ResCom: Activation Residual Compensation for Large Language Model Quantization
ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error, is proposed, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error.
Ning-Rui Yang, Bo-Han Cui, Zhuo-Wen Liu et al.
· 0 citations