Magic
Member of Technical Staff, Kernels
$275K–$550K
San Francisco +1 · On-site · Staff · Full-time · Posted Jan 24, 2024
- Location
- San Francisco
Remote - Visa sponsorship
- Visa sponsorship is available, if possible.
About the job
Designs, implements, and maintains high-performance kernels to improve throughput and latency for training and inference, especially for long-context workloads.
What you'll do
- Design, implement, deploy, and maintain kernels, with responsibility for production reliability and correctness.
- Optimize kernel performance for long-context behavior, including memory utilization, data movement, and sustained throughput.
- Evaluate porting kernels to alternative hardware and co-design with training, inference, and RL teams.
What they ask for
- Low-level programming experience targeting AI accelerators such as NVIDIA Blackwell or Google TPUs.
- Experience developing and optimizing GPU kernels using frameworks such as NCCL, MSCCLPP, CUTLASS, CuTeDSL, Triton, Quack, or Flash-Attention.
- Deep expertise in computer architecture, low-level machine optimizations, and code generation.
Nice to have
- Experience with Pallas, Mosaic, or Mojo.
- Breadth across machine learning.
Pay
$275K - $550K USD base salary
Tools
NVIDIA Blackwell, Google TPUs, NCCL, MSCCLPP, CUTLASS, CuTeDSL, Triton, Flash-Attention
AI focus
Kernel optimization supports training and inference for Magic’s long-context models and its broader effort to build safe AGI.
Model-written summary · the original posting is authoritative