Prefixing Attention Sinks can Mitigate Activation Outliers for Large Language Model Quantization

Devs

Prefixing Attention Sinks can Mitigate Activation Outliers for Large Language Model Quantization | Read Paper on Bytez