Achieve batch inference using cuda shared memory

[HosseinKD]

Hello there!
I'm new to Triton and I can't get my head around solving how to achieve batch inference.
this is my function to call inference method using a grpc client.

void TritonClient::infer(std::shared_ptr<triton::client::InferResult> &results_ptr,
                         const std::vector<unsigned char *> &inputs,
                         const std::vector<size_t> &input_byte_sizes)
{
    triton::client::InferResult *result;
    prepare_input(inputs, input_byte_sizes);
    FAIL_IF_ERR(m_client->Infer(&result, m_options, m_inputs, m_outputs), "unable to run inference");
    results_ptr.reset(result);
    results_ptr->RequestStatus();
}

here I pass a inputs vector containing addresses allocated bu cudaMalloc, the addresses are not contiguous.
I already created my triton::client::InferInput. this is my prepare_input function which I'm having trouble with:

void TritonClient::prepare_input(const std::vector<unsigned char*>& inputs,
                                  const std::vector<size_t>& input_byte_sizes) {
    unregister_input_names();

    std::vector<std::unique_ptr<TritonInput>> input_ptrs;
    for (size_t i = 0; i < inputs.size(); ++i) {
        cudaIpcMemHandle_t input_cuda_handle;
        CreateCUDAIPCHandle(&input_cuda_handle, (void*)inputs[i]);
        FAIL_IF_ERR(
            m_client->RegisterCudaSharedMemory(m_triton_registrable_input_name[i], input_cuda_handle, 0,
                                               input_byte_sizes[i]),
            "failed to register input shared memory region");

        std::unique_ptr<TritonInput> input_ptr(new TritonInput());
        FAIL_IF_ERR(input_ptr->SetSharedMemory(m_triton_registrable_input_name[i], input_byte_sizes[i], 0),
                    "unable to set shared memory for input");
        input_ptrs.push_back(std::move(input_ptr));
    }

    m_inputs.clear();
    for (const auto& ptr : input_ptrs) {
        m_inputs.push_back(ptr.get());
    }
}

but there is a several issues which I can't find a solution for. first the second time I call SetSharedMemory I get segmentation fault! second if I remove all of these and just push the same input_ptrs into the m_inputs multiple times I get error that the input already exists. I could manage something similar in python without shared memory by concatenating the inputs but I can't manage to solve this issue, any help or suggestions would be appreciated. I want to pass different gpu memory addresses with their size (which is same for all of them) for batch inference and these memories are not contiguous.