audio prefix cache

LLM prefix/KV caching

Both exploit a stable, repeated prefix to skip recomputation; the audio version operates on rendered output rather than attention state.

1 source videos