Errors, Statuses & Troubleshooting

Error Response Format

OpenAI-compatible inference (preferred)

Elastic OpenAI-style routes (/api/chat/completions, /api/embeddings, /api/models, /api/completions, /api/responses, …) return a dual shape — object error for OpenAI SDKs, plus top-level message for older parsers:

{
  "error": {
    "message": "Human readable message",
    "type": "invalid_request_error",
    "code": "optional_code",
    "param": null
  },
  "message": "Human readable message"
}

Prefer error.message or top-level message. Do not assume error is a bare string.

Legacy / other APIs

Some older or non-OpenAI surfaces still use:

{
  "error": "Human readable message"
}

Some APIs also return a machine-readable code field (for example Chronos). HTTP status codes are meaningful.

Common Errors (Elastic & Dedicated)

Dual-shape OpenAI errors include error.type (e.g. rate_limit_error on 429). Prefer that and error.message over assuming a bare string.

Chronos Errors

Full Chronos error table: Chronos → Errors.

Dedicated Endpoint Statuses

See the Dedicated Inference page for the full table.

Key ones for debugging (dashboard product language in quotes):

• building — PROVISIONING. Normal during first 2–10 minutes. Do not call yet.
• up — ONLINE. Healthy.
• hibernated — SLEEPING (scale-to-zero). First request wakes the endpoint.
• waking — WAKING UP. Brief delay; status usually updates within about a minute.
• failed — FAILED. Check Logs. Common causes: model too large for hardware, gated model without token, container crash.

Terminate removes a dedicated endpoint from the dashboard detail page; it is not an API error state. See Dedicated Inference → Terminate.

Debugging Checklist

Elastic Calls Failing

1. Verify the key is active in Settings → Keys.
2. Confirm you are using a valid display name from the catalog.
3. Try the exact same request in the dashboard playground (isolates client vs server issues).
4. Check that your client properly handles SSE (many issues are on the consumer side).
5. If unpaid, confirm free queries remain or prepaid balance is positive (Settings → Billing).
6. On 429 / 503, retry with exponential backoff. Some models may already have tried an automatic fallback; a remaining error means capacity is still exhausted.

Dedicated Endpoint Not Responding

1. Confirm status is ONLINE (up) in the dashboard — not PROVISIONING, FAILED, or mid-WAKING UP.
2. Check the Logs tab for runtime startup errors or out-of-memory conditions.
3. If you used a gated model, confirm the HF token was valid at provisioning time.
4. Try the in-dashboard playground bound to that endpoint ID — it removes network variables.
5. If tools fail only on dedicated, check prepaid balance (managed tools use prepaid, not the weekly GPU invoice).

Slow First Token / High Latency

• Large models on Elastic can have occasional cold-start latency on first request.
• Dedicated endpoints that have scaled to zero (SLEEPING) will wake on first request (WAKING UP). You'll see a clear message in the playground; allow a short delay — status usually updates within about a minute — before retrying. Cold starts are faster than a full model reload thanks to cached model weights.
• Very high max_tokens or complex prompts increase time-to-first-token.

Long requests / agent wall clock

Elastic completions (stream or non-stream) share a platform wall-clock limit of about 13 minutes per request. That is the serverless function budget for the whole call (model + tools + managed loops), not only “time to first token.”

Streaming remains available for progressive UI; it is not a way around the ~13 minute wall clock.

If timeouts persist on short requests, contact support@hypervize.tech.

Streaming Client Issues

If you are not seeing tokens:

• Make sure you are reading the response as a stream and parsing lines that start with data: .
• Do not buffer the entire response.
• Handle data: [DONE]\n\n as the terminator.
• Some frameworks (especially older fetch wrappers) have poor SSE support — consider using a dedicated library (eventsource, openai SDK, etc.).

See Code Examples for robust client patterns.

Related

• Dedicated Inference — statuses, scale-to-zero, terminate
• Billing & Payments
• Chronos
• Tool Calling

Still Stuck?

Reach out to support@hypervize.tech.