Senior Applied Research Engineer
Fundamental7 months ago
Barcelona, SpainSenior
Responsibilities
- Profile end-to-end distributed training runs to identify compute, GPU memory, and inter-GPU communication bottlenecks.
- Improve the efficiency and reliability of large-scale training jobs through architectural decisions and Triton/CUDA kernel development.
- Design and implement model scaling, parallelization, and memory optimization techniques for very large context sizes.
- Collaborate with ML researchers to diagnose architectural inefficiencies and ensure research ideas scale efficiently in practice.
- Drive productionization and serving of models, including improving inference efficiency through quantization.
- Share internal knowledge about model efficiency and optimization.
Requirements
- Strong understanding of modern machine learning architectures and large-scale training pipelines.
- Experience running distributed training jobs on multi-GPU systems.
- Advanced profiling and debugging skills across CPU, GPU, memory usage, latency, and inter-GPU communication.
- Strong programming skills in Python.
- Experience with model scaling and parallelization strategies, including tensor and pipeline parallelism.
- Preferred: familiarity with NCCL, MPI, and distributed communication primitives.
- Preferred: knowledge of PyTorch and Triton internals.
- Preferred: programming experience with C++ and CUDA.
Benefits
- Competitive compensation with salary and equity.
- Comprehensive health coverage for employees and dependents.
- Paid parental leave for all new parents, including adoptive and surrogate journeys.
- Relocation support for employees moving to one of the company's office locations.
- Mission-driven, low-ego culture that values diversity of thought, ownership, and bias toward action.
Categories
About Fundamental
Fundamental builds NEXUS, a large tabular AI model and platform for enterprise decision-making on structured data, delivered with deployment support. It sells to large companies, including Fortune 100 firms, through software licensing and forward-deployed data science services. Founded in 2024 by former DeepMind engineers and headquartered in San Francisco, the company is expanding industry use cases such as oil and gas, where it supports predictive analytics and operational planning.