diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py index 6fd62e1a6ff..a3cb4479917 100644 --- a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py @@ -8,6 +8,9 @@ import litellm from litellm._logging import verbose_proxy_logger from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.litellm_logging import use_custom_pricing_for_model +from litellm.litellm_core_utils.prompt_templates.common_utils import ( + get_content_from_model_response, +) from litellm.llms.anthropic import get_anthropic_config from litellm.llms.anthropic.chat.handler import ( ModelResponseIterator as AnthropicModelResponseIterator, @@ -136,6 +139,84 @@ class AnthropicPassthroughLoggingHandler: return model return None + @staticmethod + def _stream_was_interrupted( + all_chunks: Sequence[Union[str, bytes]], + ) -> bool: + """ + Anthropic ends a stream with ``content_block_stop`` -> ``message_delta`` + -> ``message_stop``; a client disconnect leaves the last event mid + ``content_block_delta``. Scan from the tail and decide on the first + terminal-region event, so the common completed case is O(1) rather than + re-deserializing every line of the stream. + """ + for raw in reversed(all_chunks): + text = raw.decode("utf-8") if isinstance(raw, bytes) else raw + for line in reversed(text.splitlines()): + if not line.startswith("data:"): + continue + try: + data = json.loads(line[len("data:") :].strip()) + except (json.JSONDecodeError, ValueError): + continue + if not isinstance(data, dict): + continue + etype = data.get("type") + if etype == "message_delta": + return False + if etype in ( + "content_block_delta", + "content_block_stop", + "message_start", + ): + return True + return True + + @staticmethod + def _recover_interrupted_stream_output_tokens( + response: Union[ModelResponse, TextCompletionResponse], + all_chunks: Sequence[Union[str, bytes]], + model: str, + ) -> None: + """ + An Anthropic stream interrupted before its terminal ``message_delta`` + (client disconnect) carries only the ``message_start`` ``output_tokens`` + placeholder (typically 1-3), so completion tokens and spend are + undercounted ~20x. Re-tokenize the buffered output text to recover a + realistic ``output_tokens`` for usage/cost. Completed streams are + untouched because their terminal ``message_delta`` short-circuits here. + """ + if not isinstance(response, ModelResponse): + return + if not AnthropicPassthroughLoggingHandler._stream_was_interrupted(all_chunks): + return + usage = getattr(response, "usage", None) + if usage is None: + return + output_text = get_content_from_model_response(response) + if not output_text: + return + try: + recovered_output_tokens = litellm.token_counter( + model=model, text=output_text, count_response_tokens=True + ) + except Exception: + verbose_proxy_logger.warning( + "Could not re-tokenize interrupted stream output; " + "keeping placeholder completion token count." + ) + return + if recovered_output_tokens <= (usage.completion_tokens or 0): + return + usage.completion_tokens = recovered_output_tokens + usage.total_tokens = (usage.prompt_tokens or 0) + recovered_output_tokens + # Anthropic costing reads completion_tokens_details.text_tokens, so the + # stale message_start placeholder there must be corrected too or spend + # stays undercounted even after completion_tokens is fixed. + details = getattr(usage, "completion_tokens_details", None) + if details is not None and getattr(details, "text_tokens", None) is not None: + details.text_tokens = recovered_output_tokens + @staticmethod def _create_anthropic_response_logging_payload( litellm_model_response: Union[ModelResponse, TextCompletionResponse], @@ -277,6 +358,11 @@ class AnthropicPassthroughLoggingHandler: "result": None, "kwargs": {}, } + AnthropicPassthroughLoggingHandler._recover_interrupted_stream_output_tokens( + response=complete_streaming_response, + all_chunks=all_chunks, + model=model, + ) kwargs = AnthropicPassthroughLoggingHandler._create_anthropic_response_logging_payload( litellm_model_response=complete_streaming_response, model=model, diff --git a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py index 8c3ce348a0e..fd761eb37cb 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_anthropic_passthrough_logging_handler.py @@ -752,6 +752,148 @@ class TestBuildCompleteStreamingResponseRobustness: result = self._build(chunks) assert result is not None assert result.choices[0].message.content == "The stream ends