# Scheduling policy for GPU serving pods: a PriorityClass so vLLM preempts
# lower-priority work for a GPU slot instead of queuing behind it, and a GPU-aware
# ResourceQuota so a runaway Deployment/HPA/KEDA ScaledObject in the `inference`
# namespace can't request more `nvidia.com/gpu` than the cluster actually has — cheap
# insurance now that a card can advertise more than one schedulable GPU unit (see
# ../../bootstrap/prereqs/nvidia-device-plugin-timeslicing.yaml).
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: gpu-serving
  labels:
    app.kubernetes.io/part-of: inference-platform
value: 1000000        # above default (0), below cluster add-ons' system-node-critical
globalDefault: false
description: >-
  GPU-serving workloads (vLLM). Pods at this priority preempt default-priority pods
  for a GPU node/slot; used by phase-2-capstone/gpu-node/vllm-2gpu.yaml and
  vllm-timeslice.yaml so a demo load-test job or other best-effort pod never starves
  serving of a GPU it needs.
---
# GPU capacity in `inference` is 4 schedulable nvidia.com/gpu units: 2 physical GPUs x
# the time-slicing ConfigMap's `replicas: 2`
# (../../bootstrap/prereqs/nvidia-device-plugin-timeslicing.yaml). Lower `hard` to "2" if
# running the plain (non-time-sliced) device plugin
# (../../bootstrap/prereqs/nvidia-device-plugin.yaml) instead.
#
# Extended resources (like nvidia.com/gpu) can't be overcommitted, so Kubernetes only
# accepts a `requests.` quota for them — no bare or `limits.` form. See
# https://kubernetes.io/docs/concepts/policy/resource-quotas/#quota-for-extended-resources
apiVersion: v1
kind: ResourceQuota
metadata:
  name: inference-gpu-quota
  namespace: inference
spec:
  hard:
    requests.nvidia.com/gpu: "4"