Minjang worked on the luanfujun/triton repository to enhance GPU benchmarking by making it device-independent. He refactored the do_bench function, moving cache creation logic from the host to the GPU driver backends for both Nvidia and AMD devices. This approach allowed empty cache allocation to be managed directly by the drivers, reducing host-side variability and improving the consistency of benchmarking results across different hardware. Using Cuda and Python, Minjang focused on backend development, GPU computing, and performance optimization. His work improved the reproducibility and comparability of benchmarks, enabling more reliable performance analysis and fairer cross-device evaluations.