lambda-labs_skill

Overview

This skill provides a practical guide to running ML training and inference on Lambda Labs GPU Cloud. It highlights how to launch reserved and on‑demand GPU instances, attach persistent filesystems, and run single-node or multi-node distributed jobs with simple SSH access. The content focuses on fast setup, cost-aware GPU selection, and production‑grade cluster workflows.

How this skill works

The skill explains the console, API, and CLI steps to create instances, add SSH keys, and attach persistent filesystems. It documents available GPU types, preinstalled Lambda Stack software, verification commands, and examples for launching single‑GPU, multi‑GPU, and Slurm 1‑Click clusters. It also covers networking, storage patterns, checkpointing, and cost‑optimization tactics for ML workloads.

When to use it

• You need dedicated GPUs with full SSH access and a familiar Linux environment.
• Running long training jobs that require persistent checkpoints and large datasets.
• Launching high‑performance multi‑node training (16–512 GPUs) with InfiniBand and NCCL.
• Quickly prototyping or fine‑tuning models with preinstalled Lambda Stack and JupyterLab.
• When you prefer simple pay‑per‑minute pricing with no egress fees and global regions.

Best practices

• Choose the GPU type that matches model size and throughput (A10/A6000 for inference, H100/B200 for large training).
• Attach a Lambda filesystem for datasets and checkpoints to avoid repeated downloads and data loss.
• Checkpoint frequently to enable interrupts and resumption; store checkpoints on persistent filesystems.
• Right‑size instances—start small for development and scale to multi‑GPU for final training to reduce cost.
• Use SSH tunneling for secure JupyterLab/TensorBoard access and open only required ports in the console.

Example use cases

• Fine‑tuning a 7B LLM on an 8x A100 or 1x A100 instance with checkpoints stored on a Lambda filesystem.
• Large‑scale distributed training with a 16–512 GPU 1‑Click Slurm cluster using NCCL and InfiniBand.
• Batch inference on cost‑effective A10 or A6000 instances reading models from persistent storage.
• Development and debugging: spin up a single‑GPU instance, run experiments, then scale to multi‑GPU nodes for production training.