Research Engineer/Scientist - Machine Learning RL & Optimisation (Contractor)

About Huawei Research and Development UK Limited Founded in 1987, Huawei is a leading global provider of information and communications technology (ICT) infrastructure and smart devices. We have 207,000 employees and operate in over 170 countries and regions, serving more than three billion people around the world. Our vision and mission is to bring digital to every person, home and organization for a fully connected, intelligent world.

Huawei Research and Development UK Limited Overview Huawei's vision is a fully connected, intelligent world. To achieve this, we work to inspire passion for basic research around the world. Our combined passion drives development across the global innovation value chain. Huawei has the largest Research and Development organization in the world with 96,000+ employees in research centers around the globe. In the UK, we already have design centers in Cambridge, London, Edinburgh and Ipswich.

Job Summary Research and develop large-scale machine learning systems, alignment workflows, and optimization infrastructure to advance LLM reasoning and post-training capabilities. Design and execute scaled reinforcement learning pipelines (e.g., PPO, GRPO) utilizing distributed training frameworks (verl, trl, DeepSpeed, FSDP) integrated with high-performance inference engines (vLLM). Optimize low-level training throughput, kernel performance, and memory utilization across heterogeneous hardware clusters using expressive hardware DSLs (e.g., TileLang, Triton).

Key Responsibilities

• Design and execute scaled RL finetuning workflows (e.g., PPO, GRPO) to enhance LLM reasoning, instruction-following, and alignment.
• Architect and manage large-scale distributed training experiments across multi-node GPU, optimizing for maximum throughput and hardware utilization.
• Develop and maintain training infrastructure using advanced parallelization frameworks (verl, trl, DeepSpeed, FSDP) to support rapidly evolving research needs.
• Integrate high-performance inference engines like vLLM directly into RL generation loops to reduce rollout latency and accelerate training cycles.
• Implement robust profiling and debugging pipelines to diagnose bottlenecks in GPU memory, compute, and inter-node communication.
• Collaborate with data and evaluation teams to design dense reward functions and synthetic data generation pipelines.
• Design, benchmark, and deploy highly optimized custom tensor operators (e.g., FlashAttention, GEMM) across heterogeneous hardware architectures using modern Domain-Specific Languages (DSLs) and AI compilers.

This job description is only an outline of the tasks, responsibilities and outcomes required of the role. The jobholder will carry out any other duties as may be reasonably required by his/her line manager. The job description and personal specification may be reviewed on an ongoing basis in accordance with the changing needs of Huawei Research and Development UK Limited.

Person Specification Required

• Master's or PhD (or equivalent industry research experience) in Machine Learning, Computer Science, Data Science, or a highly quantitative field with a heavy focus on Machine Learning.
• Deep proficiency in PyTorch and experience writing custom training loops and data pipelines.
• Hands-on experience with RLHF/RLVF methods (e.g., PPO, GRPO) and an understanding of policy optimization dynamics.
• Technical familiarity with at least two of the following: DeepSpeed, FSDP, verl or trl.
• Production or heavy research experience utilizing vLLM or similar high-throughput inference serving engines for generation.
• Ability to thrive in a fast-paced, iterative environment where research and production infrastructure deeply intersect.

Desired

• Proven track record of running scaled GPU experiments across multi-node clusters.
• Experience implementing next-gen alignment and reasoning paradigms, such as GRPO or Monte Carlo Tree Search (MCTS).
• Deep understanding of GPU architectures, kernels, FlashAttention, and profiling tools.
• Familiarity with cluster environments and schedulers like Slurm or Kubernetes.
• Hands-on experience developing within NPU stack or ecosystem integrations.
• Practical knowledge of hardware-expressive Domain-Specific Languages (e.g., TileLang, Triton) to optimize low-level memory placement, layout propagation, and thread-block scheduling.
• Research publications at top-tier AI/ML conferences (NeurIPS, ICLR, ICML) or a strong open-source GitHub footprint in LLM training/infrastructure.

What we offer

• 33 days annual leave entitlement per year (including UK public holidays)
• Group Personal Pension
• Life insurance
• Private medical insurance
• Medical expense claim scheme
• Employee Assistance Program
• Cycle to work scheme
• Company sports club and social events
• Additional time off for learning and development