Research Engineer - (Seed Model - Infra - Pretraining Infrastructure)

About the Team

The Seed Infrastructures team oversees the distributed training, reinforcement learning framework, high-performance inference, and heterogeneous hardware compilation technologies for AI foundation models.

Responsibilities

• Conduct research and engineering development on large-scale infrastructure to enable efficient pretraining of foundation models, multimodal LLMs, and video generation models
• Design and optimize distributed training strategies for model training, including parallelism schemes, computation and communication optimization, and throughput scaling on large GPU clusters
• Analyze performance bottlenecks in exascale training systems and propose principled, data-driven optimization methods
• Bridge cutting-edge research and large-scale production deployment by translating research ideas into scalable, real-world infrastructure solutions

Minimum Qualifications

• Bachelor's degree or above in Computer Science, Software Engineering or a related field, with 2+ years of experience in large-scale model training, distributed training systems, or GPU kernel optimization
• Solid understanding of large-scale training of LLMs and multimodal models
• Proven experience with parallelism strategies (e.g., data, pipeline, expert parallelism) on large GPU clusters, or GPU kernel performance optimization
• Strong grasp of algorithm–system co-design and cross-layer optimization for training efficiency, scalability, and reliability
• Strong programming skills and hands-on experience implementing production-grade ML systems or infrastructure

Preferred Qualifications

• Hands-on experience pretraining LLMs, multimodal, or video generation models at scale, or with mainstream distributed training frameworks (e.g., Megatron-LM, DeepSpeed, PyTorch FSDP, NCCL), or with custom kernel development using CUDA or Triton
• Publications at top-tier machine learning or systems conferences (e.g., NeurIPS, ICML, MLSys, OSDI, SOSP), or notable open-source contributions to widely used training frameworks