Troubleshooting

Use this guide to diagnose and fix common deployment issues. Start with the symptom you're seeing, then follow the steps.

What's happening?

How to Check Logs

Most issues can be diagnosed from the pod logs:

1. Go to your Dashboard
2. Click the deployment card or the ⋮ menu
3. Click View Logs
4. Look for lines with ERROR or WARNING

If logs show "Container starting up..." with a spinner, the container is not ready yet. Read the provider and startup logs to distinguish image pulling from model downloads or a failed process.

⏳ Deployment Stuck on "Starting"

Elapsed time alone cannot tell you whether startup failed. Read the current phase and the latest relevant log entry. Model downloads can continue while a GPU is already billable, so check your budget as well as progress.

• Pulling or downloading: Check whether file downloads are progressing. A new machine may need files that a previous one already had.
• Installing nodes or starting the interface: Look for the first missing dependency, failed import, or memory error.
• Interface ready: Open it from the dashboard. Readiness means the service responds; it does not prove your chosen graph or input will succeed.
• No progress or an explicit error: Save the relevant error and inspect the deployment before retrying. Stop compute if you do not want it to keep consuming credits.

Why it's slow: First-time deployments may need large model downloads and custom-node installation. Reused images or retained files can shorten a later start, but a fresh machine may need to download them again.

❌ Deployment Failed

Check your logs for these common errors:

GPU Memory Exhausted (CUDA OOM)

Log message: CUDA out of memory or OutOfMemoryError

The model needs more VRAM than your GPU has.

• Upgrade to a GPU with more VRAM (L4 24GB → A6000 48GB → A100 80GB)
• Use a smaller model variant (e.g., 8B instead of 32B)
• Reduce image resolution or batch size in your workflow

Model Not Found

Log message: Model not found or No such file .safetensors

• Check the model name is spelled correctly
• For workflows: verify model filenames match what ComfyUI expects
• For gated models: add your HuggingFace token in settings

Missing Python Package

Log message: ModuleNotFoundError or No module named

• A custom node requires a Python package not in the Docker image
• Open ComfyUI Manager → Install Missing Packages → Restart
• Or redeploy to trigger a fresh custom node installation

GPU Driver Error

Log message: CUDA error or cuDNN error

• Restart the pod to reset GPU state
• Try a different GPU type
• This is usually a transient hardware issue on the RunPod node

Process Killed (OOMKilled)

Log message: Killed or OOMKilled

• The system ran out of RAM (not VRAM — system memory)
• Use a GPU with more system RAM
• Reduce the complexity of your workflow

Disk Full

Log message: No space left on device

• The selected instance’s container disk is full — often from model downloads or generated files. Capacity depends on your configuration.
• Delete unused generated files from the ComfyUI output folder
• Check the configured disk capacity before retrying; CPU and GPU configurations can differ.

Gated Model (Access Denied)

Log message: gated model or access token required

• Some HuggingFace models require you to accept a license first
• Go to the model page on huggingface.co and accept the terms
• Add your HuggingFace token to your ModelPilot account settings

🔒 GPU Unavailable

Error: "The selected instance type is temporarily unavailable"

This means RunPod's data centers have no machines with your selected GPU available for this request. Review any offered fallback and its rate before trying another launch.

• Try a different GPU: Check the offered instance’s memory, rate, and compatibility before accepting a fallback
• Wait and retry: GPU capacity can change; inspect the current failure before retrying
• Check existing resources: A failed launch attempt does not prove that every allocated resource was removed. Use the dashboard to inspect what remains

🐌 Model Running But Slow

• Wrong GPU for the model: Large models (32B+ parameters) on a 24 GB GPU will swap to system memory. Upgrade to A6000 (48 GB) or A100 (80 GB).
• First request is slow: The model loads into GPU memory on the first request. Subsequent requests are much faster.
• High resolution / large batch: Reduce image dimensions or batch size in your ComfyUI workflow.

🔌 API Errors

Still Need Help?

If your issue isn't covered here:

• Email support@modelpilot.ai with your deployment ID and the error from your logs
• Check the FAQ for more answers