In this role, you will be responsible for the development, enablement, and performance optimization of large scale ML model training across diverse model families. This includes massive scale pre-training and post-training of LLMs with Dense and Mixture-of-Experts architectures, Multimodal models that are transformer and diffusion based, and Reinforcement Learning workloads. You will work at the intersection of ML research and high performance systems, collaborating closely with chip architects, compiler engineers, runtime engineers and AWS solution architects to deliver cost-effective, performant machine learning solutions on AWS Trainium based systems.

You will design, implement and optimize distributed training solutions for large scale ML models running on Trainium instances. A significant part of your work will involve extending and optimizing popular distributed training frameworks including FSDP (Fully-Sharded Data Parallel), torchtitan and Hugging Face libraries for the Neuron ecosystem.

A core focus of this role involves developing and optimizing mixed-precision and low-precision training techniques. You will work with BF16, FP8, and emerging numerical formats to maximize training throughput while maintaining model accuracy and convergence quality. This requires implementing precision aware training strategies, loss scaling techniques, and careful gradient management to ensure training stability across reduced precision formats. Understanding the tradeoffs between computational efficiency and numerical fidelity is essential to success in this position.

Beyond precision optimization, you will profile, analyze, and tune end-to-end training pipelines to achieve optimal performance on Trainium hardware. You will partner with hardware, compiler, and runtime teams to influence system design and unlock new capabilities. Additionally, you will work directly with AWS solution architects and customers to deploy and optimize training workloads at scale.