vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.common.embedding

Local CuTe vocabulary gather plus the ordinary framework TP all-reduce.