DuplexCadence: Exact State and Execution from a Speech Model's Declared Timelines
DuplexCadence is proposed, which explicitly declares native clocks to the runtime and derives two mutually enabling rules: demand-sized state allocation at a stable address, and exact-shape graph replay without padding, which eliminates idle memory and stabilizes tensor pointers.
Hai-Xiao Gao, Yi-Min Zheng, Lin-You Xiao et al.
· 0 citations