fix(decisions): count proxy /v1/decisions calls in least-busy's in-flight totals

litellm/decisions/call.py

@@ -557,6 +557,7 @@ def prepare_call(
             litellm_params={
                 "litellm_call_id": kwargs.get("litellm_call_id"),
                 "api_base": provider_config.get_complete_url(resolved_api_base, canonical_model),
+                **({} if (model_info := kwargs.get("model_info")) is None else {"model_info": model_info}),
             },
             custom_llm_provider=provider,
         )

tests/unit/decisions/test_main.py

@@ -6,6 +6,8 @@ import copy
 import json
 import re
 import threading
+import uuid
+from datetime import datetime
 from collections.abc import Mapping, Sequence
 from concurrent.futures import wait
 from types import MappingProxyType
@@ -296,6 +298,43 @@ async def test_router_dispatches_typesafe_decisions_without_api_base(
     assert response.answers["sentiment"].choice == "positive"
 
 
+@pytest.mark.asyncio
+async def test_least_busy_frees_a_decisions_deployment_once_a_proxy_logged_call_finishes(
+    respx_mock: respx.MockRouter,
+) -> None:
+    respx_mock.post("https://api.typesafe.ai/v1/systemone").respond(json=_RESPONSE)
+    router: Final = litellm.Router(
+        model_list=[
+            {
+                "model_name": "jev",
+                "litellm_params": {"model": "typesafe/jev-latest", "api_key": "k"},
+                "model_info": {"id": deployment_id},
+            }
+            for deployment_id in ("jev-a", "jev-b")
+        ],
+        routing_strategy="least-busy",
+    )
+
+    async def call_through_proxy_logging() -> object:
+        logging_obj, _ = litellm.utils.function_setup(
+            original_function="adecisions",
+            rules_obj=litellm.utils.Rules(),
+            start_time=datetime.now(),
+            model="jev",
+            litellm_call_id=str(uuid.uuid4()),
+        )
+        response: Final = await router.adecisions(
+            model="jev",
+            input="Is this a defect?",
+            questions=[{"type": "predicate", "name": "is_defect", "instructions": "Is this a defect?"}],
+            litellm_logging_obj=logging_obj,
+        )
+        await _drain_logging_worker()
+        return response._hidden_params["model_id"]
+
+    assert (await call_through_proxy_logging(), await call_through_proxy_logging()) == ("jev-a", "jev-a")
+
+
 def test_decisions_uses_the_same_wire_contract_for_sync_calls(respx_mock: respx.MockRouter) -> None:
     route: Final = respx_mock.post("https://api.perplexity.ai/v1/decisions").respond(json=_RESPONSE)
 

tests/unit/router_strategy/test_least_busy.py

@@ -236,9 +236,7 @@ class OpenBreakerRedis(SharedRedisCounters):
 
 
 @pytest.mark.asyncio
-async def test_an_open_circuit_breaker_falls_back_without_a_warning_per_request(
-    caplog: pytest.LogCaptureFixture,
-) -> None:
+async def test_an_open_circuit_breaker_falls_back_without_a_warning_per_request(caplog: pytest.LogCaptureFixture) -> None:
     worker: Final = _worker(OpenBreakerRedis())
 
     with caplog.at_level(logging.DEBUG, logger="LiteLLM Router"):
@@ -410,7 +408,9 @@ async def test_router_picks_least_busy_deployment_and_completion_restores_counts
         if use_async
         else router.get_available_deployment(model=ROUTER_GROUP, messages=None)
     )
-    response: Final = await router.acompletion(model=ROUTER_GROUP, messages=[{"role": "user", "content": "hi"}])
+    response: Final = await router.acompletion(
+        model=ROUTER_GROUP, messages=[{"role": "user", "content": "hi"}]
+    )
     for _ in range(10):
         await asyncio.sleep(0)
         await GLOBAL_LOGGING_WORKER.flush()