Description
We are seeking an AI Infrastructure Engineer to optimize large-scale model training efficiency and high-throughput data pipelines.
You will profile GPU bottlenecks, optimize multi-GPU distributed communication, and build scalable data loaders to support our deep learning research at scale.
Principal Responsibilities
- Optimize end-to-end model training throughput and hardware utilization across large-scale GPU clusters.
- Maintain and optimize high-performance, low-latency Data Loaders.
Qualifications/Skills Required
- Strong Python programming skills.
- Deep familiarity with PyTorch internals.
- Solid understanding of GPU architecture, including CUDA kernels, streams, events, and contexts.
- Proficient in using profiling tools to identify and resolve model compute, memory, and communication bottlenecks.
- Strong understanding of multi-GPU distributed communication principles and optimization techniques.
- Familiarity with static graph compilation technologies (e.g., torch.compile, XLA, TensorRT).
- Ability to architect and scale high-throughput Data Loaders utilizing asynchronous I/O and concurrency techniques.
Preferred Qualifications
- Familiarity with common model architectures (such as Attention mechanisms) and the ability to collaborate with researchers on co-designing more efficient model structures.
- Hands-on experience with mixed-precision training, model quantization, and convergence/accuracy alignment techniques.
For more fresh new jobs, check out yubhub.co.