The contribution of generation loop changes
Horizontal bars compare generation throughput before and after two Transformers generation loop changes.
Generation loop contribution
All layer kernels enabled in both runs · tokens per second
Before changes
After changes
020406080tok/s
Qwen3.5-4BQ4_K_M · 1.42×Qwen3.8-27BUD-Q4_K_M · 1.16×Qwen3.5-35B-A3BUD-IQ4_XS · 1.79×
49.670.413.715.933.760.2
Changes: remove an unnecessary attention mask early and defer the stopping check.