diff --git a/litellm/proxy/hooks/parallel_request_limiter_v3.py b/litellm/proxy/hooks/parallel_request_limiter_v3.py index 2f457367e66..ebc0a72f95e 100644 --- a/litellm/proxy/hooks/parallel_request_limiter_v3.py +++ b/litellm/proxy/hooks/parallel_request_limiter_v3.py @@ -5575,18 +5575,18 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): else None, ) - await self._apply_usage_operations( - self._build_success_event_pipeline_operations( - kwargs=kwargs, - response_obj=response_obj, - rate_limit_type=rate_limit_type, - ), - litellm_parent_otel_span, + usage_operations: Final = self._build_success_event_pipeline_operations( + kwargs=kwargs, + response_obj=response_obj, + rate_limit_type=rate_limit_type, ) io_token_operations: Final = self._build_io_token_reservation_ops( kwargs=kwargs, response_obj=response_obj, ) + if stash is not None: + stash.clear_reserved_tokens() + await self._apply_usage_operations(usage_operations, litellm_parent_otel_span) if io_token_operations: if isinstance(io_token_operations, list): await self.async_increment_tokens_with_ttl_preservation( @@ -5598,8 +5598,6 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): pipeline_operations=io_token_operations, parent_otel_span=litellm_parent_otel_span, ) - if stash is not None: - stash.clear_reserved_tokens() except Exception as e: verbose_proxy_logger.exception("Error in rate limit success event: %s", e) @@ -5822,6 +5820,8 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): else () ) + if stash is not None and (reserved_tokens > 0 or itpm_reserved > 0 or otpm_reserved > 0): + stash.reservation_released = True if pipeline_operations: await self.internal_usage_cache.dual_cache.async_increment_cache_pipeline_post_call( pipeline_operations, parent_otel_span=litellm_parent_otel_span @@ -5836,8 +5836,6 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): pipeline_operations=project_operations, parent_otel_span=litellm_parent_otel_span, ) - if stash is not None and (reserved_tokens > 0 or itpm_reserved > 0 or otpm_reserved > 0): - stash.reservation_released = True if stash is None and tpm_actual > 0 and _is_socket_turn(failed_call): await self._apply_usage_operations( self._build_success_event_pipeline_operations( @@ -6040,6 +6038,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): itpm_reserved, otpm_reserved, ) + stash.reservation_released = True if combined_ops: await self.internal_usage_cache.dual_cache.async_increment_cache_pipeline( increment_list=list(combined_ops), @@ -6056,7 +6055,6 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): pipeline_operations=project_ops, parent_otel_span=user_api_key_dict.parent_otel_span, ) - stash.reservation_released = True except Exception as e: verbose_proxy_logger.exception("Error releasing TPM reservation on post-call failure: %s", e) return diff --git a/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py b/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py index 7fda0b8796d..c0f7156e85c 100644 --- a/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py +++ b/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py @@ -3746,6 +3746,53 @@ async def test_a_request_logging_a_failure_and_a_success_settles_its_reservation assert int(await local_cache.async_get_cache(key=tokens_key) or 0) == used_tokens +class _YieldingIncrementInMemoryCache(InMemoryCache): + async def async_increment(self, key: str, value: float, **kwargs: object) -> float: + await asyncio.sleep(0) + return await super().async_increment(key, value, **kwargs) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("success_starts_first", [True, False]) +@pytest.mark.parametrize("failure_path", ["failure_event", "post_call_failure_hook"]) +async def test_concurrent_failure_and_success_callbacks_settle_a_reservation_once_v3( + success_starts_first: bool, failure_path: str +) -> None: + api_key: Final = hash_token("sk-concurrent-failure-and-success") + local_cache: Final = DualCache(in_memory_cache=_YieldingIncrementInMemoryCache()) + handler: Final = _PROXY_MaxParallelRequestsHandler(internal_usage_cache=InternalUsageCache(local_cache)) + tokens_key: Final = handler.create_rate_limit_keys(key="api_key", value=api_key, rate_limit_type="tokens") + await _reserve_tpm_for_owner_call(handler, local_cache, api_key, "concurrent-call") + kwargs: Final = { + "litellm_call_id": "concurrent-call", + "standard_logging_object": {"metadata": {"user_api_key_hash": api_key}}, + } + now: Final = datetime.now() + log_failure: Final = ( + partial(handler.async_log_failure_event, kwargs=dict(kwargs), response_obj=None, start_time=now, end_time=now) + if failure_path == "failure_event" + else partial( + handler.async_post_call_failure_hook, + request_data=dict(kwargs), + original_exception=Exception("stream ended"), + user_api_key_dict=UserAPIKeyAuth(api_key=api_key, tpm_limit=10_000), + ) + ) + log_success: Final = partial( + handler.async_log_success_event, + kwargs=dict(kwargs), + response_obj=ModelResponse(usage=Usage(prompt_tokens=20, completion_tokens=10, total_tokens=30)), + start_time=now, + end_time=now, + ) + + await asyncio.gather( + *((log_success(), log_failure()) if success_starts_first else (log_failure(), log_success())) + ) + + assert int(await local_cache.async_get_cache(key=tokens_key) or 0) == 30 + + @pytest.mark.asyncio async def test_post_call_failure_hook_settles_reservation_at_recovered_partial_usage_v3(): """