fix(health): look up a provider-prefixed base_model under its own provider and keep text-only operator modalities

This commit is contained in:
mateo-berri 2026-10-28 14:33:10 -07:00
parent e707374599
commit a0042f7a32
3 changed files with 28 additions and 9 deletions

View file

@ -135,10 +135,13 @@ _AUDIO_OUTPUT_PROBE_PROVIDERS: Final[frozenset[str]] = frozenset({"openai", "azu
def _health_check_requires_audio_output(
model_info: Mapping[str, object], probe_params: Mapping[str, object], mode: str | None, deployment_model: object
model_info: Mapping[str, object], probe_params: Mapping[str, object], mode: str | None, deployment_model: str | None
) -> bool:
if (mode or "chat") != "chat":
return False
operator_modalities: Final = probe_params.get("modalities")
if isinstance(operator_modalities, list) and "audio" not in operator_modalities:
return False
model: Final = probe_params.get("model")
if not isinstance(model, str):
return False
@ -152,14 +155,16 @@ def _health_check_requires_audio_output(
if provider not in _AUDIO_OUTPUT_PROBE_PROVIDERS:
return False
base_model: Final = model_info.get("base_model") or probe_params.get("base_model")
checked_model: Final = base_model if isinstance(base_model, str) and model == deployment_model else model
checked_model_prefix: Final = checked_model.partition("/")[0]
return litellm.supports_audio_output(
model=base_model if isinstance(base_model, str) and model == deployment_model else model,
custom_llm_provider=provider,
model=checked_model,
custom_llm_provider=checked_model_prefix if checked_model_prefix in _AUDIO_OUTPUT_PROBE_PROVIDERS else provider,
)
def _health_check_audio_output_params(
model_info: Mapping[str, object], probe_params: Mapping[str, object], mode: str | None, deployment_model: object
model_info: Mapping[str, object], probe_params: Mapping[str, object], mode: str | None, deployment_model: str | None
) -> Mapping[str, object]:
if not _health_check_requires_audio_output(model_info, probe_params, mode, deployment_model):
return MappingProxyType({})
@ -848,9 +853,11 @@ def update_litellm_params_for_health_check(
`audio` in the `health_check_voice` (default "alloy"), since those models reject text-only requests.
`base_model` (from `model_info`, then `litellm_params`) decides it only while the probe still targets
`deployment_model` (default: the incoming `litellm_params` model), so a `health_check_model` or a `model`
in `health_check_params` is judged on its own. Keys already set through `health_check_params` are kept
in `health_check_params` is judged on its own. Keys already set through `health_check_params` are kept,
and `modalities` set there without "audio" turns the audio fields off
"""
base_model_owner: Final = deployment_model or _PROBE_PARAMS_ADAPTER.validate_python(litellm_params).get("model")
incoming_model: Final = _PROBE_PARAMS_ADAPTER.validate_python(litellm_params).get("model")
base_model_owner: Final = deployment_model or (incoming_model if isinstance(incoming_model, str) else None)
mode: Final = resolve_health_check_mode(
model_info,
litellm_params, # any-ok: untyped router config dict

View file

@ -394,7 +394,7 @@ async def test_test_model_connection_loads_config_from_router():
mock_run_with_timeout = AsyncMock(return_value=mock_health_check_result)
# Mock _update_litellm_params_for_health_check
def mock_update_params(model_info, litellm_params, deployment_model=None):
def mock_update_params(model_info, litellm_params, deployment_model: str | None = None):
# Just return params with messages added
params = litellm_params.copy()
params["messages"] = [{"role": "user", "content": "test"}]
@ -545,7 +545,7 @@ async def test_test_model_connection_uses_model_info_id_to_disambiguate_duplicat
mock_ahealth_check = AsyncMock(return_value=mock_health_check_result)
mock_run_with_timeout = AsyncMock(return_value=mock_health_check_result)
def mock_update_params(model_info, litellm_params, deployment_model=None):
def mock_update_params(model_info, litellm_params, deployment_model: str | None = None):
params = litellm_params.copy()
params["messages"] = [{"role": "user", "content": "test"}]
return params
@ -656,7 +656,7 @@ async def test_test_model_connection_falls_back_to_deployments_zero_without_id()
mock_ahealth_check = AsyncMock(return_value=mock_health_check_result)
mock_run_with_timeout = AsyncMock(return_value=mock_health_check_result)
def mock_update_params(model_info, litellm_params, deployment_model=None):
def mock_update_params(model_info, litellm_params, deployment_model: str | None = None):
params = litellm_params.copy()
params["messages"] = [{"role": "user", "content": "test"}]
return params

View file

@ -855,6 +855,7 @@ def test_health_check_params_apply_to_non_chat_modes():
{"base_model": "gpt-4o-mini", "health_check_params": {"model": "openai/gpt-audio-mini"}},
{"model": "openai/my-text-model"},
),
({"base_model": "openai/gpt-audio-1.5"}, {"model": "azure/my-audio-deployment"}),
],
)
def test_openai_and_azure_audio_chat_probe_asks_for_audio_output(
@ -875,6 +876,8 @@ def test_openai_and_azure_audio_chat_probe_asks_for_audio_output(
({}, {"model": "azure/my-audio-deployment"}),
({}, {"model": "unmapped-model-name"}),
({"mode": "responses"}, {"model": "openai/gpt-audio-mini"}),
({"mode": "completion"}, {"model": "openai/gpt-audio-mini"}),
({"base_model": "gemini/lyria-3-clip-preview"}, {"model": "azure/my-audio-deployment"}),
(
{"base_model": "azure/gpt-audio-mini", "health_check_model": "azure/gpt-4o-mini"},
{"model": "azure/my-audio-deployment"},
@ -939,6 +942,15 @@ def test_every_audio_output_chat_model_in_the_cost_map_gets_the_audio_probe(cost
assert updated["modalities"] == ["text", "audio"]
def test_text_only_modalities_in_health_check_params_get_no_default_audio():
updated: Final = update_litellm_params_for_health_check(
{"health_check_params": {"modalities": ["text"]}}, {"model": "openai/gpt-audio-mini"}
)
assert updated["modalities"] == ["text"]
assert "audio" not in updated
def test_audio_chat_probe_uses_health_check_voice():
updated: Final = update_litellm_params_for_health_check(
{"health_check_voice": "verse"}, {"model": "openai/gpt-audio-mini"}