Worked on performance optimization for the jeejeelee/vllm repository, focusing on enhancing Mamba SSD chunk kernels to improve inference throughput and reduce latency. The main contribution involved implementing the do_not_specialize directive and refining sequence length handling across core kernel functions, targeting SSD-based deep learning workloads. Leveraged expertise in GPU programming and performance optimization, using Python to deliver kernel-level improvements without introducing major bug fixes during the period. Maintained high code quality through signed-off commits and collaborative development practices. The work addressed the need for efficient, scalable inference by optimizing critical paths in the kernel, supporting higher workload throughput for deep learning applications.