Quick overview
Job description
About the Role
Join an AI systems team rebuilding the training and inference stack for image, video, and world-model workloads. You will improve performance across GPU kernels, distributed engines, and multi-node systems, helping make demanding model workloads faster and more efficient.
What You'll Do
Optimize training and inference performance for image, video, and world-model workloads.
Profile bottlenecks at the kernel, memory, system, and cluster levels using Nsight and related tools.
Implement low-level GPU optimizations in CUDA and Triton.
Design and improve distributed training and inference engines for diffusion models.
Improve communication performance across GPUs and nodes using NCCL, RDMA, InfiniBand, RoCE, and disaggregated serving approaches.
Build benchmarks and regression tests to preserve performance gains in production.
Co-design kernels, runtimes, and models with hardware characteristics in mind.
What We're Looking For
At least 2 years of experience in deep learning training or inference systems, distributed systems, or related ML systems work, with 1 or more years of hands-on GPU performance experience in a recent role.
Experience authoring core features in an inference or training framework, such as vLLM, SGLang, TensorRT-LLM, Megatron, or a comparable system.
Direct experience writing GPU kernels for NVIDIA hardware using CUDA, CUTLASS, Triton, or PTX.
Strong machine learning and computer science fundamentals, supported by a degree in computer science or a related quantitative field.
Experience optimizing training or inference performance across GPU kernels, runtimes, or distributed execution.
Experience with PyTorch, Nsight, NCCL, and RDMA, and ideally optimization of diffusion, video, image, or other multimodal workloads.
Compensation & Benefits
Base salary: $180,000 to $260,000 annually. Visa sponsorship is available.
Location
On-site in Menlo Park, California, United States.