Skip to main content

Model cache

Use Model cache to download model weights once and reuse them in deployments. If Storage volume is empty, first create a model-cache volume.

Download a model from Hugging Face​

  1. On the Model cache page, click Download model.
  2. In Download a model, choose a Cluster and Storage volume.
  3. Leave Model source set to Hugging Face.
  4. Either type a HuggingFace repo id and Revision manually, or click Browse Hugging Face Hub.
  5. If you opened Browse HuggingFace Hub, use Search models…, select a result from the left-hand list, then click Use this model.
  6. For a private or gated model, ask the cluster administrator to save a Hugging Face token with access to that model in the cluster settings.
  7. Click Start download.

The model card shows download progress. Wait for the cache to be ready before deploying. Downloads use the token saved with the cluster; public models can download without one.

Download a model the volume already has​

If the storage volume you picked already holds that model and revision, Lens does not start the download. The dialog explains what is already there -- downloaded, queued, still downloading, or failed -- and waits for you:

  • Download anyway downloads it again. A finished or failed cache is refreshed from the source in place; a download that is still running simply continues. Either way the volume keeps one entry for the model, never a duplicate.
  • Cancel keeps what is already on the volume.

Download a model from Model Market​

  1. Click Download model.
  2. In Download a model, set Model source to Model Market.
  3. Click Browse Model Market.
  4. In Browse Model Market, use Search by model name, family, or repo... and the category tabs under Model categories to find a model, then click its model card.
  5. Back in Download a model, review the selected model details. If needed, click Change model to pick a different one.
  6. Choose the Cluster and Storage volume. Only ready volumes tagged with the Model cache purpose are eligible.
  7. Downloads use the HuggingFace token saved with the cluster (set in the cluster wizard) for private models.
  8. Click Start download.

Filter cached models​

  1. Use Target cluster to limit the page to one cluster, or leave it on All clusters.
  2. In Model cache storage, click a storage card to filter the list to that volume.
  3. Use Search by model to narrow the Cached models list.
  4. Use the type tabs such as All and any model-type tabs shown by the page.
  5. Use Rows per page and the pagination controls to move through longer result sets.

View download logs​

  1. On a model card, click the action with the tooltip View logs.
  2. In Download logs, choose a node from the dropdown if more than one node is listed.
  3. Click the button titled Refresh logs if you want an immediate refresh. The modal also refreshes automatically while it is open.
  4. Read the log output and the per-node status chips shown above it.

Retry a failed download​

  1. Find a model card that shows a failed state.
  2. Click the action with the tooltip Retry download.
  3. Watch for the page notice Retry started and monitor the card again from Progress or View logs.

If Job creation fails with a webhook x509 error, check the webhook certificate/CA and, if a proxy is configured, the control-plane proxy exclusions before retrying.

Sync cached models to new nodes​

  1. If the page shows the header action Sync new nodes (...), click it to start syncing every cached model that still needs to be copied to new nodes. Watch for the notice Sync started for N models (count of models queued), or Nothing to sync — all models are up to date if none needed it.
  2. To sync only one model, find a card that says new nodes need the model and click the action with the tooltip Sync new nodes. Watch for the notice Sync started for new nodes.

Delete a cached model​

  1. On the model card you want to remove, click the action with the tooltip Delete cached model.
  2. In Delete cached model, review the warning about the exact cache path that will be removed.
  3. Click Delete cached model to start the removal. The dialog stays open and shows Deleting while the files are being removed.
  4. Once deletion is confirmed, the list and status counts refresh and the dialog closes. If deletion or its status check fails, the dialog displays the error so you can retry.

Related page​

If you do not see a usable cache volume, register one on Storage first.

Use the downloaded model​

  1. Open the model in Model Market.
  2. Select the same Cluster and Model cache storage used for the download.
  3. Confirm the model is ready in that cache, then create the deployment.

Deployment creation requires Model Cache. Download the exact model revision in full first, and synchronize new local-disk nodes before deploying. The model server in ordinary provider-rendered deployments loads the read-only cache offline. Exact published manifests are not covered by the same enforcement; see deployment limitations. Change Hugging Face credentials and proxy settings through the cluster settings, not custom deployment environment variables.