From b5a7032eb4774b481f23359695ded8f4b1ea4835 Mon Sep 17 00:00:00 2001 From: yassin Date: Tue, 8 Sep 2026 18:42:38 +0000 Subject: [PATCH] fix(proxy): run the remaining inline token counts off the event loop Wrap the context-management editors, the end-of-stream chunk builder, acount_tokens, the compression interception hook, the passthrough interrupted-stream recovery, the A2A usage counters, and the semantic cache embedding truncation in asyncify so a multi-megabyte payload no longer stalls the worker's event loop while it is tokenized The pass-through suite now drains the process-global logging worker from an autouse conftest fixture so work queued on one test's loop cannot fire against the next test's callbacks Resolves LIT-7190 Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/a2a_protocol/main.py | 3 +- litellm/a2a_protocol/streaming_iterator.py | 5 +- litellm/caching/qdrant_semantic_cache.py | 3 +- litellm/caching/redis_semantic_cache.py | 3 +- .../compression_interception/handler.py | 3 +- .../litellm_core_utils/streaming_handler.py | 3 +- .../context_management/dispatcher.py | 9 +- .../context_management/editors/compact.py | 3 +- .../messages/streaming_iterator.py | 2 +- litellm/main.py | 4 +- .../streaming_handler.py | 13 +-- tests/pass_through_unit_tests/conftest.py | 17 +++ .../test_a2a_streaming_iterator.py | 54 ++++++++++ tests/test_litellm/a2a_protocol/test_main.py | 53 ++++++++- .../caching/test_qdrant_semantic_cache.py | 33 ++++++ .../caching/test_redis_semantic_cache.py | 29 +++++ .../test_compression_interception_handler.py | 25 +++++ .../test_streaming_handler.py | 47 ++++++++ .../context_management/test_compact.py | 31 ++++++ .../context_management/test_dispatcher.py | 32 ++++++ .../test_streaming_handler.py | 102 ++++++++++++++++++ .../test_count_tokens_public_api.py | 20 ++++ 22 files changed, 472 insertions(+), 22 deletions(-) diff --git a/litellm/a2a_protocol/main.py b/litellm/a2a_protocol/main.py index 0e8b8136c19..39600328074 100644 --- a/litellm/a2a_protocol/main.py +++ b/litellm/a2a_protocol/main.py @@ -21,6 +21,7 @@ from litellm._logging import verbose_logger, verbose_proxy_logger from litellm.a2a_protocol.streaming_iterator import A2AStreamingIterator from litellm.a2a_protocol.utils import A2ARequestUtils from litellm.constants import DEFAULT_A2A_AGENT_TIMEOUT +from litellm.litellm_core_utils.asyncify import asyncify from litellm.litellm_core_utils.litellm_logging import Logging from litellm.llms.custom_httpx.http_handler import ( get_async_httpx_client, @@ -507,7 +508,7 @@ async def asend_message( prompt_tokens, completion_tokens, _, - ) = A2ARequestUtils.calculate_usage_from_request_response( + ) = await asyncify(A2ARequestUtils.calculate_usage_from_request_response)( request=request, response_dict=response_dict, ) diff --git a/litellm/a2a_protocol/streaming_iterator.py b/litellm/a2a_protocol/streaming_iterator.py index 67db8e905e3..d936caeb75e 100644 --- a/litellm/a2a_protocol/streaming_iterator.py +++ b/litellm/a2a_protocol/streaming_iterator.py @@ -11,6 +11,7 @@ import litellm from litellm._logging import verbose_logger from litellm.a2a_protocol.cost_calculator import A2ACostCalculator from litellm.a2a_protocol.utils import A2ARequestUtils +from litellm.litellm_core_utils.asyncify import asyncify from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj if TYPE_CHECKING: @@ -99,11 +100,11 @@ class A2AStreamingIterator: # Calculate tokens from collected text input_message: Final = A2ARequestUtils.get_input_message_from_request(self.request) input_text: Final = A2ARequestUtils.extract_text_from_message(input_message) - prompt_tokens: Final = A2ARequestUtils.count_tokens(input_text) + prompt_tokens: Final = await asyncify(A2ARequestUtils.count_tokens)(input_text) # Use the last (most complete) text from chunks output_text: Final = self.collected_text_parts[-1] if self.collected_text_parts else "" - completion_tokens: Final = A2ARequestUtils.count_tokens(output_text) + completion_tokens: Final = await asyncify(A2ARequestUtils.count_tokens)(output_text) total_tokens: Final = prompt_tokens + completion_tokens diff --git a/litellm/caching/qdrant_semantic_cache.py