Staff ML Engineer, Generative Model Performance & Efficiency

Waymo is an autonomous driving technology company with the mission to be the world's most trusted driver. Since its start as the Google Self-Driving Car Project in 2009, Waymo has focused on building the Waymo Driver—The World's Most Experienced Driver™—to improve access to mobility while saving thousands of lives now lost to traffic crashes. The Waymo Driver powers Waymo’s fully autonomous ride-hail service and can also be applied to a range of vehicle platforms and product use cases. The Waymo Driver has provided over ten million rider-only trips, enabled by its experience autonomously driving over 100 million miles on public roads and tens of billions in simulation across 15+ U.S. states.

The Simulator Team at Waymo builds state-of-the-art simulations of realistic environments for testing, training, and validation of the Waymo Driver. Our team is a diverse, and collaborative group of machine learning (ML) engineers, software engineers, and ML research engineers. We develop industry-leading simulation solutions using advanced generative and reconstructive ML algorithms, to model the real world, encompassing realistic agents, roads, traffic systems, weather, and the full sensor suite (Camera, Lidar, Radar).

To accelerate the fidelity, scalability, controllability, and richness of our simulations, we are pushing the frontiers of 3D world modeling. We leverage state-of-the-art ML technologies trained on large-scale datasets to create dynamic and semantically rich virtual worlds, directly impacting the development and validation of the Waymo Driver.

In this role, you will report to a Senior Staff Engineering Manager

You will:

• Analyze model architectures and identify bottlenecks in training and inference performance (e.g., memory bandwidth, compute, communication).
• Apply and develop techniques such as quantization (e.g., FP8, INT4), pruning, knowledge distillation, and efficient attention mechanisms.
• Optimize model code for specific hardware accelerators (TPUs, GPUs), leveraging compiler features and low-level libraries (e.g., XLA).
• Experiment with different model partitioning and sharding strategies (e.g., data, tensor, pipeline parallelism, expert parallelism) to improve scalability and efficiency.
• Design and implement low-latency, high-throughput serving solutions for generative models and optimize training pipelines to reduce training time.
• Build and maintain tools for performance analysis, profiling (e.g., xprof), and debugging of ML models.

You have:

• MS or PhD in Computer Science, Machine Learning, Robotics, or a related field.
• 5+ years of experience with deep learning architectures (especially Transformers, Diffusion Models, MoEs), algorithms, and optimization techniques.
• Proficiency in JAX, Flax, and potentially TensorFlow/PyTorch.
• Expertise in using profiling tools (e.g., XProf, Perfetto, NVIDIA Nsight) to diagnose performance issues in ML workloads.
• Hands-on experience with quantization, pruning, distillation, and other model compression methods.
• Strong programming skills in Python and potentially C++, with experience in software development best practices.

We prefer:

• Knowledge of TPU and GPU architectures and how to optimize code for them.
• Familiarity with ML compilers like XLA and an understanding of how they translate high-level code to efficient hardware instructions.
• Understanding of concepts related to training and serving models across multiple devices and machines.
• Experience contributing to frameworks and libraries that improve training speed and scalability (e.g., JAX, Gemax, XManager).