ABOUT US

Deep learning systems,
optimized.

Yui AI Labs optimizes deep learning systems, from distributed training and inference to data pipelines and agent workflows. We profile workloads, identify bottlenecks, and implement improvements tailored to your models and infrastructure. Our work includes tuning parallelism and kernels for faster training, balancing resources across reinforcement learning workflows, and improving inference latency, throughput, and serving costs. We build data pipelines that keep CPUs and GPUs productive, and custom agent harnesses that connect domain knowledge, tools, and execution workflows. From performance analysis to hands-on implementation, we work with your team to measure improvements against real workloads.

Founded by Masahiro Tanaka, the company builds on 15+ years in distributed computing. He is a core contributor to DeepSpeed ↗, an open-source library with 43K+ GitHub stars and 1M+ monthly PyPI downloads ↗ (November 2026). He led work on compiler-level optimizations for distributed training (DeepCompile ↗), sequence parallelism (Ulysses ↗), and inference batch scheduling (DeepSpeed-FastGen ↗), with scheduling techniques now adopted by systems including vLLM. At Microsoft, he improved foundation model training efficiency for Phi-3, MAI-DS-R1, and internal models on 2,000+ GPUs. At Anyscale, he contributed to Ray ↗, widely used by open-source LLM reinforcement learning frameworks, and led a hybrid-parallelism framework for multimodal training on heterogeneous compute. Earlier at Japan's NICT, he developed RaSC to orchestrate NLP tools across hundreds of nodes, processing billions of web pages. His work earned the 2026 PyTorch Foundation Contributor Award ↗ (Innovator). He also co-developed RaNNC, an automatic parallelization framework that won first place at the official PyTorch Annual Hackathon 2021 ↗ in the Developer Tools & Libraries category.