Meta AI Research is at the forefront of advancing foundational and applied artificial intelligence, developing breakthroughs that power products used by billions of people and shape the future of human-computer interaction. We are seeking a Research Scientist at the Staff level (IC6) with deep expertise in TPU performance optimization, large-scale model training, and systems-level machine learning. In this role, you will lead high-impact research on model efficiency and optimization for first party models within Meta's native PyTorch stack, collaborating across research and engineering teams to drive AI capabilities that define Meta's next generation of products and platforms.

Qualifications:

Bachelor's degree in Computer Science, Computer Engineering, relevant technical field, or equivalent practical experience 8+ years of experience in machine learning systems, model optimization, or high-performance computing research Experience with TPU architecture and performance optimization, including profiling, kernel development, and memory management Experience with XLA compilation, graph optimization, and low-level performance tuning for accelerator hardware Experience developing and optimizing large-scale distributed training systems, including parallelism strategies such as data, tensor, and pipeline parallelism Experience with PyTorch and its integration with accelerator backends Experience communicating complex technical findings in writing, including technical reports, design documents, or peer-reviewed publications Experience developing custom kernels using Pallas or similar kernel authoring frameworks for TPU or GPU Demonstrated track record of transitioning performance research into deployed systems used at significant scale PhD in Computer Science, Machine Learning, Computer Architecture, or a related technical field, or equivalent depth of research experience Publication record in systems for ML venues such as MLSys, OSDI, SOSP, or related AI conferences such as NeurIPS, ICML, or ICLR Experience with Mixture of Experts (MoE) architectures and their optimization for efficient training and inference Experience optimizing production-scale models with billions of parameters