Member of Technical Staff, Kernels

Magic’s mission is to build safe AGI that accelerates humanity’s progress on the world’s most important problems. We believe the most promising path to safe AGI lies in automating research and code generation to improve models and solve alignment more reliably than humans can alone. Our approach combines frontier-scale pre-training, domain-specific RL, ultra-long context, and inference-time compute to achieve this goal.

ABOUT THE ROLE

As a Kernel Engineer, you will design, implement, and maintain high-performance kernels to optimize throughput and latency during training and inference.

Magic's long-context windows create distinct kernel optimization challenges around memory utilization, data movement, and sustained throughput.

WHAT YOU'LL WORK ON

- Design and implement kernels that support high-performance long-context behavior

- Ownership of kernel design, implementation, deployment, and production reliability

- Focus on robustness, extensive testing, and functional correctness, while pushing on performance

- Evaluate porting Magic’s compute kernels to alternative hardware options

- Co-design kernels with understanding and interaction with training, inference, and RL teams

- For a sample of our work, see Magic-Attention, presented at GTC 2026 https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s82294/

WHAT WE’RE LOOKING FOR

- Low-level programming experience targeting AI accelerators such as NVIDIA Blackwell or Google TPUs

- Develop and optimize GPU kernels in frameworks such as NCCL https://developer.nvidia.com/nccl, MSCCLPP https://github.com/microsoft/mscclpp, CUTLASS https://github.com/NVIDIA/cutlass, CuTeDSL https://docs.nvidia.com/cutlass/latest/media/docs/pythonDSL/cute_dsl.html, Triton https://github.com/triton-lang/triton, Quack https://github.com/Dao-AILab/quack, Flash-Attention https://github.com/Dao-AILab/flash-attention, and similar frameworks

- Experience in other kernel authoring frameworks such as Pallas http...