What's in this release

What’s new

Adds visual management for model files in public and private model registries.
API keys now support group management, so administrators can quickly identify the service ownership of API keys.
After a cluster is created, you can update the Kubeconfig of a Kubernetes cluster and the SSH private key of a static node cluster.
The CLI can be used to push any container image to the image registry managed by Neutree. This can be used to import custom workloads or supporting images.
You can deploy Neutree control plane in Kubernetes through NodePort and use Helm to configure a unified external access address for the model gateway, VictoriaMetrics, and Grafana.
Access logs support filtering requests by common HTTP status code suggestions or custom three-digit status codes.

Improvements

Unifies the UI visual style and improves the user experience on the model registry, model catalog, endpoint, and cluster accelerator resource pages.
Optimizes model deployment by dynamically calculating KV cache and VRAM requirements based on the model.
When a static node cluster uses the vLLM or SGLang engine and multiple accelerators are configured, if the user does not explicitly configure the tensor parallelism parameter, the system automatically sets this parameter based on the number of requested accelerators.
When a single upstream in an external endpoint is abnormal, other available upstreams can continue to serve requests.
When using the CLI to view models in a file system model registry, pagination, search, and structured output are supported, along with displaying the model alias and the source of model information fields.
The endpoint details page automatically displays or hides the Playground based on the capabilities declared by the engine version.
Re-importing a model catalog with the same name in the same workspace automatically updates the existing model catalog.
Model registries support the ModelScope type. You can view model information and README.md in a drawer on the model registry details page.
When creating or editing an endpoint, you can view the VRAM requirements, model information, and KV cache estimate declared by the model catalog in the Weights & Cache area.
Kubernetes endpoints support configuring the startup timeout through deployment_options.startup_timeout_seconds. If a container restarts more than 5 times and is still not ready, the system marks the endpoint as failed and displays the failure reason.

Resolved issues

When the CLI was used to query resources, an empty resource list returned an unstructured message, causing scripts to fail to parse the output reliably. The issue has been resolved in this release.
When the CLI --server-url parameter value contained a trailing /, the request path might have contained duplicate slashes and returned 404. The issue has been resolved in this release.
In static node clusters, custom engines that do not integrate built-in model downloading capabilities might have been incorrectly marked as Model Downloading. The issue has been resolved in this release.
After a model catalog was edited or an invalid YAML file was imported, the model catalog list or details page might have become blank. The issue has been resolved in this release.
Writing an empty Spec to a model catalog might have cleared saved configurations. The issue has been resolved in this release.
After a model catalog was deleted, the list might not have refreshed automatically based on the asynchronous deletion result. The issue has been resolved in this release.
In Kubernetes clusters, neutree-node-agent might have been scheduled to control plane nodes. The issue has been resolved in this release.
When metrics remote write was not configured for a static node cluster, node-agent was not deployed, which caused node and resource allocation information to be unavailable. The issue has been resolved in this release.
When an HTTP image registry was used, the connectivity check might have incorrectly accessed it over HTTPS, causing the check to fail. The issue has been resolved in this release.
When resources such as workspace policies or API keys were created, the workspace selector could not select workspaces outside the first page. The issue has been resolved in this release.
When a custom date range was reselected, the old start date might have been incorrectly retained. The issue has been resolved in this release.
Force-deleting an API key might have left credentials still usable when the gateway was unreachable. The issue has been resolved in this release. When deleting an API key, gateway credentials must be revoked first. If revocation fails, the record is retained and retries continue.
When a Kubernetes cluster name consisted only of digits, cluster component deployment failed, causing cluster creation to fail. The issue has been resolved in this release.
Editing an endpoint might have allowed changing its cluster, which could leave residual resources occupied in the original cluster. The issue has been resolved in this release.
When creating or editing an endpoint, the precision of the manually entered accelerator card count might not have been restricted based on the cluster type, allowing an invalid resource configuration to be submitted. The issue has been resolved in this release.
When deploying an endpoint from a model catalog, the VRAM validation prompt might have shown a display value that was inconsistent with the validation result in boundary value scenarios. The issue has been resolved in this release.
In static node clusters, the GPU metrics collection component and node-agent might not have explicitly selected the nvidia runtime, which could cause GPU access capability to be lost after the runtime environment changed. The issue has been resolved in this release.
The License API might have accepted non-empty trailing content after valid JSON and processed an invalid request body as a valid request. The issue has been resolved in this release.
When an endpoint or external endpoint name contained non-ASCII characters such as Chinese characters, access logs and usage statistics might have displayed or filtered percent-encoded names. The issue has been resolved in this release.
When cluster accelerator virtualization configurations were partially updated through the API, a misleading “Only Kubernetes clusters are supported” error might have been returned. The issue has been resolved in this release.
When an external endpoint name contained invalid characters such as Chinese characters, spaces, or uppercase letters, the external endpoint might have been created successfully but the gateway route could not be accessed. The issue has been resolved in this release.
In a single-node HAMi cluster, updating the scheduling component might have failed to complete the rolling update. The issue has been resolved in this release. This fix also prevents enabling accelerator virtualization from affecting running GPU endpoints.