Hogwild! Inference

A technical essay from an unnamed source argues that the KV cache, typically viewed as a decoding optimization, can serve as an active execution state enabling pretrained large language models to obse…