Notre avis

Guide l'utilisateur dans le lancement d'une instance GPU sur Vast.ai avec configuration complète.

Points forts

Automatise la recherche et la sélection d'offres GPU
Prend en charge les connexions SSH, Jupyter et Jupyter Lab
Permet de configurer l'image Docker, le disque, les volumes et les variables d'environnement
Gère les instances spot avec enchères

Limites

Nécessite un compte Vast.ai et l'outil CLI vastai configuré
Ne gère pas les configurations multi-nœuds complexes
Les instances interruptibles peuvent perdre les travaux en cours

Quand l'utiliser

Quand vous avez besoin d'une machine GPU rapidement pour de l'entraînement, de l'inférence ou du développement.

Quand l'éviter

Pour des serveurs persistants à long terme ou des tâches ne nécessitant pas de GPU.

Exemples

Launch a basic GPU instance

Launch a GPU instance on Vast.ai with a PyTorch image and SSH access.

Find and launch a low-cost GPU

Find a cheap GPU instance on Vast.ai with at least 16GB VRAM and 100GB disk, then launch it with Jupyter Lab.

Launch with volume and custom script

Create a Vast.ai GPU instance using the vllm image, attach an existing volume at /data, and run a startup script.

name: launch-instance description: "Launch a GPU instance on Vast.ai. Use when the user wants to create, start, or spin up a new GPU machine for training, inference, or development." argument-hint: "[offer-id or requirements]" disable-model-invocation: true allowed-tools: Bash

Launch a Vast.ai GPU Instance

Guide the user through launching a GPU instance on Vast.ai.

User Request

$ARGUMENTS

Instructions

Step 1: Determine the Offer

If the user provided an offer ID (number), use it directly.

If they described requirements, search first:

vastai search offers '<query>' -o 'dph_total'

Present top options and ask the user to pick one.

Step 2: Gather Configuration

| Setting | Flag | Default | |---------|------|---------| | Docker image | --image | Ask user (suggest pytorch/pytorch) | | Disk size | --disk | 64 GB | | Access method | --ssh / --jupyter / --jupyter-lab | --ssh | | Direct connect | --direct | No (not all machines support it) | | Startup script | --onstart-cmd | None | | Environment/ports | --env | None | | Pricing | --bid_price | On-demand (no flag) | | Label | --label | None | | Template | --template_hash | None | | Volume | --create-volume / --link-volume | None | | Mount path | --mount-path | None | | Entrypoint | --entrypoint | Container default | | Container user | --user | root |

Common images: pytorch/pytorch, nvidia/cuda:12.1.0-devel-ubuntu22.04, vllm/vllm-openai:latest, tensorflow/tensorflow:latest-gpu

Step 3: Launch

vastai create instance <OFFER_ID> \
  --image <IMAGE> \
  --disk <GB> \
  --ssh \
  [--env '<DOCKER_OPTS>'] \
  [--onstart-cmd '<SCRIPT>'] \
  [--label '<NAME>']

Alternatively, auto-select an offer:

vastai launch instance -g <GPU_NAME> -n <NUM_GPUS> -i <IMAGE> -d <DISK> --ssh

Step 4: Verify & Connect

After creation succeeds (note the new_contract instance ID):

# Wait for running status
vastai show instances

# Get SSH connection
vastai ssh-url <INSTANCE_ID>

Provide the user with the SSH command or Jupyter URL.

Step 5: Volume Setup (if needed)

# With new volume
vastai create instance <OFFER_ID> --image <IMG> --ssh --disk 64 \
  --create-volume <VOLUME_OFFER_ID> --volume-size 100 --mount-path /root/data

# With existing volume
vastai create instance <OFFER_ID> --image <IMG> --ssh --disk 64 \
  --link-volume <VOLUME_ID> --mount-path /root/data

Safety

Always confirm offer ID and $/hr cost before creating
Remind user to vastai destroy instance <ID> when done
Warn about interruptible instances losing work if preempted
For spot: use --bid_price — if bid is too low, instance won't start

Lancer une instance GPU sur Vast.ai

Recommandé pour