feat(sglang,vllm): run the router's wait-upstreams on debian:12-slim

charts/sglang/Chart.yaml

@@ -5,7 +5,7 @@ home: https://github.com/modelsphere/helm-charts
 sources:
   - https://github.com/modelsphere/helm-charts
 type: application
-version: 0.9.30
+version: 0.9.31
 appVersion: "1.0.0"
 
 # CART is vendored under charts/cart -- already present, so no `helm dependency

charts/sglang/templates/_helpers.tpl

@@ -280,3 +280,11 @@
 {{- $img := .Values.pd.router.imageOverride | default dict -}}
 {{ $img.repository | default "lightseekorg/smg" }}:{{ $img.tag | default "1.14.0" }}
 {{- end -}}
+
+{{/*
+  The PD router's wait-upstreams init container image. Takes the root context.
+*/}}
+{{- define "sglang.pdWaitUpstreamsImage" -}}
+{{- $img := .Values.pd.router.waitUpstreamsImage | default dict -}}
+{{ $img.repository | default "debian" }}:{{ $img.tag | default "12-slim" }}
+{{- end -}}

charts/sglang/templates/pd-ds.yaml

@@ -181,7 +181,7 @@ spec:
              loop just blocks until the full set shows up. */}}
       initContainers:
       - name: wait-upstreams
-        image: "{{ include "sglang.pdRouterImage" $root }}"
+        image: "{{ include "sglang.pdWaitUpstreamsImage" $root }}"
         command: ["bash", "-lc"]
         args:
         - |

charts/sglang/tests/pd_router_smg_test.yaml

@@ -17,7 +17,7 @@ set:
   pd.decode.gpus: 8
   model.localPath: /models/x
 tests:
-  - it: pd-ds router and its init container default to the pinned smg image
+  - it: pd-ds router defaults to the pinned smg image; its init container to debian
     template: templates/pd-ds.yaml
     documentSelector:
       path: metadata.name
@@ -28,7 +28,7 @@ tests:
           value: lightseekorg/smg:1.14.0
       - equal:
           path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
-          value: lightseekorg/smg:1.14.0
+          value: debian:12-slim
 
   - it: pd-ds router execs smg and keeps its state dir where non-root can write
     template: templates/pd-ds.yaml
@@ -59,7 +59,7 @@ tests:
           path: spec.template.spec.containers[?(@.name == "router")].livenessProbe.exec.command[2]
           pattern: 'file="/tmp/pd-router/\$3"'
 
-  - it: pd-ds override wins over the default
+  - it: pd-ds router override wins over the default and leaves the init container alone
     template: templates/pd-ds.yaml
     documentSelector:
       path: metadata.name
@@ -71,6 +71,9 @@ tests:
       - equal:
           path: spec.template.spec.containers[?(@.name == "router")].image
           value: registry.example.com/mirror/smg:1.14.1
+      - equal:
+          path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
+          value: debian:12-slim
 
   - it: colocate router runs smg while its engines stay on the engine image
     template: templates/pd-colocate.yaml
@@ -110,7 +113,7 @@ tests:
           value: lightseekorg/smg:1.14.0
       - equal:
           path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
-          value: lightseekorg/smg:1.14.0
+          value: debian:12-slim
 
   - it: colocate emptied override falls back to smg, not the engine image
     template: templates/pd-colocate.yaml
@@ -125,3 +128,32 @@ tests:
       - equal:
           path: spec.template.spec.containers[?(@.name == "router")].image
           value: lightseekorg/smg:1.14.0
+
+  - it: wait-upstreams image can be mirrored without touching the router
+    template: templates/pd-ds.yaml
+    documentSelector:
+      path: metadata.name
+      value: ci-router-s0
+    set:
+      pd.router.waitUpstreamsImage.repository: registry.example.com/mirror/debian
+      pd.router.waitUpstreamsImage.tag: 12-slim-pinned
+    asserts:
+      - equal:
+          path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
+          value: registry.example.com/mirror/debian:12-slim-pinned
+      - equal:
+          path: spec.template.spec.containers[?(@.name == "router")].image
+          value: lightseekorg/smg:1.14.0
+
+  - it: emptied wait-upstreams image falls back to debian
+    template: templates/pd-ds.yaml
+    documentSelector:
+      path: metadata.name
+      value: ci-router-s0
+    set:
+      pd.router.waitUpstreamsImage.repository: ""
+      pd.router.waitUpstreamsImage.tag: ""
+    asserts:
+      - equal:
+          path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
+          value: debian:12-slim

charts/sglang/values.schema.json

@@ -192,6 +192,15 @@
                 "tag": { "type": "string" }
               }
             },
+            "waitUpstreamsImage": {
+              "description": "Image of the router's wait-upstreams init container. Needs bash, getent, awk, sort, wc and sleep. Independent of imageOverride.",
+              "type": "object",
+              "additionalProperties": false,
+              "properties": {
+                "repository": { "type": "string" },
+                "tag": { "type": "string" }
+              }
+            },
             "affinityReplace": { "type": ["object", "null"] },
             "tolerationsReplace": { "type": ["array", "null"] }
           }

charts/sglang/values.yaml

@@ -238,6 +238,11 @@ pd:
     imageOverride:
       repository: "lightseekorg/smg"
       tag: "1.14.0"
+    # The wait-upstreams init container only resolves DNS, so it stays on a
+    # plain base image whatever the router runs. Set repository/tag to mirror it.
+    waitUpstreamsImage:
+      repository: "debian"
+      tag: "12-slim"
     # The router always gets a copy of the model mounted read-only at
     # model.mountPath (as /model for the tokenizer's HF lookup, the path
     # cache_aware's prefix-hashing needs). Its default scheduling is

charts/vllm/Chart.yaml

@@ -5,7 +5,7 @@ home: https://github.com/modelsphere/helm-charts
 sources:
   - https://github.com/modelsphere/helm-charts
 type: application
-version: 0.9.2
+version: 0.9.3
 appVersion: "1.0.0"
 
