Worked on the vllm-project/llm-compressor repository to deliver robust AWQ quantization support for the GLM-4.7-Flash MoE-lite model. Developed AWQ smoothing mappings for Glm4MoeLiteForCausalLM, enhancing inference stability and prediction quality under quantization. Applied architecture-aware modifications by excluding the MoE router from balance layers, addressing resource grouping bottlenecks in deep models. Validated the deployment path by quantizing with AWQ and W4A16, then running inference in vLLM to ensure coherent outputs. The work demonstrated strong proficiency in deep learning and machine learning, leveraging Python to implement and test quantization strategies for large-scale language model inference.