Worked on backend development for the ggml-org/llama.cpp repository, focusing on improving stability and reliability in SWA-only model workflows. Addressed a critical bug in the graph input path by implementing robust null-buffer checks for input tensors, preventing crashes when no non-SWA layers are present. Enhanced the set_input logic for both base and SWA paths and updated the can_reuse function to avoid dereferencing unallocated tensors, aligning with established memory guard patterns. Utilized C++ to deliver these improvements, which reduced production crashes and strengthened the safety of deploying SWA slices, particularly benefiting models such as Gemma 4.