diff --git a/litellm/constants.py b/litellm/constants.py index 1063b6ddeeb..f5acadc32ab 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -40,6 +40,7 @@ ROUTER_SETTINGS_MANAGED_OUTSIDE_CONFIG: Final[frozenset[str]] = frozenset( "router_general_settings", "ignore_invalid_deployments", "fallback_access_check", + "heuristic_v2_router_limit", } ) DEFAULT_BATCH_SIZE: Final = int(os.getenv("DEFAULT_BATCH_SIZE", 512)) diff --git a/litellm/proxy/auth/litellm_license.py b/litellm/proxy/auth/litellm_license.py index 677f1a0fdda..55bb1e3925a 100644 --- a/litellm/proxy/auth/litellm_license.py +++ b/litellm/proxy/auth/litellm_license.py @@ -16,6 +16,10 @@ if TYPE_CHECKING: from litellm.proxy._types import EnterpriseLicenseData +AUTO_ROUTER_LICENSE_FEATURE: Final = "auto_router" +HEURISTIC_V2_LICENSE_REMEDY: Final = "A LiteLLM license with the 'auto_router' feature lifts the limit." + + class LicenseCheck: """ - Check if license in env @@ -149,6 +153,19 @@ class LicenseCheck: return False return team_count > _max_teams_in_license + def heuristic_v2_router_limit(self) -> int | None: + """ + How many heuristic_v2 auto-routers this proxy may hold: unlimited (None) only when the + signed license lists the auto_router feature, otherwise one. A license verified through + the API carries no feature list, so it does not lift the limit either. + """ + if self.airgapped_license_data is None: + return 1 + allowed_features: Final = self.airgapped_license_data.get("allowed_features") + if isinstance(allowed_features, list) and AUTO_ROUTER_LICENSE_FEATURE in allowed_features: + return None + return 1 + def verify_license_without_api_request(self, public_key, license_key): try: from cryptography.hazmat.primitives import hashes @@ -179,19 +196,21 @@ class LicenseCheck: # Decode and parse the data license_data: Final = json.loads(message.decode()) - self.airgapped_license_data = EnterpriseLicenseData(**license_data) - # debug information provided in license data verbose_proxy_logger.debug("License data: %s", license_data) # Check expiration date expiration_date: Final = datetime.strptime(license_data["expiration_date"], "%Y-%m-%d") if expiration_date < datetime.now(): + self.airgapped_license_data = None return False, "License has expired" + self.airgapped_license_data = EnterpriseLicenseData(**license_data) + return True except Exception as e: + self.airgapped_license_data = None verbose_proxy_logger.debug( "litellm.proxy.auth.litellm_license.py::verify_license_without_api_request - Unable to verify License locally. - %s", e, diff --git a/litellm/proxy/management_endpoints/model_management_endpoints.py b/litellm/proxy/management_endpoints/model_management_endpoints.py index 613e726f89d..82ee33cbc39 100644 --- a/litellm/proxy/management_endpoints/model_management_endpoints.py +++ b/litellm/proxy/management_endpoints/model_management_endpoints.py @@ -13,10 +13,11 @@ model/{model_id}/update - PATCH endpoint for model update. import asyncio import datetime import json -from collections.abc import Awaitable, Mapping, Sequence +from collections.abc import AsyncGenerator, Awaitable, Callable, Mapping, Sequence +from contextlib import AbstractAsyncContextManager, asynccontextmanager from json import JSONDecodeError from types import MappingProxyType -from typing import TYPE_CHECKING, Annotated, Final, Literal, Protocol, cast +from typing import TYPE_CHECKING, Annotated, Final, Literal, Protocol, TypeVar, cast from fastapi import APIRouter, Depends, Header, HTTPException, Request, status from pydantic import BaseModel, ConfigDict, Field, ValidationError, field_validator @@ -49,6 +50,7 @@ from litellm.proxy._types import ( TeamModelDeleteRequest, UserAPIKeyAuth, ) +from litellm.proxy.auth.litellm_license import HEURISTIC_V2_LICENSE_REMEDY from litellm.proxy.auth.user_api_key_auth import user_api_key_auth from litellm.proxy.common_utils.config_sync_pubsub import ( coordination_redis_cache, @@ -96,6 +98,9 @@ from litellm.router_strategy.complexity_router import ( from litellm.router_utils.auto_router_model_naming import ( STRATEGY_ROUTER_PARAM_FIELDS, carries_complexity_router_settings, + count_heuristic_v2_routers, + heuristic_v2_limit_violation, + uses_heuristic_v2_classifier, validate_complexity_router_config_placement, validate_complexity_router_config_write, validate_strategy_router_model_write, @@ -153,6 +158,8 @@ class _ProxyModelTable(Protocol): def find_many(self, *, where: Mapping[str, object]) -> Awaitable[Sequence[_ProxyModelRow]]: ... + def create(self, *, data: Mapping[str, object]) -> Awaitable[_ProxyModelRow]: ... + def update( self, *, where: Mapping[str, object], data: Mapping[str, object] ) -> Awaitable[_ProxyModelRow | None]: ... @@ -166,6 +173,9 @@ class _TxModelTables(Protocol): litellm_proxymodeltable: _ProxyModelTable +_RowT = TypeVar("_RowT") + + class _ExistingModelRow(Protocol): @property def litellm_params(self) -> Mapping[str, object]: ... @@ -269,6 +279,66 @@ def _raise_on_strategy_router_write_violation( ) +HEURISTIC_V2_SLOT_LOCK_KEY: Final = 5_872_301 +_HEURISTIC_V2_LOCK_SQL: Final = "SELECT 1 AS locked FROM pg_advisory_xact_lock($1)" +_HEURISTIC_V2_DB_ROWS_SQL: Final = """ +SELECT count(*)::int AS held FROM "LiteLLM_ProxyModelTable" +WHERE model_id <> $1 + AND (CASE jsonb_typeof(litellm_params) WHEN 'string' THEN (litellm_params #>> '{}')::jsonb ELSE litellm_params END) + -> 'complexity_router_config' ->> 'classifier_type' = 'heuristic_v2' +""" + + +def _effective_complexity_router_config( + incoming_params: GenericLiteLLMParams | None, existing_params: GenericLiteLLMParams | None +) -> object: + """The complexity config a write leaves on the row: the incoming one when the write carries it, else the stored one.""" + incoming: Final = None if incoming_params is None else incoming_params.complexity_router_config + if incoming is not None or existing_params is None: + return incoming + return existing_params.complexity_router_config + + +@asynccontextmanager +async def _heuristic_v2_slot( + prisma_client: PrismaClient, *, effective_config: object, model_id: str | None +) -> AsyncGenerator[_ProxyModelTable, None]: + """Hand out the model table to write through while the row's claim on a heuristic_v2 slot is settled. + + A write that leaves the row on classifier_type heuristic_v2 under a limited license runs + inside one transaction that takes an advisory lock in its own statement before counting + (a statement's snapshot predates anything it locks), so pods cannot both pass the count: + the DB rows (any pod, either JSON shape) plus this proxy's config.yaml routers are judged + against the license limit and the write is refused with a 403 before it happens. The row + being edited keeps its own slot through ``model_id``. Every other write, and every write on + an unlimited license, goes through the repository table with no lock. Only the row write + itself may run inside: anything that needs a second connection (the team model bookkeeping) + must wait until the transaction has committed and the lock is released. The transaction + writes bypass the repository's publish-on-write, so the config change is published once + after commit, the way delete_team_models does. + """ + from litellm.proxy.proxy_server import _license_check, llm_router + + limit: Final = _license_check.heuristic_v2_router_limit() + if limit is None or not