Worked on the ROCm/onnxruntime repository to enhance Windows compatibility for NVIDIA CUDA/cuDNN site packages by implementing a DLL preload mechanism. Developed and verified support for preloading cudnn_engines_tensor_ir64_9.dll, addressing missing cuDNN dependencies that previously caused session failures during model inference. Focused on backend development using Python, with an emphasis on robust DLL management to improve inference reliability and deployment readiness in Windows environments. The solution ensured that workflows relying on CUDA/cuDNN could operate without PyTorch, demonstrating careful attention to Windows-specific dependency handling and patch-level integration for production-grade machine learning deployments within the ROCm/onnxruntime project.