Worked on the flash-linear-attention repository to enhance reliability for large-batch and large-sequence workloads on Blackwell GPUs. Focused on GPU programming and numerical computing using Python, addressing a critical bug involving large-offset pointer arithmetic by standardizing offsets to int64 to prevent overflow. Expanded the synthetic test suite to cover conv, norm, and gated delta kernel paths, and introduced robust regression tests to guard against overflow scenarios. Improved maintainability by refining test guards and extending coverage to 16k GDN kernels and fused norm paths, resulting in more reliable hardware-specific guardrails and enabling larger models with greater confidence across teams.