b/litellm/caching/qdrant_semantic_cache.py index c5876e993d3..058cc8a1579 100644 --- a/litellm/caching/qdrant_semantic_cache.py +++ b/litellm/caching/qdrant_semantic_cache.py @@ -21,6 +21,7 @@ from litellm.constants import ( QDRANT_VECTOR_SIZE, SEMANTIC_CACHE_EMBEDDING_TIMEOUT_SECONDS, ) +from litellm.litellm_core_utils.asyncify import asyncify from litellm.litellm_core_utils.prompt_templates.common_utils import ( get_str_from_messages, ) @@ -255,7 +256,7 @@ class QdrantSemanticCache(BaseCache): llm_router = None router: Final = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list) - embedding_input: Final = self._embedding_input(prompt, router) + embedding_input: Final = await asyncify(self._embedding_input)(prompt, router) embedding_call: Final = ( router.aembedding( model=self.embedding_model, diff --git a/litellm/caching/redis_semantic_cache.py b/litellm/caching/redis_semantic_cache.py index 9a70bfc1418..d4c815e15b7 100644 --- a/litellm/caching/redis_semantic_cache.py +++ b/litellm/caching/redis_semantic_cache.py @@ -19,6 +19,7 @@ from typing import TYPE_CHECKING, Any, Final, cast import litellm from litellm._logging import print_verbose, verbose_logger from litellm.constants import SEMANTIC_CACHE_EMBEDDING_TIMEOUT_SECONDS +from litellm.litellm_core_utils.asyncify import asyncify from litellm.litellm_core_utils.prompt_templates.common_utils import ( get_str_from_messages, ) @@ -522,7 +523,7 @@ class RedisSemanticCache(BaseCache): llm_router = None router: Final = resolve_embedding_router(self.embedding_model, llm_router, llm_model_list) - embedding_input: Final = self._embedding_input(prompt, router) + embedding_input: Final = await asyncify(self._embedding_input)(prompt, router) embedding_call: Final = ( router.aembedding( model=self.embedding_model, diff --git a/litellm/integrations/compression_interception/handler.py b/litellm/integrations/compression_interception/handler.py index 1be7a01ba3a..5352ce6b6a0 100644 --- a/litellm/integrations/compression_interception/handler.py +++ b/litellm/integrations/compression_interception/handler.py @@ -15,6 +15,7 @@ from typing_extensions import ReadOnly, TypedDict from litellm._logging import verbose_logger from litellm.compression import compress from litellm.integrations.custom_logger import CustomLogger +from litellm.litellm_core_utils.asyncify import asyncify from litellm.types.integrations.compression_interception import ( CompressionInterceptionConfig, CompressionSavingsMetadata, @@ -153,7 +154,7 @@ class CompressionInterceptionLogger(CustomLogger): self._prune_expired_cache() - compressed: Final = compress( + compressed: Final = await asyncify(compress)( messages=messages, model=model, call_type=CallTypes.anthropic_messages, diff --git a/litellm/litellm_core_utils/streaming_handler.py b/litellm/litellm_core_utils/streaming_handler.py index 6ae17bac6ff..86b6d125554 100644 --- a/litellm/litellm_core_utils/streaming_handler.py +++ b/litellm/litellm_core_utils/streaming_handler.py @@ -19,6 +19,7 @@ from typing_extensions import NotRequired, TypedDict import litellm from litellm import verbose_logger from litellm._uuid import uuid +from litellm.litellm_core_utils.asyncify import asyncify from litellm.litellm_core_utils.model_response_utils import ( is_model_response_stream_empty, ) @@ -2247,7 +2248,7 @@ class CustomStreamWrapper: if self.sent_last_chunk is True: # log the final chunk with accurate streaming values try: - complete_streaming_response = litellm.stream_chunk_builder( + complete_streaming_response = await asyncify(litellm.stream_chunk_builder)( chunks=self.chunks, messages=self.messages, logging_obj=self.logging_obj, diff --git a/litellm/llms/anthropic/experimental_pass_through/context_management/dispatcher.py b/litellm/llms/anthropic/experimental_pass_through/context_management/dispatcher.py index 902808647c0..ad33e5e0592 100644 --- a/litellm/llms/anthropic/experimental_pass_through/context_management/dispatcher.py +++ b/litellm/llms/anthropic/experimental_pass_through/context_management/dispatcher.py @@ -5,6 +5,7 @@ from collections.abc import Awaitable, Callable from typing import TYPE_CHECKING, Final, TypeAlias from litellm._logging import verbose_logger +from