diff --git a/litellm/llms/anthropic/experimental_pass_through/messages/utils.py b/litellm/llms/anthropic/experimental_pass_through/messages/utils.py index 9deff950724..242300c7b6d 100644 --- a/litellm/llms/anthropic/experimental_pass_through/messages/utils.py +++ b/litellm/llms/anthropic/experimental_pass_through/messages/utils.py @@ -6,6 +6,7 @@ from litellm.types.llms.anthropic import AnthropicMessagesRequestOptionalParams from litellm.types.llms.anthropic_messages.anthropic_response import ( AnthropicMessagesResponse, ) +from litellm.types.llms.openai import ChatCompletionSystemMessage if TYPE_CHECKING: from litellm.exceptions import ContentPolicyViolationError @@ -36,6 +37,16 @@ def safeguard_refusal_error(model: str, stop_details: Mapping[str, object]) -> " ) +def anthropic_system_to_openai_message(system: object) -> ChatCompletionSystemMessage | None: + """ + Return the Anthropic Messages top-level ``system`` (a string or a list of text + blocks) as an OpenAI-style system message, or None when the request has none. + """ + if not isinstance(system, (str, list)) or not system: + return None + return ChatCompletionSystemMessage(role="system", content=system) + + @lru_cache(maxsize=1) def _anthropic_messages_optional_param_keys() -> frozenset[str]: """ diff --git a/litellm/router.py b/litellm/router.py index f33dfbba7bf..dea9aa62729 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -197,6 +197,7 @@ from litellm.router_utils.router_callbacks.track_deployment_metrics import ( from litellm.scheduler import FlowItem, Scheduler from litellm.types.llms.openai import ( AllMessageValues, + ChatCompletionToolParam, FileTypes, OpenAIFileObject, OpenAIFilesPurpose, @@ -11762,7 +11763,7 @@ class Router: self, messages: list[dict[str, str]] | None, input: str | list | None, - instructions: str | None = None, + request_kwargs: Mapping[str, object] | None = None, ) -> int: """ Count input tokens for context-window pre-call checks. @@ -11772,9 +11773,28 @@ class Router: The Responses payload is normalized to chat messages via the shared LiteLLMCompletionResponsesConfig transform so the same token_counter path covers both API surfaces and `instructions` tokens are included in the count. + + Prompt content the message list never carries is read from `request_kwargs`: + `tools` (Chat Completions, Responses and Anthropic Messages shapes) and the + Anthropic Messages top-level `system` block. """ + from litellm.llms.anthropic.experimental_pass_through.messages.utils import ( + anthropic_system_to_openai_message, + ) + + extras: Final = request_kwargs if request_kwargs is not None else MappingProxyType({}) + raw_instructions: Final = extras.get("instructions") + instructions: Final = raw_instructions if isinstance(raw_instructions, str) else None + raw_tools: Final = extras.get("tools") + tools: Final = ( + cast(list[ChatCompletionToolParam], raw_tools) # cast-ok: token_counter formats any tool dict shape + if isinstance(raw_tools, list) and raw_tools + else None + ) + system_message: Final = anthropic_system_to_openai_message(extras.get("system")) if messages is not None: - return litellm.token_counter(messages=messages) + counted_messages: Final = (system_message, *messages) if system_message is not None else messages + return litellm.token_counter(messages=counted_messages, tools=tools) if input is not None: from openai.types.responses.response_create_params import ResponseInputParam @@ -11787,7 +11807,10 @@ class Router: input=typed_input, responses_api_request={"instructions": instructions} if instructions is not None else {}, ) - return litellm.token_counter(messages=cast(list, input_messages)) # cast-ok: transformed chat messages + return litellm.token_counter( + messages=cast(list, input_messages), # cast-ok: transformed chat messages + tools=tools, + ) raise ValueError("Either messages or input must be provided to count tokens") def _deployment_max_input_tokens(self, model: str, deployment: Mapping[str, object]) -> int | None: @@ -11833,14 +11856,13 @@ class Router: """ if messages is None and input is None: return None - raw_instructions: Final = request_kwargs.get("instructions") if request_kwargs else None try: if not self._pre_call_checks_need_token_count(model, healthy_deployments): return None return await asyncify(self._count_pre_call_check_tokens)( messages=cast(list[dict[str, str]] | None, messages), # cast-ok: forwarded to the sync counter input=cast(str | list | None, input), # cast-ok: forwarded to the sync counter - instructions=raw_instructions if isinstance(raw_instructions, str) else None, + request_kwargs=request_kwargs, ) except Exception as e: # noqa: BLE001 # best-effort: an uncountable prompt must not fail the request verbose_router_logger.error( @@ -11887,8 +11909,6 @@ class Router: _rate_limit_error = False parent_otel_span: Final = _get_parent_otel_span_from_kwargs(request_kwargs) - raw_instructions: Final = request_kwargs.get("instructions") if request_kwargs else None - instructions: Final = raw_instructions if isinstance(raw_instructions, str) else None has_countable_input: Final = messages is not None or input is not None ## get model group RPM ## @@ -11919,7 +11939,7 @@ class Router: return _returned_deployments try: input_tokens = self._count_pre_call_check_tokens( - messages=messages, input=input, instructions=instructions + messages=messages, input=input, request_kwargs=request_kwargs ) except Exception as e: verbose_router_logger.error( diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 228588d974f..f7f0d79b4fd 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -3855,7 +3855,7 @@ def