uses_heuristic_v2_classifier(effective_config): + yield _proxy_model_table(prisma_client) + return + async with prisma_client.db.tx() as tx_ctx: + tables: Final[_TxModelTables] = tx_ctx + await tx_ctx.query_raw(_HEURISTIC_V2_LOCK_SQL, HEURISTIC_V2_SLOT_LOCK_KEY) + rows: Sequence[Mapping[str, object]] = await tx_ctx.query_raw(_HEURISTIC_V2_DB_ROWS_SQL, model_id or "") + db_held: Final = rows[0].get("held") if rows else 0 + config_rows: Final = () if llm_router is None else tuple(llm_router.config_deployments()) + held: Final = (db_held if isinstance(db_held, int) else 0) + count_heuristic_v2_routers(config_rows) + violation: Final = heuristic_v2_limit_violation(held=held + 1, limit=limit) + if violation is not None: + raise HTTPException( + status_code=status.HTTP_403_FORBIDDEN, detail=f"{violation} {HEURISTIC_V2_LICENSE_REMEDY}" + ) + yield tables.litellm_proxymodeltable + await publish_config_change(redis_cache=coordination_redis_cache(), object_type="litellm_proxymodeltable") + + ENFORCE_RPM_TPM_ON_MODEL_ADD_SETTING: Final = "enforce_rpm_tpm_on_model_add" _REQUIRED_RATE_LIMIT_FIELDS: Final = ("rpm", "tpm") @@ -720,22 +790,29 @@ async def patch_model( ) requested_model_name: Final = patch_data.model_name + stored_model_name: str | None = None + + async def write_row(update_data: PrismaCompatibleUpdateDBModel) -> _ProxyModelRow | None: + nonlocal stored_model_name + stored_model_name = update_data.get("model_name") + update_data["updated_by"] = user_api_key_dict.user_id or litellm_proxy_admin_name + update_data["updated_at"] = cast(str, get_utc_datetime()) + async with _heuristic_v2_slot( + prisma_client, + effective_config=_effective_complexity_router_config( + patch_data.litellm_params, db_model.litellm_params + ), + model_id=model_id, + ) as table: + return await table.update(where={"model_id": model_id}, data=update_data) + # Handle team model updates with proper alias management - update_data: Final = await _update_team_model_in_db( + updated_model: Final = await _update_team_model_in_db( db_model=db_model, patch_data=patch_data, user_api_key_dict=user_api_key_dict, prisma_client=prisma_client, - ) - - # Add metadata about update - update_data["updated_by"] = user_api_key_dict.user_id or litellm_proxy_admin_name - update_data["updated_at"] = cast(str, get_utc_datetime()) - - # Perform partial update - updated_model: Final = await _proxy_model_table(prisma_client).update( - where={"model_id": model_id}, - data=update_data, + write_row=write_row, ) if updated_model is None: @@ -746,7 +823,6 @@ async def patch_model( param=None, ) - stored_model_name: Final = update_data.get("model_name") if ( stored_model_name is not None and stored_model_name == requested_model_name @@ -980,7 +1056,8 @@ async def _add_model_to_db( prisma_client: PrismaClient, new_encryption_key: str | None = None, should_create_model_in_db: bool = True, -) -> "prisma_models.LiteLLM_ProxyModelTable | LiteLLM_ProxyModelTable | None": + slot: AbstractAsyncContextManager[_ProxyModelTable] | None = None, +) -> "_ProxyModelRow | LiteLLM_ProxyModelTable": # encrypt litellm params # _litellm_params_dict: Final = model_params.litellm_params.dict(exclude_none=True) _original_litellm_model_name: Final = model_params.litellm_params.model @@ -998,18 +1075,20 @@ async def _add_model_to_db( if model_params.model_info.id is not None: _data["model_id"] = model_params.model_info.id _create_data: Final = cast("Mapping[str, object]", _data) # cast-ok: str-keyed json payload built just above - if should_create_model_in_db: - model_response = await ModelRepository(prisma_client).table.create(data=_create_data) - else: - model_response = LiteLLM_ProxyModelTable(**_data) - return model_response + if not should_create_model_in_db: + return LiteLLM_ProxyModelTable(**_data) + if slot is None: + return await _proxy_model_table(prisma_client).create(data=_create_data) + async with slot as table: + return await table.create(data=_create_data) async def _add_team_model_to_db( model_params: Deployment, user_api_key_dict: UserAPIKeyAuth, prisma_client: PrismaClient, -) -> "prisma_models.LiteLLM_ProxyModelTable | LiteLLM_ProxyModelTable | None": + slot: AbstractAsyncContextManager[_ProxyModelTable] | None = None, +) -> "_ProxyModelRow | LiteLLM_ProxyModelTable": """ If 'team_id' is provided, @@ -1040,6 +1119,7 @@ async def _add_team_model_to_db( model_params=model_params, user_api_key_dict=user_api_key_dict, prisma_client=prisma_client, + slot=slot, ) if original_model_name: @@ -1060,7 +1140,8 @@ async def _update_team_model_in_db( patch_data: updateDeployment, user_api_key_dict: UserAPIKeyAuth, prisma_client: PrismaClient, -) -> PrismaCompatibleUpdateDBModel: + write_row: Callable[[PrismaCompatibleUpdateDBModel], Awaitable[_RowT]], +) -> _RowT: """ Handle team model updates with proper alias management. @@ -1068,6 +1149,9 @@ async def _update_team_model_in_db( - Creates unique internal model_name and team alias - Adds model to team object - Preserves team_public_model_name for external reference + + The row is written through ``write_row`` before the team's model list is touched, so a + refused or failed write leaves the team as it was (the create path orders itself the same way). """ # Validate team_id if present in patch_data from litellm.proxy.proxy_server import premium_user @@ -1079,9 +1163,7 @@ async def _update_team_model_in_db( premium_user=premium_user, ) - # Validated before any write, beside the premium check the create path already runs - # here. The team ACL is updated below and autocommits, so a validator that raises - # further down would leave the team mutated and the deployment row never written. + # Validated before the row write, beside the premium check the create path already runs here. # # The merged view is what gets stored, so that is what has to satisfy the invariants. # Validating the patch alone rejected a partial edit of an already valid deployment: @@ -1101,7 +1183,7 @@ async def _update_team_model_in_db( # No team_id in patch, proceed with standard update if patch_team_id is None: - return update_db_model(db_model=db_model, updated_patch=patch_data) + return await write_row(update_db_model(db_model=db_model, updated_patch=patch_data)) # Determine public model name public_model_name: Final = _get_public_model_name( @@ -1120,11 +1202,14 @@ async def _update_team_model_in_db( db_team_id: Final = db_model.model_info.team_id if db_model.model_info else None is_new_team_assignment: Final = db_team_id != patch_team_id + # Team rows keep their internal UUID-based model_name; the public name lives in model_info + patch_data.model_name = f"model_name_{patch_team_id}_{uuid.uuid4()}" if is_new_team_assignment else None + row: Final = await write_row(update_db_model(db_model=db_model, updated_patch=patch_data)) + if is_new_team_assignment: await _setup_new_team_model_assignment( team_id=patch_team_id, public_model_name=public_model_name, - patch_data=patch_data, user_api_key_dict=user_api_key_dict, ) else: @@ -1132,12 +1217,11 @@ async def _update_team_model_in_db( team_id=patch_team_id, public_model_name=public_model_name, db_model=db_model, - patch_data=patch_data, user_api_key_dict=user_api_key_dict, prisma_client=prisma_client, ) - return update_db_model(db_model=db_model, updated_patch=patch_data) + return row def _get_public_model_name( @@ -1189,13 +1273,9 @@ def _get_public_model_name( async def _setup_new_team_model_assignment( team_id: str, public_model_name: str, - patch_data: updateDeployment, user_api_key_dict: UserAPIKeyAuth, ) -> None: - """Set up a new team model with unique name and team membership.""" - unique_model_name: Final = f"model_name_{team_id}_{uuid.uuid4()}" - patch_data.model_name = unique_model_name - + """Register a newly team-assigned model's public name on the team.""" await team_model_add( data=TeamModelAddRequest( team_id=team_id, @@ -1385,7 +1465,6 @@ async def _update_existing_team_model_assignment( team_id: str, public_model_name: str, db_model: Deployment, - patch_data: updateDeployment, user_api_key_dict: UserAPIKeyAuth, prisma_client: PrismaClient | None, ) -> None: @@ -1409,9 +1488,6 @@ async def _update_existing_team_model_assignment( old_public_name: Final = db_model.model_info.team_public_model_name if db_model.model_info else None if old_public_name and public_model_name != old_public_name: - # Clear user-supplied public name from patch before any early return so the - # caller does not overwrite the internal UUID-based model_name in the DB. - patch_data.model_name = None if prisma_client is None: verbose_proxy_logger.warning( "prisma_client not initialized; skipping public name update entirely to avoid orphaned entries" @@ -1459,10 +1535,6 @@ async def _update_existing_team_model_assignment( # else: old_public_name == public_model_name (no rename needed) # No team_model_add/delete calls required; public name is already registered - # Always clear patch_data.model_name to prevent caller from overwriting - # the internal UUID-based model_name in the DB with the user-supplied public name - patch_data.model_name = None - class ModelManagementAuthChecks: """ @@ -1878,18 +1950,19 @@ async def add_new_model( reload_outcome: ReconcileOutcome = ReconcileOutcome(still_desired=None, live_after=None) try: _original_litellm_model_name: Final = model_params.model_name - if model_params.model_info.team_id is None: - model_response = await _add_model_to_db( - model_params=priced_model_params, - user_api_key_dict=user_api_key_dict, - prisma_client=prisma_client, - ) - else: - model_response = await _add_team_model_to_db( - model_params=priced_model_params, - user_api_key_dict=user_api_key_dict, - prisma_client=prisma_client, - ) + add_model: Final = ( + _add_model_to_db if model_params.model_info.team_id is None else _add_team_model_to_db + ) + model_response = await add_model( + model_params=priced_model_params, + user_api_key_dict=user_api_key_dict, + prisma_client=prisma_client, + slot=_heuristic_v2_slot( + prisma_client, + effective_config=priced_model_params.litellm_params.complexity_router_config, + model_id=priced_model_params.model_info.id, + ), + ) reload_outcome = await proxy_config.add_deployment( prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj ) @@ -1903,6 +1976,8 @@ async def add_new_model( passed_model_info=priced_model_params.model_info, ) except Exception as e: + if isinstance(e, HTTPException): + raise verbose_proxy_logger.exception("Exception in add_new_model: %s", e) else: @@ -2070,10 +2145,17 @@ async def update_model( "updated_by": user_api_key_dict.user_id or LITELLM_PROXY_ADMIN_NAME, **({} if renamed_to is None else {"model_name": renamed_to}), } - model_response: Final = await _proxy_model_table(prisma_client).update( - where={"model_id": _model_id}, - data=_data, - ) + async with _heuristic_v2_slot( + prisma_client, + effective_config=_effective_complexity_router_config( + model_params.litellm_params, deployment.litellm_params + ), + model_id=_model_id, + ) as table: + model_response: Final = await table.update( + where={"model_id": _model_id}, + data=_data, + ) if renamed_to is not None: await sync_access_groups_for_renamed_model( prisma_client=prisma_client, diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 27132c90e05..96b425f2a24 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -120,6 +120,8 @@ from litellm.router_utils.add_retry_fallback_headers import ( from litellm.router_utils.auto_router_model_naming import ( STRATEGY_ROUTER_PARAM_FIELDS, carries_complexity_router_settings, + count_heuristic_v2_routers, + heuristic_v2_limit_violation, validate_complexity_router_config_placement, ) from litellm.types.utils import ( @@ -301,7 +303,7 @@ from litellm.proxy.auth.auth_utils import ( ) from litellm.proxy.auth.fallback_model_access import router_fallback_access_check from litellm.proxy.auth.handle_jwt import JWTHandler -from litellm.proxy.auth.litellm_license import LicenseCheck +from litellm.proxy.auth.litellm_license import HEURISTIC_V2_LICENSE_REMEDY, LicenseCheck from litellm.proxy.auth.model_checks import ( expand_wildcard_deployments_for_model_info, get_all_fallbacks, @@ -4316,6 +4318,19 @@ def validate_deployment_complexity_router_placement(model: Mapping[str, object]) raise ValueError(f"model {model.get('model_name', '')!r}: {violation}") +def validate_heuristic_v2_router_limit(model_list: Sequence[Mapping[str, object]], *, limit: int | None) -> None: + """ + Refuse to start when config.yaml defines more heuristic_v2 auto-routers than the license allows. + + Checked here rather than left to router registration for the same reason as the two + validators above: the proxy builds its router with `ignore_invalid_deployments=True`, so + the router's own refusal would turn the extra router into a silently missing model. + """ + violation: Final = heuristic_v2_limit_violation(held=count_heuristic_v2_routers(model_list), limit=limit) + if violation is not None: + raise ValueError(f"config.yaml model_list: {violation} {HEURISTIC_V2_LICENSE_REMEDY}") + + def pin_complexity_router_model_id(model: dict) -> None: # mutable-ok: out-param, model_info is stamped in place """ Stamps `model_info.id` from the raw litellm_params before plugin resolution swaps @@ -5721,6 +5736,7 @@ class ProxyConfig: model_list: Final = config.get("model_list", None) if model_list: router_params["model_list"] = model_list + validate_heuristic_v2_router_limit(model_list, limit=_license_check.heuristic_v2_router_limit()) print( # noqa: T201 "\033[32mLiteLLM: Proxy initialized with Config, Set models:\033[0m" ) @@ -5810,6 +5826,7 @@ class ProxyConfig: ), ignore_invalid_deployments=True, # don't raise an error if a deployment is invalid fallback_access_check=router_fallback_access_check, + heuristic_v2_router_limit=_license_check.heuristic_v2_router_limit, ) if redis_usage_cache is not None and router.cache.redis_cache is None: @@ -6270,6 +6287,7 @@ class ProxyConfig: search_tools=search_tools, ignore_invalid_deployments=True, fallback_access_check=router_fallback_access_check, + heuristic_v2_router_limit=_license_check.heuristic_v2_router_limit, ) verbose_proxy_logger.debug("updated llm_router: %s", llm_router) else: diff --git a/litellm/router.py b/litellm/router.py index cfb080a24a0..dc750941559 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -21,7 +21,7 @@ import time import traceback import weakref from collections import defaultdict -from collections.abc import AsyncGenerator, AsyncIterator, Callable, Generator, Mapping, Sequence +from collections.abc import AsyncGenerator, AsyncIterator, Callable, Generator, Iterator, Mapping, Sequence from functools import lru_cache, partial from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypeAlias, TypeVar, Union, cast @@ -117,6 +117,9 @@ from litellm.router_utils.add_retry_fallback_headers import ( from litellm.router_utils.auto_router_model_naming import ( AUTO_ROUTER_MODEL_PREFIX, classify_strategy_router_model, + count_heuristic_v2_routers, + heuristic_v2_limit_violation, + uses_heuristic_v2_classifier, ) from litellm.router_utils.batch_utils import ( _get_router_metadata_variable_name, @@ -211,6 +214,7 @@ from litellm.types.router import ( DeploymentTypedDict, FallbackAccessCheck, GuardrailTypedDict, + HeuristicV2RouterLimit, LiteLLM_Params, MockRouterTestingParams, ModelGroupInfo, @@ -683,6 +687,7 @@ class Router: background_health_check_model_groups: Sequence[str] | None = None, enable_weighted_failover: bool = False, fallback_access_check: FallbackAccessCheck | None = None, + heuristic_v2_router_limit: HeuristicV2RouterLimit | None = None, ) -> None: """ Initialize the Router class with the given parameters for caching, reliability, and routing strategy. @@ -759,6 +764,7 @@ class Router: self.set_verbose = set_verbose self.ignore_invalid_deployments = ignore_invalid_deployments + self.heuristic_v2_router_limit = heuristic_v2_router_limit self.fallback_access_check: Final = fallback_access_check self.debug_level = debug_level self.enable_pre_call_checks = enable_pre_call_checks @@ -8796,6 +8802,30 @@ class Router: """ return classify_strategy_router_model(litellm_params.model) == "complexity" + def config_deployments(self) -> Iterator[Mapping[str, object]]: + """The model_list rows that came from config.yaml rather than the DB (``model_info.db_model`` unset).""" + for deployment in self.model_list: + if not isinstance(deployment, Mapping): + continue + model_info = deployment.get("model_info") + if not (isinstance(model_info, Mapping) and model_info.get("db_model")): + yield deployment + + def heuristic_v2_router_limit_violation(self) -> str | None: + """ + Why one more heuristic_v2 router cannot join this router, or None when it can. + + Judged against every deployment currently on the model_list; an upsert pops the row being + edited first, so an edit of an existing heuristic_v2 router keeps its own slot. The limit is + resolved on every call through ``heuristic_v2_router_limit``; unset means unlimited, which + is the SDK default, and the proxy injects a resolver backed by its license. + """ + limit: Final = self.heuristic_v2_router_limit() if self.heuristic_v2_router_limit is not None else None + others: Final = count_heuristic_v2_routers( + deployment for deployment in self.model_list if isinstance(deployment, Mapping) + ) + return heuristic_v2_limit_violation(held=others + 1, limit=limit) + def init_complexity_router_deployment(self, deployment: Deployment): """ Initialize the complexity-router deployment. @@ -8813,6 +8843,10 @@ class Router: ) complexity_router_config: Final[dict | None] = deployment.litellm_params.complexity_router_config + if uses_heuristic_v2_classifier(complexity_router_config): + limit_violation: Final = self.heuristic_v2_router_limit_violation() + if limit_violation is not None: + raise ValueError(limit_violation) default_model: str | None = deployment.litellm_params.complexity_router_default_model @@ -9636,8 +9670,16 @@ class Router: raise e def _restore_deployment_after_failed_upsert(self, previous_deployment: Deployment | None, model_id: str) -> None: + """Put a deployment back the way it was before a failed upsert popped it. + + A rollback re-admits state that was already serving, so it does not go through the + heuristic_v2 ceiling a newcomer gets: with the ceiling tightened since the deployment first + registered, judging the rollback would drop a serving router over an unrelated failed edit. + """ if previous_deployment is None or self.has_model_id(model_id): return + limit_resolver: Final = self.heuristic_v2_router_limit + self.heuristic_v2_router_limit = None try: self.add_deployment(deployment=previous_deployment) verbose_router_logger.info( @@ -9652,6 +9694,8 @@ class Router: model_id, restore_error, ) + finally: + self.heuristic_v2_router_limit = limit_resolver @staticmethod def _backend_cost_map_keys(model: str, custom_llm_provider: str | None) -> tuple[str, ...]: diff --git a/litellm/router_utils/auto_router_model_naming.py b/litellm/router_utils/auto_router_model_naming.py index a8aa543d735..2efbfb5782e 100644 --- a/litellm/router_utils/auto_router_model_naming.py +++ b/litellm/router_utils/auto_router_model_naming.py @@ -10,7 +10,7 @@ the router silently dropping the deployment at load time under ``ignore_invalid_deployments``. """ -from collections.abc import Mapping, Sequence +from collections.abc import Iterable, Mapping, Sequence from dataclasses import dataclass from types import MappingProxyType from typing import Final, Literal, TypeAlias @@ -163,6 +163,38 @@ def strategy_router_dependencies( ) +def uses_heuristic_v2_classifier(complexity_router_config: object) -> bool: + """Whether this complexity config classifies with the bundled heuristic_v2 model.""" + return _mapping(complexity_router_config).get("classifier_type") == "heuristic_v2" + + +def is_heuristic_v2_router(litellm_params: Mapping[str, object]) -> bool: + """Whether this deployment is a complexity router that classifies with heuristic_v2.""" + return classify_strategy_router_model(str(litellm_params.get("model") or "")) == "complexity" and ( + uses_heuristic_v2_classifier(litellm_params.get("complexity_router_config")) + ) + + +def count_heuristic_v2_routers(deployments: Iterable[Mapping[str, object]]) -> int: + """How many of ``deployments`` (router model_list entries or config.yaml rows) are heuristic_v2 routers.""" + return sum(1 for deployment in deployments if is_heuristic_v2_router(_mapping(deployment.get("litellm_params")))) + + +def heuristic_v2_limit_violation(*, held: int, limit: int | None) -> str | None: + """Why holding ``held`` heuristic_v2 routers exceeds ``limit``, or None when it fits. + + ``limit`` None means unlimited. The message is shared by every enforcement point (config + load, model writes, router registration) and stays SDK-neutral: it names the cap and what + the caller can change; the proxy appends how its license lifts the cap. + """ + if limit is None or held <= limit: + return None + return ( + f"At most {limit} auto-router(s) with classifier_type 'heuristic_v2' can be registered but this would make " + f"{held}. Use classifier_type 'heuristic' for this router or remove an existing heuristic_v2 router." + ) + + def validate_complexity_router_config_write(complexity_router_config: Mapping[str, object] | None) -> str | None: """Reject a complexity config the router would refuse to build a deployment from. diff --git a/litellm/types/router.py b/litellm/types/router.py index 4f4df1a8d2e..7ebd50f1328 100644 --- a/litellm/types/router.py +++ b/litellm/types/router.py @@ -885,6 +885,18 @@ class FallbackAccessCheck(Protocol): async def __call__(self, *, model: str, request_kwargs: Mapping[str, object], llm_router: "Router") -> bool: ... +class HeuristicV2RouterLimit(Protocol): + """ + Resolves how many heuristic_v2 complexity routers the Router may hold right now; None means unlimited. + + The Router calls it on every registration and limit query instead of caching the answer, so the + proxy can keep the limit on its license object (re-verified on config