fix(proxy): follow the router's order for tied team patterns and annotate quota tests

This commit is contained in:
mateo-berri 2026-10-28 03:05:37 -07:00
parent 8f00e1a839
commit 15f6e551ef
2 changed files with 28 additions and 20 deletions

View file

@ -234,8 +234,11 @@ def _routable_name(llm_router: "Router | None", model: str) -> str:
return model if llm_router is None else llm_router.routable_model_group(model)
def _outranks(rival: str, regex: str) -> bool:
return PatternUtils.calculate_pattern_specificity(rival) > PatternUtils.calculate_pattern_specificity(regex)
def _patterns_tried_before(patterns: Sequence[str], regex: str) -> Sequence[str]:
if regex in patterns:
return patterns[: patterns.index(regex)]
specificity: Final = PatternUtils.calculate_pattern_specificity(regex)
return tuple(rival for rival in patterns if PatternUtils.calculate_pattern_specificity(rival) >= specificity)
def _team_pattern_serves(llm_router: "Router", team_id: str, pattern: str, model: str) -> bool:
@ -243,14 +246,14 @@ def _team_pattern_serves(llm_router: "Router", team_id: str, pattern: str, model
if re.match(regex, model) is None or _named_deployments(llm_router, model, team_id):
return False
team_router: Final = llm_router.team_pattern_routers.get(team_id)
rivals: Final = () if team_router is None else tuple(team_router.patterns)
return not any(re.match(rival, model) is not None and _outranks(rival, regex) for rival in rivals)
patterns: Final = () if team_router is None else tuple(team_router.patterns)
return not any(re.match(rival, model) is not None for rival in _patterns_tried_before(patterns, regex))
def team_write_serves(llm_router: "Router | None", team_id: str, public_name: str, model: str) -> bool:
"""Whether a request from the team for ``model`` reaches the team route named ``public_name``: the name
itself or an alias of it, or, for a `*` name, a model no named deployment shadows and no more specific
team pattern takes."""
itself or an alias of it, or, for a `*` name, a model no named deployment shadows and no team pattern the
router tries first takes."""
target: Final = _routable_name(llm_router, model)
if "*" not in public_name or llm_router is None:
return target == public_name

