Worked on performance optimization and robustness improvements in deep learning workflows, focusing on FP8 workloads and tensor operations. Delivered an FP8-optimized matrix multiplication kernel within the TritonBench repository, extending auto-tuning capabilities to support various block sizes and hardware-specific parameters using Python and Triton. Enhanced the pytorch/ao repository by refining TAO operation lowering and improving tensor type handling for CutlassSemiSparseTensor, addressing both data type conversions and quantized tensor implementations. Fixed a shape validation bug for FP8 tensors, ensuring correct dimension handling and edge case coverage. The work demonstrated depth in benchmarking, GPU computing, and matrix multiplication using C++ and Python.