load) rather than hand + over a snapshot. + """ + + def __call__(self) -> int | None: ... + + class LiteLLM_RouterFileObject(TypedDict, total=False): """ Tracking the litellm params hash, used for mapping the file id to the right model diff --git a/tests/test_litellm/proxy/auth/test_litellm_license.py b/tests/test_litellm/proxy/auth/test_litellm_license.py index 8da365cb587..1db53638070 100644 --- a/tests/test_litellm/proxy/auth/test_litellm_license.py +++ b/tests/test_litellm/proxy/auth/test_litellm_license.py @@ -2,6 +2,8 @@ import asyncio import json from unittest.mock import AsyncMock, MagicMock, patch +from cryptography.hazmat.primitives.asymmetric.rsa import RSAPublicKey + from litellm.proxy.auth.litellm_license import LicenseCheck @@ -30,3 +32,70 @@ def test_is_over_limit(): assert license_check.is_over_limit(101) is False assert license_check.is_over_limit(100) is False assert license_check.is_over_limit(99) is False + + +def test_heuristic_v2_router_limit() -> None: + """Only the signed license's auto_router feature lifts the one-router limit; an API-verified + license (no airgapped data) and an airgapped license without the feature keep it.""" + license_check = LicenseCheck() + license_check.airgapped_license_data = {"expiration_date": "2999-01-01", "allowed_features": ["auto_router"]} + assert license_check.heuristic_v2_router_limit() is None + + license_check.airgapped_license_data = { + "expiration_date": "2999-01-01", + "allowed_features": ["sso", "auto_router", "audit_logs"], + } + assert license_check.heuristic_v2_router_limit() is None + + license_check.airgapped_license_data = {"expiration_date": "2999-01-01", "allowed_features": ["sso"]} + assert license_check.heuristic_v2_router_limit() == 1 + + license_check.airgapped_license_data = {"expiration_date": "2999-01-01"} + assert license_check.heuristic_v2_router_limit() == 1 + + license_check.airgapped_license_data = None + assert license_check.heuristic_v2_router_limit() == 1 + + +def _signed_license(expiration_date: str) -> tuple[RSAPublicKey, str]: + import base64 + + from cryptography.hazmat.primitives import hashes + from cryptography.hazmat.primitives.asymmetric import padding, rsa + + private_key = rsa.generate_private_key(public_exponent=65537, key_size=2048) + message = json.dumps( + {"expiration_date": expiration_date, "user_id": "u", "allowed_features": ["auto_router"]} + ).encode() + signature = private_key.sign( + message, + padding.PSS(mgf=padding.MGF1(hashes.SHA256()), salt_length=padding.PSS.MAX_LENGTH), + hashes.SHA256(), + ) + return private_key.public_key(), base64.b64encode(message + b"." + signature).decode() + + +def test_expired_or_unreadable_license_grants_no_features() -> None: + """The verifier stores the signed payload only after the expiry check passes and clears it when a + later verify rejects the license, so a stale payload cannot keep lifting the heuristic_v2 limit.""" + license_check = LicenseCheck() + public_key, valid_key = _signed_license("2999-01-01") + assert license_check.verify_license_without_api_request(public_key=public_key, license_key=valid_key) is True + assert license_check.heuristic_v2_router_limit() is None + + _, expired_key = _signed_license("2000-01-01") + assert license_check.verify_license_without_api_request(public_key=public_key, license_key=expired_key) is not True + assert license_check.airgapped_license_data is None + assert license_check.heuristic_v2_router_limit() == 1 + + assert license_check.verify_license_without_api_request(public_key=public_key, license_key=valid_key) is True + assert license_check.verify_license_without_api_request(public_key=public_key, license_key="not-a-license") is not True + assert license_check.airgapped_license_data is None + + +def test_valid_signed_license_with_auto_router_lifts_the_limit() -> None: + license_check = LicenseCheck() + public_key, license_key = _signed_license("2999-01-01") + + assert license_check.verify_license_without_api_request(public_key=public_key, license_key=license_key) is True + assert license_check.heuristic_v2_router_limit() is None diff --git a/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py index 5fa59a85c9d..c69f8f20a13 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py @@ -28,9 +28,18 @@ from litellm.proxy.management_endpoints.model_management_endpoints import ( delete_team_models, ) from litellm.proxy.utils import PrismaClient +from litellm.router import Router from litellm.types.router import Deployment, LiteLLM_Params, updateDeployment +async def _passthrough_row(update_data): + return update_data + + +async def _write_empty_row(**kwargs): + return await kwargs["write_row"]({}) + + class MockPrismaClient: def __init__( self, @@ -1191,7 +1200,7 @@ class TestTeamModelSiblingRouting: team_id = "team_no_alias" public_name = "gpt-4.1-mini" - async def mock_add_model_to_db(model_params, user_api_key_dict, prisma_client): + async def mock_add_model_to_db(model_params, user_api_key_dict, prisma_client, slot=None): return MagicMock(model_id=str(uuid.uuid4())) mock_team_model_add = AsyncMock() @@ -1372,7 +1381,8 @@ class TestTeamModelUpdate: db_model=db_model, patch_data=patch_data, user_api_key_dict=user_api_key_dict, - prisma_client=prisma_client, # type: ignore + prisma_client=prisma_client, # type: ignore, + write_row=_passthrough_row, ) assert result.get("model_name", "").startswith("model_name_test_team_123_") @@ -1435,7 +1445,6 @@ class TestTeamModelUpdate: team_id="team_123", public_model_name="new-public-name", db_model=db_model, - patch_data=patch_data, user_api_key_dict=user_api_key_dict, prisma_client=prisma_client, # type: ignore ) @@ -1481,7 +1490,6 @@ class TestTeamModelUpdate: team_id="team_123", public_model_name="new-public-name", db_model=db_model, - patch_data=patch_data, user_api_key_dict=user_api_key_dict, prisma_client=None, ) @@ -1490,39 +1498,72 @@ class TestTeamModelUpdate: mock_delete.assert_not_called() @pytest.mark.asyncio - async def test_rename_with_prisma_none_clears_patch_model_name(self): - """Rename path must clear patch_data.model_name even when prisma is unavailable (P1).""" + async def test_a_refused_row_write_leaves_the_team_untouched(self): + """The team's model list autocommits, so it is written only after the row write succeeded: a + refused write (the heuristic_v2 slot 403, a DB error) must not leave the team listing a name + whose row never changed.""" + from fastapi import HTTPException + from litellm.proxy.management_endpoints.model_management_endpoints import ( - _update_existing_team_model_assignment, + _update_team_model_in_db, ) from litellm.types.router import ModelInfo db_model = Deployment( - model_name="model_name_team_123_uuid1", + model_name="gpt-4o", litellm_params=LiteLLM_Params(model="azure/gpt-4o-mini"), - model_info=ModelInfo( - team_id="team_123", team_public_model_name="old-public-name" + model_info=ModelInfo(), + ) + user_api_key_dict = UserAPIKeyAuth(user_id="test_user", user_role=LitellmUserRoles.PROXY_ADMIN) + events: list[str] = [] + written: dict[str, object] = {} + + def patch_data() -> updateDeployment: + return updateDeployment(model_name="team-public", model_info=ModelInfo(team_id="team_123")) + + async def refuse_row(update_data): + events.append("row") + raise HTTPException(status_code=403, detail="slot held") + + async def