Worked on enhancing the reliability of large-batch inference in the jeejeelee/vllm repository by addressing a critical bug in the Triton sampler kernels. Using Python, CUDA, and Triton, implemented a fix for an int32 offset overflow by casting row offsets and program IDs to int64, which resolved CUDA memory access errors encountered during speculative decoding with large batch sizes. To ensure ongoing stability, added regression tests targeting large GPU configurations, guarding against future regressions. This work improved the maintainability and scalability of the Model Runner V2 Spec Decode pipeline, reducing production crashes and supporting robust deployment in demanding backend environments.