Worked on stabilizing GPU-accelerated deep learning workloads, focusing on reliability and performance optimization across ROCm/aiter and ping1jing2/sglang repositories. Addressed kernel-level crashes by reverting problematic Triton GEMM kernel configurations and tuning block size and stage parameters, reducing core dumps and improving GEMM workload stability. In ping1jing2/sglang, fixed a critical crash in MTP FP4/FP8 dispatch and introduced environment variable-based configurability for NEXTN dispatch, enabling safer experimentation and robust fallback behavior. Demonstrated proficiency in Python development, configuration management, and GPU programming, with disciplined version control and clear documentation practices, contributing to more resilient and maintainable machine learning infrastructure.