Skip to content

[Batch] Template and profile registry never called#2626

Description

@CarolWinddd

馃悰 Describe the bug

_load_batch_k8s_context in aibrix/metadata/app.py builds the CoreV1Api/AppsV1Api k8s clients under --enable-k8s-support but never instantiates the TemplateRegistry or ProfileRegistry from the fixed-name ConfigMaps, so both stay None. Any batch that references a model template by name (e.g. {"model_template": {"name": "mock-vllm"}}, no inline spec) then fails in JobManifestRenderer with RenderError("template registry is not configured"), surfaced as inference_failed on the batch.

Steps to Reproduce

  1. Install via Helm/VKE instructions and apply samples/batch/batch_v1alpha1_{batch_profiles,model_deployment_templates}.yaml.
  2. Submit a batch with aibrix.model_template.name = "mock-vllm" / profile.name = "prod-24h" and runtime.target = "KubernetesJob" (no inline spec), referencing only the ConfigMap entries by name.
  3. Poll the batch. Expected before fix: errors.data[0].message == "template registry is not configured".

Expected behavior

MDS startup, under --enable-k8s-support, reads POD_NAMESPACE, builds the two registries and loads before serving.

Environment

AIBrix version: released v0.7.0
Development environment: Kubernetes

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions