Developed and integrated a native QH64 fp8 persistent decode kernel for the ROCm/aiter repository, targeting the MI300X (gfx942) platform. Leveraging C++, CUDA/HIP, and Python, the work replaced the previous folded QH16 path by introducing optimized kernel binaries and updating the framework for native metadata and batching support. The engineering effort included critical correctness fixes for output writeback, paged-tail processing, and variable-length decoding, ensuring robust handling across diverse batch and context sizes. Extensive validation demonstrated reliable performance, achieving up to 367 TFLOPS. This feature improved throughput and reliability for large-scale inference workloads, enhancing the machine learning infrastructure.