Fast implementation of BERT inference directly on NVIDIA (CUDA, CUBLAS) and Intel MKL - View it on GitHub