litellm.litellm_core_utils.asyncify import asyncify from litellm.types.llms.anthropic import AppliedEdit from .constants import CLEAR_TOOL_USES_EDIT_TYPE, COMPACT_EDIT_TYPE @@ -82,9 +83,9 @@ async def apply_context_management( """Run edits in order; return a single ``PolyfillResult``. The dispatcher is async so async editors (``compact_20260112``) can - ``await`` the configured summarization model. Sync editors are called - inline — ``inspect.iscoroutinefunction`` decides how each editor is - invoked. + ``await`` the configured summarization model. Sync editors run in a + worker thread so their token counts stay off the event loop; + ``inspect.iscoroutinefunction`` decides how each editor is invoked. """ edits: Final = _normalize_spec(context_management_spec) if not edits: @@ -121,7 +122,7 @@ async def apply_context_management( user_api_key_auth=user_api_key_auth, ) if editor_is_async - else editor( + else await asyncify(editor)( model=model, messages=current_messages, tools=tools, diff --git a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py index 050ab67c86c..fb6a1c40253 100644 --- a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py +++ b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py @@ -20,6 +20,7 @@ from typing_extensions import NotRequired, ReadOnly, TypedDict, Unpack import litellm from litellm._logging import verbose_logger +from litellm.litellm_core_utils.asyncify import asyncify from litellm.types.llms.anthropic import ( AppliedEdit, CompactionBlock, @@ -1157,7 +1158,7 @@ async def apply_compact_20260112( # Phase B: threshold check. try: - current_tokens = _count_effective_tokens( + current_tokens = await asyncify(_count_effective_tokens)( model=model, effective_messages=effective_messages, # ``augmented_system`` already carries the prior compaction summary diff --git a/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py b/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py index 7d01aee5d98..98c5c6d6d4e 100644 --- a/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py +++ b/litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py @@ -678,7 +678,7 @@ class BaseAnthropicMessagesStreamingIterator: """ from litellm.proxy.pass_through_endpoints.streaming_handler import PassThroughStreamingHandler - PassThroughStreamingHandler.schedule_stream_failure_logging( + await PassThroughStreamingHandler.schedule_stream_failure_logging( litellm_logging_obj=self.litellm_logging_obj, endpoint_type=EndpointType.ANTHROPIC, request_body=self.request_body, diff --git a/litellm/main.py b/litellm/main.py index 75b7f7f10a5..bfbbbba2110 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -67,7 +67,7 @@ from litellm.constants import ( ) from litellm.exceptions import LiteLLMUnknownProvider from litellm.integrations.custom_logger import CustomLogger -from litellm.litellm_core_utils.asyncify import run_async_function +from litellm.litellm_core_utils.asyncify import asyncify, run_async_function from litellm.litellm_core_utils.audio_utils.utils import ( calculate_request_duration, get_audio_file_for_health_check, @@ -9076,7 +9076,7 @@ async def acount_tokens( fallback_messages = messages or [] if system and fallback_messages: fallback_messages = [{"role": "system", "content": system}] + fallback_messages - local_count: Final = litellm.token_counter( + local_count: Final = await asyncify(litellm.token_counter)( model=model, messages=fallback_messages, tools=tools, diff --git a/litellm/proxy/pass_through_endpoints/streaming_handler.py b/litellm/proxy/pass_through_endpoints/streaming_handler.py index 4be0235adbb..debda4321ef 100644 --- a/litellm/proxy/pass_through_endpoints/streaming_handler.py +++ b/litellm/proxy/pass_through_endpoints/streaming_handler.py @@ -8,6 +8,7 @@ import httpx import litellm from litellm._logging import verbose_proxy_logger +from litellm.litellm_core_utils.asyncify import asyncify from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from litellm.proxy._types import PassThroughEndpointLoggingResultValues @@ -60,7 +61,7 @@ class PassThroughStreamingHandler: litellm_logging_obj._update_completion_start_time(completion_start_time=datetime.now()) @staticmethod - def schedule_stream_failure_logging( + async