dsx-ai-factory/run-ai

Run:ai is the AI workload orchestration platform for Kubernetes. This monorepo contains the control plane services, in-cluster Kubernetes operator/controllers, shared libraries and clients, the installer, and the local dev environment.

What it solves

NVIDIA Run:ai addresses the challenge of managing and sharing expensive GPU resources across teams and projects in a Kubernetes environment. It prevents resource contention and underutilization by providing a centralized platform for scheduling, governing, and monitoring AI workloads.

How it works

Run:ai acts as an orchestration layer that sits on top of Kubernetes. It consists of a central control plane for management and cluster-specific services that extend native Kubernetes scheduling. It uses advanced techniques like fair-share scheduling, preemption, and fractional GPU allocation to ensure that AI training, inference, and development workloads receive the resources they need based on defined organizational policies.

Who it’s for

This platform is designed for organizations and platform engineers managing shared GPU infrastructure across multiple teams, departments, or Kubernetes clusters who need to enforce resource quotas and improve overall GPU utilization.

Highlights

  • AI-aware GPU scheduling including fair-share, preemption, and bin-packing.
  • Dynamic, fractional GPU allocation to maximize hardware efficiency.
  • Centralized multi-cluster management and observability for GPU workloads.
  • Role-based access control and governance to enforce organizational policies.
  • Seamless integration with standard Kubernetes workflows and tools like Kubeflow, Ray, and KServe.

Related

  • Project
    jdtoscano94/NABLA-SciML

    A unified framework for Scientific Machine Learning (SciML) providing implementations of PINNs, DeepONets, and KANs for studying complex physical systems.

  • Project
    facebookresearch/silk

    SiLK is a self-supervised framework for learning image keypoints that emphasizes simplicity and flexibility while achieving state-of-the-art results on benchmarks. It eliminates the need for manual annotations and provides pre-trained models and tools for training, evaluation, and deployment.

  • Project
    JordiCorbilla/stock-prediction-deep-neural-learning

    A research laboratory for building and evaluating reproducible deep-learning models for financial time-series forecasting, focusing on out-of-sample performance and realistic baseline comparisons.

  • Project
    KarhouTam/FL-bench

    FL‑bench is an open‑source federated‑learning benchmark that lets you run dozens of classic and recent FL algorithms on many image datasets. With a single command you can switch methods, adjust hyper‑parameters via YAML or CLI, and scale training using Ray. It includes ready‑made models, dataset partitioning tools, and monitoring support (TensorBoard/Visdom), and is extensible for custom algorithms, models, or data.