Worked on model optimization and backend improvements for large language models, focusing on quantization and precision alignment. In the vllm-project/llm-compressor repository, developed a data-free input embedding quantization workflow using Python and PyTorch, enabling memory-efficient deployment of large-vocabulary models without requiring calibration data. Demonstrated near-lossless accuracy across multiple architectures and ensured compatibility with vLLM’s Embedding module. In jeejeelee/vllm, implemented FP32 precision support for the language model head, aligning training and inference for generation models and supporting reinforcement learning workflows. Added comprehensive validation tests and collaborated with cross-team reviewers to ensure production readiness and reproducibility in model generation tasks.