def schedule_stream_failure_logging( litellm_logging_obj: LiteLLMLoggingObj, endpoint_type: EndpointType, request_body: dict[str, object], @@ -68,7 +69,7 @@ class PassThroughStreamingHandler: exception: Exception, stream_context: PassThroughStreamContext | None = None, ) -> None: - PassThroughStreamingHandler._record_partial_usage_for_failure( + await asyncify(PassThroughStreamingHandler._record_partial_usage_for_failure)( litellm_logging_obj=litellm_logging_obj, endpoint_type=endpoint_type, request_body=request_body, @@ -222,7 +223,7 @@ class PassThroughStreamingHandler: verbose_proxy_logger.error("Error in chunk_processor: %s", e) if response.status_code < 400: logging_scheduled = True - PassThroughStreamingHandler.schedule_stream_failure_logging( + await PassThroughStreamingHandler.schedule_stream_failure_logging( litellm_logging_obj=litellm_logging_obj, endpoint_type=endpoint_type, request_body=resolved_request_body, @@ -274,7 +275,7 @@ class PassThroughStreamingHandler: ( standard_logging_response_object, kwargs, - ) = PassThroughStreamingHandler._build_passthrough_logging_result( + ) = await asyncify(PassThroughStreamingHandler._build_passthrough_logging_result)( litellm_logging_obj=litellm_logging_obj, passthrough_success_handler_obj=passthrough_success_handler_obj, url_route=url_route, @@ -316,8 +317,8 @@ class PassThroughStreamingHandler: Synchronous, CPU-bound reconstruction of the standard logging payload from collected raw SSE bytes. Extracted from _route_streaming_logging_to_handler so the per-endpoint dispatch can - be unit-tested in isolation. Still invoked synchronously on the event - loop; an off-loop dispatch is a future change, not part of this PR. + be unit-tested in isolation. The async callers run it in a worker + thread so the token counts inside stay off the event loop. """ all_chunks: Final = PassThroughStreamingHandler._convert_raw_bytes_to_str_lines(raw_bytes) standard_logging_response_object: PassThroughEndpointLoggingResultValues | None = None diff --git a/tests/pass_through_unit_tests/conftest.py b/tests/pass_through_unit_tests/conftest.py index e6e98f790e8..df8196f1785 100644 --- a/tests/pass_through_unit_tests/conftest.py +++ b/tests/pass_through_unit_tests/conftest.py @@ -1,6 +1,8 @@ +import asyncio import pytest +from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from tests._vcr_conftest_common import ( # noqa: E402,F401 VerboseReporterState, @@ -45,6 +47,21 @@ def _vcr_outcome_gate(request, vcr): record_vcr_outcome(request, vcr) +@pytest.fixture(autouse=True) +async def _drain_logging_worker(): + """ + The logging queue is bound to the running loop, so anything left queued when a test's loop + goes away is carried onto the next loop and fires against that test's callbacks. + """ + GLOBAL_LOGGING_WORKER.start() + try: + await asyncio.wait_for(GLOBAL_LOGGING_WORKER.flush(), timeout=10) + except asyncio.TimeoutError: + pass + await GLOBAL_LOGGING_WORKER.stop() + yield + + def pytest_configure(config): _verbose_state.remember_pluginmanager(config) reset_vcr_diag_dir() diff --git a/tests/test_litellm/a2a_protocol/test_a2a_streaming_iterator.py b/tests/test_litellm/a2a_protocol/test_a2a_streaming_iterator.py index d86cbb94a91..2603d135dce 100644 --- a/tests/test_litellm/a2a_protocol/test_a2a_streaming_iterator.py +++ b/tests/test_litellm/a2a_protocol/test_a2a_streaming_iterator.py @@ -100,3 +100,57 @@ async def test_custom_logger_only_never_submits_sync_success_handler(monkeypatch assert recorder.async_hook_fired is True assert recording_executor.submitted_for(logging_obj) == [] + + +class _AgentChunk: + def __init__(self, text: str): + self._text = text + + def model_dump(self, mode: str, exclude_none: bool) -> dict: + return {"result": {"kind": "message", "role": "agent", "parts": [{"kind": "text", "text": self._text}]}} + + +@pytest.mark.asyncio +async def test_stream_completion_counts_tokens_off_the_event_loop(monkeypatch): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + warm_tokenizer("gpt-5.6-luna") + monkeypatch.setattr(litellm, "success_callback", []) + monkeypatch.setattr(litellm, "_async_success_callback", []) + logging_obj = LitellmLogging( + model="a2a/test-agent", + messages=[{"role": "user", "content": "hi"}], + stream=True, + call_type="a2a_send_message_streaming", + start_time=time.time(), + litellm_call_id="lit-7190-test", + function_id="lit-7190-test", + ) + + async def _stream(): + yield _AgentChunk(text * 100) + + iterator = A2AStreamingIterator( + stream=_stream(), + request=SimpleNamespace( + params=SimpleNamespace(message={"role": "user", "parts": [{"kind": "text", "text": text * 100}]}) + ), + logging_obj=logging_obj, + agent_name="test-agent", + ) + + async def drain() -> int: + return len([chunk async for chunk in iterator]) + + yielded, took, lags = await timed_with_loop_lags(drain) + + assert yielded == 1 + usage = logging_obj.model_call_details["usage"] + assert usage.prompt_tokens > 100_000 + assert usage.completion_tokens > 100_000 + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/a2a_protocol/test_main.py b/tests/test_litellm/a2a_protocol/test_main.py index 8850a2eca6c..318b40138ed 100644 --- a/tests/test_litellm/a2a_protocol/test_main.py +++ b/tests/test_litellm/a2a_protocol/test_main.py @@ -1,5 +1,7 @@ """Tests for litellm/a2a_protocol/main.py non-streaming send behavior.""" +import asyncio + import httpx import pytest @@ -13,7 +15,8 @@ from a2a.compat.v0_3.types import ( ) import litellm -from litellm.a2a_protocol.main import _send_message, _stream_messages, create_a2a_client +from litellm.integrations.custom_logger import CustomLogger +from litellm.a2a_protocol.main import _send_message, _stream_messages, asend_message, create_a2a_client from litellm.caching.llm_caching_handler import LLMClientCache from litellm.constants import DEFAULT_A2A_AGENT_TIMEOUT from litellm.llms.custom_httpx.http_handler import ( @@ -413,3 +416,51 @@ async def test_the_pooled_a2a_client_arrives_with_cookie_persistence_disabled(is assert dict(handler.client.cookies) == {}, "the pooled A2A client kept an upstream's cookie" await handler.close() + + +class _UsageRecorder(CustomLogger): + def __init__(self): + super().__init__() + self.logged = asyncio.Event() + self.payload = None + + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time): + self.payload = kwargs["standard_logging_object"] + self.logged.set() + + +@pytest.mark.asyncio +async def test_asend_message_counts_usage_off_the_event_loop(monkeypatch): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + warm_tokenizer("gpt-5.6-luna") + recorder = _UsageRecorder() + monkeypatch.setattr(litellm, "callbacks", [recorder]) + monkeypatch.setattr(litellm, "success_callback", [recorder]) + monkeypatch.setattr(litellm, "_async_success_callback", [recorder]) + + reply = _conv.pb2_v10.StreamResponse() + reply.message.message_id = "reply-1" + reply.message.role = _conv.pb2_v10.Role.ROLE_AGENT + reply.message.parts.add().text = text * 100 + request = SendMessageRequest( + id="r1", + params=MessageSendParams( + message={"messageId": "m1", "role": "user", "parts": [{"kind": "text", "text": text * 100}]} + ), + ) + + response, took, lags = await timed_with_loop_lags( + lambda: asend_message(a2a_client=_FakeClient(reply), request=request) + ) + + assert response.id == "r1" + await asyncio.wait_for(recorder.logged.wait(), timeout=10) + assert recorder.payload["prompt_tokens"] > 100_000 + assert recorder.payload["completion_tokens"] > 100_000 + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/caching/test_qdrant_semantic_cache.py b/tests/test_litellm/caching/test_qdrant_semantic_cache.py index e07578dd7e5..a0a9b71787c 100644 --- a/tests/test_litellm/caching/test_qdrant_semantic_cache.py +++ b/tests/test_litellm/caching/test_qdrant_semantic_cache.py @@ -1026,3 +1026,36 @@ def test_qdrant_semantic_cache_defaults_embedding_timeout(): cache = QdrantSemanticCache.__new__(QdrantSemanticCache) assert cache.embedding_timeout == SEMANTIC_CACHE_EMBEDDING_TIMEOUT_SECONDS assert SEMANTIC_CACHE_EMBEDDING_TIMEOUT_SECONDS < 60 + + +@pytest.mark.asyncio +async def test_qdrant_async_embedding_truncates_off_the_event_loop(monkeypatch): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + from litellm.caching.qdrant_semantic_cache import QdrantSemanticCache + + warm_tokenizer("sem-embed") + cache = QdrantSemanticCache.