diff --git a/litellm/llms/ollama/chat/transformation.py b/litellm/llms/ollama/chat/transformation.py index 59020b058a2..5c4fd4e1607 100644 --- a/litellm/llms/ollama/chat/transformation.py +++ b/litellm/llms/ollama/chat/transformation.py @@ -341,6 +341,8 @@ class OllamaChatConfig(BaseConfig): ) response_json: Final = raw_response.json() + if "error" in response_json: + raise _structured_error(f"{response_json['error']}") ## RESPONSE OBJECT _done_reason: Final = map_finish_reason(response_json.get("done_reason") or "stop") @@ -393,9 +395,10 @@ class OllamaChatConfig(BaseConfig): model_response.created = int(time.time()) model_response.model = "ollama_chat/" + model prompt_tokens = response_json.get("prompt_eval_count", litellm.token_counter(messages=messages)) - completion_tokens: Final = response_json.get( - "eval_count", - litellm.token_counter(text=response_json["message"]["content"]), + completion_tokens: Final = ( + response_json["eval_count"] + if "eval_count" in response_json + else litellm.token_counter(text=response_json["message"]["content"]) ) setattr( model_response, @@ -424,6 +427,14 @@ class OllamaChatConfig(BaseConfig): ) +def _structured_error(message: str) -> OllamaError: + return OllamaError( + message=message, + status_code=400, + headers={"Content-Type": "application/json"}, + ) + + def _done_chunk_usage(chunk: Mapping[str, object]) -> ChatCompletionUsageBlock | None: prompt_eval_count: Final = chunk.get("prompt_eval_count") eval_count: Final = chunk.get("eval_count") @@ -482,6 +493,9 @@ class OllamaChatCompletionResponseIterator(BaseModelResponseIterator): """ from litellm.types.utils import Delta, StreamingChoices + if "error" in chunk: + raise _structured_error(f"{chunk['error']}") + # process tool calls - if complete function arg - add id to tool call tool_calls: Final = chunk["message"].get("tool_calls") if tool_calls is not None: @@ -558,5 +572,3 @@ class OllamaChatCompletionResponseIterator(BaseModelResponseIterator): status_code=400, headers={"Content-Type": "application/json"}, ) - except Exception as e: - raise e diff --git a/tests/unit/llms/ollama/test_ollama_chat_transformation.py b/tests/unit/llms/ollama/test_ollama_chat_transformation.py index 98ca5aee407..55a205ed2d3 100644 --- a/tests/unit/llms/ollama/test_ollama_chat_transformation.py +++ b/tests/unit/llms/ollama/test_ollama_chat_transformation.py @@ -18,6 +18,7 @@ from litellm.llms.ollama.chat.transformation import ( OllamaChatCompletionResponseIterator, ) +from litellm.llms.ollama.common_utils import OllamaError from litellm.types.llms.openai import AllMessageValues from litellm.utils import get_llm_provider, get_optional_params @@ -661,6 +662,88 @@ class TestOllamaFinishReasonLength: ), f"Expected 'tool_calls' when tool_calls were streamed earlier, got '{done_result.choices[0].finish_reason}'" +OLLAMA_TOOL_CALL_PARSE_ERROR: Final = ( + "error parsing tool call: raw='{\"command\":\"grep -n flag\"]}', err=invalid character ']' after object key:value pair" +) + + +class TestOllamaStructuredErrorPayload: + def _transform(self, ollama_response: dict[str, object]) -> ModelResponse: + mock_response: Final = MagicMock() + mock_response.json.return_value = ollama_response + mock_response.text = json.dumps(ollama_response) + + model_response: Final = ModelResponse() + model_response.choices = [Choices(message=Message(content=""), index=0)] + + return OllamaChatConfig().transform_response( + model="gpt-oss:120b", + raw_response=mock_response, + model_response=model_response, + logging_obj=MagicMock(), + request_data={}, + messages=[{"role": "user", "content": "grep for flag"}], + optional_params={}, + litellm_params={}, + encoding=None, + api_key=None, + json_mode=False, + ) + + def test_streaming_error_chunk_surfaces_ollama_message(self): + iterator: Final = OllamaChatCompletionResponseIterator(streaming_response=iter([]), sync_stream=True) + + with pytest.raises(OllamaError) as exc_info: + iterator.chunk_parser({"error": OLLAMA_TOOL_CALL_PARSE_ERROR}) + + assert exc_info.value.message == OLLAMA_TOOL_CALL_PARSE_ERROR + assert exc_info.value.status_code == 400 + + def test_non_streaming_error_payload_surfaces_ollama_message(self): + with pytest.raises(OllamaError) as exc_info: + self._transform({"error": OLLAMA_TOOL_CALL_PARSE_ERROR}) + + assert exc_info.value.message == OLLAMA_TOOL_CALL_PARSE_ERROR + assert exc_info.value.status_code == 400 + + def test_streaming_error_chunk_does_not_mask_earlier_tool_call_state(self): + iterator: Final = OllamaChatCompletionResponseIterator(streaming_response=iter([]), sync_stream=True) + tool_chunk: Final = { + "model": "gpt-oss:120b", + "message": { + "role": "assistant", + "content": "", + "tool_calls": [{"function": {"name": "shell", "arguments": {"command": "ls"}}}], + }, + "done": False, + } + iterator.chunk_parser(tool_chunk) + + with pytest.raises(OllamaError) as exc_info: + iterator.chunk_parser({"error": "llama runner process has terminated: exit status 2"}) + + assert exc_info.value.message == "llama runner process has terminated: exit status 2" + + def test_usage_uses_eval_count_when_message_has_no_content(self): + result: Final = self._transform( + { + "model": "gpt-oss:120b", + "message": { + "role": "assistant", + "tool_calls": [{"function": {"name": "shell", "arguments": {"command": "ls"}}}], + }, + "done": True, + "done_reason": "stop", + "prompt_eval_count": 11, + "eval_count": 7, + } + ) + + assert result.usage.prompt_tokens == 11 + assert result.usage.completion_tokens == 7 + assert result.usage.total_tokens == 18 + + class TestOllamaReasoningContentStreaming: """Test that reasoning_content is properly extracted from all thinking chunks."""