View file

@ -1207,7 +1207,7 @@ def test_update_quota_existing_quotaless_model_rejected():
def _enforced_new_project_mocks(
monkeypatch, team_models: list[str], llm_router: object, quota_setting: object = True
monkeypatch: pytest.MonkeyPatch, team_models: list[str], llm_router: object, quota_setting: object = True
) -> None:
from litellm.proxy._types import LiteLLM_TeamTable
from litellm_enterprise.proxy.management_endpoints import project_endpoints as pe
@ -1650,14 +1650,14 @@ def _quota_error(
def test_quota_policy_exempts_provider_and_passthrough_name_but_keeps_vertex_strict():
assert _quota_error(["runway-gen4-turbo", "jev-latest"]) is None
error = _quota_error(["gemini-3.8-flash"], rpm={"gemini-3.8-flash": 10})
error: Final = _quota_error(["gemini-3.8-flash"], rpm={"gemini-3.8-flash": 10})
assert error is not None
assert "gemini-3.8-flash (model_tpm_limit)" in error
def test_quota_policy_model_pattern_overrides_its_provider_per_limit():
assert _quota_error(["veo-3.1-lite-generate-001"], rpm={"veo-3.1-lite-generate-001": 5}) is None
error = _quota_error(["veo-3.1-lite-generate-001"], tpm={"veo-3.1-lite-generate-001": 5000})
error: Final = _quota_error(["veo-3.1-lite-generate-001"], tpm={"veo-3.1-lite-generate-001": 5000})
assert error is not None
assert "veo-3.1-lite-generate-001 (model_rpm_limit)" in error
assert "model_tpm_limit" not in error
@ -1685,7 +1685,7 @@ def test_quota_policy_rejects_a_non_positive_limit_set_on_an_exempt_model(
def test_quota_policy_error_names_each_model_with_the_limits_it_misses():
error = _quota_error(["gemini-3.8-flash", "veo-3.1-lite-generate-001", "runway-gen4-turbo"])
error: Final = _quota_error(["gemini-3.8-flash", "veo-3.1-lite-generate-001", "runway-gen4-turbo"])
assert error is not None
assert "gemini-3.8-flash (model_rpm_limit and model_tpm_limit)" in error
assert "veo-3.1-lite-generate-001 (model_rpm_limit)" in error
@ -1716,7 +1716,7 @@ def test_quota_policy_precedence_is_model_then_pattern_then_provider_then_defaul
def test_quota_policy_model_entry_matches_names_literally_apart_from_star():
setting = {"models": {"claude-opus-5-5[1m]": [], "gpt.5": []}}
setting: Final = {"models": {"claude-opus-5-5[1m]": [], "gpt.5": []}}
assert _quota_error(["claude-opus-5-5[1m]"], setting=setting) is None
assert _quota_error(["claude-opus-5-51"], setting=setting) is not None
assert _quota_error(["gptx5"], setting=setting) is not None
@ -1726,7 +1726,7 @@ def test_quota_policy_model_on_several_providers_needs_every_required_limit():
def mixed_providers(_model: str) -> ModelRoute:
return ModelRoute(providers=frozenset({"runwayml", "vertex_ai"}))
error = _quota_error(["video-pool"], rpm={"video-pool": 10}, route_of=mixed_providers)
error: Final = _quota_error(["video-pool"], rpm={"video-pool": 10}, route_of=mixed_providers)
assert error is not None
assert "video-pool (model_tpm_limit)" in error
@ -1786,9 +1786,9 @@ def test_update_quota_policy_resolves_providers_for_the_resulting_models():
)
from litellm.proxy.management_helpers.model_quota_policy import parse_model_quota_setting
policy = parse_model_quota_setting(_CUSTOMER_QUOTA_POLICY)
policy: Final = parse_model_quota_setting(_CUSTOMER_QUOTA_POLICY)
assert policy is not None
existing = types.SimpleNamespace(models=["runway-gen4-turbo"], metadata={})
existing: Final = types.SimpleNamespace(models=["runway-gen4-turbo"], metadata={})
assert (
_raise_on_missing_project_model_quota_on_update(
UpdateProjectRequest(project_id="p", description="unrelated change"),
@ -1876,7 +1876,7 @@ def test_quota_policy_rejects_an_unknown_provider_id():
parse_model_quota_setting({"providers": {"vertex_ai": ["rpm"], "runway": []}})
def test_quota_policy_accepts_a_provider_registered_at_runtime(monkeypatch):
def test_quota_policy_accepts_a_provider_registered_at_runtime(monkeypatch: pytest.MonkeyPatch):
from litellm.proxy.management_helpers.model_quota_policy import parse_model_quota_setting
monkeypatch.setattr(litellm, "provider_list", [*litellm.provider_list, "qa-custom-llm"])
@ -2013,6 +2013,8 @@ def _team_pattern_router() -> "litellm.Router":
{"model_name": "video-named", "litellm_params": dict(_VERTEX_PARAMS)},
_team_deployment("video-*", "t1", {"model": "runwayml/*"}),
_team_deployment("video-veo-*", "t2", {"model": "vertex_ai/veo-*", "vertex_project": "qa-project"}),
_team_deployment("clip-*x", "t3", {"model": "runwayml/*"}),
_team_deployment("clip-x*", "t4", {"model": "vertex_ai/*", "vertex_project": "qa-project"}),
],
model_group_alias={"media": "media-latest"},
)
@ -2030,6 +2032,9 @@ def _team_pattern_router() -> "litellm.Router":
pytest.param("video-*", "video-named", False, id="a-named-deployment-shadows-the-pattern"),
pytest.param("video-*", "image-gen-1", False, id="a-name-the-pattern-misses"),
pytest.param("audio-*", "audio-gen-1", True, id="a-pattern-the-write-adds"),
pytest.param("clip-*x", "clip-xx", True, id="the-first-registered-of-two-tied-team-patterns"),
pytest.param("clip-x*", "clip-xx", False, id="a-tied-team-pattern-registered-first-takes-it"),
pytest.param("cli*-xx", "clip-xx", False, id="a-new-pattern-tied-with-a-registered-one"),
],
)
def test_team_write_serves_the_project_names_a_team_request_routes_through_it(
@ -2070,8 +2075,8 @@ def test_team_pattern_write_reads_each_served_name_by_the_model_it_calls(model:
@pytest.mark.asyncio
async def test_new_project_quota_policy_reads_providers_from_router_deployments(monkeypatch):
team_models = ["gemini-3.8-flash", "veo-3.1-lite-generate-001", "runway-gen4-turbo", "jev-latest"]
async def test_new_project_quota_policy_reads_providers_from_router_deployments(monkeypatch: pytest.MonkeyPatch):
team_models: Final = ["gemini-3.8-flash", "veo-3.1-lite-generate-001", "runway-gen4-turbo", "jev-latest"]
_enforced_new_project_mocks(
monkeypatch, team_models=team_models, llm_router=_customer_router(), quota_setting=_CUSTOMER_QUOTA_POLICY
)
@ -2089,7 +2094,7 @@ async def test_new_project_quota_policy_reads_providers_from_router_deployments(
@pytest.mark.asyncio
async def test_new_project_rejects_malformed_quota_setting_instead_of_skipping_it(monkeypatch):
async def test_new_project_rejects_malformed_quota_setting_instead_of_skipping_it(monkeypatch: pytest.MonkeyPatch):
_enforced_new_project_mocks(
monkeypatch,
team_models=["runway-gen4-turbo"],
@ -2105,7 +2110,7 @@ async def test_new_project_rejects_malformed_quota_setting_instead_of_skipping_i
@pytest.mark.asyncio
async def test_new_project_checks_permission_before_reading_the_quota_setting(monkeypatch):
async def test_new_project_checks_permission_before_reading_the_quota_setting(monkeypatch: pytest.MonkeyPatch):
from litellm_enterprise.proxy.management_endpoints import project_endpoints as pe
_enforced_new_project_mocks(
@ -2124,7 +2129,7 @@ async def test_new_project_checks_permission_before_reading_the_quota_setting(mo
@pytest.mark.asyncio
async def test_new_project_quota_policy_ignores_model_entries_for_a_team_model_name(monkeypatch):
async def test_new_project_quota_policy_ignores_model_entries_for_a_team_model_name(monkeypatch: pytest.MonkeyPatch):
router: Final = litellm.Router(
model_list=[
{