Software Engineer, ML Training Optimization

Job description

Zoox is on a mission to reimagine transportation and ground-up build autonomous robotaxis that are safe, reliable, clean, and enjoyable for everyone. We are still in the early stages of deploying our robotaxis on public roads, and it is a great time to join Zoox and have a significant impact in executing this mission. The ML Platform team at Zoox plays a crucial role in enabling innovations in large-scale Foundation models, VLMs, and VLAs to make autonomous driving as seamless as possible. **The Opportunity** Are you excited to drive our ML Training Optimization initiatives and make the models that power autonomous driving trains faster, cheaper, and at greater scale? You will work with SOTA accelerators and cutting-edge techniques in distributed training, parallelism strategies, mixed-precision training and RL, especially for VLA and VLM models that are multi-modal — partnering closely with all the Autonomy teams within Zoox (Perception, Prediction, Planner, Simulation, Collision Avoidance) and pushing the boundaries of how ML training is practiced at Zoox. We build and operate the base layer of ML tools, model development, and training systems that our applied research teams use for in- and off-vehicle ML use cases. You will work alongside a team of strong software engineers and act as a force multiplier for our internal customers. This team has many growth opportunities as we expand our robotaxi deployments and venture into new ML domains. If you want to learn more about our stack behind autonomous driving, please look [here](https://zoox.com/autonomy/). If you want to learn more about our ML Infrastructure, here is one of our [past talks](https://www.youtube.com/watch?v=j-LpPv9vBvk) at re:Invent. ### In this role, you will: - Design, implement, and operate cutting-edge ML training performance optimization techniques — distributed training (data/tensor/pipeline parallelism), mixed-precision, gradient checkpointing, and GPU kernels — to scale our VLM, VLA, and Foundational models training efficiency. - Work on RL framework and RL performance optimization for driving foundation model - Profile and debug training throughput bottlenecks across the compute, compiler, framework, and cluster-networking stack. - Collaborate closely with ML researchers, data engineers, and hardware engineers to define requirements and align on architectural decisions for large-scale training infrastructure. - Implement custom kernels and build benchmarks to track and validate training-performance optimizations across the stack. ### Qualifications Note: You do not have to meet all the requirements below to be considered for this position: - 4+ years of total experience, including 2+ years working on large-scale model training platforms. - Experience with distributed training frameworks (Megatron-LM, TorchTitan), efficiently utilizing GPU/accelerator clusters at scale. - Experience with RL framework such as Verl/MILES and RL algorithms(GRPO, PPO) - Experience with collective communication libraries (e.g., NCCL) and optimizing communication/compute overlap in multi-node training. - Experience using profiling tools (NVIDIA Nsight, PyTorch Profiler, or XLA/JAX profiling) to identify and resolve training bottlenecks. - Proficient in Python or C++. - Experience with ML compilers, performance tooling, or hardware/software co-design (e.g., GPU/TPU/accelerator architectures).