AI Training/Inference Acceleration Algorithm Engineer

About the Role

We are collaborating with a leading global information and communication technology company to find an experienced AI Training/Inference Acceleration Algorithm Engineer. This role focuses on advancing AI training and inference acceleration algorithms within the Agentic AI and Multimodal domains, aiming to enhance performance on upcoming AI computing platforms.

Key Responsibilities

• Lead development of AI acceleration algorithms tailored for next-generation compute architectures, ensuring optimal efficiency and resource utilization.
• Implement and iteratively optimize acceleration algorithms within proprietary AI frameworks and acceleration libraries, guided by live performance data.
• Monitor and anticipate trends in training and inference technologies, such as long-sequence processing and sparsity, to innovate algorithms that sustain competitive advantages.

Qualifications & Expertise

• Advanced degree (Master's or PhD) in Computer Science, AI, Electronic Engineering, or closely related field.
• Deep understanding of Large Language Models (LLMs) and Mixture of Experts (MoE) model architectures with hands-on experience in large-scale training, fine-tuning, or inference deployment.
• Proven skills in AI acceleration methods including zero-redundancy optimizers, distributed parallelism strategies (tensor, pipeline, stage, vector, data parallelism), communication compression, and memory management.
• Experience with high-performance attention mechanisms and KV-cache compression; expertise in model quantization and sparsity-aware acceleration techniques.
• Proficiency in Python programming and familiarity with major deep learning frameworks and acceleration libraries; background in low-level kernel optimizations or operator tuning is highly advantageous.
• Strong knowledge of hardware-software co-design principles, including critical hardware characteristics such as memory bandwidth and compute cycles, especially for extremely large models (100 billion parameters and above).
• Demonstrated ability to optimize AI models in distributed and heterogeneous compute environments by leveraging hardware-specific features for improved algorithmic performance.