Worked on ROCm/aiter and sgl-project/sglang repositories to deliver new features focused on GPU performance and deployment reliability. Developed DSv4-Flash shape support for FMoE, enabling efficient handling of expanded token buckets and improving runtime efficiency for Flash MoE calls. Enhanced ROCm decoding by optimizing FP8 scale handling, removing redundant operations to boost performance. Automated nightly testing for Miles on ROCm using GitHub Actions, Docker, and Bash, streamlining CI/CD and reducing manual effort. Leveraged Python, CUDA programming, and PyTorch to validate performance on MI350X GPUs, ensuring robust deployment paths and efficient data processing for deep learning workloads.