 # CART is vendored under charts/cart -- already present, so no `helm dependency

charts/vllm/templates/_helpers.tpl

@@ -301,3 +301,11 @@
 {{- $img := .Values.pd.router.imageOverride | default dict -}}
 {{ $img.repository | default "vllm/vllm-router" }}:{{ $img.tag | default "v0.1.16.1" }}
 {{- end -}}
+
+{{/*
+  The PD router's wait-upstreams init container image. Takes the root context.
+*/}}
+{{- define "vllm.pdWaitUpstreamsImage" -}}
+{{- $img := .Values.pd.router.waitUpstreamsImage | default dict -}}
+{{ $img.repository | default "debian" }}:{{ $img.tag | default "12-slim" }}
+{{- end -}}

charts/vllm/templates/pd-ds.yaml

@@ -215,7 +215,7 @@ spec:
              loop just blocks until the full set shows up. */}}
       initContainers:
       - name: wait-upstreams
-        image: "{{ include "vllm.pdRouterImage" $root }}"
+        image: "{{ include "vllm.pdWaitUpstreamsImage" $root }}"
         command: ["bash", "-lc"]
         args:
         - |

charts/vllm/tests/pd_router_test.yaml

@@ -18,7 +18,7 @@ set:
   pd.decode.gpus: 8
   model.localPath: /models/x
 tests:
-  - it: pd-ds router and its init container default to the pinned vllm-router image
+  - it: pd-ds router defaults to the pinned vllm-router image; its init container to debian
     template: templates/pd-ds.yaml
     documentSelector:
       path: metadata.name
@@ -29,7 +29,7 @@ tests:
           value: vllm/vllm-router:v0.1.16.1
       - equal:
           path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
-          value: vllm/vllm-router:v0.1.16.1
+          value: debian:12-slim
 
   - it: pd-ds router execs vllm-router and keeps its state dir where non-root can write
     template: templates/pd-ds.yaml
@@ -60,7 +60,7 @@ tests:
           path: spec.template.spec.containers[?(@.name == "router")].livenessProbe.exec.command[2]
           pattern: 'file="/tmp/pd-router/\$3"'
 
-  - it: pd-ds override wins over the default
+  - it: pd-ds router override wins over the default and leaves the init container alone
     template: templates/pd-ds.yaml
     documentSelector:
       path: metadata.name
@@ -72,6 +72,9 @@ tests:
       - equal:
           path: spec.template.spec.containers[?(@.name == "router")].image
           value: registry.example.com/mirror/vllm-router:v0.1.17
+      - equal:
+          path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
+          value: debian:12-slim
 
   - it: colocate router runs vllm-router while its engines stay on the engine image
     template: templates/pd-colocate.yaml
@@ -108,7 +111,7 @@ tests:
           value: vllm/vllm-router:v0.1.16.1
       - equal:
           path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
-          value: vllm/vllm-router:v0.1.16.1
+          value: debian:12-slim
 
   - it: colocate emptied override falls back to vllm-router, not the engine image
     template: templates/pd-colocate.yaml
@@ -123,3 +126,32 @@ tests:
       - equal:
           path: spec.template.spec.containers[?(@.name == "router")].image
           value: vllm/vllm-router:v0.1.16.1
+
+  - it: wait-upstreams image can be mirrored without touching the router
+    template: templates/pd-ds.yaml
+    documentSelector:
+      path: metadata.name
+      value: ci-router-s0
+    set:
+      pd.router.waitUpstreamsImage.repository: registry.example.com/mirror/debian
+      pd.router.waitUpstreamsImage.tag: 12-slim-pinned
+    asserts:
+      - equal:
+          path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
+          value: registry.example.com/mirror/debian:12-slim-pinned
+      - equal:
+          path: spec.template.spec.containers[?(@.name == "router")].image
+          value: vllm/vllm-router:v0.1.16.1
+
+  - it: emptied wait-upstreams image falls back to debian
+    template: templates/pd-ds.yaml
+    documentSelector:
+      path: metadata.name
+      value: ci-router-s0
+    set:
+      pd.router.waitUpstreamsImage.repository: ""
+      pd.router.waitUpstreamsImage.tag: ""
+    asserts:
+      - equal:
+          path: spec.template.spec.initContainers[?(@.name == "wait-upstreams")].image
+          value: debian:12-slim

charts/vllm/values.schema.json

@@ -231,6 +231,15 @@
                 "tag": { "type": "string" }
               }
             },
+            "waitUpstreamsImage": {
+              "description": "Image of the router's wait-upstreams init container. Needs bash, getent, awk, sort, wc and sleep. Independent of imageOverride.",
+              "type": "object",
+              "additionalProperties": false,
+              "properties": {
+                "repository": { "type": "string" },
+                "tag": { "type": "string" }
+              }
+            },
             "affinityReplace": { "type": ["object", "null"] },
             "tolerationsReplace": { "type": ["array", "null"] }
           }

charts/vllm/values.yaml

@@ -312,6 +312,11 @@ pd:
     imageOverride:
       repository: "vllm/vllm-router"
       tag: "v0.1.16.1"
+    # The wait-upstreams init container only resolves DNS, so it stays on a
+    # plain base image whatever the router runs. Set repository/tag to mirror it.
+    waitUpstreamsImage:
+      repository: "debian"
+      tag: "12-slim"
     # The router always gets a copy of the model mounted read-only at
     # model.mountPath (used by the tokenizer's HF lookup). Its default
     # scheduling is pinned: required podAffinity to THIS slice's prefill