__new__(QdrantSemanticCache) + cache.embedding_model = "sem-embed" + cache.embedding_max_input_tokens = 5 + cache.embedding_timeout = 5 + + router = MagicMock() + router.get_configured_token_limits.return_value = (8191, None) + router.aembedding = AsyncMock(return_value={"data": [{"embedding": [0.1, 0.2]}]}) + monkeypatch.setitem( + sys.modules, + "litellm.proxy.proxy_server", + _router_proxy_module(router, "sem-embed"), + ) + + response, took, lags = await timed_with_loop_lags(lambda: cache._get_async_embedding(text * 100)) + + assert response["data"][0]["embedding"] == [0.1, 0.2] + assert _token_count("sem-embed", router.aembedding.call_args.kwargs["input"]) == 5 + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/caching/test_redis_semantic_cache.py b/tests/test_litellm/caching/test_redis_semantic_cache.py index df990c43530..6ea5f1e0007 100644 --- a/tests/test_litellm/caching/test_redis_semantic_cache.py +++ b/tests/test_litellm/caching/test_redis_semantic_cache.py @@ -1472,3 +1472,32 @@ def test_redis_semantic_cache_defaults_embedding_timeout(): cache = RedisSemanticCache.__new__(RedisSemanticCache) assert cache.embedding_timeout == SEMANTIC_CACHE_EMBEDDING_TIMEOUT_SECONDS assert SEMANTIC_CACHE_EMBEDDING_TIMEOUT_SECONDS < 60 + + +@pytest.mark.asyncio +async def test_redis_async_embedding_truncates_off_the_event_loop(monkeypatch): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + from litellm.caching.redis_semantic_cache import RedisSemanticCache + + warm_tokenizer("sem-embed") + cache = RedisSemanticCache.__new__(RedisSemanticCache) + cache.embedding_model = "sem-embed" + cache.embedding_max_input_tokens = 5 + cache.embedding_timeout = 5 + + router = MagicMock() + router.get_configured_token_limits.return_value = (8191, None) + router.aembedding = AsyncMock(return_value={"data": [{"embedding": [0.1, 0.2]}]}) + _proxy_with_router(monkeypatch, router, "sem-embed") + + embedding, took, lags = await timed_with_loop_lags(lambda: cache._get_async_embedding(text * 100)) + + assert embedding == [0.1, 0.2] + assert _token_count("sem-embed", router.aembedding.call_args.kwargs["input"]) == 5 + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/integrations/compression_interception/test_compression_interception_handler.py b/tests/test_litellm/integrations/compression_interception/test_compression_interception_handler.py index bc4dccc7d70..e66cd654f93 100644 --- a/tests/test_litellm/integrations/compression_interception/test_compression_interception_handler.py +++ b/tests/test_litellm/integrations/compression_interception/test_compression_interception_handler.py @@ -523,3 +523,28 @@ async def test_pre_call_hook_no_compression_records_no_savings(monkeypatch): await logger.async_pre_call_deployment_hook(kwargs=kwargs, call_type=CallTypes.anthropic_messages) assert "compression_savings" not in litellm_metadata + + +@pytest.mark.asyncio +async def test_pre_call_hook_counts_tokens_off_the_event_loop(): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + model = "anthropic/claude-fable-5" + warm_tokenizer(model) + logger = CompressionInterceptionLogger(compression_trigger=10_000_000) + messages = [{"role": "user", "content": text * 100}] + kwargs = {"model": model, "messages": messages} + + result, took, lags = await timed_with_loop_lags( + lambda: logger.async_pre_call_deployment_hook(kwargs=kwargs, call_type=CallTypes.anthropic_messages) + ) + + assert result is not None + assert result["messages"] is messages + assert "tools" not in result + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py index 0aa73833677..f16e24bb120 100644 --- a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py +++ b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py @@ -4875,3 +4875,50 @@ class TestStableStreamingResponseId: ) wrapper.response_id = "chatcmpl-from-provider" assert wrapper.model_response_creator().id == "chatcmpl-from-provider" + + +@pytest.mark.asyncio +async def test_async_stream_without_usage_counts_tokens_off_the_event_loop(): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + model = "gpt-5.6-luna" + warm_tokenizer(model) + messages = [{"role": "user", "content": text * 100}] + content_chunks = [_make_chunk(text) for _ in range(100)] + stop_chunk = ModelResponseStream( + id="test", + created=1741037890, + model=model, + choices=[StreamingChoices(index=0, delta=Delta(content=""), finish_reason="stop")], + ) + logging_obj = Logging( + model=model, + messages=messages, + stream=True, + call_type="acompletion", + start_time=time.time(), + litellm_call_id="12345", + function_id="1245", + ) + wrapper = CustomStreamWrapper( + completion_stream=ModelResponseListIterator(model_responses=content_chunks + [stop_chunk]), + model=model, + custom_llm_provider="openai", + logging_obj=logging_obj, + stream_options={"include_usage": True}, + ) + + async def consume() -> list[ModelResponseStream]: + return [chunk async for chunk in wrapper] + + chunks, took, lags = await timed_with_loop_lags(consume) + + assert "".join(chunk.choices[0].delta.content or "" for chunk in chunks) == text * 100 + assert chunks[-1].usage.prompt_tokens > 100_000 + assert chunks[-1].usage.completion_tokens > 100_000 + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py index 28c82fdf528..7660a8649b5 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py @@ -2605,3 +2605,34 @@ def test_build_summary_messages_keeps_midturn_system_correction_in_place(): assert summary_messages[0]["content"] == "caller system prompt" assert summary_messages[2]["content"] == "use the corrected result" assert summary_messages[-1]["content"] == "summarize the conversation" + + +async def test_threshold_check_counts_tokens_off_the_event_loop(monkeypatch): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + from litellm.llms.anthropic.experimental_pass_through.context_management.constants import ( + COMPACT_SUMMARY_MODEL_SETTING_KEY, + ) + from litellm.proxy.proxy_server import general_settings + + monkeypatch.setitem(general_settings, COMPACT_SUMMARY_MODEL_SETTING_KEY, "claude-haiku-4-5") + warm_tokenizer(MODEL) + messages = [{"role": "user", "content": text * 100}, *_simple_messages()] + result, took, lags = await timed_with_loop_lags( + lambda: apply_compact_20260112( + model=MODEL, + messages=messages, + tools=None, + system=None, + edit_spec={"type": "compact_20260112", "trigger": {"type": "input_tokens", "value": 10_000_000}}, + ) + ) + + assert result.messages == messages + assert result.compaction_block is None + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_dispatcher.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_dispatcher.py index 50c72cfe8d0..a21c22cf5fa 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_dispatcher.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_dispatcher.py @@ -129,3 +129,35 @@ async def test_malformed_edit_entries_are_skipped(): ) assert result.applied_edits == [] assert result.messages == messages + + +async def test_sync_editor_counts_tokens_off_the_event_loop(): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + warm_tokenizer(MODEL) + messages = [{"role": "user", "content": text * 100}, *_history_with_two_tool_pairs()] + + result, took, lags = await timed_with_loop_lags( + lambda: apply_context_management( + model=MODEL, + messages=messages, + tools=None, + system=None, + context_management_spec={ + "edits": [ + { + "type": "clear_tool_uses_20250919", + "trigger": {"type": "input_tokens", "value": 10_000_000}, + } + ] + }, + ) + ) + + assert result.messages == messages + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_streaming_handler.py b/tests/test_litellm/proxy/pass_through_endpoints/test_streaming_handler.py index dd9fbd9161f..e91b7ef970c 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_streaming_handler.