skills.rest
yale-som-hpc
yale-som-hpcOFFICIAL
Explore all skills by this author →
GitHub

using-gpus

Align GPU allocations with active CUDA workloads on SLURM clusters.

5|1|Updated Apr 29, 2026
Search Tags:#gpu#cuda#hpc#resource-management#slurm#nvidia-smi#gpu-scheduling
One-click install
npx skills add https://github.com/yale-som-hpc/claude-code-marketplace --skill using-gpus
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: using-gpus
Source: https://github.com/yale-som-hpc/claude-code-marketplace/tree/main/plugins/hpc/skills/using-gpus
Command: npx skills add https://github.com/yale-som-hpc/claude-code-marketplace --skill using-gpus

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

GPU resources on the Yale SOM HPC cluster are scarce and expensive to waste. Idle GPU allocations block other users; this skill ensures GPUs are requested only when the code actively uses them, reducing wasted compute time and avoiding unnecessary resource contention.

Core Features & Use Cases

  • Enforces on-demand GPU provisioning by aligning GPU requests with active CUDA workloads (e.g., PyTorch, TensorFlow, JAX, RAPIDS).
  • Guides CPU-GPU separation, testing, and monitoring to quickly identify idle allocations and cancel them.
  • Provides practical sbatch guidance and best practices for incremental GPU usage in ML pipelines and research jobs.

Quick Start

Tell Claude Code to apply the GPU-on-demand policy to your project and validate with a small GPU workload.

Frequently Asked Questions about using-gpus

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

FAQPage Schema
How do I know if my PyTorch or TensorFlow job is leaving a GPU idle?▼

Diagnose idle GPUs by monitoring your allocation with nvidia-smi to check active CUDA utilization, ensuring your PyTorch, TensorFlow, or JAX workloads are actively processing data.

Does this GPU scheduling approach work with RAPIDS and JAX workloads?▼

Yes, the GPU-on-demand policy applies to common ML workloads including RAPIDS, JAX, PyTorch, and TensorFlow, aligning resource requests with active CUDA processing on SLURM clusters.

How do I cancel an idle GPU allocation on a SLURM cluster?▼

Apply guardrails to automatically cancel idle GPU allocations by monitoring nvidia-smi output, releasing scarce compute resources back to the HPC scheduler when no active CUDA work is detected.

What is the best way to start using GPUs incrementally for ML pipelines?▼

Start with a one-GPU-first approach to validate your ML pipeline with a small workload, then scale up GPU requests only after confirming active CUDA utilization and efficient resource management.

Related Skills

Explore More Data & Analytics Skills→
niznik-devniznik-dev
16

Run Experiment

Orchestrate SLURM fine-tuning jobs with torchtune and evaluate with inspect-ai.

Community
Advanced
qcmuuqcmuu

lambda-labs-gpu-cloud

Provision Lambda Labs GPU instances and Slurm clusters for ML workloads.

Community
Intermediate
LvcidPsycheLvcidPsyche
7

system_monitor

Monitor local Linux server CPU, RAM, and GPU usage via standard utilities.

Community
Basic
yale-som-hpcyale-som-hpc
5

self-diagnosing-resource-use

Diagnose Slurm job CPU, memory, and GPU utilization using sacct, seff, and nvidia-smi.

Official
Intermediate