Research Intern (Inference, Summer 2027)
Together AI
- San Francisco, United States
- Internship
- Intern/New Grad
- Posted
Join the Inference Research team as a Research Intern for Summer 2027. You will work on building efficient, scalable, and reliable serving systems for large foundation models. Your projects will focus on co-designing software, algorithms, and models to lower the cost and latency of modern AI systems. You will conduct rigorous experiments, communicate progress to the team, and document findings in scientific publications. This internship offers competitive benefits and the opportunity to work alongside industry-leading engineers and researchers.
Understanding of model optimization techniques and hardware acceleration approaches
Responsibilities
- Participer à la conception et à la mise en œuvre d'optimisations croisées entre les modèles, les systèmes et le matériel, en mettant l'accent sur des domaines tels que la conception du cache KV et les architectures de service à grande échelle.
- Concevoir et réaliser des expériences rigoureuses pour valider les hypothèses, et communiquer les plans, les progrès et les résultats des projets à l'équipe.
- Documenter les résultats dans des publications scientifiques et des articles de blog, et contribuer à l'avancement de l'IA ouverte et transparente.
Requirements
- Experience with deep learning frameworks (PyTorch, JAX, etc.)
Strong programming skills in Python
- Strong knowledge of Machine Learning and Deep Learning fundamentals
- Currently pursuing a final year of Bachelor's, Master's, or Ph.D. degree in Computer Science, Electrical Engineering, or a related field
- Familiarity with Transformer architectures and recent developments in foundation models
Preferred
- Prior research experience in foundation models, efficient machine learning, or ML systems
- Publications at leading conferences in machine learning or systems (i.e., MLSys, ICLR)
- Experience with CUDA programming (for kernel development)
- Contributions to open-source machine learning projects
