# vLLM Deployment used by the e2e_gpu integration spec.
#
# Renders a single-replica vLLM serve, configured to reach the LMCache
# DaemonSet via the operator-produced <engine>-connection ConfigMap.
# Placeholders are substituted by the spec at apply time so the test
# can target whichever namespace/engine/model the run was parameterised
# with:
#
#   __NAMESPACE__       — engine namespace
#   __ENGINE_NAME__     — LMCacheEngine metadata.name (also Service / ConfigMap name)
#   __MODEL__           — Hugging Face model id (e.g., Qwen/Qwen2.5-0.5B)
#   __VLLM_IMAGE__      — container image (defaults to lmcache/vllm-openai:latest,
#                         the bundled vLLM + LMCache build)
#
# Requirements baked in (matching the operator's LMCache pod template
# and the constraints of cudaIpcOpenMemHandle-based KV handoff):
#   - hostIPC: true                — share /dev/shm with the LMCache pod
#   - runtimeClassName: nvidia     — same as LMCache so they co-schedule
#                                    on GPU nodes
#   - nodeSelector: gpu.present    — pin to a GPU node so the service's
#                                    internalTrafficPolicy=Local routes
#                                    to the on-node LMCache pod
#   - nvidia.com/gpu: 1            — consume one GPU device
#   - --no-enable-prefix-caching   — REQUIRED: APC must be off so prompt
#                                    re-runs miss vLLM's local KV and
#                                    fall through to LMCache lookup
#   - --kv-transfer-config         — inlined from the operator-produced
#                                    ConfigMap by the entrypoint wrapper
#                                    (vllm serve takes a JSON string,
#                                    not a path)
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-smoke
  namespace: __NAMESPACE__
  labels:
    app: vllm-smoke
spec:
  replicas: 1
  selector:
    matchLabels:
      app: vllm-smoke
  template:
    metadata:
      labels:
        app: vllm-smoke
    spec:
      hostIPC: true
      runtimeClassName: nvidia
      nodeSelector:
        nvidia.com/gpu.present: "true"
      containers:
        - name: vllm
          image: __VLLM_IMAGE__
          imagePullPolicy: IfNotPresent
          # vllm serve --kv-transfer-config takes the JSON connector
          # config as a single argv token (string), so we read the
          # mounted ConfigMap file inline via `cat`. Going through a
          # shell wrapper keeps the JSON source of truth in the
          # operator-produced ConfigMap rather than duplicated here.
          #
          # --load-format dummy: skip downloading the model weights
          # from Hugging Face. vLLM still fetches the model's metadata
          # (config.json, tokenizer.json) but generates random weights
          # in-process. The test asserts on KV-cache hit counters, not
          # on generation quality, so dummy weights are fine and we
          # save several GB of download + ~5 min of test runtime.
          command: ["/bin/sh", "-c"]
          args:
            - >-
              exec vllm serve "__MODEL__"
              --port 8000
              --no-enable-prefix-caching
              --enforce-eager
              --load-format dummy
              --gpu-memory-utilization 0.5
              --kv-transfer-config "$(cat /etc/lmcache/kv-transfer-config.json)"
          env:
            # MP transport requires single-process scheduler so the
            # connector lives in the same proc as the engine core.
            - name: VLLM_ENABLE_V1_MULTIPROCESSING
              value: "0"
            # Determinism: same prompt -> same chunk hashes across runs
            # so the cache lookup on the second request actually hits.
            - name: PYTHONHASHSEED
              value: "0"
            - name: NVIDIA_VISIBLE_DEVICES
              value: "all"
            - name: NVIDIA_DRIVER_CAPABILITIES
              value: "all"
          ports:
            - name: http
              containerPort: 8000
              protocol: TCP
          resources:
            limits:
              nvidia.com/gpu: 1
          readinessProbe:
            httpGet:
              path: /v1/models
              port: 8000
            # /v1/models is ready only after the model is loaded onto
            # the GPU — a strong signal that the LMCache connector
            # also negotiated successfully (otherwise vllm would have
            # crashed during init).
            initialDelaySeconds: 30
            periodSeconds: 10
            failureThreshold: 60
          volumeMounts:
            - name: lmcache-connection
              mountPath: /etc/lmcache
              readOnly: true
      volumes:
        - name: lmcache-connection
          configMap:
            # __ENGINE_NAME__-connection is owned by the LMCacheEngine
            # so its lifecycle is tied to the CR — the test asserts on
            # GC of this same ConfigMap in lifecycle_smoke_test.go.
            name: __ENGINE_NAME__-connection
            items:
              - key: kv-transfer-config.json
                path: kv-transfer-config.json