[fix]修复eagle 创建cu_num_tokens类型错误问题 See merge request dcutoolkit/deeplearing/vllm!203
update the default values of VLLM_USE_TRITON_CAT and VLLM_USE_LIGHT_OP to True
fix: w4a8 marlin 中 weight重排接入lightop算子 See merge request dcutoolkit/deeplearing/vllm!202
fix: 优化w4a8 marlin 中 weight重排耗时 See merge request dcutoolkit/deeplearing/vllm!200
V0.9.2 dev fth See merge request dcutoolkit/deeplearing/vllm!199
fix precision issue in mtp See merge request dcutoolkit/deeplearing/vllm!198
fix bugs in zero overhead and tbo See merge request dcutoolkit/deeplearing/vllm!197