with [DONE]" +class TestInterruptedStreamOutputTokenRecovery: + """ + When an Anthropic pass-through stream is interrupted (client disconnect) + before the terminal ``message_delta``, the only usage signal is the + ``message_start`` ``output_tokens`` placeholder (typically 1-3), so + completion tokens and spend are undercounted ~20x. The handler must + re-tokenize the buffered ``content_block_delta`` text to recover a + realistic ``output_tokens``; completed streams must stay untouched. + """ + + @staticmethod + def _sse(event, data): + return f"event: {event}\ndata: {json.dumps(data)}\n\n".encode() + + _MODEL = "claude-3-5-haiku-20241022" + _OUTPUT_TEXT = ( + "The history of computing spans centuries, beginning with mechanical " + "calculators and the abacus, advancing through Charles Babbage's " + "analytical engine, Ada Lovelace's first algorithm, Alan Turing's " + "theoretical machine, and the electronic computers of the twentieth " + "century that gave rise to the modern information age." + ) + + def _interrupted_chunks(self, *, placeholder_output_tokens: int = 2): + from litellm.proxy.pass_through_endpoints.streaming_handler import ( + PassThroughStreamingHandler, + ) + + words = self._OUTPUT_TEXT.split(" ") + frames = [ + self._sse( + "message_start", + { + "type": "message_start", + "message": { + "id": "msg_interrupted", + "type": "message", + "role": "assistant", + "model": self._MODEL, + "content": [], + "stop_reason": None, + "stop_sequence": None, + "usage": { + "input_tokens": 29, + "output_tokens": placeholder_output_tokens, + }, + }, + }, + ), + self._sse( + "content_block_start", + { + "type": "content_block_start", + "index": 0, + "content_block": {"type": "text", "text": ""}, + }, + ), + ] + for i, word in enumerate(words): + text = word if i == 0 else " " + word + frames.append( + self._sse( + "content_block_delta", + { + "type": "content_block_delta", + "index": 0, + "delta": {"type": "text_delta", "text": text}, + }, + ) + ) + # Client disconnects here: no content_block_stop / message_delta / + # message_stop are ever received. + return list(PassThroughStreamingHandler._convert_raw_bytes_to_str_lines(frames)) + + def _completed_chunks(self, *, final_output_tokens: int = 80): + chunks = self._interrupted_chunks() + chunks.append( + "data: " + + json.dumps( + { + "type": "message_delta", + "delta": {"stop_reason": "end_turn", "stop_sequence": None}, + "usage": {"output_tokens": final_output_tokens}, + } + ) + ) + chunks.append('data: {"type": "message_stop"}') + return chunks + + def _run(self, all_chunks): + logging_obj = MagicMock() + logging_obj.model_call_details = {"model": self._MODEL, "stream": True} + logging_obj.litellm_call_id = "test-call-id" + logging_obj.litellm_params = {} + logging_obj.get_router_model_id.return_value = None + + return AnthropicPassthroughLoggingHandler._handle_logging_anthropic_collected_chunks( + litellm_logging_obj=logging_obj, + passthrough_success_handler_obj=MagicMock(), + url_route="/anthropic/v1/messages", + request_body={"model": self._MODEL, "stream": True}, + endpoint_type="messages", + start_time=datetime.now(), + all_chunks=all_chunks, + end_time=datetime.now(), + ) + + def test_interrupted_stream_retokenizes_buffered_output(self): + import litellm + + placeholder = 2 + result = self._run( + self._interrupted_chunks(placeholder_output_tokens=placeholder) + ) + usage = result["result"].usage + + expected = litellm.token_counter( + model=self._MODEL, + text=self._OUTPUT_TEXT, + count_response_tokens=True, + ) + + assert expected > placeholder * 5 + assert usage.completion_tokens == expected + assert usage.completion_tokens > placeholder + assert usage.total_tokens == usage.prompt_tokens + expected + # Anthropic spend is priced off completion_tokens_details.text_tokens; if the + # placeholder leaks through here, cost stays undercounted even though + # completion_tokens looks right. + assert usage.completion_tokens_details.text_tokens == expected + + def test_completed_stream_keeps_message_delta_tokens(self): + final = 80 + result = self._run(self._completed_chunks(final_output_tokens=final)) + usage = result["result"].usage + + # Terminal message_delta present: recovery must not fire; the authoritative + # provider count is preserved verbatim. + assert usage.completion_tokens == final + + + class TestStreamFalseDeduplication: """ Regression tests for the duplicate-callback bug where a streaming pass-through