AMD模型优化工程师（推理&训练）Model Optimization Engineer (Inference & Training)

任职要求

* Strong software engineering in Python and C/C++. * Practical experience with PyTorch/JAX and building/extending deep learning frameworks. * Hands‑on CUDA and/or ROCm development; experience writing or optimizing GPU kernels. * Experience with Triton (kernel development/optimization) is highly desired. * Proven experience with model optimization techniques, especially low‑bitwidth quantization and other compression methods. * Familiarity with GenAI inference engines and optimizations (e.g., vLLM, SGLang, xDiT, continuous batching, speculative decoding). * Skilled at profiling and performance debugging across stack layers (operator → model → framework → hardware). PREFERRED QUALIFICATIONS * Publications or contributions in model optimization / ML systems are a strong plus. * Experience with distributed traini…

工作职责

THE ROLE We are looking for a hands‑on Engineer to design, implement, and optimize AI model training and inference solutions for AMD platforms. The role focuses on end‑to‑end performance and accuracy improvements at the framework, model, and operator levels, with strong emphasis on low‑bitwidth quantization, model compression, and real‑world deployment. You will work closely with AMD hardware and software teams, support customers, and contribute to open‑source projects and inference/training frameworks. KEY RESPONSIBILITIES * Design, implement, and optimize inference and training pipelines for AMD GPUs/accelerators at the framework, model, and operator levels. * Lead research and development of model optimization algorithms: low‑bitwidth quantization, pruning/sparsity, compression, efficient attention mechanisms, and lightweight architectures. * Implement and tune CUDA/ROCm/Triton kernels for critical operators; profile and eliminate performance bottlenecks. * Integrate and optimize models for PyTorch/JAX and common distributed training/inference stacks (Torchtitan, Megatron, DeepSpeed, HF Transformers, etc.). * Reduce latency and increase throughput for large‑model inference (e.g., batching strategies, caching, speculative decoding). * Contribute to and/or maintain open‑source inference/training tools, ensuring production readiness and community adoption. * Provide technical support and guidance to customers and internal teams to achieve target accuracy and performance on AMD platforms. TECHNICAL