Troubleshooting
The usual suspects, in the order people hit them.
macOS: Gatekeeper
Release builds are signed with a Developer ID and notarized by Apple, so a normal install shows no warning. If you see "CozyGPU" cannot be opened because the developer cannot be verified or "CozyGPU" is damaged and can't be opened, you most likely have a beta or a build that was moved between machines:
- Right-click the app in Applications, choose Open, then Open again in the dialog.
- Or clear the quarantine flag in Terminal:
xattr -d com.apple.quarantine /Applications/CozyGPU.app - Still blocked? Download the installer again from Download and compare the checksum.
Windows: SmartScreen
Windows shows Windows protected your PC for installers signed with a certificate that has not yet built up reputation with Microsoft. New apps go through this; the warning fades as installs accumulate. Click More info, check the publisher name, then Run anyway. If the dialog shows Unknown publisher instead of a name, stop and download again: the installer is always signed.
The first run is slow
The first job on a new machine can take 10 to 30 minutes while RunPod pulls the model (about 10 for image models, up to 30 for video) image to a worker (measured 11 minutes for Z-Image Turbo, about 10 for Qwen Image 2.1, about 15 expected for MiniMax H3). The app shows the warm-up job's state so you can tell it is working. This happens once per endpoint. Afterwards:
| Model | Warm | Cold start after idle |
|---|---|---|
| Z-Image Turbo | about 5 s | about 25 s |
| Qwen Image 2.1 | about 25 s | about 90 s |
| MiniMax H3 | about 2 min per 5 s clip | about 4 min |
Occasionally RunPod places a worker on a host that has not cached the image, and the pull repeats. FlashBoot is on for every endpoint to keep this rare. If a job sits in queue for much longer than the numbers above, check the endpoint in the RunPod console: the GPU pool may have no capacity in your datacenters right now.
RunPod balance and spend limit
- Jobs fail immediately with a billing error when your RunPod balance is zero. Top up in the console under Billing; the app does not retry on its own.
- Set a spend limit under Billing in RunPod if you want a ceiling. Hitting it stops workers mid-job; the job shows as failed in the library and you are charged only for the seconds used.
- The app only ever runs one worker per model. Costs are computed from the job's execution time and the pool price and shown per job and per month in Settings → Compute; RunPod's own invoice is authoritative.
Idle endpoints
Each model gets one serverless endpoint named cozygpu-<model>. Workers stop ten seconds after the last job and an endpoint with no running workers costs nothing. You do not need to delete endpoints between sessions; if you do delete one in the console, the app recreates it on the next job (and pays the first-run pull again). Network volumes, if you created any through LoRAs or large inputs, are billed monthly by RunPod while they exist; delete them in the console if you stop using those features.
"Volumes missing" or S3 errors
Custom LoRAs and inputs over 7 MB need the RunPod S3 API key and at least one network volume. The app offers to create the volumes when it first needs them. If the S3 key is rejected, create a new one under Settings → S3 API Keys in RunPod and paste both parts; see Keys.
Logs and support
The service log is logs/service.log in the app's data folder (Settings → About → Open logs). Send the relevant part to support@cozygpu.com with your OS and app version; the log never contains your keys.