Software Engineer, GPU Kernels
Builds performance-critical GPU kernels (attention, matrix multiplication, mixture-of-experts, low-precision FP8/FP4 compute) for River AI's training and inference infrastructure. Day to day: profiling workloads, writing kernels with CUDA/Triton/CUTLASS in C++ and Python, and integrating optimizations into training and serving runtimes.
- Compensation
- $200K–$420K / yr
- Location
- Palo Alto, CA
- Work setup
- Not specified
- Seniority
- Not specified
- Employment
- Not specified
- Experience
- Not specified
- Category
- Software engineering
- Posted
Skills
Must have
- Bachelor's degree in Computer Science, Computer Engineering, or equivalent practical experience
- Experience optimizing GPU kernels with CUDA, Triton, CUTLASS, CuTe, or comparable tools
- Strong understanding of GPU architecture, memory hierarchies, and parallel execution
- Proficiency in C++ and Python
- Strong foundations in linear algebra, floating-point arithmetic, and numerical computing
- Strong debugging and profiling skills, with a collaborative approach to engineering
Nice to have
- Experience optimizing for NVIDIA Blackwell or Hopper GPUs
- Work on attention, mixture-of-experts kernels, grouped GEMMs, or low-rank adapters
- Experience implementing backward passes and validating gradients
- Familiarity with FP8, FP4, and quantized weight layouts
- Experience integrating custom operators into PyTorch, SGLang, vLLM, or similar frameworks
- Open-source contributions or a track record of shipping substantial kernel optimizations
Full description
At River AI, our mission is to create personal AI owned and shaped by each individual. To achieve this, we are rewriting the entire stack from scratch: personal hardware for local inference, bespoke training infrastructure, next-generation UIs, and frontier deep learning research.
Who we are
We are scientists, engineers, and builders from the industry's top tech companies and AI labs. We bring a proven track record of scaling consumer systems for hundreds of millions of users and architecting the pre-training infrastructure behind today's frontier models.
About the Role
We are looking for exceptional GPU kernel engineers to build the compute primitives behind River’s training and inference infrastructure. Your goal is to make large models faster to train and more efficient to serve.
You will own performance-critical operations, including attention, matrix multiplication, mixture-of-experts execution, and low-precision computation. Working closely with researchers and systems engineers, you will identify bottlenecks, implement kernels, validate correctness, and bring improvements into production.
What You’ll Do
- Build fast GPU kernels for attention, matrix multiplication, expert routing, and related operations.
- Optimize memory access, tiling, and synchronization to make efficient use of GPU hardware.
- Develop FP8, FP4, and mixed-precision kernels while preserving numerical correctness.
- Accelerate fine-tuning and RL through optimized adapters, backward passes, and fused operations.
- Profile real workloads and integrate improvements into training and inference runtimes.
- Build reproducible benchmarks that verify correctness, gradients, and performance.
Skills & QualificationsMinimum Qualifications:
- Bachelor’s degree in Computer Science, Computer Engineering, or equivalent practical experience.
- Experience optimizing GPU kernels with CUDA, Triton, CUTLASS, CuTe, or comparable tools.
- Strong understanding of GPU architecture, memory hierarchies, and parallel execution.
- Proficiency in C++ and Python.
- Strong foundations in linear algebra, floating-point arithmetic, and numerical computing.
- Strong debugging and profiling skills, with a collaborative approach to engineering.
Preferred Qualifications: (We encourage you to apply even if you don't meet all of these)
- Experience optimizing for NVIDIA Blackwell or Hopper GPUs.
- Work on attention, mixture-of-experts kernels, grouped GEMMs, or low-rank adapters.