Worked on PyTorch integration within the FBGEMM repository, focusing on enhancing meta-device compatibility and inference-mode robustness. Developed fake tensor support for the fbgemm::all_to_one_device operator, enabling it to return correctly shaped and typed empty tensors on the meta device, which improves library integration and deployment workflows. Addressed a bug in jagged_to_padded_dense by registering it with CompositeImplicitAutograd, ensuring accurate behavior when Autograd is disabled during inference. Utilized C++, Python, and GPU computing skills to deliver these improvements, which increased reliability for shape and type inference and streamlined the deployment process for PyTorch-based machine learning models.