Worked on the tensorflow/tensorflow repository to enhance the CPU backend, focusing on OneDnnFusion and OneDnnFusionThunk components. Developed an asynchronous threadpool for OneDnnFusionThunk using C++ to enable concurrent execution of fusion tasks, updating the Invoke method to return AsyncValueRef for improved event management and runtime efficiency. Addressed a critical bug in OneDnnFusion by ensuring unique logical tensor IDs for parameters, which corrected matmul fusion accuracy and performance. The work emphasized asynchronous programming, performance optimization, and rigorous testing, directly improving the stability and efficiency of deep neural network fusion paths for models leveraging OneDnn-enabled fusion.