accept_row(update_data): + events.append("row") + written.update(update_data) + return update_data + + async def team_add(**_): + events.append("team_model_add") + + with ( + patch( # test-quality-ok: the team auth check needs a live DB; the write order is what is under test + "litellm.proxy.management_endpoints.model_management_endpoints.ModelManagementAuthChecks.allow_team_model_action", + AsyncMock(return_value=True), ), - ) - patch_data = updateDeployment( - model_name="new-public-name", - model_info=ModelInfo(team_id="team_123"), - ) - user_api_key_dict = UserAPIKeyAuth( - user_id="test_user", - user_role=LitellmUserRoles.PROXY_ADMIN, - ) + patch("litellm.proxy.proxy_server.premium_user", True), # test-quality-ok: team models are premium-gated through a proxy global with no injection seam + patch( # test-quality-ok: the team list write is the collaborator whose ordering is asserted + "litellm.proxy.management_endpoints.model_management_endpoints.team_model_add", + side_effect=team_add, + ), + ): + with pytest.raises(HTTPException): + await _update_team_model_in_db( + db_model=db_model, + patch_data=patch_data(), + user_api_key_dict=user_api_key_dict, + prisma_client=MockPrismaClient(team_exists=True), # type: ignore + write_row=refuse_row, + ) + assert events == ["row"] - await _update_existing_team_model_assignment( - team_id="team_123", - public_model_name="new-public-name", - db_model=db_model, - patch_data=patch_data, - user_api_key_dict=user_api_key_dict, - prisma_client=None, - ) - - assert patch_data.model_name is None + await _update_team_model_in_db( + db_model=db_model, + patch_data=patch_data(), + user_api_key_dict=user_api_key_dict, + prisma_client=MockPrismaClient(team_exists=True), # type: ignore + write_row=accept_row, + ) + assert events == ["row", "row", "team_model_add"] + assert str(written["model_name"]).startswith("model_name_team_123_") + assert "team-public" in str(written["model_info"]) @pytest.mark.asyncio async def test_rename_handles_legacy_string_model_info(self): @@ -1574,7 +1615,6 @@ class TestTeamModelUpdate: team_id="team_123", public_model_name="new-public-name", db_model=db_model, - patch_data=patch_data, user_api_key_dict=user_api_key_dict, prisma_client=prisma_client, # type: ignore ) @@ -1614,7 +1654,8 @@ class TestTeamModelUpdate: db_model=db_model, patch_data=patch_data, user_api_key_dict=user_api_key_dict, - prisma_client=prisma_client, # type: ignore + prisma_client=prisma_client, # type: ignore, + write_row=_passthrough_row, ) assert "403" in str(exc_info.value) @@ -1900,7 +1941,8 @@ class TestTeamModelUpdate: db_model=db_model, patch_data=patch_data, user_api_key_dict=user_api_key_dict, - prisma_client=prisma_client, # type: ignore + prisma_client=prisma_client, # type: ignore, + write_row=_passthrough_row, ) # team ACL must not be touched on a no-op edit @@ -4311,6 +4353,321 @@ class TestStrategyRouterWriteValidation: is None ) + @staticmethod + def _live_router_holding_one_heuristic_v2(limit: int | None) -> Router: + return Router( + model_list=[ + {"model_name": "gpt-4o-mini", "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "k"}}, + { + "model_name": "held-v2", + "litellm_params": { + "model": "auto_router/complexity_router", + "complexity_router_config": {"classifier_type": "heuristic_v2", "tiers": {"SIMPLE": "gpt-4o-mini"}}, + }, + "model_info": {"id": "held-id"}, + }, + ], + heuristic_v2_router_limit=lambda: limit, + ) + + class _FakeTx: + """Stands in for a prisma transaction: records the raw statements and exposes the model table.""" + + def __init__(self, db_held: int) -> None: + self.db_held = db_held + self.raw_calls: list[tuple[str, tuple[object, ...]]] = [] + self.litellm_proxymodeltable = MagicMock(create=AsyncMock(), update=AsyncMock()) + + async def query_raw(self, sql: str, *args: object) -> list[dict[str, object]]: + self.raw_calls.append((sql, args)) + return [{"held": self.db_held}] if "count(*)" in sql else [] + + async def __aenter__(self) -> "TestStrategyRouterWriteValidation._FakeTx": + return self + + async def __aexit__(self, *exc: object) -> None: + return None + + class _FakeDb: + """Stands in for prisma_client: the plain client and the transaction it opens are told apart by identity.""" + + def __init__(self, db_held: int, existing_row: object = None) -> None: + self.db = self + self.tx_obj = TestStrategyRouterWriteValidation._FakeTx(db_held) + self.litellm_proxymodeltable = MagicMock( + create=AsyncMock(), update=AsyncMock(), find_unique=AsyncMock(return_value=existing_row) + ) + + def tx(self) -> "TestStrategyRouterWriteValidation._FakeTx": + return self.tx_obj + + _V2 = {"classifier_type": "heuristic_v2", "tiers": {"SIMPLE": "gpt-4o-mini"}} + _V1 = {"classifier_type": "heuristic", "tiers": {"SIMPLE": "gpt-4o-mini"}} + + @pytest.mark.parametrize( + "incoming,existing,expected", + [ + (_V2, None, _V2), + (_V2, _V1, _V2), + (None, _V1, _V1), + (None, None, None), + ("no-config", _V2, _V2), + ], + ) + def test_effective_complexity_router_config( + self, incoming: object, existing: object, expected: object + ) -> None: + """A write is judged on the config it leaves on the row: the incoming one when it carries one, else the stored one.""" + from litellm.proxy.management_endpoints.model_management_endpoints import ( + _effective_complexity_router_config, + ) + from litellm.types.router import updateLiteLLMParams + + incoming_params = None if incoming is None else updateLiteLLMParams( + complexity_router_config=None if incoming == "no-config" else incoming + ) + existing_params = None if existing is None else updateLiteLLMParams(complexity_router_config=existing) + assert _effective_complexity_router_config(incoming_params, existing_params) == expected + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "limit,effective_config,db_held,config_holds_one,model_id,expected", + [ + (1, _V2, 1, False, None, "refused"), + (1, _V2, 0, True, None, "refused"), + (1, _V2, 0, False, None, "reserved"), + (1, _V2, 0, False, "held-id", "reserved"), + (2, _V2, 1, False, None, "reserved"), + (1, _V1, 5, True, None, "plain"), + (1, None, 5, True, None, "plain"), + (None, _V2, 5, True, None, "plain"), + ], + ) + async def test_heuristic_v2_slot_matrix( + self, + limit: int | None, + effective_config: object, + db_held: int, + config_holds_one: bool, + model_id: str | None, + expected: str, + ) -> None: + """The slot is claimed inside a locked transaction only for a heuristic_v2 write under a limit; the DB rows + (other pods included) plus config.yaml routers decide, the row being edited is excluded through the SQL + parameter, and every other write runs on the plain client with no lock.""" + from fastapi import HTTPException + + from litellm.proxy.management_endpoints.model_management_endpoints import ( + HEURISTIC_V2_SLOT_LOCK_KEY, + _heuristic_v2_slot, + ) + + fake = self._FakeDb(db_held) + live_router = self._live_router_holding_one_heuristic_v2(limit) if config_holds_one else None + with ( + patch("litellm.proxy.proxy_server._license_check.heuristic_v2_router_limit", lambda: limit), # test-quality-ok: the guard reads the proxy license singleton with no injection seam + patch("litellm.proxy.proxy_server.llm_router", live_router), # test-quality-ok: the guard reads the proxy router global with no injection seam + patch( # test-quality-ok: the cross-pod publish is the side effect