Vast.ai Common Errors

Overview

Quick reference for the most common Vast.ai errors across CLI, REST API, and instance operations. Vast.ai uses HTTP status codes for API errors and instance status strings for machine-level issues.

Prerequisites

• Vast.ai CLI installed (pip install vastai)
• API key configured

Instructions

API Errors

Instance Status Errors

Common CLI Errors

# Error: "No offers found"
# Cause: Filters too restrictive
# Fix: Relax filters
vastai search offers 'num_gpus=1 rentable=true' --limit 5  # broader search

# Error: "Insufficient funds"
# Fix: Check balance and add credits
vastai show user --raw | python3 -c "import sys,json; d=json.load(sys.stdin); print(f'Balance: \${d[\"balance\"]:.2f}')"

# Error: "Instance creation failed"
# Fix: Try a different offer or smaller disk
vastai create instance OFFER_ID --image ubuntu --disk 10

Docker Image Errors

# Error: Instance stuck in "loading" for >10 minutes
# Cause: Very large Docker image or slow host internet
# Fix: Use smaller base images or pre-cached templates
# Good: pytorch/pytorch:2.2.0-cuda12.1-cudnn8-runtime (4GB)
# Bad:  custom-image-with-everything:latest (30GB)

# Error: "Container exited immediately"
# Cause: Bad entrypoint or missing dependencies
# Fix: Check logs
vastai logs INSTANCE_ID --tail 50

SSH Connection Errors

# Error: "Connection refused" after instance shows "running"
# Cause: SSH server not yet started inside container
# Fix: Wait 30-60 seconds after status changes to running

# Error: "Permission denied (publickey)"
# Cause: SSH key not uploaded to Vast.ai
# Fix: Upload at cloud.vast.ai > Account > SSH Keys

# Error: "Host key verification failed"
# Fix: Use -o StrictHostKeyChecking=no for ephemeral instances
ssh -p PORT -o StrictHostKeyChecking=no root@HOST

GPU/CUDA Errors

# Error: "CUDA not available" inside container
# Cause: Docker image missing CUDA or driver mismatch
# Fix: Use NVIDIA-provided images with matching CUDA version
# Check host CUDA: vastai show instance ID --raw | jq '.cuda_max_good'

# Error: "CUDA out of memory"
# Cause: Model too large for GPU VRAM
# Fix: Search for more VRAM or reduce batch size
vastai search offers 'gpu_ram>=48 num_gpus=1' --order dph_total

Output

• Identified error with matching resolution
• Corrected configuration or command
• Verification that the fix resolves the issue

Error Handling

Resources

• Vast.ai Documentation
• Vast.ai Status Page
• CLI Reference

Next Steps

For comprehensive debugging, see vastai-debug-bundle.

Examples

Instance won't start: Run vastai show instance ID --raw | jq '{actualstatus, statusmsg}' to get the status message. If error, destroy and reprovision on a different host.

Billing surprise: Run vastai show instances to check for forgotten running instances. Destroy any you're not using to stop charges.