About the role
What You’ll Do
* Drive down wall-clock time to convergence by profiling and eliminating bottlenecks across the foundation model training stack stack, from data pipelines to GPU kernels
* Design, build, and optimize distributed training systems (PyTorch) for multi-node GPU clusters, ensuring scalability, robustness, and high utilization
* Implement efficient low-level code (CUDA, cuDNN, Triton, custom kernels) and integrate it seamlessly into high-level training frameworks
* Optimize workloads for hardware efficiency: CPU/GPU compute balance, memory management, data throughput, and networking
* Develop monitoring and debugging tools for large-scale runs, enabling rapid diagnosis of performance regressions and failures
What You’ll Bring
* Deep experience in distributed systems, ML infrastructure, or high-performance computing (8+ years)
* Production-grade expertise in Python
* Low-level performance mastery: CUDA/cuDNN/Triton, CPU–GPU interactions, data movement, and kernel optimization
* Scaling at the frontier: experience with PyTorch and training jobs using data, context, pipeline, and model parallelism
* System-level mindset with a track record of tuning hardware–software interactions for maximum utilization