diff --git a/litellm/llms/openai/chat/guardrail_translation/handler.py b/litellm/llms/openai/chat/guardrail_translation/handler.py index b245394e1c0..7ea98fc5ce7 100644 --- a/litellm/llms/openai/chat/guardrail_translation/handler.py +++ b/litellm/llms/openai/chat/guardrail_translation/handler.py @@ -696,11 +696,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation): ModelResponse, stream_chunk_builder( chunks=[ # mutable-ok: callee takes a list - response.model_copy( - update=MappingProxyType( - {"choices": tuple(choice for choice in response.choices if choice.index == index)} - ) - ) + OpenAIChatCompletionsHandler._narrowed_to_choice(response, index) for response in responses_so_far ], logging_obj=litellm_logging_obj, @@ -710,16 +706,16 @@ class OpenAIChatCompletionsHandler(BaseTranslation): for index in choice_indices ) (_, base_response), *_ = rebuilt_by_index - return base_response.model_copy( - update=MappingProxyType( - { - "choices": tuple( - rebuilt.choices[0].model_copy(update=MappingProxyType({"index": index})) - for index, rebuilt in rebuilt_by_index - ) - } - ) - ) + stitched_choices: Final = [ # mutable-ok: choices is a List field; a tuple there breaks model_dump round-trips + rebuilt.choices[0].model_copy(update=MappingProxyType({"index": index})) + for index, rebuilt in rebuilt_by_index + ] + return base_response.model_copy(update=MappingProxyType({"choices": stitched_choices})) + + @staticmethod + def _narrowed_to_choice(response: "ModelResponseStream", index: int) -> "ModelResponseStream": + narrowed: Final = [choice for choice in response.choices if choice.index == index] # mutable-ok: List field + return response.model_copy(update=MappingProxyType({"choices": narrowed})) def build_stream_error_items( self, diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_moderations.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_moderations.py index 88b4ac7172a..c7adefe9886 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_moderations.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_moderations.py @@ -369,6 +369,7 @@ async def test_openai_moderation_guardrail_streaming_safe_content(): chunk1.choices[0].delta = MagicMock() chunk1.choices[0].delta.content = "Hello " chunk1.choices[0].finish_reason = None + chunk1.choices[0].index = 0 chunk2 = MagicMock() chunk2.model = "gpt-4" @@ -376,6 +377,7 @@ async def test_openai_moderation_guardrail_streaming_safe_content(): chunk2.choices[0].delta = MagicMock() chunk2.choices[0].delta.content = "world" chunk2.choices[0].finish_reason = None + chunk2.choices[0].index = 0 # Last chunk with finish_reason chunk3 = MagicMock() @@ -384,6 +386,7 @@ async def test_openai_moderation_guardrail_streaming_safe_content(): chunk3.choices[0].delta = MagicMock() chunk3.choices[0].delta.content = "!" chunk3.choices[0].finish_reason = "stop" + chunk3.choices[0].index = 0 for chunk in [chunk1, chunk2, chunk3]: yield chunk @@ -480,6 +483,7 @@ async def test_openai_moderation_guardrail_streaming_harmful_content(): chunk1.choices[0].delta = MagicMock() chunk1.choices[0].delta.content = "This is " chunk1.choices[0].finish_reason = None + chunk1.choices[0].index = 0 # Last chunk - with finish_reason to signal end of stream chunk2 = MagicMock() @@ -488,6 +492,7 @@ async def test_openai_moderation_guardrail_streaming_harmful_content(): chunk2.choices[0].delta = MagicMock() chunk2.choices[0].delta.content = "harmful content" chunk2.choices[0].finish_reason = "stop" + chunk2.choices[0].index = 0 for chunk in [chunk1, chunk2]: yield chunk diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_openai_moderation_streaming.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_openai_moderation_streaming.py index cb6772977ec..16f04073fae 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_openai_moderation_streaming.