Contributed to deep learning infrastructure by enabling a new MORI execution path for unquantized mixture of experts models in the jeejeelee/vllm repository, supporting the AITER backend to dispatch raw BF16 and FP16 data without scale adjustments. This work involved updating quantization and dispatch logic in Python, enhancing model flexibility and performance potential. Additionally, improved developer guidance for vLLM on AMD ROCm platforms by co-authoring documentation updates in reStructuredText, detailing new AITER attention backends and performance variables for AMD Instinct GPUs. The contributions focused on AI model optimization, GPU programming, and performance tuning, supporting production adoption and streamlined workflows.