Skip to content

Mokka

Simulate your GPU infrastructure on CPU nodes.

Mokka

Mokka turns any Kubernetes cluster into a multi-GPU environment for testing. It implements the NVIDIA driver interfaces that GPU software talks to, so the device plugin, the DRA driver, the GPU Operator and nvidia-smi all behave as though real hardware were present. No physical NVIDIA GPU is required.

  • Quick Start

    Install into a KIND cluster and see simulated GPUs in five minutes.

    Get started

  • Architecture

    How the mock NVML library, the CGo bridge and the GPU profiles fit together.

    Read the design

  • Demos

    Six runnable walkthroughs, from a standalone install to NVSentinel health monitoring.

    Browse demos

  • Configuration

    Every YAML knob: profiles, topology, failure injection and dynamic metrics.

    See the reference

Try it

kind create cluster --name mokka

helm install nvml-mock oci://ghcr.io/nvidia/k8s-test-infra/chart/nvml-mock \
    --namespace mokka --create-namespace

Use make cluster-create instead of kind create cluster when you need the CDI-enabled Kind node image — that is what the device plugin, DRA driver and GPU Operator paths run on. make cluster-delete tears it down.

The Helm chart guide has the full walkthrough for each consumer, including the device plugin, the DRA driver, the GPU Operator and a multi-node heterogeneous fleet.

How it fits together

Mokka architecture

Components

Component Description Status
Mock NVML (libnvidia-ml.so) 400 NVML C API exports (111 with configurable behavior, 289 stubs), YAML-configurable GPU profiles Production
nvidia-smi Real binary with RPATH patch, backed by mock NVML Production
Helm Chart DaemonSet deployment with 7 GPU profiles Production
CDI Injection Container Device Interface specs for GPU Operator Production

GPU profiles

Profile GPU Name VRAM Architecture
gb300 GB300 NVL 288 GiB Blackwell Ultra
gb200 GB200 192 GiB Blackwell
b200 B200 192 GiB Blackwell
h100 H100 80GB HBM3 80 GiB Hopper
a100 A100-SXM4-40GB 40 GiB Ampere
l40s L40S 48 GiB Ada Lovelace
t4 Tesla T4 16 GiB Turing

Tested consumers

Consumer Role Status
NVIDIA Device Plugin nvidia.com/gpu extended resources Tested
NVIDIA DRA Driver Dynamic Resource Allocation Tested
NVIDIA GPU Operator Full stack device plugin, GFD and validator Tested
GPU Feature Discovery Node labeling from NVML Tested

Integrations

Integration Description
fake-gpu-operator Run:ai's K8s-level GPU simulation plus nvml-mock driver fidelity