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_streaming_handler.py @@ -7,6 +7,7 @@ import pytest import litellm from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from litellm.proxy.pass_through_endpoints.llm_provider_handlers.vertex_passthrough_logging_handler import ( VertexPassthroughLoggingHandler, ) @@ -128,3 +129,104 @@ def test_vertex_generate_content_payload_prices_gemini_urls_at_gemini_rates(): assert result["kwargs"]["response_cost"] == pytest.approx(GEMINI_COST) assert logging_obj.model_call_details["custom_llm_provider"] == "gemini" + + +def _interrupted_anthropic_stream(model: str, output_text: str) -> list[bytes]: + def sse(event: str, data: dict) -> bytes: + return f"event: {event}\ndata: {json.dumps(data)}\n\n".encode() + + message_start = { + "type": "message_start", + "message": { + "id": "msg_interrupted", + "type": "message", + "role": "assistant", + "model": model, + "content": [], + "stop_reason": None, + "stop_sequence": None, + "usage": {"input_tokens": 29, "output_tokens": 2}, + }, + } + block_start = {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}} + delta = {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": output_text}} + return [ + sse("message_start", message_start), + sse("content_block_start", block_start), + sse("content_block_delta", delta), + ] + + +@pytest.mark.asyncio +async def test_interrupted_anthropic_stream_recovers_output_tokens_off_the_event_loop(): + from unittest.mock import AsyncMock + + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + model = "claude-fable-5" + warm_tokenizer(model) + logging_obj = _logging_obj() + logging_obj.model_call_details = {"model": model, "stream": True} + logging_obj.litellm_params = {} + logging_obj.get_router_model_id.return_value = None + logging_obj.dispatch_success_handlers = AsyncMock() + + _, took, lags = await timed_with_loop_lags( + lambda: PassThroughStreamingHandler._route_streaming_logging_to_handler( + litellm_logging_obj=logging_obj, + passthrough_success_handler_obj=PassThroughEndpointLogging(), + url_route="/anthropic/v1/messages", + request_body={"model": model, "stream": True}, + endpoint_type=EndpointType.ANTHROPIC, + start_time=datetime.now(), + raw_bytes=_interrupted_anthropic_stream(model, text * 100), + end_time=datetime.now(), + model=model, + ) + ) + + logging_obj.dispatch_success_handlers.assert_awaited_once() + logged_usage = logging_obj.dispatch_success_handlers.await_args.kwargs["result"].usage + assert logged_usage.completion_tokens > 100_000 + assert_loop_stayed_free(took, lags) + + +@pytest.mark.asyncio +async def test_failed_anthropic_stream_records_partial_usage_off_the_event_loop(): + from unittest.mock import AsyncMock + + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + model = "claude-fable-5" + warm_tokenizer(model) + logging_obj = _logging_obj() + logging_obj.model_call_details = {"model": model, "stream": True} + logging_obj.litellm_params = {} + logging_obj.get_router_model_id.return_value = None + logging_obj.dispatch_failure_handlers = AsyncMock() + + _, took, lags = await timed_with_loop_lags( + lambda: PassThroughStreamingHandler.schedule_stream_failure_logging( + litellm_logging_obj=logging_obj, + endpoint_type=EndpointType.ANTHROPIC, + request_body={"model": model, "stream": True}, + raw_bytes=_interrupted_anthropic_stream(model, text * 100), + exception=RuntimeError("upstream closed the stream"), + ) + ) + await GLOBAL_LOGGING_WORKER.flush() + + logging_obj.dispatch_failure_handlers.assert_awaited_once() + partial_usage = logging_obj.record_partial_usage_for_failure.call_args.kwargs["usage"] + assert partial_usage.completion_tokens > 100_000 + assert_loop_stayed_free(took, lags) diff --git a/tests/test_litellm/test_count_tokens_public_api.py b/tests/test_litellm/test_count_tokens_public_api.py index 86c33c3e8f7..2918d0aa522 100644 --- a/tests/test_litellm/test_count_tokens_public_api.py +++ b/tests/test_litellm/test_count_tokens_public_api.py @@ -155,3 +155,23 @@ def test_acount_tokens_no_api_key_falls_back(monkeypatch): # Should fall back to local tokenizer since no API key assert result.total_tokens > 0 assert result.tokenizer_type == "local_tokenizer" + + +async def test_acount_tokens_local_fallback_counts_off_the_event_loop(): + from tests.large_text import text + from tests.test_litellm.litellm_core_utils.event_loop_lag import ( + assert_loop_stayed_free, + timed_with_loop_lags, + warm_tokenizer, + ) + + model = "together_ai/meta-llama/Llama-3-8b-chat-hf" + warm_tokenizer(model) + + result, took, lags = await timed_with_loop_lags( + lambda: litellm.acount_tokens(model=model, messages=[{"role": "user", "content": text * 100}]) + ) + + assert result.tokenizer_type == "local_tokenizer" + assert result.total_tokens > 100_000 + assert_loop_stayed_free(took, lags)