Merge df6126edaa into 9ffb82e139

litellm/llms/ollama/chat/transformation.py

@@ -442,6 +442,7 @@ class OllamaChatCompletionResponseIterator(BaseModelResponseIterator):
     started_reasoning_content: bool = False
     finished_reasoning_content: bool = False
     seen_tool_calls: bool = False
+    _next_tool_call_index: int = 0
 
     def _is_function_call_complete(self, function_args: str | dict) -> bool:
         if isinstance(function_args, dict):
@@ -489,6 +490,8 @@ class OllamaChatCompletionResponseIterator(BaseModelResponseIterator):
             if tool_calls is not None:
                 self.seen_tool_calls = True
                 for tool_call in tool_calls:
+                    tool_call["index"] = self._next_tool_call_index
+                    self._next_tool_call_index += 1
                     function_args = tool_call.get("function").get("arguments")
                     if function_args is not None and len(function_args) > 0:
                         is_function_call_complete = self._is_function_call_complete(function_args)

tests/unit/llms/ollama/test_ollama_chat_transformation.py

@@ -1044,6 +1044,72 @@ class TestOllamaToolCallTransformation:
         assert tool_msg["tool_call_id"] == "call_abc123"
 
 
+class TestOllamaStreamingParallelToolCallIndex:
+    """Regression for https://github.com/BerriAI/litellm/issues/33678.
+
+    Ollama streams each parallel tool call in its own chunk without an OpenAI-style top-level
+    index. Every delta used to carry index 0, so index-accumulating clients appended the
+    arguments of every later call onto the first one.
+    """
+
+    @staticmethod
+    def _chunk(*tool_calls: dict) -> dict:
+        return {
+            "model": "qwen3:8b",
+            "message": {"role": "assistant", "content": "", "tool_calls": list(tool_calls)},
+            "done": False,
+        }
+
+    @staticmethod
+    def _parsed_tool_calls(iterator: OllamaChatCompletionResponseIterator, *chunks: dict) -> list:
+        return [
+            tool_call
+            for chunk in chunks
+            for tool_call in (iterator.chunk_parser(chunk).choices[0].delta.tool_calls or [])
+        ]
+
+    def test_tool_calls_in_separate_chunks_get_distinct_sequential_indices(self):
+        iterator = OllamaChatCompletionResponseIterator(streaming_response=iter([]), sync_stream=True)
+        chunks = [
+            self._chunk({"function": {"name": "bash", "arguments": {"command": command}}})
+            for command in ("date", "uname -a", "whoami")
+        ]
+
+        tool_calls = self._parsed_tool_calls(iterator, *chunks)
+
+        assert [tool_call.index for tool_call in tool_calls] == [0, 1, 2]
+        assert [json.loads(tool_call.function.arguments) for tool_call in tool_calls] == [
+            {"command": "date"},
+            {"command": "uname -a"},
+            {"command": "whoami"},
+        ]
+
+    def test_tool_calls_in_one_chunk_get_distinct_indices(self):
+        iterator = OllamaChatCompletionResponseIterator(streaming_response=iter([]), sync_stream=True)
+        chunk = self._chunk(
+            {"function": {"name": "bash", "arguments": {"command": "date"}}},
+            {"function": {"name": "bash", "arguments": {"command": "whoami"}}},
+        )
+
+        tool_calls = self._parsed_tool_calls(iterator, chunk)
+
+        assert [tool_call.index for tool_call in tool_calls] == [0, 1]
+
+    def test_index_counter_continues_across_chunks(self):
+        iterator = OllamaChatCompletionResponseIterator(streaming_response=iter([]), sync_stream=True)
+        chunks = [
+            self._chunk(
+                {"function": {"name": "bash", "arguments": {"command": "date"}}},
+                {"function": {"name": "bash", "arguments": {"command": "uname -a"}}},
+            ),
+            self._chunk({"function": {"name": "bash", "arguments": {"command": "whoami"}}}),
+        ]
+
+        tool_calls = self._parsed_tool_calls(iterator, *chunks)
+
+        assert [tool_call.index for tool_call in tool_calls] == [0, 1, 2]
+
+
 class TestOllamaStreamingUsage:
     @staticmethod
     def _parse(chunk: dict) -> ModelResponseStream: