During April 2025, contributed to the ROCm/FBGEMM repository by enabling support for the permute multi-embedding function in Torch export, specifically targeting LPV embeddings. This work involved registering the function for graph mode lowering and implementing an FP16 reference kernel to accelerate LPV embedding processing. Leveraging C++, PyTorch, and deep learning expertise, the changes allowed LPV models to utilize enhanced embedding workloads, resulting in improved inference throughput. The technical approach focused on expanding embedding layer capabilities without introducing new bugs, demonstrating a strong understanding of embedding architectures and performance optimization within machine learning frameworks using both C++ and Python.