### What this PR does / why we need it? While using the LLM Compressor quantization tool from the VLLM community to generate quantized weights, the VLLM Ascend engine needs to be adapted to support the compressed tensors quantization format. 1. Support Moe model W4A8 dynamic weight. - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/bde38c11df0ea066a740efe9b77fff5418be45df --------- Signed-off-by: LHXuuu <scut_xlh@163.com> Signed-off-by: menogrey <1299267905@qq.com> Co-authored-by: menogrey <1299267905@qq.com>