Training

Work on foundation model training performance by profiling and removing bottlenecks across the training stack, from data pipelines to GPU kernels. Build and optimize distributed training systems for multi-node GPU clusters using PyTorch, and implement efficient low-level components with CUDA, cuDNN, Triton, and custom kernels. Improve hardware utilization through workload and memory optimizations, and create monitoring/debugging tools to quickly diagnose performance regressions and failures.