Commit History
-
inference: Qwen3.5 support (VL unwrap, Gated-DeltaNet LoRA targets, linear_attn walk)
Verified signaturef91a401
-
Phase A serving speedup: zero-copy state-cache expand (no deepcopy per decision)
Verified signature5c02833 -
Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)
Verified signature5ca406a -
Serving-path speedup: state-KV LRU cache + suffix diet (see inference/bench_serve.py)
Verified signature68ae554
-
Add inference/: minimal self-contained Typical inference package
Verified signature002c12c