Performance Engineer; Inference, Training & GPU

About World Labs:

We build foundational world models that can perceive, generate, reason, and interact with the 3D world — unlocking AI's full potential through spatial intelligence by transforming seeing into doing, perceiving into reasoning, and imagining into creating. We believe spatial intelligence will unlock new forms of storytelling, creativity, design, simulation, and immersive experiences across both virtual and physical worlds. We bring together a world‑class team, united by a shared curiosity, passion, and deep backgrounds in technology — from AI research to systems engineering to product design — creating a tight feedback loop between our cutting‑edge research and products that empower our users.

We are looking for a Performance Engineer to make World Labs’ models train and serve as fast as the hardware allows.

Running large generative world models at scale is a novel systems problem. You will find the bottlenecks — in kernels, in the serving path, in the training loop, in how we use our GPUs — and eliminate them. Your ownership is technical and concrete: the throughput you unlock, the latency you cut, the utilization you win back, and the correctness you hold while doing it.

You will work up and down the stack, from low‑level tensor and kernel optimization to fleet‑wide serving efficiency, in close partnership with the researchers whose models you are accelerating.

This is a hands‑on, individual‑contributor role. You will profile, design, build, and ship code directly.

• Optimize inference and serving end to end — latency, throughput, batching, caching, and scheduling — to serve our models efficiently at production scale.
• Write and tune GPU kernels (CUDA, Triton) for hot paths; drive kernel fusion, memory- and bandwidth-bound optimization, and low-precision (FP8/INT8) execution.
• Optimize training throughput and GPU utilization: parallelism strategies, communication/compute overlap, mixed precision, and eliminating pipeline stalls.
• Build performance models, profiling workflows, and observability that make throughput, latency, cost, utilization, and their tradeoffs legible across the stack.
• Own numerical correctness across precision, kernel, and hardware changes — treating correctness as part of performance, not separate from it.
• Partner with researchers to product ionize models for serving and to make experiments run faster and more reliably.
• Where needed, work on the distributed systems that training and inference run on — but the core of the job is squeezing the most out of every GPU.

You should excel at the fundamentals below — we index on inference, serving, GPU optimization, and training performance. Distributed‑systems breadth is welcome, but secondary.

• Strong performance-engineering foundations: profiling, roofline analysis, latency/throughput optimization, and disciplined root-cause investigation.
• Deep GPU programming and optimization experience (CUDA and/or Triton) — kernel-level tuning, memory hierarchy, and bandwidth optimization at scale.
• Hands‑on experience optimizing inference and serving for large models: batching, KV/prompt caching, quantization, and low‑latency, high‑throughput sampling.
• Hands‑on experience optimizing training performance: parallelism, distributed communication, mixed/low precision, and utilization.
• Working knowledge of ML framework internals (PyTorch and/or JAX; torch.compile, XLA, or similar compiler paths).
• Strong proficiency in Python, with the ability to drop into C++/CUDA (and Rust or Go) as the work demands.
• High‑ownership mindset — you measure yourself by throughput shipped and latency cut, not tickets closed.

• Experience at an AI lab or ML-native company, optimizing systems used directly by researchers and product ionizing research code.
• Low‑precision and numerics depth: FP8/INT8 quantization, mixed-precision, and detecting numerical regressions across hardware platforms.
• Distributed systems for large-scale training and inference — collective communication (NCCL), interconnects (NVLink), model and tensor parallelism, and fault tolerance. A strong plus, but not a substitute for the core…