Worked on backend development and deep learning model reliability across the sglang repositories, focusing on critical bug fixes rather than feature additions. Addressed a key issue in JustinTong0323/sglang by correcting the initialization logic for DP attention and tensor parallelism in the BailingMoEModel, ensuring that word embeddings are properly configured based on the enable_dp_attention flag. In kvcache-ai/sglang, refactored the Mambaish model’s key-value cache logic to accurately count only layers within specified boundaries, improving inference stability. Leveraged Python and distributed systems expertise to enhance model implementation, prioritizing code traceability, reproducibility, and robust training workflows in complex environments.