tabctx is starting: serving {{ join ", " .Values.models }} on {{ if gt (int .Values.gpus) 0 }}{{ .Values.gpus }} GPU(s){{ else }}the CPU{{ end }}.
The first start downloads the model weights and compiles kernels (minutes);
`kubectl get pods -w` until Ready. A restart drops every fitted table: clients fit again.

Try it:

  kubectl port-forward svc/{{ include "tabctx.fullname" . }} {{ .Values.service.port }}:{{ .Values.service.port }}
  curl localhost:{{ .Values.service.port }}/v1/models
{{- if .Values.requireTenant }}

Every /v1/tabctx request needs a tenant header (requireTenant=true), e.g.
  curl localhost:{{ .Values.service.port }}/v1/tabctx/datasets -H 'x-tabctx-tenant-id: acme'
tabctx trusts that header as sent: put the Service behind a gateway that
authenticates callers and sets it.
{{- end }}

Docs: https://vectorinstitute.github.io/tabctx/
