[Model-Gateway: grpc]: create tokenizer with chat template (#17052)

Signed-off-by: jeff.ye <jeff.ye@novita.ai>

sgl-model-gateway/src/core/job_queue.rs

@@ -578,7 +578,7 @@ impl JobQueue {
                                 tokenizer_path: None,
                                 reasoning_parser: None,
                                 tool_parser: None,
-                                chat_template: None,
+                                chat_template: router_config.chat_template.clone(),
                                 bootstrap_port: None,
                                 health_check_timeout_secs: router_config.health_check.timeout_secs,
                                 health_check_interval_secs: router_config
@@ -649,7 +649,7 @@ impl JobQueue {
                         tokenizer_path: None,
                         reasoning_parser: None,
                         tool_parser: None,
-                        chat_template: None,
+                        chat_template: router_config.chat_template.clone(),
                         bootstrap_port,
                         health_check_timeout_secs: router_config.health_check.timeout_secs,
                         health_check_interval_secs: router_config.health_check.check_interval_secs,

sgl-model-gateway/src/core/steps/worker/local/register_tokenizer.rs

@@ -29,6 +29,13 @@ impl StepExecutor<LocalWorkerWorkflowData> for RegisterTokenizerStep {
             .actual_workers
             .as_ref()
             .ok_or_else(|| WorkflowError::ContextValueNotFound("workers".to_string()))?;
+        // Get chat_template: worker config > global router config
+        let chat_template = context
+            .data
+            .config
+            .chat_template
+            .clone()
+            .or_else(|| app_context.router_config.chat_template.clone());
 
         for worker in workers.iter() {
             let model_id = worker.model_id().to_string();
@@ -55,12 +62,17 @@ impl StepExecutor<LocalWorkerWorkflowData> for RegisterTokenizerStep {
 
             // Load tokenizer with thread safe lock
             let tokenizer_path_owned = tokenizer_path.clone();
+            let template = chat_template.clone();
             if let Err(e) = app_context
                 .tokenizer_registry
-                .load(&tokenizer_id, &model_id, &source, || async move {
+                .load(&tokenizer_id, &model_id, &source, move || {
-                    factory::create_tokenizer_async(&tokenizer_path_owned)
+                    let path = tokenizer_path_owned;
+                    let tmpl = template;
+                    async move {
+                        factory::create_tokenizer_async_with_chat_template(&path, tmpl.as_deref())
                             .await
                             .map_err(|e| e.to_string())
+                    }
                 })
                 .await
             {