Worked on the intel-xpu-backend-for-triton repository, focusing on optimizing RDNA kernel configurations to enhance GPT-OSS throughput on AMD GPUs. Developed and implemented a matmul_ogs configuration that reduces register spilling, specifically targeting large inference workloads with batch sizes greater than 512. This optimization led to measurable throughput improvements for GPT-OSS 20B on AMD Navi GPUs, increasing requests per second across multiple hardware configurations. The work involved GPU programming, performance optimization, and rigorous testing, all conducted in Python. The approach emphasized practical performance gains for real-world batch sizes, addressing bottlenecks in end-to-end inference and contributing to more efficient model deployment.