Troubleshooting
Most issues fall into one of three buckets: auth, launch, or runtime. This page groups the error codes you will see in logs and the dashboard, and the recovery path for each.
Always include the request id
The platform tags every response with an
X-Request-Id header. Capture it when you file a support ticket; it is the fastest way for the team to find the trace in the platform logs.Auth errors
| Code | Meaning |
|---|---|
| AUTH_MISSING_TOKEN | The Authorization header is missing. Add `Authorization: Bearer <key>` to the request. |
| AUTH_INVALID_TOKEN | The token does not match any key. Double-check the prefix and the secret; copy the value from your secret manager again if needed. |
| AUTH_REVOKED_TOKEN | The key was revoked. Create a new key and roll the change out to all clients. |
Launch errors
| Code | Meaning |
|---|---|
| BILLING_BELOW_MINIMUM | Account balance is below the launch threshold. Top up and retry. |
| GPU_INCOMPATIBLE | The model requires more VRAM than the selected GPU provides. Pick a larger GPU or a smaller model. |
| SMART_NO_MATCH | No provider satisfies the smart constraints. Loosen the max price, region, or capacity tier. |
| PROVIDER_TIMEOUT | The provider did not allocate capacity in time. Stop the instance and start it again; the scheduler will re-route. |
Runtime errors
| Code | Meaning |
|---|---|
| OOM_KILLED | The runtime exceeded available VRAM. Reduce the context size or pick a larger GPU. |
| CTX_TOO_LARGE | The request exceeded the configured context size. The platform returns a 413 with the limit. |
| FIREWALL_BLOCKED | A firewall in block mode matched a rule. The response includes the matched rule slugs. |
| JUDGE_UNREACHABLE | The firewall evaluator instance is not running. Start the judge or switch the firewall to audit mode. |
General recovery
- Use the dashboard: most failures expose a remediation hint on the instance or balancer detail page.
- Inspect the failure_reason: the platform stores a stable enum plus a one-line human explanation.
- For runtime errors, switch to a smaller context and re-run.
- For launch errors, check the billing balance and the smart constraints before retrying.
- For auth errors, regenerate the key and update every client in a single deploy.
- If you cannot recover after 15 minutes, capture the request id (X-Request-Id header) and open a support ticket.