Member of Technical Staff Research Engineer

You will improve the performance, stability, and reliability of large-scale multimodal model training. You will profile training systems, implement GPU-level optimizations, investigate distributed-training failures, validate numerical behavior, and build benchmarking tools that connect systems improvements to model-quality outcomes.

Responsibilities

• Improve performance, reliability, and numerical stability of production training runs
• Profile full training steps across model code, kernels, data loading, communication, optimization, checkpointing, and memory
• Implement and validate GPU-level optimizations
• Advance low-precision training and assess quality tradeoffs
• Translate architecture changes into efficient training implementations
• Debug distributed-training failures and throughput regressions
• Build benchmarking and profiling harnesses
• Turn repeated training failures into improved abstractions and tools

Requirements

• Experience with large-scale training systems
• Strong PyTorch knowledge and ability to modify low-level training code
• Knowledge of distributed training, including FSDP, parallelism, activation checkpointing, NCCL, and compute-communication overlap
• Experience improving throughput, memory use, or stability in real training runs
• Experience profiling GPU workloads
• Understanding of low-precision training and quantization tradeoffs
• Ability to partner with researchers on ablations and measurements

Benefits

• Equity
• Coverage of reasonable travel costs for required in-person time