llm-d-inference-scheduler/deploy/components/overlays/simulator/kustomization.yaml at f74e4661bcb001463a389d41ddd29ec9c5e4c309 · mayabar/llm-d-inference-scheduler · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
# ------------------------------------------------------------------------------
# Simulator Overlay
#
# Adds simulator-specific configuration to vLLM deployments:
#   - --mode=${VLLM_SIM_MODE} arg on Decode (vllm-d) only — controls simulator
#     response mode (echo: returns input, random: returns random sentences)
#   - --zmq-endpoint on Decode only — publishes KV cache events to the EPP
#   - vLLM Render sidecar on all deployments (Decode, Prefill, Encode)
#   - KV cache args on Decode for cache-aware scheduling tests
#
# Uses JSON patches (op: add) to INSERT into args arrays without replacing
# them, avoiding conflicts with scenario-specific strategic merge patches.
#
# Usage: include as a Kustomize component in scenario overlays:
#   components:
#   - ../../../components/overlays/simulator/
# ------------------------------------------------------------------------------
apiVersion: kustomize.config.k8s.io/v1alpha1
kind: Component

patches:
# --- Decode deployment (vllm-d) ---
# Has initContainers (routing-sidecar), so insert tokenizer at index 0
- target:
    kind: Deployment
    name: vllm-d
  patch: |
    - op: add
      path: /spec/template/spec/containers/0/args/0
      value: "--mode=${VLLM_SIM_MODE}"
    - op: add
      path: /spec/template/spec/containers/0/args/1
      value: "--enable-kvcache=${KV_CACHE_ENABLED}"
    - op: add
      path: /spec/template/spec/containers/0/args/2
      value: "--kv-cache-size=1024"
    - op: add
      path: /spec/template/spec/containers/0/args/3
      value: "--block-size=16"
    - op: add
      path: /spec/template/spec/containers/0/args/4
      value: "--zmq-endpoint=tcp://${EPP_NAME}.${NAMESPACE}.svc.cluster.local:5557"