Fix tokenizer issue with chat template

nemo_deploy/llm/megatronllm_deployable.py

@@ -233,6 +233,14 @@ def apply_chat_template(self, messages, add_generation_prompt=True):
                 # Some tokenizers might not have bos_token, use empty string as fallback
                 bos_token = ""
 
+            # Try to get eos_token - many chat templates reference it
+            eos_token = None
+            try:
+                eos_token = self.mcore_tokenizer._tokenizer.eos_token
+            except AttributeError:
+                # Some tokenizers might not have eos_token, use empty string as fallback
+                eos_token = ""
+
             # Check if chat_template is None or empty
             if tokenizer_chat_template is None:
                 raise ValueError(
@@ -251,6 +259,7 @@ def apply_chat_template(self, messages, add_generation_prompt=True):
         rendered_output = template.render(
             messages=messages,
             bos_token=bos_token,
+            eos_token=eos_token,
             add_generation_prompt=add_generation_prompt,
         )
 

tests/unit_tests/deploy/test_megatronllm_deployable.py

@@ -466,6 +466,52 @@ def test_apply_chat_template(deployable):
         template_mock.render.assert_called_once()
 
 
+@pytest.mark.run_only_on("GPU")
+def test_apply_chat_template_passes_eos_token(deployable):
+    """Test that eos_token is forwarded to template.render."""
+    deployable.mcore_tokenizer._tokenizer.eos_token = "<custom_eos>"
+    messages = [{"role": "user", "content": "Hello"}]
+
+    template_mock = MagicMock()
+    template_mock.render.return_value = "rendered"
+
+    with patch("nemo_deploy.llm.megatronllm_deployable.Template", return_value=template_mock):
+        deployable.apply_chat_template(messages)
+
+    call_kwargs = template_mock.render.call_args[1]
+    assert call_kwargs["eos_token"] == "<custom_eos>"
+    assert call_kwargs["bos_token"] == "<bos>"
+
+
+@pytest.mark.run_only_on("GPU")
+def test_apply_chat_template_eos_token_fallback(deployable):
+    """Test eos_token falls back to empty string when tokenizer lacks the attribute."""
+    del deployable.mcore_tokenizer._tokenizer.eos_token
+    messages = [{"role": "user", "content": "Hello"}]
+
+    template_mock = MagicMock()
+    template_mock.render.return_value = "rendered"
+
+    with patch("nemo_deploy.llm.megatronllm_deployable.Template", return_value=template_mock):
+        deployable.apply_chat_template(messages)
+
+    call_kwargs = template_mock.render.call_args[1]
+    assert call_kwargs["eos_token"] == ""
+
+
+@pytest.mark.run_only_on("GPU")
+def test_apply_chat_template_renders_template_using_eos_token(deployable):
+    """Regression test: chat templates that reference eos_token must render without UndefinedError."""
+    deployable.mcore_tokenizer._tokenizer.chat_template = (
+        "{{ bos_token }}{% for m in messages %}{{ m['content'] }}{{ eos_token }}{% endfor %}"
+    )
+    messages = [{"role": "user", "content": "Hello"}]
+
+    rendered = deployable.apply_chat_template(messages)
+
+    assert rendered == "<bos>Hello<eos>"
+
+
 @pytest.mark.run_only_on("GPU")
 def test_remove_eos_token(deployable):
     """Test EOS token removal covering all code paths."""