How to convert AWQ matmul to onnxruntime MatmulNBits

[fuyao2024]

Since the AWQ quantization model requires the NV GPU, we can't run it on CPU. I want to convert the model from pytorch to ONNX which can be run on CPU. One issue is to understand the 4 bits weights packing. How are the 4 bits packed for NV CUDA inference?