Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(vlm): scope LiteLLM thinking param to DashScope providers only (#958)

Deepak Dev Panwar54c36c8ffe2f
diff --git a/openviking/models/vlm/backends/litellm_vlm.py b/openviking/models/vlm/backends/litellm_vlm.py
index f868b8626..4ee8c8921 100644
--- a/openviking/models/vlm/backends/litellm_vlm.py
+++ b/openviking/models/vlm/backends/litellm_vlm.py
@@ -17,7 +17,7 @@ from typing import Any, Dict, List, Optional, Union
import litellm
from litellm import acompletion, completion
-from ..base import VLMBase, VLMResponse, ToolCall
+from ..base import ToolCall, VLMBase, VLMResponse
logger = logging.getLogger(__name__)
@@ -196,7 +196,14 @@ class LiteLLMVLMProvider(VLMBase):
else:
return {"type": "image_url", "image_url": {"url": image}}
- def _build_kwargs(self, model: str, messages: list, tools: Optional[List[Dict[str, Any]]] = None, tool_choice: Optional[str] = None) -> dict[str, Any]:
+ def _build_kwargs(
+ self,
+ model: str,
+ messages: list,
+ tools: Optional[List[Dict[str, Any]]] = None,
+ tool_choice: Optional[str] = None,
+ thinking: bool = False,
+ ) -> dict[str, Any]:
"""Build kwargs for LiteLLM call."""
kwargs: dict[str, Any] = {
"model": model,
@@ -223,6 +230,13 @@ class LiteLLMVLMProvider(VLMBase):
kwargs["tools"] = tools
kwargs["tool_choice"] = tool_choice or "auto"
+ # Only send enable_thinking to DashScope-compatible providers
+ provider = self._detected_provider or detect_provider_by_model(model)
+ if provider == "dashscope":
+ extra = kwargs.get("extra_body", {})
+ extra["enable_thinking"] = thinking
+ kwargs["extra_body"] = extra
+
return kwargs
def _parse_tool_calls(self, message) -> List[ToolCall]:
@@ -236,11 +250,7 @@ class LiteLLMVLMProvider(VLMBase):
args = json.loads(args)
except json.JSONDecodeError:
args = {"raw": args}
- tool_calls.append(ToolCall(
- id=tc.id,
- name=tc.function.name,
- arguments=args
- ))
+ tool_calls.append(ToolCall(id=tc.id, name=tc.function.name, arguments=args))
return tool_calls
def _build_vlm_response(self, response, has_tools: bool) -> Union[str, VLMResponse]:
@@ -282,7 +292,7 @@ class LiteLLMVLMProvider(VLMBase):
else:
kwargs_messages = [{"role": "user", "content": prompt}]
- kwargs = self._build_kwargs(model, kwargs_messages, tools, tool_choice)
+ kwargs = self._build_kwargs(model, kwargs_messages, tools, tool_choice, thinking=thinking)
t0 = time.perf_counter()
response = completion(**kwargs)
@@ -306,7 +316,7 @@ class LiteLLMVLMProvider(VLMBase):
else:
kwargs_messages = [{"role": "user", "content": prompt}]
- kwargs = self._build_kwargs(model, kwargs_messages, tools, tool_choice)
+ kwargs = self._build_kwargs(model, kwargs_messages, tools, tool_choice, thinking=thinking)
last_error = None
for attempt in range(max_retries + 1):
@@ -315,7 +325,8 @@ class LiteLLMVLMProvider(VLMBase):
response = await acompletion(**kwargs)
elapsed = time.perf_counter() - t0
self._update_token_usage_from_response(
- response, duration_seconds=elapsed,
+ response,
+ duration_seconds=elapsed,
)
return self._build_vlm_response(response, has_tools=bool(tools))
except Exception as e:
@@ -349,7 +360,7 @@ class LiteLLMVLMProvider(VLMBase):
content.append({"type": "text", "text": prompt})
kwargs_messages = [{"role": "user", "content": content}]
- kwargs = self._build_kwargs(model, kwargs_messages, tools)
+ kwargs = self._build_kwargs(model, kwargs_messages, tools, thinking=thinking)
t0 = time.perf_counter()
response = completion(**kwargs)
@@ -379,7 +390,7 @@ class LiteLLMVLMProvider(VLMBase):
content.append({"type": "text", "text": prompt})
kwargs_messages = [{"role": "user", "content": content}]
- kwargs = self._build_kwargs(model, kwargs_messages, tools)
+ kwargs = self._build_kwargs(model, kwargs_messages, tools, thinking=thinking)
t0 = time.perf_counter()
response = await acompletion(**kwargs)
@@ -388,7 +399,9 @@ class LiteLLMVLMProvider(VLMBase):
return self._build_vlm_response(response, has_tools=bool(tools))
def _update_token_usage_from_response(
- self, response, duration_seconds: float = 0.0,
+ self,
+ response,
+ duration_seconds: float = 0.0,
) -> None:
"""Update token usage from response."""
if hasattr(response, "usage") and response.usage:
diff --git a/tests/unit/test_vlm_thinking_param.py b/tests/unit/test_vlm_thinking_param.py
new file mode 100644
index 000000000..8628df682
--- /dev/null
+++ b/tests/unit/test_vlm_thinking_param.py
@@ -0,0 +1,90 @@
+# Copyright (c) 2026 Beijing Volcano Engine Technology Co., Ltd.
