About the position
As a Software Engineer within the Autopilot AI Infrastructure team, you will work on reinforcing, optimizing, and scaling our infrastructure components supporting AI research activities for Autopilot and the Tesla Bot. At the core of our autonomy capabilities are neural networks that the research team is designing to train on very large amounts of data, across large-scale GPU clusters and our supercomputer Dojo. Robustly training these models at scale and in the shortest amount of time is critical to our mission. We are building and improving the in-house distributed training framework used by the research team to train production models, ensuring good ergonomics and flexibility for experimentation while providing good stability and performance.
Responsibilities
- Write robust Python software code in our machine learning training repository while applying best software practices to support the research team
- Increase the reliability of our training jobs by debugging and root causing failures across thousands of nodes and implementing fixes to prevent future failures
- Improve our training framework to support new training paradigms and experimentation methods
- Build and improve our monitoring/observability infra to quickly debug cluster and training application issues
- Profile and identify performance bottlenecks of training software in our training cluster
- Coordinate with the supercomputing team managing the training cluster to maintain high availability and job throughput
- Contribute across all parts of the AI training software stack as required by the dynamic needs of AI research
Requirements
- Practical programming experience in Python and/or C/C++
- Experience working with ML training frameworks (ideally PyTorch)
- Demonstrated experience scaling neural network training jobs across many GPUs
- Experience with parallel programming concepts and primitives
- Experience profiling and optimizing CPU-GPU interactions (pipelining computation with data transfers, etc)
- Proficient in system-level software, in particular hardware-software interactions and resource utilization
- Understanding of state-of-the-art deep learning concepts
- Experience programming in CUDA/Triton and/or NCCL internals
