The change extends the pluggable layer framework to support custom encoder implementations, specifically for visual‑encoding models like Qwen2. It allows runtime substitution of the decoder without ...