+# SPDX-License-Identifier: Apache-2.0
+"""Tests for LiteLLM thinking parameter scoping to DashScope providers."""
+
+from openviking.models.vlm.backends.litellm_vlm import LiteLLMVLMProvider
+
+
+class TestLiteLLMThinkingParam:
+ """Test that enable_thinking is only sent to DashScope-compatible providers."""
+
+ def _make_provider(self, model: str, **extra_config) -> LiteLLMVLMProvider:
+ config = {"model": model, "provider": "litellm", **extra_config}
+ return LiteLLMVLMProvider(config)
+
+ def test_litellm_dashscope_thinking_disabled(self):
+ """DashScope model with thinking=False should set enable_thinking=False in extra_body."""
+ vlm = self._make_provider("qwen-plus")
+ model = vlm._resolve_model("qwen-plus")
+ messages = [{"role": "user", "content": "hello"}]
+
+ kwargs = vlm._build_kwargs(model, messages, thinking=False)
+
+ assert "extra_body" in kwargs
+ assert kwargs["extra_body"]["enable_thinking"] is False
+
+ def test_litellm_dashscope_thinking_enabled(self):
+ """DashScope model with thinking=True should set enable_thinking=True in extra_body."""
+ vlm = self._make_provider("qwen-plus")
+ model = vlm._resolve_model("qwen-plus")
+ messages = [{"role": "user", "content": "hello"}]
+
+ kwargs = vlm._build_kwargs(model, messages, thinking=True)
+
+ assert "extra_body" in kwargs
+ assert kwargs["extra_body"]["enable_thinking"] is True
+
+ def test_litellm_non_dashscope_no_thinking_field(self):
+ """Non-DashScope model should NOT have enable_thinking in extra_body."""
+ vlm = self._make_provider("gpt-4o")
+ model = vlm._resolve_model("gpt-4o")
+ messages = [{"role": "user", "content": "hello"}]
+
+ kwargs = vlm._build_kwargs(model, messages, thinking=False)
+
+ extra_body = kwargs.get("extra_body", {})
+ assert "enable_thinking" not in extra_body
+
+ def test_litellm_non_dashscope_thinking_true_no_field(self):
+ """Non-DashScope model with thinking=True should still NOT have enable_thinking."""
+ vlm = self._make_provider("gpt-4o")
+ model = vlm._resolve_model("gpt-4o")
+ messages = [{"role": "user", "content": "hello"}]
+
+ kwargs = vlm._build_kwargs(model, messages, thinking=True)
+
+ extra_body = kwargs.get("extra_body", {})
+ assert "enable_thinking" not in extra_body
+
+ def test_litellm_anthropic_no_thinking_field(self):
+ """Anthropic model should NOT have enable_thinking in extra_body."""
+ vlm = self._make_provider("claude-3-opus")
+ model = vlm._resolve_model("claude-3-opus")
+ messages = [{"role": "user", "content": "hello"}]
+
+ kwargs = vlm._build_kwargs(model, messages, thinking=False)
+
+ extra_body = kwargs.get("extra_body", {})
+ assert "enable_thinking" not in extra_body
+
+ def test_litellm_ollama_no_thinking_field(self):
+ """Ollama model should NOT have enable_thinking in extra_body."""
+ vlm = self._make_provider("ollama/llama3")
+ model = vlm._resolve_model("ollama/llama3")
+ messages = [{"role": "user", "content": "hello"}]
+
+ kwargs = vlm._build_kwargs(model, messages, thinking=True)
+
+ extra_body = kwargs.get("extra_body", {})
+ assert "enable_thinking" not in extra_body
+
+ def test_dashscope_detected_via_model_name(self):
+ """Provider detection via model name 'dashscope' keyword should trigger enable_thinking."""
+ vlm = self._make_provider("dashscope/custom-model")
+ model = vlm._resolve_model("dashscope/custom-model")
+ messages = [{"role": "user", "content": "hello"}]
+
+ kwargs = vlm._build_kwargs(model, messages, thinking=True)
+
+ assert "extra_body" in kwargs
+ assert kwargs["extra_body"]["enable_thinking"] is True