> See the full index of documentation pages at https://docs.zettabyte.com/llms.txt.

# Common issues

> Find your symptom, then apply the fix. Grouped by resources, storage, serving, training, images, and environment.

Source: https://docs.zettabyte.com/docs/troubleshooting/common-issues

Last updated: 2026-07-21

This page is indexed by symptom. Scan for the line that matches what you are seeing, then follow the fix and the linked page for the full procedure.

## Resources and billing

### An instance keeps billing while it sits idle

Billing is per allocation: you are charged from the moment a resource is allocated to you until you release it, whether or not a task is running. Release the resource when you are done to stop the charge. See GPU instances and How billing works.

### I cannot stop charges without losing the instance

There is currently no pause for an instance. Releasing is the only way to stop charges, and it returns the capacity to the pool. Save anything you need to a PVC or S3 first, because data not on persistent storage is lost on release. See Storage and PVCs.

### A task is queued and will not start

The capacity you requested is not free right now, or your task priority is low. Check the resource queue and your task's **Priority** (1 to 6). See The resource queue and Run a training job.

## Storage and data

### My outputs or data disappeared after the task ended

Anything not written to a persistent volume (a PVC) or to S3 is deleted when the task ends. Ephemeral and in-container data do not survive. Mount a PVC and write your checkpoints and weights there. See Storage and PVCs.

### I cannot mount two volumes on one task

Mount directories must be unique within a task. Give each volume a distinct mount path. See Storage and PVCs.

## Deploying and calling models

### The endpoint is stuck on Pending

The replicas are still scheduling, or capacity is short. Wait until Scaling shows ready over target (for example `1/1`). If it stays Pending, check the resource queue. See Deploy a model and The resource queue.

### The inference call returns 401 or the wrong service

A `401` usually means the API key lacks the **Inference Service** permission or has expired. A wrong or empty response usually means a malformed `Service-Name` header: it must be your tenant ID, a dot, then the endpoint's job alias name (`TENANT_ID.JOB_ALIAS_NAME`). See Call your endpoint and API keys.

## Training

### A distributed (PyTorch or MPI) task fails immediately

The framework is set to **PyTorch** or **MPI** without multi-node resources. Use the **Simple** framework for single-node jobs, or allocate more nodes for a genuinely distributed run. See Run a training job.

### Training hangs during data preprocessing

This usually means too many worker processes are spawned. Lower `Data_Num_Proc` to a single-digit value (for example `4` or `8`) instead of the default, which matches the CPU core count.

### TorchDynamo compilation errors

Disable compilation by setting these environment variables to `1`, for example in your task or IDE **Environment Variables**:

```bash
TORCHDYNAMO_DISABLE=1
TORCH_COMPILE_DISABLE=1
```

## Container images

### I cannot find the control to create a container image

> [!NOTE]
> **Known issue being fixed**
> A front-end issue can hide the create-image control in some browser states. This is being fixed. If you hit it, contact support rather than relying on a workaround. See Container images.

## Environment and downloads

### Slow model downloads

Slowness usually comes from instability in the Hugging Face `xet` protocol or gateway throttling. Pre-download the model with `huggingface-cli download <model-id>`, or allow the required domains in your network policy.

### Jupyter WebSocket connection failures

These come from cross-origin restrictions. Add the `--NotebookApp.allow_origin` flag (set to your frontend service URL, or `'*'`) to the command that launches Jupyter. See Notebooks and IDEs.

### Network connection problems

Check that your connection is stable and review any proxy or VPN settings. Make sure outbound ports `80` and `443` are open. Throttling can be temporary, so retry after a few minutes.