Research Engineer (LLM Performance)

This is an exciting opportunity for you to contribute to frontier research at the intersection of AI and drug design. Working in a highly creative, iterative environment, you will join the model performance and scaling team, where you will partner with scientists and engineers to scale foundational models that will transform the biopharmaceutical world as we know it. You will draw upon your existing engineering experience whilst learning from those around you, to apply novel techniques and ideas to newly encountered model and systems performance optimizations, as well as machine learning, computational biology and chemistry problems.What you will do

• Implement and optimize LLM post-training methods at scale on frontier models.
• Collaborate with research teams to translate new methods into production-ready systems.
• Relentlessly prioritize and execute on performance optimization opportunities.
• Evaluate and deploy frameworks for supervised fine-tuning, reinforcement learning and LLM evaluation.
• Diagnose and fix performance bottlenecks and communication overhead in distributed training and inference systems.
• Deploy low-precision methods to balance performance with accuracy, impacting real world drug design programs.
  
  Significant experience with large scale distributed training of LLMs.
• Experience with deep learning ML frameworks (either JAX or PyTorch).
• Knowledge of parallelism strategies and collective communication libraries (e.g. NCCL).
• Good understanding of GPU architectures. Reasoning about performance concepts is more important than writing kernels from scratch
• Excellent collaboration skills., Experience with general LLM serving stacks.
• Knowledge of XLA, Triton, Pallas, CUDA or similar accelerator DSLs / compilers.
• Experience with optimising ML accuracy using low-precision formats.
• Prior experience building, deploying and maintaining production systems on GCP.
• Interest in chemistry and biology.
  
  Isomorphic Labs (IsoLabs) was launched in 2021 to advance human health by building on and beyond the Nobel-winning AlphaFold system. Since then, our interdisciplinary team of drug discovery experts and machine learning specialists has built powerful new predictive and generative AI models that accelerate scientific discovery at digital speed.
Our name comes from the belief that there is an underlying symmetry between biology and information science. By harnessing AI's powerful capabilities, we can use it to model complex biological phenomena to help design novel molecules, anticipate how drugs will perform and develop innovative medicines to treat and cure some of the world's most devastating diseases. We have built a world-leading drug design engine comprising AI models that are capable of working across multiple therapeutic areas and drug modalities. We are continually innovating on model architecture and developing cutting-edge capabilities to advance rational drug design. Every day, and with each new breakthrough, we're getting closer to the promise of digital biology, and achieving our ambitious mission to one day solve all disease with the help of AI., We are guided by our shared values. It's not about finding people who think and act in the same way. These values help to guide our work and will continue to strengthen it. Thoughtful Thoughtful at Iso is about curiosity, creativity and care. It is about good people doing good, rigorous and future-making science every single day. Brave Brave at Iso is about fearlessness, but it's also about initiative and integrity. The scale of the challenge demands nothing less. Determined Determined at Iso is the way we pursue our goal. It's a confidence in our hypothesis, as well as the urgency and agility needed to deliver on it. Because disease won't wait, so neither should we. Together Together at Iso is about connection, collaboration across fields and catalytic relationships. It's knowing that transformation is a group project, and remembering that what we're doing will have a real impact on real people everywhere.