Research Engineer, Large-Scale Training

Together AI•San Francisco, CA

1d•$200,000 - $290,000

About The Position

The Model Shaping team at Together AI focuses on developing products and research for tailoring open foundation models to specific applications. They build services that enable machine learning developers to select optimal models for their tasks and enhance these models with domain-specific data. Additionally, they create new methods for efficient model training and evaluation, drawing from machine learning, natural language processing, and ML systems. As a Research Engineer on the Scaling Team within Model Shaping, you will translate advanced research in efficient foundation model training into robust, high-performance systems. This involves profiling and optimizing Together's training infrastructure, identifying performance bottlenecks across the entire stack, and implementing state-of-the-art techniques from research literature and internal scientists into production environments. Your contributions will directly influence the fine-tuning experience for Together's customers. You will be responsible for quickly integrating newly released open-source models onto the Model Shaping platform, ensuring efficient and reliable training across various customer workloads. Collaborating closely with Research Scientists, you will also develop the experimental infrastructure necessary to accelerate research and facilitate the reliable deployment of validated ideas at scale.

Requirements

Demonstrated ability to independently take ambiguous performance or infrastructure problems from investigation through deployment.
Strong programming skills in Python and PyTorch, with an emphasis on writing efficient, maintainable code.
Hands-on experience training or fine-tuning large neural networks in multi-GPU or multi-node environments.
Solid understanding of ML systems fundamentals, including GPU architecture, mixed-precision training, and distributed training paradigms such as data, tensor, pipeline, or expert parallelism.
Strong communication skills and the ability to collaborate effectively with both researchers and engineers.
Passion for staying current with advances in AI research and applying them to real-world systems.
Excitement about translating cutting-edge research into production systems that deliver customer impact.

Nice To Haves

Experience writing optimized NVIDIA GPU kernels using CUDA or Triton, or implementing communication collectives with technologies such as NCCL or NVSHMEM.
Experience with large-scale training frameworks such as FSDP, DeepSpeed, Megatron-LM, or custom distributed training systems.
Experience optimizing distributed training for compute efficiency, memory efficiency, or scalability.
Experience running and managing large-scale GPU experiments, including scheduling, monitoring, and fault tolerance.
Contributions to widely used open-source ML or ML systems projects.
Experience building or operating ML products or managed services used by external customers.

Responsibilities

Design, implement, and optimize core components of Together's large-scale training infrastructure.
Integrate new model architectures, validate training correctness and convergence, and optimize performance for production fine-tuning workloads.
Profile distributed training workloads to identify and eliminate bottlenecks across compute, memory, and communication.
Design and execute experiments to validate performance hypotheses and benchmark new approaches against state-of-the-art methods.
Partner closely with Research Scientists to productionize novel training methods and contribute to publications and open-source releases.
Rapidly enable support for newly released open-source foundation models on the Together platform.
Build and maintain experimental infrastructure that accelerates research while ensuring production-quality reliability and scalability.