under test; redis is not configured here + "litellm.proxy.management_endpoints.model_management_endpoints.publish_config_change", + new=AsyncMock(), + ) as published, + ): + if expected == "refused": + with pytest.raises(HTTPException) as exc_info: + async with _heuristic_v2_slot(fake, effective_config=effective_config, model_id=model_id): + pass + assert exc_info.value.status_code == 403 + assert "At most 1 auto-router" in str(exc_info.value.detail) + assert "'auto_router' feature lifts the limit" in str(exc_info.value.detail) + return + async with _heuristic_v2_slot(fake, effective_config=effective_config, model_id=model_id) as tables: + handle = tables + if expected == "plain": + await handle.create(data={}) + fake.litellm_proxymodeltable.create.assert_awaited_once_with(data={}) + assert fake.tx_obj.raw_calls == [] + return + assert handle is fake.tx_obj.litellm_proxymodeltable + published.assert_awaited_once_with(redis_cache=None, object_type="litellm_proxymodeltable") + (lock_sql, lock_params), (_count_sql, count_params) = fake.tx_obj.raw_calls + assert "pg_advisory_xact_lock($1)" in lock_sql and "count" not in lock_sql + assert lock_params == (HEURISTIC_V2_SLOT_LOCK_KEY,) + assert count_params == (model_id or "",) + + @pytest.mark.asyncio + async def test_team_model_bookkeeping_runs_after_the_slot_is_released(self) -> None: + """team_model_add needs a second pool connection, so it must run only after the slot transaction + (and its advisory lock) has closed; a pool-sized burst of team creates would otherwise stall on the + lock holder waiting for a connection the waiters are occupying.""" + from contextlib import asynccontextmanager + + from litellm.proxy.management_endpoints.model_management_endpoints import _add_team_model_to_db + from litellm.types.router import ModelInfo + + events: list[str] = [] + created = MagicMock(model_id="row-1") + + @asynccontextmanager + async def slot(): + events.append("slot-enter") + yield MagicMock(create=AsyncMock(return_value=created)) + events.append("slot-exit") + + async def team_model_add(**_: object) -> None: + events.append("team_model_add") + + deployment = Deployment( + model_name="public-v2", + litellm_params=LiteLLM_Params(model="auto_router/complexity_router", complexity_router_config=self._V2), + model_info=ModelInfo(id="row-1", team_id="team-1"), + ) + with ( + patch( # test-quality-ok: params are encrypted with the proxy master key, which this test does not configure + "litellm.proxy.management_endpoints.model_management_endpoints.encrypt_value_helper", + lambda value, new_encryption_key=None: value, + ), + patch( # test-quality-ok: the team list write is the collaborator whose ordering is asserted + "litellm.proxy.management_endpoints.model_management_endpoints.team_model_add", + side_effect=team_model_add, + ), + ): + result = await _add_team_model_to_db( + model_params=deployment, + user_api_key_dict=UserAPIKeyAuth(user_id="admin", user_role=LitellmUserRoles.PROXY_ADMIN), + prisma_client=MagicMock(), + slot=slot(), + ) + + assert result is created + assert events == ["slot-enter", "slot-exit", "team_model_add"] + + @pytest.mark.asyncio + async def test_add_new_model_refuses_a_second_heuristic_v2_router_before_the_db_write(self) -> None: + from litellm.proxy._types import ProxyException + from litellm.proxy.management_endpoints.model_management_endpoints import ( + add_new_model, + ) + + admin = UserAPIKeyAuth(user_id="admin", user_role=LitellmUserRoles.PROXY_ADMIN) + fake = self._FakeDb(db_held=1) + + with ( + patch("litellm.proxy.proxy_server.prisma_client", fake), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.llm_router", None), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.store_model_in_db", True), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.premium_user", True), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server._license_check.heuristic_v2_router_limit", lambda: 1), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch( # test-quality-ok: prior auth check needs a live DB; only the license limit is under test + "litellm.proxy.management_endpoints.model_management_endpoints.ModelManagementAuthChecks.can_user_make_model_call", + new=AsyncMock(return_value=None), + ), + patch( # test-quality-ok: params are encrypted before the slot is entered; no master key in this test + "litellm.proxy.management_endpoints.model_management_endpoints.encrypt_value_helper", + lambda value, new_encryption_key=None: value, + ), + ): + with pytest.raises(ProxyException) as exc_info: + await add_new_model( + model_params=Deployment( + model_name="second-v2", + litellm_params=LiteLLM_Params(model="auto_router/complexity_router", complexity_router_config=self._V2), + ), + user_api_key_dict=admin, + ) + assert exc_info.value.code == "403" + assert "At most 1 auto-router" in str(exc_info.value.message) + fake.tx_obj.litellm_proxymodeltable.create.assert_not_awaited() + fake.litellm_proxymodeltable.create.assert_not_awaited() + + @pytest.mark.asyncio + async def test_patch_model_refuses_switching_another_router_to_heuristic_v2(self) -> None: + """patch_model relays HTTPException as-is, so the license refusal reaches the client as a plain 403.""" + from fastapi import HTTPException + + from litellm.proxy.management_endpoints.model_management_endpoints import ( + patch_model, + ) + from litellm.types.router import updateLiteLLMParams + + model_id = "other-id" + admin = UserAPIKeyAuth(user_id="admin", user_role=LitellmUserRoles.PROXY_ADMIN) + fake = self._FakeDb(db_held=1) + + with ( + patch("litellm.proxy.proxy_server.prisma_client", fake), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.llm_router", None), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.store_model_in_db", True), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.premium_user", True), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server._license_check.heuristic_v2_router_limit", lambda: 1), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch( # test-quality-ok: the write must be refused before this DB step runs + "litellm.proxy.management_endpoints.model_management_endpoints.get_db_model", + new=AsyncMock(return_value=self._db_complexity_router(model_id)), + ), + patch( # test-quality-ok: prior auth check needs a live DB; only the license limit is under test + "litellm.proxy.management_endpoints.model_management_endpoints.ModelManagementAuthChecks.can_user_make_model_call", + new=AsyncMock(return_value=None), + ), + patch( # test-quality-ok: the helper's team bookkeeping needs a live DB; the row writer it is handed is what is under test + "litellm.proxy.management_endpoints.model_management_endpoints._update_team_model_in_db", + new=AsyncMock(side_effect=_write_empty_row), + ), + ): + with pytest.raises(HTTPException) as exc_info: + await patch_model( + model_id=model_id, + patch_data=updateDeployment(litellm_params=updateLiteLLMParams(complexity_router_config=self._V2)), + user_api_key_dict=admin, + ) + assert exc_info.value.status_code == 403 + fake.tx_obj.litellm_proxymodeltable.update.assert_not_awaited() + fake.litellm_proxymodeltable.update.assert_not_awaited() + + @pytest.mark.asyncio + async def test_update_model_refuses_switching_another_router_to_heuristic_v2(self) -> None: + from litellm.proxy._types import ProxyException + from