Research Scientist (Generative Modeling)

Advanced Micro Devices, Inc•San Francisco, CA

About The Position

The role is described as a hands-on individual-contributor position focused on improving the performance of large generative world models in training and production serving. The engineer will work across the stack, from GPU kernels and tensor optimization to fleet-wide serving efficiency, in close partnership with researchers. Success is measured through increased throughput, reduced latency, improved utilization, and preserved correctness. The posting identifies the position as being at World Labs, while the surrounding employer information refers to AMD.

Requirements

  • Strong performance-engineering foundations, including profiling, roofline analysis, latency/throughput optimization, and root-cause investigation.
  • Deep GPU programming and optimization experience with CUDA and/or Triton, including kernel tuning, memory hierarchy, and bandwidth optimization at scale.
  • Hands-on experience optimizing inference and serving for large models, including batching, KV/prompt caching, quantization, and low-latency, high-throughput sampling.
  • Hands-on experience optimizing training performance, including parallelism, distributed communication, mixed/low precision, and utilization.
  • Working knowledge of ML framework internals such as PyTorch and/or JAX, and compiler paths such as torch.compile or XLA.
  • Strong proficiency in Python and ability to work in C++/CUDA, with Rust or Go as the work demands.
  • High-ownership mindset focused on shipped throughput and reduced latency.

Nice To Haves

  • Experience at an AI lab or ML-native company, optimizing systems used directly by researchers and productionizing research code.
  • Depth in low-precision numerics, including FP8/INT8 quantization, mixed precision, and detecting numerical regressions across hardware platforms.
  • Distributed-systems experience for large-scale training and inference, including NCCL, NVLink, model/tensor parallelism, and fault tolerance.
  • Experience serving generative, diffusion, video, or 3D/spatial models.
  • Experience with multiple accelerator types, such as GPU plus TPU or Trainium, and partnering with hardware vendors.
  • Experience building performance-modeling and observability frameworks for GPU utilization and cost.

Responsibilities

  • Optimize inference and serving end to end, including latency, throughput, batching, caching, and scheduling.
  • Write and tune GPU kernels using CUDA and/or Triton, including kernel fusion, memory- and bandwidth-bound optimization, and low-precision execution.
  • Improve training throughput and GPU utilization through parallelism strategies, communication/compute overlap, mixed precision, and pipeline-stall reduction.
  • Build performance models, profiling workflows, and observability for throughput, latency, cost, utilization, and related tradeoffs.
  • Maintain numerical correctness across precision, kernel, and hardware changes.
  • Partner with researchers to productionize models for serving and make experiments faster and more reliable.
  • Work on distributed systems for training and inference when needed, while focusing primarily on GPU performance.

Benefits

  • AMD benefits are referenced, but no specific benefits are listed in the posting.
© 2026 Teal Labs, Inc
Privacy PolicyTerms of Service