Research Engineer, GPU Performance

As a Research Engineer, GPU performance, you will sit at the heart of that spectrum: your low-level work directly determines how fast our frontier models can be trained and served. Working hand-in-hand with Research Scientists, you'll make every GPU count.

Mistral AI is hiring an expert in pushing large language model's GPU utilization and performance to their limits.. In this role, you will own the performance of our models end-to-end: writing the low-level code that extracts the maximum capacity from high-end hardware, and rethinking model architectures so they run efficiently at scale.

What You Will Do

• Develop low-level code that takes full advantage of high-end GPUs and maxes out their capacity — from custom kernels to memory-layout and communication optimisations.

• Rethink parts of the generative model architecture to make them more suitable for efficient inference, in collaboration with Research Scientists.

• Integrate low-level efficient code into a high-level MLOps framework, so that optimisations benefit every training run and deployment rather than a single experiment.

• Profile, benchmark and squeeze performance out of large-scale training and inference on thousands of GPUs.

• Model LLM execution on various GPUs with high fidelity to assess hardware usage and build serving performance estimates.

• High technical competence for writing custom low-level kernels and pushing GPUs to their limits.

• High expertise in the distributed computation infrastructure of current-generation GPU clusters.

• Overall understanding of the field of generative AI; knowledge of, or interest in, fine-tuning and using language models for applications.

• Strong software-design instincts: testing, code review, CI/CD.

• Self-starter, low-ego, collaborative.