Developed TurboQuant ROCm backend support and ensured flash attention compatibility for the jeejeelee/vllm repository, expanding TurboQuant’s functionality to ROCm-enabled GPUs. The work focused on backend development and GPU programming using Python, addressing integer overflow issues in the ROCm path to stabilize numeric computations and prevent edge-case failures. By routing backend operations and integrating flash attention mechanisms, the implementation improved throughput and reliability for TurboQuant workloads on ROCm hardware. This feature reduced production risk for ROCm deployments and consolidated changes in a focused commit, demonstrating depth in deep learning and machine learning engineering within a complex, hardware-accelerated environment.