# KServe InferenceService — the same Qwen2.5-1.5B model, served the KServe-native way
# via the built-in huggingface ServingRuntime (vLLM backend) instead of a hand-written
# Deployment. Compare operational overhead + numbers against vllm-plain.yaml.
#
# Mode: RawDeployment (annotation below) — plain K8s Deployment + Service + HPA, NO
# Knative/Istio. Serverless mode would pull in a service mesh that depends on the
# cross-node pod overlay, which is down on this WSL2 cluster (see troubleshooting log).
#
# runtimeClassName: nvidia is required for GPU access under k3s. gpu_memory_utilization
# is held to 0.80 so it fits the desktop 3060 Ti (~6.7 GB free); the laptop is cordoned
# during the KServe pass so the predictor + webhooks co-locate with the API server.
apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: qwen
  namespace: inference
  annotations:
    serving.kserve.io/deploymentMode: RawDeployment
spec:
  predictor:
    runtimeClassName: nvidia
    model:
      modelFormat:
        name: huggingface
      args:
        - --model_id=Qwen/Qwen2.5-1.5B-Instruct
        - --gpu_memory_utilization=0.80
        - --max_model_len=8192
      resources:
        # Both limits AND requests set explicitly: the kserve-huggingfaceserver runtime
        # defaults memory limit to 2Gi, and K8s rejects requests > limits. (A plain
        # Deployment doesn't surprise you this way — you write the whole pod spec.)
        limits:
          cpu: "4"
          memory: 12Gi
          nvidia.com/gpu: "1"
        requests:
          cpu: "1"
          memory: 4Gi
          nvidia.com/gpu: "1"