Training, Process Management Engineer

As a Training Runtime: Process Management Engineer, you will work on the software that ties thousands of computers together and exposes them as a unified system. This system has to serve individual researchers running multiple parallel experiments, as well as our largest training runs spanning 100’s of thousands and even millions of machines and accelerators. This requires easy to use, introspectable systems that can promote a fast debugging and development cycle, as well as relentless optimization for scale while maintaining stability and performance throughout. You will work primarily in Rust, building high-performance asynchronous systems with a strong emphasis on performance, correctness, and scalability. Working at this scale and at the frontier of AI development poses novel challenges. Out-of-the-box approaches often don’t work. The problems you will be working on are highly ambiguous and require strong design judgment as well as proficient execution to advance the state of our infrastructure. We’re looking for people who love optimizing an end-to-end platform, understanding high-performance architectures to maximize both local and distributed performance across our supercomputers. We’re looking for engineers excited by the rapid pace of responding to the dynamic and evolving needs of our training runtime and compute stack.

Responsibilities

Work across our Python and Rust stack; Design, build, and maintain software to orchestrate and monitor machine learning workloads on our largest supercomputers; Profile and optimize our software stack to support computation orchestration at frontier scale; Improve reliability, observability, and fault tolerance for long-running jobs; Debug complex distributed systems issues across large clusters; Respond to the changing shapes and needs of the ML systems to enable our researchers

Requirements

Have experience developing distributed systems (not just operating them); Enjoy understanding how large systems behave and fail at scale; Care deeply about performance, correctness, and reliability; Have strong software engineering skills and are proficient in Python and Rust or another systems programming language (e.g. C++); Have solid Linux knowledge, and are comfortable with systems-level debugging, performance analysis, and memory profiling; Are comfortable and experienced working and developing asynchronous and concurrent systems; Like high-ownership environments with light process and strong engineering agency

Benefits

Relocation assistance