About the role:

Token Factory is a part of Nebius Cloud, one of the world’s largest GPU clouds, running tens of thousands of GPUs. We are building an inference platform that makes every kind of foundation model — text, vision, audio, and emerging multimodal architectures — fast, reliable, and effortless to deploy at massive scale.

Responsibilities:

• Develop and optimize low-level kernels and runtime components for AI inference

• Improve performance of inference engines GPU platforms

• Profile and debug system-level and hardware-level performance issues

• Integrate support for new hardware architectures (Hopper, Blackwell, Rubin)

• Collaborate with ML and backend teams to optimize end-to-end execution

Required Qualifications:

• Strong proficiency in C++, OR expertise in GPU programming with a focus on low-level high-performance coding and memory management

• Experience in GPU programming or systems-level software development, e.g. operating system internals, kernel modules, or device drivers

• Hands-on experience with profiling and debugging tools to identify performance issues on both CPUs and GPUs, and the ability to optimize code based on those findings.

• Solid understanding of CPU/GPU architecture and memory hierarchy

Preferred Qualifications:

• Experience with GPU computing programming: CUDA, ROCm, CUTLASS, Cute, ThunderKittens, Triton, Pallas, Mosaic GPU

• Familiarity with ML inference runtimes (e.g. TensorRT, TVM)

• Knowledge of Linux internals, drivers, or compiler toolchains

• Experience with tools like perf, VTune, Nsight, or ROCm profiler

• Familiarity with popular inference engines (e.g. such as vLLM, sglang, TGI)

We conduct coding interviews as part of the process.