Worked on the flash-linear-attention repository to enhance stability and memory safety for CUDA kernels targeting high-SM GPUs. Focused on resolving illegal memory access in the LayerNorm backward kernel by separating tensor dimensions from per-program work bounds and introducing robust grid-size safeguards. Used Python and CUDA to implement fixes that prevent out-of-bounds access, regardless of hardware configuration. Expanded regression test coverage to include a range of token counts and parameter sizes, improving reliability for deep learning workloads. Collaborated closely with other contributors, emphasizing clear commit messages and comprehensive testing to reduce production risk and ensure long-term maintainability of GPU programming code.