Last released Apr 28, 2026
FADE: Frequency-Adaptive Decay Encoding — attention-aware tiered KV cache compression for LLM inference.
Supported by