Skip to main content
Use this page when an endpoint cannot start, is only partially running, or an API request fails.

A location is sold out

A location marked Sold out cannot accept another endpoint at that moment. If you are choosing your Inference API location for the first time, select another available location or try again later. If your Inference API location has already been established, new endpoints inherit that location in the current production version. In that case, try another available model or variant in the same location, reduce the replica count where possible, or wait for capacity to become available.
Per-model location selection is not available yet. The console is the source of truth for current regional capacity.

I chose Global but the endpoint shows a specific region

This is expected. Choosing Global lets Hivenet assign an available region. After deployment, the endpoint details show the region where it was actually placed.

A model or variant is unavailable

Model and variant availability depends on your selected location and current capacity. Try one of these options:
  • Choose another available model or variant.
  • Reduce the replica count if the console offers a lower count.
  • Try again later after capacity has been released.

The organization does not have enough credits

The console can block a new deployment or restart when the organization does not have enough credits. Add credits when the option is available, or ask the organization Owner to top up the balance if you do not have billing permissions. Inference API and Compute resources use the same organization balance.

The endpoint is still Starting

The console says starting a model takes about 90 seconds. While the endpoint is Starting, it is not ready for normal traffic yet. Wait for Running before testing your integration. If it remains in Starting well beyond the expected period, note the endpoint name, model, region, and approximate start time before contacting Support.

The endpoint is Partially running

Partially running means fewer replicas are serving than requested. For example, 1/3 means one replica is serving out of three requested. The endpoint can still accept traffic, but available throughput is lower. You can:
  1. Use Retry when the action is available.
  2. Reduce the replica count if your workload can run with less capacity.
  3. Wait for capacity to recover.
  4. Stop the endpoint if you do not want it serving with reduced capacity.

The endpoint is in Error

Error means no healthy replica is currently serving requests. Start with these checks:
  1. Use Retry if the action is available.
  2. Check that the organization still has enough credits.
  3. Check whether the selected model configuration still has capacity in the current location.
  4. If the problem continues, collect the endpoint name, region, model, replica count, and approximate failure time before contacting Support.

The endpoint shows Retrying

After you select Retry, the endpoint can temporarily show Retrying while Hivenet attempts to recover failed serving replicas. Wait for the state to settle. A successful retry returns the endpoint to Running or Partially running, depending on how many replicas recover.

Retry is unavailable or keeps failing

The Retry action appears only when it applies to the current endpoint state. Retry does not start a deliberately stopped endpoint. Use Start for an endpoint that is Stopped. If Retry repeatedly fails, the underlying capacity or serving problem may still exist. Wait and try again later, reduce the requested replica count where possible, or contact Support if the endpoint remains unusable.

A replica change cannot be saved

Capacity can change while you are editing an endpoint. If the requested count is no longer available, choose the maximum currently offered by the console or try the change later. A failed resize leaves the existing endpoint configuration in place.

A stopped endpoint will not start

Starting a stopped endpoint requires:
  • Enough organization credits
  • Enough current capacity for its existing configuration
Stopping preserves the endpoint configuration, not the underlying GPU capacity. If capacity is unavailable, wait and try again later.

The API key is rejected

Check that:
  • You are using an Inference API key, not a Compute public API token.
  • The key is Active, not disabled or revoked.
  • The key belongs to the same organization as the endpoint.
  • The request sends the key as Authorization: Bearer ....
If the key was lost, revoked, or exposed, create a new one and update the application. See Create and manage Inference API keys.

The endpoint URL does not work

Copy the URL from endpoint details or from the copy control in the Deployments list rather than reconstructing it manually. The endpoint URL already includes the API version path, such as /v1. Do not add another /v1. For the documented Chat Completions interface, the direct request path is:
Also confirm that the endpoint is Running or Partially running. Stopped and terminated endpoints do not serve requests.

The request cannot find the model

Do not assume the endpoint name or shorter catalog title is the model identifier expected by the API. Copy the exact model value from the endpoint’s generated Quick-start snippets.

Requests are slower than expected

Check:
  • Whether the endpoint is Partially running.
  • How many replicas are currently serving compared with the requested count.
  • Whether the request uses long input, large output, or another expensive workload shape.
  • The inter-token latency and output-speed metrics in endpoint details when they are available.
The requests-per-minute figure shown for a model variant is a planning estimate for that serving configuration, not a guarantee for every request shape.

A terminated endpoint still appears

Termination permanently shuts down the endpoint, but its record can remain in the Deployments list. Use Clear to remove an endpoint that is already Terminated. Clear does not stop or terminate an active endpoint. It only removes the terminated record.

What to send Support

If you still need help, include:
  • Organization name
  • Endpoint name
  • Model and variant
  • Location
  • Endpoint status
  • Requested and active replica count, when visible
  • Approximate time the problem started
  • What action or request you were trying to perform
Do not send your full API-key secret.

Next steps