feat(aws): add --tool-call-parser and --chat-template flags for RHEL AI

Enable vLLM tool calling by adding --tool-call-parser and --chat-template
flags. When --tool-call-parser is set, --enable-auto-tool-choice is
automatically added. Chat template filename is resolved to /opt/app-root/template/.
Updated Tekton task template with new params.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: are-ces

cmd/mapt/cmd/aws/hosts/rhelai.go

@@ -72,6 +72,8 @@ func getRHELAICreate() *cobra.Command {
 					HFToken:          viper.GetString(params.RhelAIHFToken),
 					APIKey:           viper.GetString(params.RhelAIAPIKey),
 					AutoStart:        viper.IsSet(params.RhelAIAutoStart),
+					ToolCallParser:   viper.GetString(params.RhelAIToolCallParser),
+					ChatTemplate:     viper.GetString(params.RhelAIChatTemplate),
 					ExposePorts:      viper.GetIntSlice(params.RhelAIExposePorts),
 				})
 		},
@@ -87,6 +89,8 @@ func getRHELAICreate() *cobra.Command {
 	flagSet.StringP(params.RhelAIAPIKey, "", "", params.RhelAIAPIKeyDesc)
 	flagSet.Bool(params.RhelAIAutoStart, false, params.RhelAIAutoStartDesc)
 	flagSet.IntSlice(params.RhelAIExposePorts, nil, params.RhelAIExposePortsDesc)
+	flagSet.StringP(params.RhelAIToolCallParser, "", "", params.RhelAIToolCallParserDesc)
+	flagSet.StringP(params.RhelAIChatTemplate, "", "", params.RhelAIChatTemplateDesc)
 	flagSet.StringP(params.Timeout, "", "", params.TimeoutDesc)
 	params.AddComputeRequestFlags(flagSet)
 	params.AddSpotFlags(flagSet)

cmd/mapt/cmd/params/params.go

@@ -131,6 +131,10 @@ const (
 	RhelAIAutoStartDesc       string = "automatically configure and start RHAIIS after provisioning"
 	RhelAIExposePorts         string = "expose-ports"
 	RhelAIExposePortsDesc     string = "comma-separated list of ports to expose through the load balancer and security group (e.g. 8000,8080)"
+	RhelAIToolCallParser      string = "tool-call-parser"
+	RhelAIToolCallParserDesc  string = "enable tool calling with the specified parser (e.g. llama3_json, hermes, mistral)"
+	RhelAIChatTemplate        string = "chat-template"
+	RhelAIChatTemplateDesc    string = "chat template jinja filename (e.g. tool_chat_template_llama3.2_json.jinja)"
 
 	// Serverless
 	Timeout        string = "timeout"

pkg/provider/aws/action/rhel-ai/rhelai.go

@@ -46,6 +46,8 @@ type rhelAIRequest struct {
 	hfToken          *string
 	apiKey           *string
 	autoStart        bool
+	toolCallParser   *string
+	chatTemplate     *string
 	exposePorts      []int
 }
 
@@ -85,6 +87,8 @@ func Create(mCtxArgs *mc.ContextArgs, args *apiRHELAI.RHELAIArgs) (err error) {
 		hfToken:          &args.HFToken,
 		apiKey:           &args.APIKey,
 		autoStart:        args.AutoStart,
+		toolCallParser:   &args.ToolCallParser,
+		chatTemplate:     &args.ChatTemplate,
 		exposePorts:      args.ExposePorts}
 	if args.Spot != nil {
 		r.spot = args.Spot.Spot
@@ -373,6 +377,15 @@ func (r *rhelAIRequest) rhaiisSetupScript() string {
 			` && sudo sed -i 's|--model .*|--model %s \\|' %s/install.conf`,
 			*r.model, confDir)
 	}
+	if len(*r.toolCallParser) > 0 {
+		toolArgs := fmt.Sprintf(`--enable-auto-tool-choice \\\n     --tool-call-parser %s`, *r.toolCallParser)
+		if len(*r.chatTemplate) > 0 {
+			toolArgs += fmt.Sprintf(` \\\n     --chat-template /opt/app-root/template/%s`, *r.chatTemplate)
+		}
+		script += fmt.Sprintf(
+			` && sudo sed -i 's|--max-model-len.*|--max-model-len 4096 \\\n     %s|' %s/install.conf`,
+			toolArgs, confDir)
+	}
 	if len(*r.apiKey) > 0 {
 		script += fmt.Sprintf(
 			" && sudo sed -i '/\\[Install\\]/i Environment=VLLM_API_KEY=%s' %s/install.conf",

