add configurable gateway replicas use gateway.gatewayParameters.replicas

Signed-off-by: nicole-lihui <nicole.li@daocloud.io>

Author: nicole-lihui

charts/llm-d-infra/README.md

@@ -122,6 +122,7 @@ Kubernetes: `>= 1.28.0-0`
 | gateway.gatewayParameters.istio | Istio-specific parameters rendered into the gateway ConfigMap when gateway.provider resolves to "istio" | object | `{"accessLogging":true}` |
 | gateway.gatewayParameters.istio.accessLogging | For istio to include access logging or not | bool | `true` |
 | gateway.gatewayParameters.logLevel | Log level for provider-managed gateway data plane components | string | `"warn"` |
+| gateway.gatewayParameters.replicas | Number of replicas for the gateway | int | `1` |
 | gateway.gatewayParameters.resources | Resource requests/limits <br /> Ref: https://kubernetes.io/docs/concepts/configuration/manage-resources-containers/#resource-requests-and-limits-of-pod-and-container | object | `{"limits":{"cpu":"2","memory":"1Gi"},"requests":{"cpu":"100m","memory":"128Mi"}}` |
 | gateway.labels | Additional labels provided to the Gateway resource | object | `{}` |
 | gateway.listeners | Set of listeners exposed via the Gateway, also propagated to the Ingress if enabled | list | `[{"allowedRoutes":{"namespaces":{"from":"All"}},"name":"default","port":80,"protocol":"HTTP"}]` |

charts/llm-d-infra/templates/gateway-infrastructure/agentgatewayparameters.yaml

@@ -47,4 +47,7 @@ spec:
       {{- if .Values.gateway.service.loadBalancerClass }}
       loadBalancerClass: {{ .Values.gateway.service.loadBalancerClass | quote }}
       {{- end }}
+  deployment:
+    spec:
+      replicas: {{ .Values.gateway.gatewayParameters.replicas }}
 {{- end}}

charts/llm-d-infra/templates/gateway-infrastructure/configmap.yaml

@@ -22,6 +22,7 @@ data:
   {{- if eq $provider "istio" }}
   deployment: |
     spec:
+      replicas: {{ .Values.gateway.gatewayParameters.replicas }}
       template:
         spec:
           containers:

charts/llm-d-infra/values.schema.json

@@ -241,6 +241,12 @@
                             "required": [],
                             "title": "logLevel"
                         },
+                        "replicas": {
+                            "default": "1",
+                            "description": "Number of replicas for the gateway",
+                            "required": [],
+                            "title": "replicas"
+                        },
                         "resources": {
                             "description": "ResourceRequirements describes the compute resource requirements.",
                             "properties": {

charts/llm-d-infra/values.schema.tmpl.json

@@ -241,6 +241,12 @@
               "required": [],
               "title": "logLevel"
             },
+            "replicas": {
+              "default": "1",
+              "description": "Number of replicas for the gateway",
+              "required": [],
+              "title": "replicas"
+            },
             "resources": {
               "$ref": "https://raw.githubusercontent.com/yannh/kubernetes-json-schema/master/master/_definitions.json#/definitions/io.k8s.api.core.v1.ResourceRequirements",
               "description": "Resource requests/limits \u003cbr /\u003e Ref: https://kubernetes.io/docs/concepts/configuration/manage-resources-containers/#resource-requests-and-limits-of-pod-and-container",

charts/llm-d-infra/values.yaml

@@ -98,6 +98,9 @@ gateway:
     # -- Log level for provider-managed gateway data plane components
     logLevel: warn
 
+    # -- Number of replicas for the gateway
+    replicas: 1
+
     # @schema
     # $ref: https://raw.githubusercontent.com/yannh/kubernetes-json-schema/master/master/_definitions.json#/definitions/io.k8s.api.core.v1.ResourceRequirements
     # @schema