Skip to main content

Common issues

Quick index of user-reported symptoms. Each entry lists the likely cause and a concrete fix.

API calls​

401 Unauthorized​

Symptom: API calls return {"error": "unauthorized"}.

Causes and fixes:

  • Missing Authorization: Bearer eco_... header → add it
  • Typo in the key → copy-paste from the API Keys page
  • Key was revoked → create a new one
  • Using a key from a different account → check the account switcher in the console

402 Payment Required — insufficient_balance​

Cause: Your balance can't cover the request or the credit hold.

Fix: Stop resources you don't need to free held credit, or add balance.

402 Payment Required — model_not_priced​

Cause: Model ID is registered but has no pricing set yet.

Fix: Ping #ecolink-support with the model ID.

404 Not Found on a model​

Causes:

  • Typo in the model ID
  • For user inference: missing :<instance_id> suffix on the model field
  • For user inference: your instance was terminated (check Compute → Model Instances)

Fix: verify the ID. For platform models, call GET /platform/models to list what's deployed.

429 Too Many Requests​

Fix: back off with exponential delay, respect retry-after, batch requests where possible. See Rate limits.

503 Service Unavailable​

Cause: No healthy pod for the requested model at this moment.

Fix: retry after a few seconds. If it persists for minutes on a single model, report in #ecolink-support.

My API call is very slow​

Possible causes:

  • Cold-start on a platform model that scaled down while idle — first request after a quiet period takes ~15–30s extra
  • Long max_tokens — naturally slower completion
  • The router picked a further-away region (check x-ecolink-region)

Fix: For user inference that you control, keep at least one replica running to avoid cold-start. For platform models, warm-pod persistence is automatic.

GPU instances​

Instance stuck in pending for > 15 minutes​

Likely cause: image pull is slow, or the pod came up but the platform's watcher missed the Ready event.

What happens automatically: a 5-minute watchdog reconciles stuck-pending rows. If a pod is actually running in K8s, the status heals. If no pod exists, status flips to failed.

Fix: wait up to 20 min. If still stuck, terminate and relaunch. If consistently stuck with the same image, it may be huge (> 10 GB) — try a slimmer variant. Report the instance ID in #ecolink-support.

Instance says running but the terminal won't open​

Likely cause: the pod was preempted or died after the DB updated but before the terminal could attach.

Fix: refresh the detail page. If status is now preempted, wait for recovery. If failed or terminated, launch a new instance.

Duration expired and the instance auto-terminated before I extended​

Cause: estimated_duration_hours is a hard commitment. You can't recover the same instance.

Fix: launch a fresh one with the same config. If you had a cloud drive attached, reattach it — your files are still there.

insufficient credit balance on launch​

Cause: the credit hold for duration × GPU count × rate exceeds your balance.

Fix: shorter duration, fewer GPUs, or add credit.

GPU count exceeds GPUs per node​

Cause: more GPUs requested than a single node in the region has.

Fix: lower GPU count, or pick a different region.

The instance is up but my service never started​

Cause: the init script failed. On failure the script stops, the startup command is skipped on purpose, and the pod stays up with SSH working so you can look at it.

Fix: open the terminal or SSH in, read the pod log above the failure, then fix the script with Edit config on the instance detail page and redeploy. Remember the script re-runs on every pod start, so install into a mounted volume and guard with [ -d ... ] ||.

a redeploy is already in progress​

Cause: a previous Edit config is still rebuilding the pod.

Fix: wait for it. The marker clears automatically after 15 minutes if the rebuild never completed, and the request is accepted again. See Editing a running instance.

My template launched but the drive is empty​

Cause: templates record storage as a shape to create, not as the original volume. Launching one gives you new, empty volumes at the same sizes and mount paths — a cloud drive can only belong to one instance at a time.

Fix: expected behaviour. Put the download step in the template's init script so the new volume fills itself. See Compute templates.

Storage​

Can't delete a cloud drive​

Cause: it's attached to an active instance.

Fix: terminate the instance, then delete the drive.

Shared filesystem shows in_use but I think it's detached​

Cause: in_use = "at least one instance has it mounted." Another of your instances may still hold it.

Fix: check Compute → GPU Instances for instances with the filesystem attached. Terminate any leftover ones.

My instance terminated and my notebooks / model checkpoints are gone​

Cause: files in the container's writable filesystem are lost at termination.

Prevention: attach a cloud drive or shared filesystem at launch. Save anything you want to keep into its mount path (e.g. /workspace, /shared).

User inference​

Instance stuck in pending for > 20 minutes​

Likely cause: image pull OR model download (large HuggingFace repo).

Check: detail page shows per-region status pending → loading → running. Stuck at pending = image pull slow. Stuck at loading = model download or container init slow.

Fix: for 70B+ models, loading can legitimately take 15–30 min. If still stuck after 30 min, terminate and relaunch.

Requests time out / 504 Gateway Timeout​

Cause: container isn't responding to /health or is hanging.

Fix: confirm the container works locally with docker run --gpus all. Bind to 0.0.0.0:<service_port>, not 127.0.0.1. Make sure /health is implemented.

402 on my own user inference instance​

Cause: balance hit $0 on the previous 24h cycle; instance was terminated.

Fix: check Model Instances — it'll show stopped with reason credit_depleted. Restore balance and launch again.

Teams and invites​

The invitee never got the email​

Cause: email delivery for invites isn't always reliable right now.

Workaround: open the Users page → find the pending invite → copy the invite link → send it via Slack or another channel. The link works identically to an emailed one.

I accepted an invite but don't see the account in my switcher​

Cause: need to be logged into EcoLink before clicking the link.

Fix: log in at console.ecohash.com, then paste the invite URL into your browser. The new account will appear in the top-left switcher.

Still stuck?​

When reporting in #ecolink-support, include:

  • What you were trying to do
  • Exact error message + HTTP status
  • x-ecolink-request-id header (for API issues)
  • Resource ID (for console issues)
  • Approximate timestamp

Your privacy choices

Essential cookies are always on. You can change these choices at any time.