ul-smf-cache-compressionliventruth
ADDEDUL-SMF is a KV cache compression framework for long-context Transformer inference, claiming up to 384x compression via FSQ and 16D latent mapping. The core compression logic requires a proprietary closed-source binary.