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

Despite recent advances in LLM quantization, activation quantization remains to be challenging due to the activation outliers. Conventional remedies, e.g., mixing precisions for different channels, introduce extra overhead and reduce the speedup. In this work, we develop a simple yet effective strategy to facilitate per-tensor activation quantization by preventing the generation of problematic tokens. Precisely, we propose a method to find a set of key-value cache, coined CushionCache, which mitigates outliers in subsequent tokens when inserted as a prefix. CushionCache works in two steps: First, we greedily search for a prompt token sequence that minimizes the maximum activation values in subsequent tokens. Then, we further tune the token cache to regularize the activations of subsequent tokens to be more quantization-friendly. The proposed method successfully addresses activation outliers of LLMs, providing a substantial performance boost for per-tensor activation quantization methods. We thoroughly evaluate our method over a wide range of models and benchmarks and find that it significantly surpasses the established baseline of per-tensor W8A8 quantization and can be seamlessly integrated with the recent activation quantization method.

MaskedKD: Efficient Distillation of Vision Transformers with Masked Images

Knowledge distillation is a popular and effective regularization technique for training lightweight models, but it also adds significant overhead to the training cost. The drawback is most pronounced when we use large-scale models as our teachers, such as vision transformers (ViTs). We present MaskedKD, a simple yet effective method for reducing the training cost of ViT distillation. MaskedKD masks a fraction of image patch tokens fed to the teacher to save the teacher inference cost. The tokens to mask are determined based on the last layer attention score of the student model, to which we provide the full image. Without requiring any architectural change of the teacher or making sacrifices in the student performance, MaskedKD dramatically reduces the computations and time required for distilling ViTs. We demonstrate that MaskedKD can save up to $50\%$ of the cost of running inference on the teacher model without any performance drop on the student, leading to approximately $28\%$ drop in the teacher and student compute combined.