Distributed Training & Inference Optimization Engineer (LLM) - GPU Optimization Department (GPUOD)
At a glance
Mid-level AI Engineering role at Rakuten. Tokyo · full-time.
Pay not stated AI Engineering salary
aimlroles summary, based on the employer's posting.
What you'll do
- Tune LLM training frameworks to use GPU capacity better and shorten training runs
- Find distributed-training slowdowns in communication, NCCL behavior and CUDA kernel performance
- Apply quantization, batching and KV caching to make LLM serving faster and handle more requests
- Work with infrastructure teams on GPU scheduling, resource distribution and recovery for large training jobs
- Build benchmarks for training speed, memory use and inference latency, then use them to guide improvements
What you bring
- At least three years optimizing GPU-based machine learning training and inference, preferably involving LLMs or large deep-learning models
- Strong command of PyTorch, DeepSpeed, FSDP or Megatron-LM, including techniques for distributed training
- Practical knowledge of LLM inference methods such as quantization, pruning or KV caching
- A bachelor's degree or above in computer science, engineering or a related discipline
Who this fits
This mid-level role suits an engineer focused on distributed LLM training and inference performance across GPU clusters. Advanced English is required.
From the employer
Job Description:
Business Overview
AI & Data Division (AIDD) spearheads data science & AI initiatives by leveraging data from Rakuten Group. We build a platform for large-scale field experimentations using cutting-edge technologies to provide critical insights that enable faster and better and faster contribution for our business. Our division boasts an international culture created by talented employees from around the world. Following the strategic vision “Rakuten as a data-driven membership company”, AIDD is expanding its data & AI related activities across multiple Rakuten Group companies.
Department Overview
GPU Optimization Department (GPUOD) is responsible for the strategic management, optimization, and governance of Rakuten's company-wide AI infrastructure, ensuring high-performance, cost-efficient utilization of compute resources for machine learning workloads. We oversee a large-scale hybrid infrastructure spanning thousands of accelerators, including the latest Hopper and upcoming Blackwell architectures.
As a central enabler for AI innovation, we:
- Optimize compute resource allocation across on-premises and multi-cloud environments, maximizing efficiency for training and inference workloads.
- Manage hybrid orchestration of diverse accelerator resources, ensuring seamless scalability and cost-effective deployment.
- Develop and enhance frameworks for large-scale distributed training, with special focus on LLMs and generative AI.
- Optimize inference performance through model optimization techniques and system-level acceleration.
- Collaborate with internal teams to deliver scalable, high-availability inference services tailored to business needs.
- Continuously evaluate next-generation hardware solutions, including specialized AI chips optimized for LLM workloads.
- By effectively managing both conventional and specialized compute resources across on-premises and cloud environments, our team ensures Rakuten's AI ecosystem remains at the forefront of performance, reliability, and cost-efficiency.
Position:
Why We Hire
- Work on cutting-edge LLM training & inference optimization at scale.
- Directly impact Rakuten’s AI infrastructure by improving efficiency and reducing costs.
- Collaborate with global AI/ML teams on high-impact challenges.
- Opportunity to research and implement state-of-the-art GPU optimizations.
Position Details
As a GPU Training & Inference Optimization Engineer, you will focus on maximizing the performance, efficiency, and scalability of LLM training and inference workloads on Rakuten’s GPU clusters. You will deeply optimize training frameworks (e.g., PyTorch, DeepSpeed, FSDP) and inference engines (e.g., vLLM, TensorRT-LLM, Triton, SGLang), ensuring Rakuten’s AI models run at peak efficiency.
This role requires strong expertise in GPU-accelerated ML frameworks, distributed training, and inference optimization, with a focus on reducing training time, improving GPU utilization, and minimizing inference latency.
Key Responsibilities
- Optimize LLM training frameworks (e.g., PyTorch, DeepSpeed, Megatron-LM, FSDP) to maximize GPU utilization and reduce training time.
- Profile and optimize distributed training bottlenecks (e.g., NCCL issues, CUDA kernel efficiency, communication overhead).
- Implement and tune inference optimizations (e.g., quantization, dynamic batching, KV caching) for low-latency, high-throughput LLM serving (vLLM, TensorRT-LLM, Triton, SGLang).
- Collaborate with infrastructure teams to improve GPU cluster scheduling, resource allocation, and fault tolerance for large-scale training jobs.
- Develop benchmarking tools to measure and improve training throughput, memory efficiency, and inference latency.
- Research and apply cutting-edge techniques (e.g., mixture-of-experts, speculative decoding) to optimize LLM performance.
Mandatory Qualifications:
- 3+ years of hands-on experience in GPU-accelerated ML training & inference optimization, preferably for LLMs or large-scale deep learning models.
- Deep expertise in PyTorch, DeepSpeed, FSDP, or Megatron-LM, with experience in distributed training optimizations.
- Strong knowledge of LLM inference optimizations (e.g., quantization, pruning, KV caching, continuous batching).
- Bachelor’s or higher degree in Computer Science, Engineering, or related field.
Desired Qualifications:
- Proficiency in CUDA, Triton kernel, NVIDIA tools (Nsight, NCCL), and performance profiling (e.g., PyTorch Profiler, TensorBoard).
- Experience with LLM-specific optimizations (e.g., FlashAttention, PagedAttention, LoRA, speculative decoding).
- Familiarity with Kubernetes (K8s) for GPU workloads (e.g., KubeFlow, Volcano).
- Contributions to open-source ML frameworks (e.g., PyTorch, DeepSpeed, vLLM).
- Experience with inference serving frameworks (e.g., vLLM, TensorRT-LLM, Triton, Hugging Face TGI).
#engineer #applicationsengineer #AI #aianddatadiv
Languages:
English (Overall - 3 - Advanced)