About the job
Our models and workflows require performance work that generic frameworks don’t solve. You’ll design and ship custom CUDA kernels, profile at the hardware level, and integrate research ideas into production code that delivers measurable speedups in real pipelines (training, post-training, and inference). Our team is small, fast-moving, and high-ownership. We're looking for someone who finds joy in memory hierarchies, tensor cores, and profiler output.
Responsibilities
Write high-performance GPU kernels for our novel model architectures
Integrate kernels into PyTorch pipelines (custom ops, extensions, dispatch, benchmarking)
Profile and optimize training and inference workflows to eliminate bottlenecks
Build correctness tests and numerics checks
Build/maintain performance benchmarks and guardrails to prevent regressions
Collaborate closely with researchers to turn promising ideas into shipped speedups
Qualifications
Minimum
Authored custom CUDA kernels (not only calling cuDNN/cuBLAS)
Strong understanding of GPU architecture and performance: memory hierarchy, warps, shared memory/register pressure, bandwidth vs compute limits
Proficiency with low-level profiling (Nsight Systems/Compute) and performance methodology
Strong C/C++ skills
Preferred
CUTLASS experience and tensor core utilization strategies
Triton kernel experience and/or PyTorch custom op integration
Experience building benchmark harnesses and perf regression tests