The company builds systems that train and run image, video, and world models at high performance.
The work centers on raising speed and efficiency across GPU kernels, distributed engines, and multi-node systems.
Responsibilities
Tune system and GPU speed for both training and inference jobs.
Use Nsight and similar tools to find slow spots in kernels, memory, systems, and clusters.
Write low-level speedups in CUDA and Triton.
Build and enhance distributed training and inference engines used for diffusion models.
Make communication faster between GPUs and nodes using NCCL, RDMA, InfiniBand, RoCE, and disaggregated serving.
Create benchmark and regression tooling so performance gains hold in production.
Shape kernels, runtimes, and models together around what the hardware can do.
Requirements
You have 2 or more years working on deep learning training or inference systems, distributed systems, or similar ML systems.
You have at least 1 year of recent, hands-on work on ML systems or GPU performance.
You have built core features in an inference or training framework such as vLLM, SGLang, TensorRT-LLM, or Megatron, not only deployed or integrated them.
You have written CUDA, CUTLASS, Triton, or PTX kernels directly on NVIDIA GPUs.
You have a strong grounding in machine learning and computer science, with a degree in computer science or a related quantitative field.
You have improved inference or training performance across GPU kernels, runtimes, or distributed execution. Experience with diffusion, video, image, or multimodal workloads is relevant.
Conditions
Base salary runs from $180,000 to $260,000 per year.
Visa sponsorship is available.
The role is on-site in Menlo Park, California, United States.