test_pre_call_checks_counts_responses_instructions_tokens(monkeypatch): input_only_tokens = router._count_pre_call_check_tokens(messages=None, input=short_input) with_instructions_tokens = router._count_pre_call_check_tokens( - messages=None, input=short_input, instructions=long_instructions + messages=None, input=short_input, request_kwargs={"instructions": long_instructions} ) assert with_instructions_tokens > input_only_tokens @@ -3871,6 +3871,164 @@ def test_pre_call_checks_counts_responses_instructions_tokens(monkeypatch): ) +_OVERSIZED_TOOL_DESCRIPTION = "look up the answer in the knowledge base. " * 40 + + +@pytest.mark.parametrize( + "prompt_kwargs, tool", + [ + pytest.param( + {"messages": [{"role": "user", "content": "hi"}]}, + { + "type": "function", + "function": { + "name": "lookup", + "description": _OVERSIZED_TOOL_DESCRIPTION, + "parameters": {"type": "object", "properties": {"q": {"type": "string"}}}, + }, + }, + id="chat_completions_tool", + ), + pytest.param( + {"input": "hi"}, + { + "type": "function", + "name": "lookup", + "description": _OVERSIZED_TOOL_DESCRIPTION, + "parameters": {"type": "object", "properties": {"q": {"type": "string"}}}, + }, + id="responses_tool", + ), + pytest.param( + {"messages": [{"role": "user", "content": "hi"}]}, + { + "name": "lookup", + "description": _OVERSIZED_TOOL_DESCRIPTION, + "input_schema": {"type": "object", "properties": {"q": {"type": "string"}}}, + }, + id="anthropic_messages_tool", + ), + ], +) +def test_pre_call_checks_counts_tool_definition_tokens(monkeypatch, prompt_kwargs, tool): + """ + Tool definitions are sent to the model as prompt tokens but never appear in + `messages` or `input`. A request whose prompt alone fits the context window but + whose prompt plus `tools` exceeds it must be rejected before dispatch, for the + Chat Completions, Responses and Anthropic Messages tool shapes alike. + """ + router = litellm.Router( + model_list=[ + {"model_name": "m", "litellm_params": {"model": "gpt-3.5-turbo"}}, + ], + enable_pre_call_checks=True, + ) + deployments = [ + {"litellm_params": {"model": "gpt-3.5-turbo"}, "model_info": {"id": "d1"}}, + ] + + prompt_only_tokens = router._count_pre_call_check_tokens( + messages=prompt_kwargs.get("messages"), input=prompt_kwargs.get("input") + ) + monkeypatch.setattr( + router, "get_router_model_info", lambda **kwargs: {"max_input_tokens": prompt_only_tokens} + ) + + assert len(router._pre_call_checks(model="m", healthy_deployments=deployments, **prompt_kwargs)) == 1 + with pytest.raises(litellm.ContextWindowExceededError): + router._pre_call_checks( + model="m", + healthy_deployments=deployments, + request_kwargs={"tools": [tool]}, + **prompt_kwargs, + ) + + +@pytest.mark.parametrize( + "system", + [ + pytest.param("You are a meticulous assistant. " * 40, id="system_string"), + pytest.param( + [{"type": "text", "text": "You are a meticulous assistant. " * 40}], + id="system_blocks", + ), + ], +) +def test_pre_call_checks_counts_anthropic_system_tokens(monkeypatch, system): + """ + The Anthropic Messages API carries the system prompt as a top-level `system` field, + not as a message. Its tokens reach the model, so a request whose `messages` fit but + whose `messages` plus `system` exceed the context window must be rejected. + """ + router = litellm.Router( + model_list=[ + {"model_name": "m", "litellm_params": {"model": "gpt-3.5-turbo"}}, + ], + enable_pre_call_checks=True, + ) + deployments = [ + {"litellm_params": {"model": "gpt-3.5-turbo"}, "model_info": {"id": "d1"}}, + ] + messages = [{"role": "user", "content": "hi"}] + + messages_only_tokens = router._count_pre_call_check_tokens(messages=messages, input=None) + monkeypatch.setattr(router, "get_router_model_info", lambda **kwargs: {"max_input_tokens": messages_only_tokens}) + + assert len(router._pre_call_checks(model="m", healthy_deployments=deployments, messages=messages)) == 1 + with pytest.raises(litellm.ContextWindowExceededError): + router._pre_call_checks( + model="m", + healthy_deployments=deployments, + messages=messages, + request_kwargs={"system": system}, + ) + + +@pytest.mark.asyncio +async def test_aanthropic_messages_enforces_context_window_with_system_and_tools(): + """ + End-to-end router regression for /v1/messages: a request whose only oversized + content lives in the top-level `system` field or in `tools` must trip the pre-call + context-window check instead of being dispatched (the deployment uses mock_response, + so reaching the provider handler would return a response rather than raise). + """ + router = litellm.Router( + model_list=[ + { + "model_name": "small-ctx", + "litellm_params": {"model": "anthropic/claude-3-5-haiku-20241022", "mock_response": "hi"}, + "model_info": {"max_input_tokens": 20}, + } + ], + enable_pre_call_checks=True, + ) + messages = [{"role": "user", "content": "hi"}] + + response = await router.aanthropic_messages(model="small-ctx", messages=messages, max_tokens=5) + assert response is not None + + with pytest.raises(litellm.ContextWindowExceededError): + await router.aanthropic_messages( + model="small-ctx", + messages=messages, + max_tokens=5, + system="You are a meticulous assistant. " * 40, + ) + with pytest.raises(litellm.ContextWindowExceededError): + await router.aanthropic_messages( + model="small-ctx", + messages=messages, + max_tokens=5, + tools=[ + { + "name": "lookup", + "description": _OVERSIZED_TOOL_DESCRIPTION, + "input_schema": {"type": "object", "properties": {"q": {"type": "string"}}}, + } + ], + ) + + def test_count_pre_call_check_tokens_across_api_surfaces(): """ _count_pre_call_check_tokens must count tokens from chat `messages`, a Responses