Missing symbol "litert::CompiledModel::CreateInputOutputBuffer"

[abtiwary]

Hi,
I have been trying to build an Android application that uses Litert-Next (release 2.0.2a1).

I have tried to follow the instructions on https://ai.google.dev/edge/litert/next/android_cpp but i've run into some issues, for example this symbol has been missing when i call "CreateInputBuffer(...)" on the compiled model (the compilation of the model works fine by the way): "ld.lld: error: undefined symbol: litert::CompiledModel::CreateInputOutputBuffer".

So far, i've tried the following approaches:

• build the binaries using bazel
• use the precompiled binaries from here https://mvnrepository.com/artifact/com.google.ai.edge.litert/litert/2.0.2a1, as well as the header files from https://github.com/google-ai-edge/LiteRT/archive/refs/tags/v2.0.2a1.tar.gz

In both cases, using nm to search for the symbols, the above symbol seems to be missing (there are other symbols that are missing too). Is this expected, given the alpha nature of the project, or am i missing something glaringly obvious?

Thanks!

[gaikwadrahul8]

Hi, @abtiwary
I apologize for the delay in my response, Thank you for filing such a detailed and well investigated issue. You've correctly identified that the symbol is missing. The root cause is an API change in the LiteRT-Next v2.0.2a1 alpha release. The function litert::CompiledModel::CreateInputOutputBuffer that the linker is looking for doesn't exist in the public C++ API for that version. As this is an alpha version, APIs are still evolving for better performance and clarity.

The correct approach is to create input and output buffers separately using CreateInputBuffers() and CreateOutputBuffers() as shown in the official Android C++ guide and available public C++ API . Please update your C++ Code to call the two separate functions on your compiled_model object something like below :

// Create input and output buffers using the correct separate functions
LITERT_ASSIGN_OR_RETURN(auto input_buffers,
                        compiled_model.CreateInputBuffers());
LITERT_ASSIGN_OR_RETURN(auto output_buffers,
                        compiled_model.CreateOutputBuffers());

// Now you can proceed to run inference
// compiled_model.Run(input_buffers, output_buffers);

To ensure the linker can find the library correctly please also verify that your app/src/main/cpp/CMakeLists.txt is configured to link against the liblitert_native.so from the AAR dependency. Here is a recommended configuration for linking the precompiled library:

# Add the LiteRT library as an IMPORTED target
add_library(litert SHARED IMPORTED)

# Set the location of the pre-built.so file from the AAR
set_target_properties(litert PROPERTIES IMPORTED_LOCATION
    ${CMAKE_BINARY_DIR}/../../../../build/intermediates/stripped_native_libs/${CMAKE_BUILD_TYPE}/out/lib/${ANDROID_ABI}/liblitert_native.so)

# Link your application's native library against LiteRT
target_link_libraries(
    your-native-lib-name  # Change this to your library's name
    litert
    android
    log
)

By updating your C++ code to use the correct API and ensuring your CMake is configured as above the linker will be able to find the required symbols and resolve the error. Please let us know if this solves the issue for you. We appreciate your feedback as we work on stabilizing the LiteRT-Next API. If I've missed something here please let me know. Thank you for your cooperation and patience.

[abtiwary]

Hi @gaikwadrahul8 , thank you for taking the time to respond! I'm going to try the approach that you've described above.

[abtiwary]

Hi @gaikwadrahul8 , apologies for the delay in getting back to you but this still fails for me - now in 2.0.2 (i got the AAR from here https://mvnrepository.com/artifact/com.google.ai.edge.litert/litert/2.0.2 , and the source files (headers) from here https://github.com/google-ai-edge/LiteRT/releases).

I noticed that the header file litert_compiled_model.h still contains the following code:

...
  // A helper function to create input tensor buffers for the given signature.
  // It uses BufferRequirements and RankedTensorType to create the input tensor
  // buffers.
  Expected<std::vector<TensorBuffer>> CreateInputBuffers(
      absl::string_view signature_name) const {
    LITERT_ASSIGN_OR_RETURN(size_t signature_index,
                            model_.GetSignatureIndex(signature_name));
    return CreateInputOutputBuffers(signature_index, /*is_input=*/true);
  }

  // A helper function to creates the input tensor buffers for the given
  // signature. It uses BufferRequirements and RankedTensorType to create the
  // input tensor buffers.
  Expected<std::vector<TensorBuffer>> CreateInputBuffers(
      size_t signature_index) const {
    return CreateInputOutputBuffers(signature_index, /*is_input=*/true);
  }

  // A helper function to creates the input tensor buffers for the default
  // signature. It uses BufferRequirements and RankedTensorType to create the
  // input tensor buffers.
  Expected<std::vector<TensorBuffer>> CreateInputBuffers() const {
    return CreateInputOutputBuffers(/*signature_index=*/0, /*is_input=*/true);
  }
...

So yeah, it's still looking for that symbol, which i cannot find in the dynamic libraries.

(ld.lld: error: undefined symbol: litert::CompiledModel::CreateInputOutputBuffers(unsigned long, bool) const)