[router] Configure experimental sgl-router via CLI flags instead of a config file (#27073)

Signed-off-by: Kangyan Zhou <zky314343421@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

experimental/sgl-router/tests/e2e/k8s_integration/manifests/router-cluster-scoped.yaml

@@ -20,9 +20,23 @@ spec:
         - name: router
           image: sgl-router:e2e
           imagePullPolicy: Never
+          # Configured entirely via CLI flags. No --service-discovery-namespace
+          # means a cluster-wide EndpointSlice watch (all namespaces); the
+          # `cross-ns-test=true` selector term scopes it to this test's workers.
           args:
-            - "--config"
-            - "/etc/config/router-cluster.toml"
+            - "--host"
+            - "0.0.0.0"
+            - "--port"
+            - "8091"
+            - "--model-id"
+            - "tiny"
+            - "--tokenizer-path"
+            - "/etc/tokenizer/tiny.json"
+            - "--policy"
+            - "round_robin"
+            - "--service-discovery"
+            - "--selector"
+            - "app=sglang,cross-ns-test=true"
           ports:
             - containerPort: 8091
               name: http
@@ -38,13 +52,6 @@ spec:
               port: 8091
             initialDelaySeconds: 5
             periodSeconds: 10
-          volumeMounts:
-            - name: config
-              mountPath: /etc/config
-      volumes:
-        - name: config
-          configMap:
-            name: sgl-router-cluster-config
 ---
 apiVersion: v1
 kind: Service

experimental/sgl-router/tests/e2e/k8s_integration/manifests/router.yaml

@@ -18,9 +18,32 @@ spec:
         - name: router
           image: sgl-router:e2e
           imagePullPolicy: Never
+          # Configured entirely via CLI flags. K8s EndpointSlice discovery
+          # watches `app=sglang` pods in this namespace. The aggressive
+          # circuit breaker (threshold 1, 5s cool-down) lets a terminating
+          # pod's connection-refused immediately drop it from the candidate
+          # set — the reconciliation tests scale workers rapidly and depend
+          # on fast eviction to absorb the churn.
           args:
-            - "--config"
-            - "/etc/config/router.toml"
+            - "--host"
+            - "0.0.0.0"
+            - "--port"
+            - "8090"
+            - "--model-id"
+            - "tiny"
+            - "--tokenizer-path"
+            - "/etc/tokenizer/tiny.json"
+            - "--policy"
+            - "round_robin"
+            - "--cb-threshold"
+            - "1"
+            - "--cb-cool-down-secs"
+            - "5"
+            - "--service-discovery"
+            - "--service-discovery-namespace"
+            - "sgl-router-test"
+            - "--selector"
+            - "app=sglang"
           ports:
             - containerPort: 8090
               name: http
@@ -36,13 +59,6 @@ spec:
               port: 8090
             initialDelaySeconds: 5
             periodSeconds: 10
-          volumeMounts:
-            - name: config
-              mountPath: /etc/config
-      volumes:
-        - name: config
-          configMap:
-            name: sgl-router-config
 ---
 apiVersion: v1
 kind: Service

experimental/sgl-router/tests/e2e/k8s_integration/setup.sh

@@ -141,38 +141,9 @@ log "Waiting for fake-worker rollout..."
 kubectl --context "${CONTEXT}" -n "${NAMESPACE}" rollout status deployment/fake-worker --timeout=120s
 
 # ---------------------------------------------------------------------------
-# Step 6: Create sgl-router ConfigMap with k8s discovery pointing at the
-#         namespace where fake-worker pods live.
-# ---------------------------------------------------------------------------
-log "Creating sgl-router-config ConfigMap..."
-ROUTER_CONFIG="[server]
-host = \"0.0.0.0\"
-port = 8090
-
-[[models]]
-id = \"tiny\"
-tokenizer_path = \"/etc/tokenizer/tiny.json\"
-policy = \"round_robin\"
-# Aggressive breaker so a terminating pod's connection-refused
-# immediately excludes it from the next request's candidate set —
-# the reconciliation tests scale workers rapidly and depend on
-# fast worker eviction to absorb the churn.
-circuit_breaker = { threshold = 1, cool_down_secs = 5 }
-
-[discovery]
-backend = \"k8s\"
-
-[discovery.k8s]
-namespace = \"${NAMESPACE}\"
-label_selector = \"app=sglang\""
-
-kubectl --context "${CONTEXT}" -n "${NAMESPACE}" create configmap sgl-router-config \
-    --from-literal=router.toml="${ROUTER_CONFIG}" \
-    --dry-run=client -o yaml \
-    | kubectl --context "${CONTEXT}" apply -f -
-
-# ---------------------------------------------------------------------------
-# Step 7: Deploy sgl-router
+# Step 6: Deploy sgl-router. It is configured entirely via CLI flags in
+#         router.yaml — k8s EndpointSlice discovery watches `app=sglang`
+#         pods in the sgl-router-test namespace (where fake-worker lives).
 # ---------------------------------------------------------------------------
 log "Deploying sgl-router..."
 kubectl --context "${CONTEXT}" apply -f "${MANIFESTS_DIR}/router.yaml"

experimental/sgl-router/tests/e2e/k8s_integration/test_cross_namespace.py

@@ -149,49 +149,9 @@ def cluster_scoped_router(k8s_cluster):
     _ensure_namespace(EXTRA_NAMESPACE)
     _ensure_service_in_ns(EXTRA_NAMESPACE)
 
-    # ConfigMap for the cluster-scoped router: empty namespace = watch all
-    cluster_config = """[server]
-host = "0.0.0.0"
-port = 8091
-
-[[models]]
-id = "tiny"
-tokenizer_path = "/etc/tokenizer/tiny.json"
-policy = "round_robin"
-
-[discovery]
-backend = "k8s"
-
-[discovery.k8s]
-namespace = ""
-label_selector = "app=sglang,cross-ns-test=true"
-"""
-    _kubectl(
-        "create",
-        "configmap",
-        "sgl-router-cluster-config",
-        f"--from-literal=router-cluster.toml={cluster_config}",
-        "-n",
-        NAMESPACE,
-        "--dry-run=client",
-        "-o",
-        "yaml",
-        check=True,
-    )
-    # pipe through apply
-    proc = _kubectl(
-        "create",
-        "configmap",
-        "sgl-router-cluster-config",
-        f"--from-literal=router-cluster.toml={cluster_config}",
-        "-n",
-        NAMESPACE,
-        "--dry-run=client",
-        "-o",
-        "yaml",
-    )
-    _apply_from_stdin(proc.stdout)
-
+    # The cluster-scoped router is configured via CLI flags in
+    # router-cluster-scoped.yaml: no --service-discovery-namespace (watch
+    # all namespaces) and --selector app=sglang,cross-ns-test=true.
     _kubectl("apply", "-f", str(router_manifest))
 
     # The cluster-scoped router's /readyz blocks on registry-not-empty, so