fix(count_tokens): let an explicit null drop_params replace the deployment's on the count like inference

The count now checks whether the request carries drop_params instead of whether its value is None, so a caller's null keeps summary and other fields on a drop_params deployment the way Router inference does. The test covers the null and the absent key and moves to claude-opus-5-5
This commit is contained in:
mateo-berri 2026-11-06 22:16:52 -08:00
parent a7b0bdc13c
commit 54ec2c8330
2 changed files with 17 additions and 11 deletions

View file

@ -15303,13 +15303,13 @@ _DEPLOYMENT_LITELLM_PARAMS: Final = TypeAdapter(Mapping[str, object])
def _deployment_with_request_drop_params(
deployment: Mapping[str, object] | None, drop_params: object
deployment: Mapping[str, object] | None, sent_params: Mapping[str, object]
) -> Mapping[str, object] | None:
if drop_params is None:
if "drop_params" not in sent_params:
return deployment
resolved_deployment: Final = deployment or {}
litellm_params: Final = _DEPLOYMENT_LITELLM_PARAMS.validate_python(resolved_deployment.get("litellm_params") or {})
return {**resolved_deployment, "litellm_params": {**litellm_params, "drop_params": drop_params}}
return {**resolved_deployment, "litellm_params": {**litellm_params, "drop_params": sent_params["drop_params"]}}
def _counted_request_messages(
@ -16151,7 +16151,7 @@ async def _count_request_tokens(
model_to_use = _model
sent_params: Final[Mapping[str, object]] = {**_router_default_litellm_params(llm_router), **request_litellm_params}
count_deployment: Final = _deployment_with_request_drop_params(deployment, sent_params.get("drop_params"))
count_deployment: Final = _deployment_with_request_drop_params(deployment, sent_params)
conversation_empty: Final = not messages and not contents
if prompt and not conversation_empty and not deployment_drops_params(count_deployment):
raise ProxyException(

View file

@ -2191,11 +2191,17 @@ async def test_anthropic_count_tokens_on_a_claude_deployment_sends_the_thinking_
@pytest.mark.asyncio
@pytest.mark.parametrize(
("deployment_drop_params", "request_drop_params"),
((True, "maybe"), (True, False), (None, " true ")),
("deployment_drop_params", "request_params"),
(
(True, {"drop_params": "maybe"}),
(True, {"drop_params": False}),
(True, {"drop_params": None}),
(True, {}),
(None, {"drop_params": " true "}),
),
)
async def test_anthropic_count_tokens_reads_the_request_drop_params_over_the_deployments_like_inference(
monkeypatch: pytest.MonkeyPatch, deployment_drop_params: bool | None, request_drop_params: JsonValue
monkeypatch: pytest.MonkeyPatch, deployment_drop_params: bool | None, request_params: Mapping[str, JsonValue]
) -> None:
monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
thinking: Final[dict[str, JsonValue]] = {"type": "enabled", "budget_tokens": 2000, "summary": "auto"}
@ -2206,19 +2212,19 @@ async def test_anthropic_count_tokens_reads_the_request_drop_params_over_the_dep
inference_route: Final = respx.post("https://api.anthropic.com/v1/messages").mock(
return_value=httpx.Response(200, json=_CLAUDE_MESSAGE_REPLY)
)
await _claude_router("anthropic/claude-opus-4-7", deployment_params).aanthropic_messages(
await _claude_router("anthropic/claude-opus-5-5", deployment_params).aanthropic_messages(
model="claude-agent",
messages=[{"role": "user", "content": "Weather in SF?"}],
max_tokens=4096,
thinking=thinking,
drop_params=request_drop_params,
**request_params,
)
sent: Final = json.loads(inference_route.calls.last.request.content)
response, (count,) = await _count_on_a_claude_deployment(
monkeypatch,
"anthropic/claude-opus-4-7",
{"thinking": thinking, "drop_params": request_drop_params},
"anthropic/claude-opus-5-5",
{"thinking": thinking, **request_params},
lambda _: httpx.Response(200, json={"input_tokens": 7}),
deployment_params=deployment_params,
)