Cuda failure: an illegal memory access was encountered

Hi Nvidia Team,

I have Implemented the Custom plugin for the Einsum operator in TensorRT. But When I am loading the plugin during the Conversion from ONNX to TRT I am getting an issue as Cuda failure: illegal memory access was encountered. It is occuring in enqueue().

Enqueue() code snippet:

```
int EinsumPlugin::enqueue(const nvinfer1::PluginTensorDesc* inputDesc,
    const nvinfer1::PluginTensorDesc* outputDesc, const void* const* inputs, void* const* outputs, void* workspace,
    cudaStream_t stream)
    {
    std::cout<<"In Enqueue \n";
     cublasHandle_t mCublas;
     cublasCreate(&mCublas);
     float onef{1.0f}, zerof{0.0f};
     cublasSetStream(mCublas, stream);
     if(equation=="nct, ncp -> ntp")
       {

           cublasSgemmBatched(mCublas, CUBLAS_OP_N,CUBLAS_OP_T, 
                                  M, K,L,&onef, 
                                  reinterpret_cast<const float *const *>(inputs[1]), M, 
                                  reinterpret_cast<const float *const *>(inputs[0]), K, 
                                  &zerof, 
                                  reinterpret_cast<float *const *>(outputs[0]), 
                                  M,N);


        }

     else if(equation=="ntg, ncg -> nct")
        {
           cublasSgemmBatched(mCublas, CUBLAS_OP_T,CUBLAS_OP_N, 
                                  L, O, K, &onef, 
                                  static_cast<const float*>(inputs[0]), K, 
                                  static_cast<const float*>(inputs[1]), K, 
                                 static_cast<float*>(outputs[0]), L, N);

        
        }
    return 0;
}
```

May I know the reason for the Issue?

I am attaching the screenshot of the log below.

Thanks in advance,
Darshan

one of your cublas calls is making an illegal memory access, perhaps.

You may get additional clues by running your code with the compute sanitizer