LLM推理优化工程师
Moonshot AI · 北京市
Apply on Moonshot AI's siteTailor my resume
Posted
Apr 1, 2026Last verified
48 min agoEmployment
Full-timeLevel
MidFunction
AI/ML EngineeringAI role
Open 7 months. Roles open this long are often filled, or kept open all year. Check the date on the employer's site before you apply.
核心职责
训练-推理协同优化
- 推动 Quantized-aware Training (QAT) 和 Quantized Training 的工程化落地,研发W8A8、W4A16等低比特训练方案,实现训练与推理精度-效率的最优平衡
- 设计并实现 Speculative Decoding(投机解码)及其变体(如EAGLE,DFlash等)的全链路方案,包括Draft模型训练、接受率优化与系统调度策略
- 建立从训练到推理的端到端优化链路,确保相关训练方案在推理侧获得可度量的性能收益
任职要求
- 计算机科学、电子工程或相关专业本科及以上学历,3年以上高性能计算/深度学习系统相关经验
- 扎实的分布式系统功底:深入理解多机并行训练/推理相关技术,具备大规模训练框架(Megatron-LM)及主流推理框架(vLLM,SGLang)的开发能力。
- 以下两项至少具备其一:
- 掌握低比特推理与量化技术:熟悉GPTQ、AWQ、SmoothQuant等后量化方案,理解训练感知量化的实现细节与数值稳定性处理,熟悉MXFP4,NVFP4等新兴低比特量化格式。
- 深入理解投机解码原理:掌握Draft模型与Target模型的协同机制,了解EAGLE3、DFlash等当前主流方法的训练、推理实现方法。
优先考虑:
- 具备LLM算法基础,熟悉大模型训练相关算法与技术
- 熟悉CUDA编程模型,熟练使用Nsight Systems/Compute进行性能剖析,能够手写高性能CUDA Kernel
- 在LLM 算法/体系结构/系统领域相关顶级会议或期刊上有论文发表
Attention Head lists this role as published on Moonshot AI's own job board. Applications go to the employer.
Similar roles at other employers
- AI Developer — Interconnect Hardware FrontendNVIDIA · China, ShanghaiNov 16, 2026
- AI Application Engineer, Product EngineeringWestern Digital · Shenzhen, Guangdong Province, ChinaNov 13, 2026
- 泊车感知算法工程师_XCBosch Group · Shanghai, Shanghai, ChinaNov 10, 2026
- AI & ML Engineer | BankingCapco · China - Hong KongNov 6, 2026
- Machine Learning Engineer — AI Agent PlatformVirtuos · Shanghai, ChinaNov 5, 2026
All mid-level AI/ML Engineering roles in China →
More roles at Moonshot AI
- Nov 20, 2026
- Nov 4, 2026
- Nov 2, 2026
- Oct 20, 2026
- Sep 18, 2026
- Sep 18, 2026
- Sep 18, 2026
- Sep 4, 2026