5.0-linux-arm64-gpu.jar".

System information

• OS Platform and Distribution : Linux Ubuntu 20.04 arm64
• TensorFlow installed from (source or binary): source
• TensorFlow version: 2.10.1
• Java version (i.e., the output of java -version): openjdk version "11.0.20"
• Java command line flags (e.g., GC parameters):
• Installed from Maven Central?:
• Bazel version (if compiling from source): 5.4.1
• GCC/Compiler version (if compiling from source): 9.4.0
• CUDA/cuDNN version: CUDA11.4+CuDNN8.6.0
• GPU model and memory:

But when I run the model , the error is occured. Here is the error log:
"2023-08-01 09:57:03.793190: E external/org_tensorflow/tensorflow/stream_executor/cuda/cuda_dnn.cc:377] Loaded runtime CuDNN library: 8.6.0 but source was compiled with: 8.9.0. CuDNN library needs to have matching major version and equal or higher minor version. If using a binary install, upgrade your CuDNN library. If building from sources, make sure the library loaded at runtime is compatible with the version specified during compile configuration."

I have tried upgrade CuDNN to 8.9.0 and run again, but it's run failed with error-"tensorflow/core/framework/op_kernel.cc:1780] OP_REQUIRES failed at conv_ops.cc:1143 : NOT_FOUND: No algorithm worked! Error messages:". I think this may be the environment doesn't match.

So how can I resolve this problem? I think this is a way to fix the problem that is to change the cuDNN compile version, but I can't find any info about this.
P.S.: I run the model successfully in my local machine(windows11 x86_64...)

Best Regards

Bắt đầu bằng cách tái hiện quá trình build và chạy model của mvn install java-0.5.0 trên môi trường Jetson được báo cáo, đồng thời so sánh phiên bản cuDNN được sử dụng trong quá trình biên dịch với phiên bản runtime. Được xem là hoàn tất khi GPU JAR được tạo có thể chạy model thành công với các phiên bản CUDA và cuDNN tương thích, hoặc khi cấu hình biên dịch cần thiết được ghi lại.