AI Inference Acceleration Algorithm Expert

Beijing Foreign Enterprise Management Consultants Co.,Ltd.; Singapore; On-site; SGD 180,000 - 260,000; Full time

Job summary

Huawei is seeking an AI Training/Inference Acceleration Algorithm Expert to lead R&D for accelerating AI training and inference across Agentic AI and multimodal domains. The role focuses on optimizing for next-gen compute architectures and integrating advancements into proprietary frameworks.

The candidate will drive end-to-end implementations, optimize memory and compute, and anticipate trends such as long-sequence modeling and sparsity to sustain a competitive edge.

Qualifications

• Masters or PhD in CS, AI, EE or related field.
• Strong background in LLMs and MoE models.
• Experience with large-scale model training and inference.
• Hands-on AI acceleration tech incl. zero-redundancy optimizers and distributed parallelism.
• Expertise in attention kernels, KV-cache compression, and quantization.
• Proficiency in Python and major DL frameworks; kernel optimization is a plus.
• Understanding of hardware-software co-design and memory bandwidth considerations.

Responsibilities

• Lead research and development of AI training and inference acceleration algorithms for Agentic AI and Multimodal domains.
• Drive end-to-end implementation within proprietary AI frameworks and acceleration libraries.
• Oversee technical foresight and pre-plan for trends like long-sequence modeling and sparsity.

Skills

AI acceleration

LLM architectures

MoE models

Python

Distributed training

Hardware optimization

Quantization

Performance tuning

Tensor frameworks

Education

Master’s or PhD in CS/AI/EE

Tools

CUDA

cuDNN

MKL

Job description

On behalf of Huawei, a world-renowned information and communication technology company, we are seeking passionate and talented individuals to join our team as AI Training/Inference Acceleration Algorithm Expert.

Job Description:

• Lead the research and development of AI training and inference acceleration algorithms for Agentic AI and Multimodal domains, specifically optimized for next-generation AI compute architectures to maximize compute efficiency and utilization.
• Drive the end-to-end implementation of acceleration algorithms within proprietary AI frameworks and acceleration libraries, ensuring seamless integration and continuous iterative optimization based on real-world performance metrics.
• Oversee technical insight and foresight within the training/inference domain, identifying emerging trends such as long-sequence modeling and sparsity to pre-plan and develop cutting-edge algorithms that ensure the sustained competitive advantage of AI computing platforms.

Skills / Qualifications:

• Master’s or PhD degree in Computer Science, Artificial Intelligence, Electronic Engineering, or a related technical discipline.
• Solid technical foundation in mainstream Large Language Model (LLM) architectures and Mixture of Experts (MoE) models, with proven experience in large-scale model training, fine-tuning, or inference deployment.
• Strong hands-on experience in AI acceleration technologies, including zero-redundancy optimizer techniques, distributed parallel strategies (TP/PP/SP/VP/DP), communication compression, and memory optimization.
• Technical expertise in high-performance attention kernels and KV-cache compression is essential, alongside proficiency in model weight/activation quantization and sparsity-aware acceleration.
• High proficiency in Python and deep familiarity with leading deep learning frameworks and high-performance acceleration libraries; expertise in hardware-level kernel optimization or native operator tuning is highly preferred.
• Deep understanding of hardware-software co-design, with knowledge of hardware characteristics (memory bandwidth, compute cycles, and interconnects) and a track record of optimizing performance for models with 100B+ parameters.
• Proven ability to perform complex AI model tuning and optimization in distributed or heterogeneous computing environments, translating hardware-specific features into significant algorithmic performance gains.