v1.5.0

1.5.0 (2026-04-08)

Features

• kv-cache: bypass compression on full attention layers in sliding window models (#85) (86daf48), closes #77
• kv-cache: per-layer head_dim support for heterogeneous architectures (#83) (795f904), closes #75

Bug Fixes

• init: warn when transformers<5.5 — Gemma 3/4 require newer version (#80) (8c75ba6), closes #78
• verify: per-layer head_dim in verification harness (#84) (e585680), closes #73
• vllm: padded slot size for hybrid model page alignment (#81) (899f714), closes #76