[https://nvbugs/6958731][fix] Use the tokenizer chat template when the processor has none (#20880)

Signed-off-by: chenfeiz0326 <203214996+chenfeiz0326@users.noreply.github.com>
Co-authored-by: chenfeiz0326 <203214996+chenfeiz0326@users.noreply.github.com>

Author: chenfeiz0326

tensorrt_llm/inputs/utils.py

@@ -644,11 +644,15 @@ def resolve_hf_chat_template(
             return templates[chat_template]
         return chat_template
 
-    # 2. If tool is not provided, use the processor's default chat template
-    if not tools and processor and hasattr(processor, 'chat_template'):
-        return processor.chat_template
-
-    # 3. If tool is provided, use the tool
+    # 2. Without tools, prefer the processor's chat template when it has one.
+    #    A processor without one (some keep the template on their tokenizer
+    #    only) falls through to the tokenizer below.
+    if not tools and processor:
+        processor_template = getattr(processor, "chat_template", None)
+        if processor_template is not None:
+            return processor_template
+
+    # 3. Use the tokenizer's chat template, which accounts for the tools
     try:
         return tokenizer.get_chat_template(chat_template, tools=tools)
     except Exception:

tests/unittest/inputs/test_chat_template_dispatch.py

@@ -460,6 +460,49 @@ def get_chat_template(self, *_args, **_kwargs):
             )
 
 
+class TestProcessorWithoutChatTemplate:
+    """A processor that has no chat template must not hide the tokenizer's."""
+
+    @staticmethod
+    def _tokenizer(template):
+        class JinjaTokenizer:
+            chat_template = template
+
+            def get_chat_template(self, chat_template=None, tools=None):
+                return self.chat_template
+
+            def apply_chat_template(self, **kwargs):
+                # Return the template it was asked to render, so the result names the winner.
+                return kwargs["chat_template"]
+
+        return JinjaTokenizer()
+
+    @staticmethod
+    def _render(tokenizer, processor):
+        return apply_chat_template(
+            model_type="test_string_model",
+            tokenizer=tokenizer,
+            processor=processor,
+            conversation=[ConversationMessage(role="user", content="hello", media=[])],
+            add_generation_prompt=True,
+            mm_placeholder_counts=[{}],
+        )
+
+    def test_falls_back_to_tokenizer_template(self):
+        class Processor:
+            chat_template = None
+
+        tokenizer = self._tokenizer("{{ 'tokenizer template' }}")
+        assert self._render(tokenizer, Processor()) == "{{ 'tokenizer template' }}"
+
+    def test_processor_template_is_still_preferred(self):
+        class Processor:
+            chat_template = "{{ 'processor template' }}"
+
+        tokenizer = self._tokenizer("{{ 'tokenizer template' }}")
+        assert self._render(tokenizer, Processor()) == "{{ 'processor template' }}"
+
+
 class TestServingChatTemplateGather:
     """Cover the asyncio.gather integration in the serving chat-template paths."""