Skip to content

The agent engineering platform.

fix(xai): drop unsupported `stop` parameter (#38335)

Mason Daugherty673ce8c0913b
diff --git a/libs/partners/xai/langchain_xai/chat_models.py b/libs/partners/xai/langchain_xai/chat_models.py
index aa78391768..7192dad7b2 100644
--- a/libs/partners/xai/langchain_xai/chat_models.py
+++ b/libs/partners/xai/langchain_xai/chat_models.py
@@ -41,6 +41,23 @@ def _get_default_model_profile(model_name: str) -> ModelProfile:
return default.copy()
+def _model_rejects_stop(model_name: str) -> bool:
+ """Whether an *unprofiled* xAI model rejects the `stop` parameter.
+
+ Used only as a fallback when no model profile is available; profiled models
+ defer to their `reasoning_output` flag, which is authoritative. The flag and
+ the name cannot be reconciled by string matching alone: the API accepts
+ `stop` for `grok-4.20-0309-non-reasoning` but rejects it for
+ `grok-4-fast-non-reasoning`, despite both containing `non-reasoning`.
+
+ Every current `grok-3`, `grok-4`, and `grok-code-fast` model that is not in
+ the generated profiles rejects `stop`, so the fallback drops it for those
+ families. Dropping a `stop` the API would reject degrades more gracefully
+ than letting the request fail outright.
+ """
+ return model_name.startswith(("grok-3", "grok-4")) or "grok-code-fast" in model_name
+
+
class ChatXAI(BaseChatOpenAI): # type: ignore[override]
r"""ChatXAI chat model.
@@ -426,6 +443,7 @@ class ChatXAI(BaseChatOpenAI): # type: ignore[override]
"""
openai_api_key: SecretStr | None = None
+
openai_api_base: str | None = None
model_config = ConfigDict(
@@ -546,6 +564,30 @@ class ChatXAI(BaseChatOpenAI): # type: ignore[override]
def _resolve_model_profile(self) -> ModelProfile | None:
return _get_default_model_profile(self.model_name) or None
+ def _get_request_payload(
+ self,
+ input_: LanguageModelInput,
+ *,
+ stop: list[str] | None = None,
+ **kwargs: Any,
+ ) -> dict:
+ payload = super()._get_request_payload(input_, stop=stop, **kwargs)
+ if payload.get("stop") is not None:
+ # xAI rejects `stop` for reasoning models. The model profile is
+ # authoritative when available (it correctly distinguishes models
+ # that only differ from each other by profile, not by name); the
+ # name-based check is a fallback for aliases absent from the
+ # generated profiles.
+ model_profile = self._resolve_model_profile()
+ rejects_stop = (
+ bool(model_profile.get("reasoning_output"))
+ if model_profile is not None
+ else _model_rejects_stop(self.model_name)
+ )
+ if rejects_stop:
+ payload.pop("stop", None)
+ return payload
+
def _stream(self, *args: Any, **kwargs: Any) -> Iterator[ChatGenerationChunk]:
"""Route to Chat Completions or Responses API."""
if self._use_responses_api({**kwargs, **self.model_kwargs}):
diff --git a/libs/partners/xai/tests/unit_tests/test_chat_models.py b/libs/partners/xai/tests/unit_tests/test_chat_models.py
index 0ecc4a88bb..66c949c3f3 100644
--- a/libs/partners/xai/tests/unit_tests/test_chat_models.py
+++ b/libs/partners/xai/tests/unit_tests/test_chat_models.py
@@ -84,6 +84,52 @@ def test_chat_xai_api_base_from_env(monkeypatch: pytest.MonkeyPatch) -> None:
assert llm.xai_api_base == "http://env.example.test/v1"
+ "model",
+ [
+ # Profiled reasoning models (`reasoning_output=True`).
+ "grok-4.3",
+ "grok-4.20-0309-reasoning",
+ # Unprofiled families that the live API rejects `stop` on. `grok-4`
+ # base and `grok-4-fast-non-reasoning` lack the substring "reasoning"
+ # yet still reject `stop`; `grok-code-fast` is a separate family.
+ "grok-3",
+ "grok-3-mini",
+ "grok-4",
+ "grok-4-0709",
+ "grok-4-fast-reasoning",
+ "grok-4-fast-non-reasoning",
+ "grok-code-fast-1",
+ ],
+)
+def test_reasoning_model_payload_drops_stop(model: str) -> None:
+ llm = ChatXAI(
+ model=model,
+ api_key=SecretStr("test-api-key"),
+ stop_sequences=["END"],
+ )
+
+ payload = llm._get_request_payload("hello")
+
+ assert "stop" not in payload
+
+
+def test_non_reasoning_model_payload_keeps_stop() -> None:
+ # `grok-4.20-0309-non-reasoning` is profiled with `reasoning_output=False`
+ # and the live API accepts `stop` for it, even though its name contains
+ # "non-reasoning" like the unprofiled `grok-4-fast-non-reasoning` that does
+ # not. The profile must take precedence over the name-based fallback.
+ llm = ChatXAI(
+ model="grok-4.20-0309-non-reasoning",
+ api_key=SecretStr("test-api-key"),
+ stop_sequences=["END"],
+ )
+
+ payload = llm._get_request_payload("hello")
+
+ assert payload["stop"] == ["END"]
+
+
def test_function_dict_to_message_function_message() -> None:
content = json.dumps({"result": "Example #1"})
name = "test_function"