Worked on performance and reliability updates for the NVIDIA/warp repository, focusing on backend development and memory management using C++, CUDA, and Python. Addressed memory safety in Warp kernel computations by fixing out-of-bounds reads and optimizing shared memory planning, which improved stability and correctness in both forward and backward training paths. Enhanced tile_matmul by resolving data corruption issues with non-dense, strided operands and introducing robust fallbacks for toolchain edge cases. Improved gradient registration reliability and expanded unit testing to cover negative indices and memory usage scenarios, resulting in more deterministic results and smoother CI across CPU and CUDA execution paths.