ML Performance Engineer: Optimizing 1T+ LLMs with Equity
Posted Updated
Description
Lila Sciences is building a Scientific Superintelligence platform where AI and automation co-evolve to accelerate discovery. We seek a ML Research Engineer to maximize iteration speed for mid- and post-training runs on large LLMs using Nvidia GPUs.
You will optimize GPU/cluster performance, manage large-scale data, and refine RL pipelines while ensuring maintainable, reusable tooling for the research team.
Profile GPU compute performance: kernels, long-context handling, memory footprint. Optimize GPU cluster performance: collective/MoE, comm/compute overlap, RDMA/IB/EFA usage. Handle large data: manage mid-training data across storage and tokenization. Improve RL pipeline: optimize async RL cycle, environment interfacing, weight sync. Set performance standards: ensure configs are tested and reusable by researchers. Communicate throughput/accuracy tradeoffs to establish optimal points. BS/MS in CS or related quantitative field with 4+ years software/ML engineering; or PhD with 1+ year research. Experience with LLM inference libraries such as vLLM and Megatron. Experience running LLM inference and/or training at scale on multi-node GPU clusters. Familiarity with core concepts in reinforcement learning. Proven track record of clean APIs, well-tested code and thoughtful abstractions. Hands-on experience maintaining production libraries used by other engineers/researchers. Strong CI/CD skills: integration tests, build pipelines, and automated quality gates. Passion for agentic systems and cutting-edge LLM tool use.Skills
Sign in to read and join the discussion.