pkg/target/host/rhelai/api.go

@@ -19,7 +19,9 @@ type RHELAIArgs struct {
 	Timeout   string
 	Model     string
 	HFToken   string
-	APIKey      string
-	AutoStart   bool
-	ExposePorts []int
+	APIKey         string
+	AutoStart      bool
+	ToolCallParser string
+	ChatTemplate   string
+	ExposePorts    []int
 }

tkn/infra-aws-rhel-ai.yaml

@@ -155,6 +155,12 @@ spec:
     - name: expose-ports
       description: Comma-separated list of ports to expose through the load balancer and security group (e.g. 8000,8080).
       default: ""
+    - name: tool-call-parser
+      description: Enable tool calling with the specified parser (e.g. llama3_json, hermes, mistral). Automatically adds --enable-auto-tool-choice.
+      default: ""
+    - name: chat-template
+      description: Chat template jinja filename (e.g. tool_chat_template_llama3.2_json.jinja).
+      default: ""
 
     # Network params
     - name: service-endpoints
@@ -317,6 +323,12 @@ spec:
           if [[ "$(params.expose-ports)" != "" ]]; then
             cmd+="--expose-ports '$(params.expose-ports)' "
           fi
+          if [[ "$(params.tool-call-parser)" != "" ]]; then
+            cmd+="--tool-call-parser '$(params.tool-call-parser)' "
+          fi
+          if [[ "$(params.chat-template)" != "" ]]; then
+            cmd+="--chat-template '$(params.chat-template)' "
+          fi
           cmd+="--tags '$(params.tags)' "
         fi
 

tkn/infra-azure-rhel-ai.yaml

@@ -85,6 +85,12 @@ spec:
     - name: disk-size
       description: Disk size in GB for the cloud instance
       default: "200"
+    - name: gpus
+      description: Number of GPUs for the cloud instance (valid marketplace values are 1, 2, 4, 8)
+      default: "8"
+    - name: gpu-manufacturer
+      description: GPU manufacturer name for instance filtering (e.g. NVIDIA, AMD)
+      default: ""
     - name: compute-sizes
       description: Comma seperated list of sizes for the machines to be requested. If set this takes precedence over compute by args
       default: "Standard_ND96is_MI300X_v5,Standard_ND96isr_MI300X_v5"
@@ -229,6 +235,12 @@ spec:
           if [[ "$(params.compute-sizes)" != "" ]]; then
             cmd+="--compute-sizes '$(params.compute-sizes)' "
           fi
+          if [[ "$(params.gpus)" != "" ]]; then
+            cmd+="--gpus '$(params.gpus)' "
+          fi
+          if [[ "$(params.gpu-manufacturer)" != "" ]]; then
+            cmd+="--gpu-manufacturer '$(params.gpu-manufacturer)' "
+          fi
           if [[ "$(params.marketplace)" == "true" ]]; then
             cmd+="--marketplace "
             cmd+="--accelerator '$(params.accelerator)' "

tkn/template/infra-aws-rhel-ai.yaml

@@ -155,6 +155,12 @@ spec:
     - name: expose-ports
       description: Comma-separated list of ports to expose through the load balancer and security group (e.g. 8000,8080).
       default: ""
+    - name: tool-call-parser
+      description: Enable tool calling with the specified parser (e.g. llama3_json, hermes, mistral). Automatically adds --enable-auto-tool-choice.
+      default: ""
+    - name: chat-template
+      description: Chat template jinja filename (e.g. tool_chat_template_llama3.2_json.jinja).
+      default: ""
 
     # Network params
     - name: service-endpoints
@@ -317,6 +323,12 @@ spec:
           if [[ "$(params.expose-ports)" != "" ]]; then
             cmd+="--expose-ports '$(params.expose-ports)' "
           fi
+          if [[ "$(params.tool-call-parser)" != "" ]]; then
+            cmd+="--tool-call-parser '$(params.tool-call-parser)' "
+          fi
+          if [[ "$(params.chat-template)" != "" ]]; then
+            cmd+="--chat-template '$(params.chat-template)' "
+          fi
           cmd+="--tags '$(params.tags)' "
         fi