Nice, similar point! I agree it works fine for full attention. But for Gated DeltaNet on Qwen3.5, vLLM only supports caching at block boundaries. So we developed a new one. Also, we optimize the CUDA graph usage and can keep answers bit-identical at any load.