Worked on enhancing deep learning infrastructure by improving benchmarking and inference stability in open-source repositories. In kvcache-ai/sglang, added missing arguments to the speculative decoding benchmark using Python scripting, which increased measurement accuracy and enabled more reliable performance tuning. Addressed a bug in the same benchmark to stabilize results and prevent failures. In sgl-project/mini-sglang, fixed a reshaping issue in the LM head all-gather path for multi-batch inputs, leveraging CUDA and PyTorch to ensure inference reliability for batch sizes greater than one. Maintained code quality by aligning with repository standards and ensuring continuous integration checks passed after each change.