[OpenAI] Drop empty assistant turns for mistral_common tokenizers (#35915)

mistral_common rejects an assistant turn carrying neither content nor tool calls, while other chat templates ignore it, so an OpenAI-compatible request that works elsewhere failed on Mistral models. Drop those turns before templating; turns with tool_calls, multimodal content, or real text are untouched, and a trailing assistant message is already consumed upstream as the continue_final_message prefix.

python/sglang/srt/utils/hf_transformers/mistral_utils.py

@@ -628,9 +628,48 @@ def patch_mistral_common_tokenizer(tokenizer):
                 adapted.append(msg)
         return adapted
 
+    def _assistant_content_is_empty(content):
+        if content is None:
+            return True
+        if isinstance(content, str):
+            return not content.strip()
+        if isinstance(content, list):
+            return all(
+                isinstance(part, dict)
+                and part.get("type") in ("text", "input_text")
+                and not str(part.get("text") or "").strip()
+                for part in content
+            )
+        return False
+
+    def _drop_empty_assistant_messages(messages):
+        """Drop assistant turns with neither content nor tool calls, which
+        mistral_common rejects while other chat templates ignore them. A trailing
+        assistant turn is consumed upstream as the continue_final_message prefix,
+        so this cannot drop a prefill.
+        """
+        if not isinstance(messages, (list, tuple)):
+            return messages
+
+        kept = []
+        for msg in messages:
+            if isinstance(msg, (list, tuple)):
+                kept.append(_drop_empty_assistant_messages(msg))
+                continue
+            if (
+                isinstance(msg, dict)
+                and msg.get("role") == "assistant"
+                and not msg.get("tool_calls")
+                and _assistant_content_is_empty(msg.get("content"))
+            ):
+                continue
+            kept.append(msg)
+        return kept
+
     def _safe_apply_chat_template(messages, **kwargs):
         kwargs.pop("add_generation_prompt", None)
         messages = _adapt_placeholder_messages_for_mistral_common(messages)
+        messages = _drop_empty_assistant_messages(messages)
         return tokenizer._orig_apply_chat_template(messages, **kwargs)
 
     tokenizer.apply_chat_template = _safe_apply_chat_template

test/registered/unit/tokenizer/test_mistral_empty_assistant.py

@@ -0,0 +1,86 @@
+import unittest
+
+from sglang.srt.utils.hf_transformers.mistral_utils import (
+    patch_mistral_common_tokenizer,
+)
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=3, suite="base-a-test-cpu")
+
+
+def _user(text):
+    return {"role": "user", "content": text}
+
+
+def _assistant(**fields):
+    return {"role": "assistant", **fields}
+
+
+class _MistralCommonStub:
+    """Stands in for MistralCommonBackend; the class name gates the patch."""
+
+    def __init__(self):
+        self.seen = None
+        self.chat_template = "x"
+
+    def apply_chat_template(self, messages, **kwargs):
+        self.seen = messages
+        return []
+
+    def add_special_tokens(self, *args, **kwargs):
+        return 0
+
+    def convert_tokens_to_ids(self, value):
+        return 0
+
+    def decode(self, *args, **kwargs):
+        return ""
+
+    def batch_decode(self, *args, **kwargs):
+        return []
+
+
+class TestDropEmptyAssistantMessages(unittest.TestCase):
+    def _roles_passed_through(self, messages):
+        tokenizer = patch_mistral_common_tokenizer(_MistralCommonStub())
+        tokenizer.apply_chat_template(messages)
+        return [msg["role"] for msg in tokenizer.seen]
+
+    def test_empty_assistant_turn_is_dropped(self):
+        for content in ("", "   ", None, [], [{"type": "text", "text": ""}]):
+            with self.subTest(content=content):
+                roles = self._roles_passed_through(
+                    [_user("a"), _assistant(content=content), _user("b")]
+                )
+
+                self.assertEqual(roles, ["user", "user"])
+
+    def test_nonempty_assistant_turn_is_kept(self):
+        for content in (
+            "hi",
+            [{"type": "text", "text": "hi"}],
+            [
+                {"type": "text", "text": ""},
+                {"type": "image_url", "image_url": {"url": "https://x/y.png"}},
+            ],
+        ):
+            with self.subTest(content=content):
+                roles = self._roles_passed_through(
+                    [_user("a"), _assistant(content=content), _user("b")]
+                )
+
+                self.assertEqual(roles, ["user", "assistant", "user"])
+
+        roles = self._roles_passed_through(
+            [
+                _user("a"),
+                _assistant(content="", tool_calls=[{"id": "call_1"}]),
+                _user("b"),
+            ]
+        )
+
+        self.assertEqual(roles, ["user", "assistant", "user"])
+
+
+if __name__ == "__main__":
+    unittest.main()