[Bugfix][Frontend] Reject an empty conversation in every renderer

tests/renderers/test_empty_conversation.py

@@ -0,0 +1,63 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""An empty conversation must be rejected whatever the renderer.
+
+The HF tokenizer refuses one inside ``apply_chat_template``, but renderers with
+their own chat encoding (e.g. DeepSeek V3.2/V4/V4.1) render it to a bare BOS
+prompt, from which the model generates unrelated text.
+"""
+
+import pytest
+
+from vllm.exceptions import VLLMValidationError
+from vllm.renderers.base import BaseRenderer
+from vllm.renderers.params import ChatParams, TokenizeParams
+
+
+class _Rendered(Exception):
+    """Raised by the stub once a conversation reaches ``render_messages``."""
+
+
+def _renderer() -> BaseRenderer:
+    class _StubRenderer(BaseRenderer):
+        def __init__(self):
+            self.tokenizer = None
+
+        def render_messages(self, messages, params):
+            raise _Rendered
+
+    return _StubRenderer()
+
+
+def _render(renderer, conversations):
+    return renderer.render_chat(
+        conversations, ChatParams(), TokenizeParams(max_total_tokens=4096)
+    )
+
+
+@pytest.mark.parametrize(
+    "conversations",
+    [[[]], [[{"role": "user", "content": "hi"}], []]],
+    ids=["only", "one_of_two"],
+)
+def test_render_chat_rejects_an_empty_conversation(conversations):
+    with pytest.raises(VLLMValidationError, match="empty conversation"):
+        _render(_renderer(), conversations)
+
+
+@pytest.mark.asyncio
+async def test_render_chat_async_rejects_an_empty_conversation():
+    with pytest.raises(VLLMValidationError, match="empty conversation"):
+        await _renderer().render_chat_async(
+            [[]], ChatParams(), TokenizeParams(max_total_tokens=4096)
+        )
+
+
+@pytest.mark.parametrize(
+    "conversation",
+    [[{"role": "user", "content": "hi"}], [{"role": "system", "content": "Be brief."}]],
+    ids=["user", "system_only"],
+)
+def test_render_chat_renders_a_non_empty_conversation(conversation):
+    with pytest.raises(_Rendered):
+        _render(_renderer(), [conversation])

vllm/renderers/base.py

@@ -1339,6 +1339,19 @@ async def render_cmpl_async(
             )
         )
 
+    @staticmethod
+    def _check_conversations_not_empty(
+        conversations: Sequence[list["ChatCompletionMessageParam"]],
+    ) -> None:
+        # the HF tokenizer refuses an empty conversation in apply_chat_template,
+        # but renderers with their own chat encoding render it to a bare BOS
+        # prompt. Refuse it here so every renderer answers the same way.
+        if any(not conversation for conversation in conversations):
+            raise VLLMValidationError(
+                "Cannot apply chat template to an empty conversation. "
+                "Provide at least one message."
+            )
+
     def render_chat(
         self,
         conversations: Sequence[list["ChatCompletionMessageParam"]],
@@ -1348,6 +1361,7 @@ def render_chat(
         prompt_extras: dict[str, Any] | None = None,
         skip_mm_cache: bool = False,
     ):
+        self._check_conversations_not_empty(conversations)
         arrival_time = time.time()
 
         if tok_params is None:
@@ -1388,6 +1402,7 @@ async def render_chat_async(
         prompt_extras: dict[str, Any] | None = None,
         skip_mm_cache: bool = False,
     ):
+        self._check_conversations_not_empty(conversations)
         arrival_time = time.time()
 
         if tok_params is None: