Staff/Sr. ML Compute Efficiency Engineer

Scaling machine learning workloads across thousands of GPUs and TPUs creates challenges that few engineers ever encounter. In Apple's Machine Learning Platform Technologies organization, we build the infrastructure that powers large-scale ML training and inference workloads, bringing together expertise in distributed systems, machine learning infrastructure, and high-performance computing. As a performance engineer in the ML Compute Efficiency team, you'll tackle ambiguous systems challenges, identify inefficiencies and build solutions that maximize accelerator utilization, reduce idle and fragmented capacity, and minimize recovery periods. This includes analyzing accelerator performance, digging into various parallelism techniques, and refining workload scheduling and orchestration across the compute fleet.Characterize ML workload behavior through profiling, benchmarks and metrics. Dive into unfamiliar codebases to prototype changes, evaluate tradeoffs, and build production-ready solutions. Design systems for efficient recovery from failures and preemptions. Create tools to identify and alert bottlenecks across applications and frameworks. Use workload-driven insights to influence next-generation hardware selection and procurement decisions. Collaborate closely with ML researchers and infrastructure engineers to address inefficiencies. Drive impact through hands-on contribution and mentorship.Experience with large-scale distributed systems for AI/ML workloads running on GPUs or TPUs. Strong software engineering skills with experience developing and optimizing training frameworks (e.g. PyTorch, JAX) using C/C++ or Python. Experience working on cross-functional projects with ML research and infrastructure teams. Familiarity with model architectures and various training techniques. Bachelor's degree in Computer Science or equivalent experience, with 7+ years of industry experience.Have a track record of delivering transformative performance improvements on large scale infrastructure. Ability to analyze ambiguous, distributed systems problems and articulate both high-level strategic metrics and underlying technical complexity.