Job Responsibilities

1. Responsible for building the underlying training and inference infrastructure for large models (Large Language Models / Multimodal Models), including GPU/CPU cluster running environment, communication layer and execution framework integration and optimization.
2. Design, develop and optimize deep learning operators (Operators / Kernels), covering CUDA/ROCm, C++, Python, etc., to improve model training and inference performance.
3. Participate in the performance tuning of distributed training and inference systems, including data parallelism, model parallelism, pipeline parallelism, and hybrid parallelism strategies.
4. Optimize computation graph, memory usage and communication efficiency, reduce GPU memory occupancy and cross-node communication costs.
5. Perform performance analysis, low-level optimization, and custom extension for mainstream deep learning frameworks (such as PyTorch, TensorFlow, JAX).
6. Collaborate closely with platform engineering, MLOps and model development teams to integrate operators and system optimization results into the actual training and inference processes.
7. Participate in the technical evaluation, adaptation, and performance validation of new hardware platforms (such as next-generation GPUs, accelerators, and interconnect technologies).