Worked on enhancing the kvcache-ai/sglang repository by implementing low-bit quantization support for neural processing unit (NPU) frameworks. Focused on enabling w4a8 quantization with activation-aware clipping, the work introduced robust initialization and processing paths for weights, accommodating both clipped and unclipped activations. This approach allows for more efficient inference on NPUs by reducing bit-width while maintaining model accuracy. Leveraging deep learning and machine learning expertise, the solution was developed in Python and centered on quantization techniques. The contribution addressed the need for flexible quantization workflows, supporting advanced hardware acceleration and improving the adaptability of the NPU framework.