Member of Technical Staff, Performance & GPU Kernels

ATBF Labs builds the inference engine for production AI. Every token a model serves in production runs through an inference stack, and that stack decides the latency, the cost, and the reliability of the product sitting on top of it. We build ours from first principles: custom GPU kernels, a purpose-built runtime, and a distributed serving layer that holds its tail latency under real load. We are a small team with a high bar, shipping to production from day one.

Key responsibilities

• Write and optimize GPU kernels (CUDA, Triton, CUTLASS) for attention, GEMMs, and quantized paths.
• Profile with Nsight / CUPTI to find kernel- and memory-level bottlenecks, then close them.
• Implement mixed-precision and quantization schemes (FP8, INT4, and beyond) with measured quality bounds.
• Optimize communication for multi-GPU and multi-node serving over NVLink and RDMA (InfiniBand, RoCE).
• Co-design model execution with the runtime and research teams for hardware efficiency.
• Debug numerical instabilities that only appear at scale, on a fraction of requests.

• Deep understanding of GPU architecture, parallel programming, and the memory hierarchy.
• Proficiency in CUDA (or ROCm) and a GPU profiler (Nsight, nvprof, CUPTI).
• Experience with performance-critical model execution and PyTorch internals.
• A track record of measurable speedups, not just refactors.

• Experience with ML compilers (torch.compile, Triton, XLA, TVM).
• Experience optimizing LLMs, VLMs, or video models for inference.
• Familiarity with low-precision numerics and quantization-aware tradeoffs.
• Contributions to open-source HPC or ML-systems projects.

• Implement a fused attention kernel for a novel variant and beat the reference by a measurable margin.
• Ship an FP8 path and run experiments to find the optimal speed/quality tradeoff.
• Optimize RDMA communication patterns to remove a stall in multi-node decode.