How do I submit a job using this skill?

You can use standard Slurm commands like sbatch or the recommended ssubmit wrapper, which simplifies parameters for GPUs, CPUs, memory, and time limits.

Can I monitor my job's resource usage in real-time?

Yes, the skill provides links to specialized Grafana dashboards for cluster overviews, workload monitoring, and job-level resource consumption stats.

Which GPU architectures are supported?

The skill includes specific configurations for Oracle OKE clusters (NVIDIA H100 GPUs) and DO DOKS clusters (NVIDIA H200 GPUs).

How does data access work across clusters?

Both clusters use JuiceFS for unified data access at /data0/ or /data/srp/, maintaining consistent directory structures and permissions across development and cluster environments.

What is the best way to run interactive GPU sessions?

Use the 'sapptainer' command provided by the skill to start an interactive job with specific CPU, memory, and GPU requirements within a chosen container image.

Slurm GPU Cluster Management

Name: Slurm GPU Cluster Management
Author: SerendipityOneInc

bySerendipityOneInc

•

データサイエンスとML

Manages GPU-accelerated workloads on Slurm clusters using Apptainer containers for training and inference.

概要

This skill streamlines the workflow for developers interacting with SRP's Slurm clusters, specifically optimized for H100 and H200 GPU workloads. It provides comprehensive guidance for submitting jobs via sbatch or the ssubmit wrapper, managing Apptainer container environments, and configuring multi-node distributed training. By integrating cluster monitoring, unified JuiceFS data access patterns, and automated Feishu notifications, this skill helps users troubleshoot failures, monitor resource utilization, and maintain efficient high-performance computing operations directly within Claude Code.

主な機能

Configuration for multi-node distributed training (NCCL, Torch Distributed)
Native Apptainer container management and interactive session support
1 GitHub stars
Simplified job submission using sbatch and ssubmit wrappers
GPU resource monitoring for H100 and H200 clusters
Real-time log analysis and Slurm job troubleshooting

ユースケース

Managing batch inference jobs and data processing tasks at scale
Debugging Slurm job failures and optimizing resource allocation
Training and fine-tuning large-scale AI models on high-performance GPU clusters

概要

主な機能

Configuration for multi-node distributed training (NCCL, Torch Distributed)
Native Apptainer container management and interactive session support
1 GitHub stars
Simplified job submission using sbatch and ssubmit wrappers
GPU resource monitoring for H100 and H200 clusters
Real-time log analysis and Slurm job troubleshooting

ユースケース

Managing batch inference jobs and data processing tasks at scale
Debugging Slurm job failures and optimizing resource allocation
Training and fine-tuning large-scale AI models on high-performance GPU clusters