Worked on GPU infrastructure for jeejeelee/vllm and kvcache-ai/Mooncake, focusing on hardware compatibility and benchmarking. In jeejeelee/vllm, implemented a GPU count validation feature to ensure disaggregated prefill tasks run reliably on AMD Instinct GPUs, reducing runtime errors and supporting hardware-agnostic execution. For kvcache-ai/Mooncake, contributed to enabling AMD CDNA4 GPU support by integrating ROCm and HIP, and enhanced benchmarking tools to improve performance measurement and CI visibility for AMD workloads. The work involved C++, Shell scripting, and GPU management, with an emphasis on robust platform support and accurate benchmarking for evolving GPU architectures in production environments.