Enhance chunk parsing for Ollama streaming responses

Updated the chunk_parser method to return a ModelResponseStream when handling 'thinking' field content, allowing UIs to render reasoning information. Adjusted tests to verify the new behavior, ensuring that reasoning content is correctly returned in the response.

litellm/llms/ollama/completion/transformation.py

@@ -24,6 +24,8 @@ from litellm.types.utils import (
     ModelResponse,
     ModelResponseStream,
     ProviderField,
+    StreamingChoices,
+    Delta,
 )
 
 from ..common_utils import OllamaError, _convert_image
@@ -423,7 +425,7 @@ class OllamaTextCompletionResponseIterator(BaseModelResponseIterator):
     ) -> Union[GenericStreamingChunk, ModelResponseStream]:
         return self.chunk_parser(json.loads(str_line))
 
-    def chunk_parser(self, chunk: dict) -> GenericStreamingChunk:
+    def chunk_parser(self, chunk: dict) -> Union[GenericStreamingChunk, ModelResponseStream]:
         try:
             if "error" in chunk:
                 raise Exception(f"Ollama Error - {chunk}")
@@ -460,13 +462,15 @@ class OllamaTextCompletionResponseIterator(BaseModelResponseIterator):
                     usage=None,
                 )
             elif "thinking" in chunk and not chunk["response"]:
-                # Handle GPT-OSS models that include 'thinking' field with empty response
-                # These are intermediate chunks that don't contain user-facing content
-                return GenericStreamingChunk(
-                    text="",
-                    is_finished=is_finished,
-                    finish_reason="",
-                    usage=None,
+                # Return reasoning content as ModelResponseStream so UIs can render it
+                thinking_content = chunk.get("thinking") or ""
+                return ModelResponseStream(
+                    choices=[
+                        StreamingChoices(
+                            index=0,
+                            delta=Delta(reasoning_content=thinking_content),
+                        )
+                    ]
                 )
             else:
                 raise Exception(f"Unable to parse ollama chunk - {chunk}")

tests/test_litellm/llms/ollama/test_ollama_completion_transformation.py

@@ -14,7 +14,7 @@ from litellm.llms.ollama.completion.transformation import (
     OllamaConfig,
     OllamaTextCompletionResponseIterator,
 )
-from litellm.types.utils import Message, ModelResponse
+from litellm.types.utils import Message, ModelResponse, ModelResponseStream
 
 
 class TestOllamaConfig:
@@ -178,11 +178,10 @@ class TestOllamaTextCompletionResponseIterator:
 
         result = iterator.chunk_parser(chunk_with_thinking)
 
-        # Should return empty text and not be finished
-        assert result["text"] == ""
-        assert result["is_finished"] is False
-        assert result["finish_reason"] == ""
-        assert result["usage"] is None
+        # Should return a ModelResponseStream with reasoning content
+        assert isinstance(result, ModelResponseStream)
+        assert result.choices and result.choices[0].delta is not None
+        assert getattr(result.choices[0].delta, "reasoning_content") == "User"
 
     def test_chunk_parser_normal_response(self):
         """Test that normal response chunks still work."""