fix: add chat_template_kwargs alias for compatibility (#5112)

Co-authored-by: Claude <noreply@anthropic.com>

lib/llm/src/http/service/openai.rs

@@ -2191,8 +2191,7 @@ mod tests {
             "model": "test-model",
             "add_special_tokens": true,
             "documents": ["doc1"],
-            "chat_template": "custom",
-            "chat_template_kwargs": {"key": "val"}
+            "chat_template": "custom"
         }"#;
 
         let request: NvCreateChatCompletionRequest = serde_json::from_str(json).unwrap();
@@ -2205,11 +2204,6 @@ mod tests {
         );
         assert!(request.unsupported_fields.contains_key("documents"));
         assert!(request.unsupported_fields.contains_key("chat_template"));
-        assert!(
-            request
-                .unsupported_fields
-                .contains_key("chat_template_kwargs")
-        );
 
         let result = validate_chat_completion_fields_generic(&request);
         assert!(result.is_err());
@@ -2221,7 +2215,6 @@ mod tests {
             assert!(msg.contains("add_special_tokens"));
             assert!(msg.contains("documents"));
             assert!(msg.contains("chat_template"));
-            assert!(msg.contains("chat_template_kwargs"));
         }
     }
 

lib/llm/src/protocols/openai/chat_completions.rs

@@ -44,7 +44,12 @@ pub struct NvCreateChatCompletionRequest {
     pub nvext: Option<NvExt>,
 
     /// Extra args to pass to the chat template rendering context
-    #[serde(default, skip_serializing_if = "Option::is_none")]
+    /// Also accepts "chat_template_kwargs" as an alias for compatibility
+    #[serde(
+        default,
+        skip_serializing_if = "Option::is_none",
+        alias = "chat_template_kwargs"
+    )]
     pub chat_template_args: Option<std::collections::HashMap<String, serde_json::Value>>,
 
     /// Runtime media decoding parameters.

lib/llm/tests/test_chat_template_args.rs

+// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+use dynamo_llm::protocols::openai::chat_completions::NvCreateChatCompletionRequest;
+
+#[test]
+fn test_both_fields_fails() {
+    // Test that when both are present, serde will fail with duplicate field error
+    // This test documents that behavior
+    let json_with_both = r#"{
+        "model": "test-model",
+        "messages": [],
+        "chat_template_args": {
+            "enable_thinking": true
+        },
+        "chat_template_kwargs": {
+            "enable_thinking": false
+        }
+    }"#;
+
+    // This will fail with duplicate field error
+    let result: Result<NvCreateChatCompletionRequest, _> = serde_json::from_str(json_with_both);
+    assert!(result.is_err());
+}
+
+#[test]
+fn test_chat_template_kwargs_alias() {
+    // Test that chat_template_kwargs is accepted as an alias for chat_template_args
+    let json_with_kwargs = r#"{
+        "model": "test-model",
+        "messages": [],
+        "chat_template_kwargs": {
+            "enable_thinking": false
+        }
+    }"#;
+
+    let request: NvCreateChatCompletionRequest = serde_json::from_str(json_with_kwargs).unwrap();
+    assert!(request.chat_template_args.is_some());
+    assert_eq!(
+        request.chat_template_args.unwrap().get("enable_thinking"),
+        Some(&serde_json::json!(false))
+    );
+}
+
+#[test]
+fn test_chat_template_args() {
+    // Test that chat_template_args still works as the primary field name
+    let json_with_args = r#"{
+        "model": "test-model",
+        "messages": [],
+        "chat_template_args": {
+            "enable_thinking": true
+        }
+    }"#;
+
+    let request: NvCreateChatCompletionRequest = serde_json::from_str(json_with_args).unwrap();
+    assert!(request.chat_template_args.is_some());
+    assert_eq!(
+        request.chat_template_args.unwrap().get("enable_thinking"),
+        Some(&serde_json::json!(true))
+    );
+}
+
+// TODO: Add template rendering test that verifies chat_template_args/chat_template_kwargs
+// values are actually passed to the Jinja template context during rendering.
+// This would require setting up PromptFormatter with ChatTemplate/ContextMixins.