Fathom Optimizes Sparse Decoding for Offloaded KV Caches

Fathom implements per-query read depth for sparse decoding, allowing queries to decide how many bits of a 4-bit K cache to read via reverse water-filling. On a Qwen3-8B model with one million tokens, it is 1.67x faster in GPU time than existing sparsity methods like Double Sparsity.