Research Engineer (Seed Model - Infra - Training Performance & Core Kernels)

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

• Design, implement, and optimize core training kernels for large-scale foundation models, including Attention, GEMM, and MoE operators, with a focus on both forward and backward passes.
• Analyze hardware performance bottlenecks through micro-benchmarking, profiling, and performance modeling, and translate findings into kernel-level optimizations.
• Implement low-precision computation paths and ensure numerical correctness and training stability.
• Collaborate with team members to bring up and optimize training kernels on new hardware architectures and integrate them into production training frameworks.

• Bachelor's degree or above in Computer Science, Electrical Engineering, or a related field.
• Strong proficiency in C/C++ and Python; solid foundations in computer architecture and parallel programming.
• Hands-on experience implementing and optimizing high-performance GPU kernels (e.g., Attention, GEMM, MoE) using at least one low-level programming stack such as CUDA / CUTLASS, Triton, or HIP.
• Ability to identify compute, memory, and communication bottlenecks and explain optimization decisions with measured data.

• Experience optimizing kernels for training workloads, including backward passes and integration with training frameworks such as Megatron.
• Experience with low-precision kernel implementation and numerical accuracy validation.
• Experience porting and optimizing kernels across different hardware generations or accelerator architectures.
• Familiarity with Triton compiler internals or ML compiler backends.