Hermit Dave

Data-aware Weight Quantization of models. The RTN is further trained using data generated from BF16 model or 8bit RTN quant if BF16 is not feasible.