I implemented a KV cache compression module for Grouped Query Attention (GQA) running on an RTX 5070 (12GB VRAM). Here’s what I found from combining two recent techniques: PackKV (arXiv 2512.24449v2, IPDPS 2026): Token-wise 4-bit quantization with encode-aware repacking. Claims 153.2% higher K cache and 179.6% higher V cache compression vs prior SOTA by packing two int4 tokens into one uint8 byte.