Research Engineer — Training Performance & ML Compilation (Torch Compile) - Seed Infra

About the Team The Seed Infrastructures team oversees the distributed training, reinforcement learning framework, high-performance inference, and heterogeneous hardware compilation technologies for AI foundation models.

Responsibilities - Optimize training performance for large-scale foundation models through compiler-level techniques, including graph optimization, operator fusion, and kernel generation. - Develop and extend ML compilation capabilities based on the PyTorch compilation stack (e.g. FX, Dynamo, Inductor) to improve training efficiency across heterogeneous GPU platforms. - Design and optimize high-performance GPU kernels for training workloads. - Conduct performance profiling and analysis of large-scale training jobs; identify and resolve bottlenecks in collaboration with research and infrastructure teams.

Minimum Qualification(s) - Bachelor's degree or above in Computer Science, Electrical Engineering, or a related field. - Strong proficiency in C/C++ and Python; solid foundations in algorithms, data structures, and systems programming. - Hands-on experience in training-side performance optimization for deep learning workloads. - Hands-on experience writing and optimizing GPU kernels (e.g., CUDA, Triton). - Experience with the PyTorch compilation stack, meeting at least one of the following: direct experience using, debugging, or extending Inductor or FX; proficiency in Triton kernel development; or solid experience with PyTorch computation graph work (graph optimization, graph capture, operator fusion).

Preferred Qualification(s) - Experience with TorchDynamo or bytecode-level program transformation. - Experience with Triton compiler internals or other ML compiler backends (e.g., MLIR, LLVM). - Contributions to related open-source projects (e.g., PyTorch, Triton, FlashAttention). - Publications in relevant venues (e.g., MLSys, OSDI, ASPLOS).