Worked on stabilizing the Gated DeltaNet kernel within the fla-org/flash-linear-attention repository, focusing on high-end GPU environments. Addressed a critical bug affecting kernel execution on H100 GPUs when the vector dimension was set to 64, ensuring reliable performance across these configurations. Improved kernel correctness by excluding autotuning for num_warps set to 8 specifically on Hopper architectures, which enhanced stability for targeted hardware. Utilized CUDA and Python to implement and validate these changes, applying GPU programming and performance optimization techniques. The work demonstrated careful attention to hardware-specific issues and contributed to the robustness of the kernel in production environments.