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/openai/test_openai_moderation_streaming.py @@ -42,6 +42,7 @@ async def test_openai_moderation_guardrail_streaming_latency(): choice.delta.content = content # Last chunk gets finish_reason choice.finish_reason = "stop" if i == len(chunks_data) - 1 else None + choice.index = 0 chunk.choices = [choice] yield chunk @@ -122,6 +123,7 @@ async def test_openai_moderation_guardrail_streaming_harmful_content(): choice.delta.content = content # Last chunk gets finish_reason choice.finish_reason = "stop" if i == len(chunks_data) - 1 else None + choice.index = 0 chunk.choices = [choice] yield chunk @@ -224,6 +226,7 @@ async def test_openai_moderation_streaming_end_of_stream_request_data_passthroug choice.delta = MagicMock() choice.delta.content = content choice.finish_reason = "stop" if i == len(chunks_data) - 1 else None + choice.index = 0 chunk.choices = [choice] yield chunk diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py index 5f3c09d9195..8bd9dc0df8a 100644 --- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py +++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py @@ -1836,6 +1836,22 @@ def _rewritten_model_response(response: Any) -> litellm.ModelResponse: return litellm.ModelResponse(**payload) +def _two_choice_stream_chunks() -> List[Any]: + return [ + litellm.ModelResponseStream(choices=[{"index": 0, "delta": {"content": "hello "}, "finish_reason": None}]), + litellm.ModelResponseStream(choices=[{"index": 1, "delta": {"content": "bonjour "}, "finish_reason": None}]), + litellm.ModelResponseStream(choices=[{"index": 0, "delta": {"content": "world"}, "finish_reason": "stop"}]), + litellm.ModelResponseStream(choices=[{"index": 1, "delta": {"content": "monde"}, "finish_reason": "stop"}]), + ] + + +def _rewritten_every_choice(response: Any) -> litellm.ModelResponse: + payload = response.model_dump() + for choice in payload["choices"]: + choice["message"]["content"] = "[REWRITTEN] " + choice["message"]["content"] + return litellm.ModelResponse(**payload) + + def test_streamable_post_call_pipelines_keeps_hook_guardrails_and_drops_iterator_only( make_user_api_key_auth, monkeypatch, caplog ): @@ -1984,6 +2000,39 @@ async def test_streaming_iterator_hook_runs_legacy_hook_and_delivers_its_rewrite assert _warnings(caplog) == [] +@pytest.mark.asyncio +async def test_streaming_iterator_hook_delivers_legacy_hook_rewrite_on_every_choice( + proxy_logging, make_user_api_key_auth, monkeypatch, caplog +): + seen: Dict[str, Any] = {} + guardrail = _legacy_hook_stream_guardrail(seen, rewrite=_rewritten_every_choice) + monkeypatch.setattr(litellm, "callbacks", [guardrail]) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False) + data = _post_call_pipeline_data(stream=True) + chunks = _two_choice_stream_chunks() + auth = make_user_api_key_auth(request_route="/v1/chat/completions") + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + await proxy_logging.pre_call_hook(user_api_key_dict=auth, data=data, call_type="completion", guardrails_only=True) + delivered = [ + item + async for item in proxy_logging.async_post_call_streaming_iterator_hook( + user_api_key_dict=auth, response=_async_chunk_iter(chunks), request_data=data + ) + ] + + assert [choice.message.content for choice in seen["response"].choices] == ["hello world", "bonjour monde"] + assert [id(item) for item in delivered] == [id(chunk) for chunk in chunks] + assert [(item.choices[0].index, item.choices[0].delta.content) for item in delivered] == [ + (0, "[REWRITTEN] hello world"), + (1, "[REWRITTEN] bonjour monde"), + (0, ""), + (1, ""), + ] + assert [item.choices[0].finish_reason for item in delivered] == [None, None, "stop", "stop"] + assert _warnings(caplog) == [] + + @pytest.mark.asyncio async def test_streaming_iterator_hook_releases_stream_untouched_when_legacy_hook_returns_none( proxy_logging, make_user_api_key_auth, monkeypatch