Worked on the FlagOpen/FlagGems repository to optimize the polar operator for hardware-accelerated inference, focusing on compatibility with Ascend NPU and robust handling of bfloat16 data. Refactored the operator to use separate contiguous tensors for real and imaginary components, addressing hardware constraints and enabling a CPU round-trip for complex result assembly. Strengthened dtype reliability by explicitly casting bfloat16 to float32 before processing, ensuring numerical stability. Validated the solution across fp16, fp32, and bf16 with comprehensive accuracy tests, including edge cases. Utilized Python, PyTorch, and TensorFlow to deliver a feature that enhances cross-dtype reliability and hardware compatibility in deep learning workflows.