Skip to main content

prepare_hybrid_decode

Function prepare_hybrid_decode 

Source
pub fn prepare_hybrid_decode(prompt: &str)
Expand description

Prepare hybrid hints before a decode turn (safe no-ops when inactive). Call after any default publish_query_tensor so route/query are not wiped. FastVerify skips mid-decode hybrid work (post-turn verify owns quality).

Do not enable route masks under mode=cuda alone. That used to OR in prefer_tensor_core_gemm(), which published random prompt-hash KV masks and destroyed attention (measured garbage decode 2026-07-24: repeated Ċ tokens). Hybrid routing is for quant-graph / explicit QUALIA_HYBRID_ROUTE=1 only.