Performance Engineer (Inference, Training & GPU)

Highlights

• Make large generative world models train and serve as fast as hardware allows.
• Own throughput unlocked, latency cut, and utilization won back end to end.
• Help power Marble turning text, images, and video into navigable 3D worlds.

What they offer

• Equity
• Health, dental and vision benefits
• 401K benefits
• Flexible PTO
• Paid parental leave
• Relocation support

Ideal candidate

A high-ownership performance obsessive who lives in profilers and loves squeezing every GPU. You partner closely with researchers and measure yourself by shipped throughput and latency wins while holding correctness.

Requirements

• Strong performance-engineering foundations including profiling and roofline analysis
• Deep GPU programming and optimization experience with CUDA and/or Triton
• Experience optimizing inference and serving for large models
• Experience optimizing training performance and GPU utilization
• Working knowledge of PyTorch and/or JAX internals including torch.compile or XLA
• Strong Python proficiency with ability to work in C++/CUDA