CUDA Engineer

Design and optimize low-level CUDA code powering transformer inference workloads. Build custom CUDA kernels, profile and remove bottlenecks across occupancy, memory throughput, and warp divergence, and apply techniques like kernel fusion, memory-hierarchy tuning, quantization-aware and mixed-precision arithmetic. Implement efficient caching for autoregressive decoding, benchmark against baselines, and maintain internal CUDA libraries with tests to ensure performance and correctness.