Overview:
In this role you drive performance for large-scale foundation-model training across thousands of accelerators. You will partner with modeling, infrastructure, simulation, data-mining and robotics teams to set the performance direction and turn wins into reusable defaults. You will extend observability tooling and build automation to reproduce and verify jobs at scale. This is a hands-on leadership role focused on accelerating AI and robotics workloads in production, shaping the speed and reliability of XPENG’s end-to-end autonomous systems. Compensation / Benefits fun, supportive and engaging environment infrastructures and computational resources opportunity to work on cutting-edge technologies impact on transportation revolution competitive compensation package snacks, lunches, dinners, and fun activities
Responsibilities:
Own training performance at scale, optimizing data path, kernels, parallelism and inter-node communication Provide cross-team technical leadership to address bottlenecks across modeling, infrastructure, simulation, data-mining and robotics Create modular, reusable components to replace one-off fixes and serve multiple models/workloads Enhance observability across the stack with low instrumentation overhead for production fleets Develop agentic automation to reproduce jobs, verify parity and separate real regressions from noise
Key requirements:
Master's or PhD in CS, CE or EE, or equivalent practical experience with 8+ years in performance engineering Hands-on experience optimizing GPU training or inference workloads in multi-node production environments Proven track record diagnosing performance issues in production workloads and delivering verifiable fixes Strong understanding of computer architecture and high-performance computing Proficient in C++ and Python with solid data structures and algorithms cross-functional collaboration problem-solving mindset clear technical communication CUDA or Triton kernel development and tuning PyTorch proficiency with insight into performance (graph breaks, host overhead, synchronization, memory layout) experience with scalable training across thousands of accelerators and multi-node failure modes