skills.rest
Cloud-Byte-Consulting
Cloud-Byte-ConsultingOFFICIAL
Explore all skills by this author →
GitHub

gpu-sharing-advisor

Recommends MIG, MPS, or time-slicing configurations for sharing NVIDIA GPUs on Kubernetes.

Updated Jul 16, 2026
Search Tags:#kubernetes#nvidia#mps#gpu utilization#mig#time-slicing#gpu sharing
One-click install
npx skills add https://github.com/Cloud-Byte-Consulting/plugins --skill gpu-sharing-advisor-cloud-byte-consulting
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: gpu-sharing-advisor
Source: https://github.com/Cloud-Byte-Consulting/plugins/tree/main/gpu-research-platform/skills/gpu-sharing-advisor
Command: npx skills add https://github.com/Cloud-Byte-Consulting/plugins --skill gpu-sharing-advisor-cloud-byte-consulting

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Whole-GPU allocation in Kubernetes leaves expensive NVIDIA GPUs mostly idle when workloads are small or bursty, and teams lack a clear method for choosing between MIG, MPS, and time-slicing or producing the correct device plugin configuration. ## Core Features & Use Cases - Sharing technique selection: Compares MIG, MPS, and time-slicing across isolation, fault tolerance, overhead, and hardware support, with decision rules per workload pattern. - Utilization waste analysis: Computes memory and compute waste from DCGM metrics (DCGM_FI_DEV_GPU_UTIL, DCGM_FI_DEV_FB_USED) to justify sharing before recommending anything. - Concrete configuration output: Produces NVIDIA GPU Operator MIG strategies (single vs mixed) and time-slicing ConfigMaps with replica counts sized from model footprints. - Use Case: A research cluster has ten replicas of a distilled 1B model each reserving a whole 24 GB L4. The skill quantifies the 92% memory waste, recommends time-slicing with 10 replicas, and outputs the ConfigMap plus verification commands. ## Quick Start Ask the assistant to evaluate whether your GPU workloads should use MIG, MPS, or time-slicing and to generate the NVIDIA device plugin configuration for your cluster.

Frequently Asked Questions about gpu-sharing-advisor

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I share one NVIDIA GPU across multiple Kubernetes pods?▼

Use MIG for hardware-isolated partitions on Ampere+ GPUs, time-slicing for oversubscription on any CUDA GPU, or MPS for cooperative multi-process workloads. Time-slicing is configured via a device plugin ConfigMap that advertises multiple virtual replicas of nvidia.com/gpu.

MIG vs time-slicing: which GPU sharing method should I use?▼

Choose MIG when you need memory protection, fault isolation, or strict tenant boundaries, since it partitions the GPU at hardware level. Choose time-slicing for non-critical dev/test and small-model inference where maximum packing matters more than isolation.

Does MIG work on L4 or older NVIDIA GPUs?▼

No, MIG requires Ampere-generation or newer hardware such as A100, H100, or H200. On L4, L40S, or pre-Ampere cards, time-slicing is the only sharing option available through the NVIDIA device plugin.

Can I request half a GPU in a Kubernetes pod spec?▼

No, GPU resources must be set in limits as whole integers and cannot be fractional. A fraction of a GPU is expressed by requesting a MIG slice such as nvidia.com/mig-1g.18gb or one virtual replica from a time-sliced device.

When should GPUs not be shared between workloads?▼

Training and fine-tuning jobs that saturate the GPU should get exclusive whole devices, since sharing only adds contention. Sharing across untrusted tenants is also unsafe with MPS or time-slicing because neither provides memory or fault isolation.

Related Skills

Explore More Software Engineering Skills→
coreweavecoreweave
7

setup-sunk-gpu-monitoring

Set up GPU and Slurm metrics collection on a SUNK cluster.

Official
Advanced
vesvietvesviet

setup-gpu-finops

Configure GPU telemetry, DCGM metrics, and Kubecost to attribute AI compute costs per namespace.

Community
Advanced
TzJ2006TzJ2006
1

slurm-gpu

Parse sinfo output to map GPU resources across Slurm cluster partitions.

Community
Advanced
BagelHoleBagelHole
46

gpu-kubernetes-operations

Manage GPU-backed Kubernetes clusters for AI inference and training.

Community
Advanced