Simple, zero overhead way to compress model, KV cache via Low-Rank Decompositionjeffreywong20.github.io1 ポイント·投稿者 thw20·2 か月前·0 コメント
thw20·3 か月前·議論Good work! This is very interesting. Here's a related work that construct low-rank approximation for attention: https://arxiv.org/abs/2505.12942.Maybe the idea of Query calibration matrix Rxx is of interest to the author!
thw20·4 か月前·議論The up to date paper documenting and analysing the observation is now available on arxiv!
thw20·4 か月前·議論This project reveals an interesting phenomena, where LLM converts semantic non-informative tokens to attention sinks through middle layer MLP.The converted sinks are termed secondary attention sinks as they are weaker then BOS attention sinks.This might be related to layer specialisation in LLM!
Maybe the idea of Query calibration matrix Rxx is of interest to the author!