Changes to move to new design

[asm582]

PR to implement the inferno controller using controller-runtime based on community review. This has a dummy optimizer implementation.

[atantawi]

Tested successfully by running on a local kind cluster.
Ran both the vllme loadgen and the variantautoscaling controller locally, external to the cluster.
Since, currently, the Prometheus address is hard-coded in the code, two changes had to be made:

• vllm_emulator: in client.py base_url="http://localhost:8000/v1"
• inferno-autoscaler: in variantautoscaling_controller.go Address: "http://localhost:9090"

/lgtm

[asm582]

Tested successfully by running on a local kind cluster. Ran both the vllme loadgen and the variantautoscaling controller locally, external to the cluster. Since, currently, the Prometheus address is hard-coded in the code, two changes had to be made:

• vllm_emulator: in client.py base_url="http://localhost:8000/v1"
• inferno-autoscaler: in variantautoscaling_controller.go Address: "http://localhost:9090"

/lgtm

Thanks, let me create an issue about it.

[atantawi]

Tested successfully on local cluster.

[clubanderson]

Hey — just a heads up that the NumReplicas int field definition in OptimizedAlloc ended up causing a subtle issue downstream (#731). Because it's a bare int without omitempty, the CRD generator marks it as required. This became a problem when later PRs introduced MergeFrom patches (#460) and scale-to-zero (#585) — the JSON merge patch omits the zero value, and the API server rejects the partial object.

The idiomatic Kubernetes pattern would be NumReplicas *int32 with json:"numReplicas,omitempty" (similar to how HPA defines minReplicas). We've got a short-term fix in PR #721 and will look at aligning the type definition as a follow-up. Just noting it here for context!