Worked on stabilizing FP4 quantization paths and improving test reliability for the jeejeelee/vllm repository, focusing on the MoE (Mixture of Experts) module. Addressed a critical bug in the MXFP4 fused MoE test by replacing a manual interleave implementation with official library functions, ensuring correct weight and scale layout for SM90 mixed-input GEMM operations. This update, implemented using CUDA, PyTorch, and Python, enhanced the accuracy of FP4 quantization and reduced flaky test results. The work contributed to more reliable MoE-related changes in production workflows, emphasizing robust testing practices and careful integration of library-supported solutions.