Used to decide how many elements of a tensor will share the same quantization group. Higher values have better performance but less quality. Default is 0, meaning it will decide the group size based ...