litellm.proxy.management_endpoints.model_management_endpoints import ( + update_model, + ) + from litellm.types.router import ModelInfo, updateLiteLLMParams + + model_id = "other-id" + admin = UserAPIKeyAuth(user_id="admin", user_role=LitellmUserRoles.PROXY_ADMIN) + + existing_row = MagicMock() + existing_row.model_dump.return_value = { + "model_name": "my-auto-router", + "litellm_params": { + "model": "auto_router/complexity_router", + "complexity_router_config": {"tiers": {"SIMPLE": "gpt-4o-mini"}}, + }, + "model_info": {"id": model_id}, + } + existing_row.litellm_params = existing_row.model_dump.return_value["litellm_params"] + fake = self._FakeDb(db_held=1, existing_row=existing_row) + + with ( + patch("litellm.proxy.proxy_server.prisma_client", fake), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.llm_router", None), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.store_model_in_db", True), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server.premium_user", True), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch("litellm.proxy.proxy_server._license_check.heuristic_v2_router_limit", lambda: 1), # test-quality-ok: endpoint reads proxy server globals with no injection seam + patch( # test-quality-ok: prior auth check needs a live DB; only the license limit is under test + "litellm.proxy.management_endpoints.model_management_endpoints.ModelManagementAuthChecks.can_user_make_model_call", + new=AsyncMock(return_value=None), + ), + ): + with pytest.raises(ProxyException) as exc_info: + await update_model( + model_params=updateDeployment( + litellm_params=updateLiteLLMParams(complexity_router_config=self._V2), + model_info=ModelInfo(id=model_id), + ), + user_api_key_dict=admin, + ) + assert exc_info.value.code == "403" + fake.tx_obj.litellm_proxymodeltable.update.assert_not_awaited() + fake.litellm_proxymodeltable.update.assert_not_awaited() + @pytest.mark.asyncio async def test_update_model_rejects_prefix_strip(self): from litellm.proxy._types import ProxyException diff --git a/tests/test_litellm/proxy/management_endpoints/test_ptu_model_settings.py b/tests/test_litellm/proxy/management_endpoints/test_ptu_model_settings.py index a1c38d26b9d..d35b77f732c 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_ptu_model_settings.py +++ b/tests/test_litellm/proxy/management_endpoints/test_ptu_model_settings.py @@ -45,6 +45,10 @@ from litellm.types.router import ( from litellm.types.utils import Usage +async def _passthrough_row(update_data): + return update_data + + def test_model_info_accepts_valid_ptu_fields(): info = ModelInfo( id="x", @@ -385,6 +389,7 @@ class TestTeamModelUpdateValidatesBeforeWriting: patch_data=patch_data, user_api_key_dict=MagicMock(), prisma_client=MagicMock(), + write_row=_passthrough_row, ) return result, touched @@ -914,6 +919,7 @@ class TestPtuDeploymentsAreNotBilledPerToken: patch_data=patch, user_api_key_dict=UserAPIKeyAuth(user_id="a", user_role=LitellmUserRoles.PROXY_ADMIN), prisma_client=MagicMock(), + write_row=_passthrough_row, ) assert exc.value.status_code == 400 diff --git a/tests/test_litellm/proxy/proxy_server/test_proxy_config.py b/tests/test_litellm/proxy/proxy_server/test_proxy_config.py index 1ab18639fff..dcfad8f6815 100644 --- a/tests/test_litellm/proxy/proxy_server/test_proxy_config.py +++ b/tests/test_litellm/proxy/proxy_server/test_proxy_config.py @@ -28,6 +28,7 @@ from litellm.proxy.proxy_server import ( resolve_routing_plugins, validate_deployment_complexity_router_placement, validate_deployment_max_agentic_loops, + validate_heuristic_v2_router_limit, ) from .conftest import normalize @@ -193,6 +194,120 @@ def test_validate_deployment_complexity_router_placement_leaves_valid_deployment assert model["litellm_params"] == litellm_params +def _heuristic_v2_row(model_name: str, classifier_type: str = "heuristic_v2") -> dict[str, object]: + return { + "model_name": model_name, + "litellm_params": { + "model": "auto_router/complexity_router", + "complexity_router_config": {"classifier_type": classifier_type, "tiers": {"SIMPLE": "gpt-4o-mini"}}, + }, + } + + +def test_validate_heuristic_v2_router_limit_refuses_to_start_over_the_limit() -> None: + """Same reason as the two validators above: the proxy router swallows registration errors, so + an over-limit config.yaml must fail here instead of booting with a silently missing router.""" + with pytest.raises(ValueError, match=re.escape("At most 1 auto-router")) as exc_info: + validate_heuristic_v2_router_limit( + [_heuristic_v2_row("a"), _heuristic_v2_row("b"), _heuristic_v2_row("c", "heuristic")], limit=1 + ) + assert "'auto_router' feature lifts the limit" in str(exc_info.value) + + +@pytest.mark.parametrize( + "model_list,limit", + [ + ([_heuristic_v2_row("a"), _heuristic_v2_row("b")], None), + ([_heuristic_v2_row("a"), _heuristic_v2_row("c", "heuristic")], 1), + ([{"model_name": "gpt-4o", "litellm_params": {"model": "gpt-4o"}}], 1), + ], +) +def test_validate_heuristic_v2_router_limit_leaves_configs_within_the_limit_alone( + model_list: list[dict[str, object]], limit: int | None +) -> None: + assert validate_heuristic_v2_router_limit(model_list, limit=limit) is None + + +_TWO_HEURISTIC_V2_ROUTERS_YAML = ( + "model_list:\n" + " - model_name: gpt-4o-mini\n" + " litellm_params:\n" + " model: openai/gpt-4o-mini\n" + " api_key: k\n" + " - model_name: v2-a\n" + " litellm_params:\n" + " model: auto_router/complexity_router\n" + " complexity_router_config:\n" + " classifier_type: heuristic_v2\n" + " tiers: {SIMPLE: gpt-4o-mini}\n" + " - model_name: v2-b\n" + " litellm_params:\n" + " model: auto_router/complexity_router\n" + " complexity_router_config:\n" + " classifier_type: heuristic_v2\n" + " tiers: {SIMPLE: gpt-4o-mini}\n" + "router_settings:\n" + " heuristic_v2_router_limit: 99\n" +) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("license_limit", [1, None]) +async def test_ProxyConfig_load_config_takes_the_heuristic_v2_limit_from_the_license_only( + tmp_path, monkeypatch, license_limit: int | None +) -> None: + """`router_settings.heuristic_v2_router_limit` is managed outside config.yaml: an operator + cannot grant the entitlement by editing the config, and a licensed proxy boots both routers.""" + f = tmp_path / "c.yaml" + f.write_text(_TWO_HEURISTIC_V2_ROUTERS_YAML) + monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None) + monkeypatch.setattr("litellm.proxy.proxy_server.store_model_in_db", False) + monkeypatch.delenv("LITELLM_CONFIG_BUCKET_NAME", raising=False) + monkeypatch.setattr( + "litellm.proxy.proxy_server._license_check.heuristic_v2_router_limit", lambda: license_limit + ) + + if license_limit is None: + router, _model_list, _general_settings = await ProxyConfig().load_config( + router=None, config_file_path=str(f) + ) + assert router.heuristic_v2_router_limit is not None + assert router.heuristic_v2_router_limit() is None + assert sorted(router.complexity_routers) == ["v2-a", "v2-b"] + return + + with pytest.raises(ValueError, match=re.escape("config.yaml model_list: At most 1 auto-router")): + await ProxyConfig().load_config(router=None, config_file_path=str(f)) + + +@pytest.mark.asyncio +async def test_ProxyConfig_load_config_router_refuses_a_db_heuristic_v2_router_beyond_the_license( + tmp_path, monkeypatch +) -> None: + """config.yaml holds the one allowed heuristic_v2 router; a second one arriving later from the DB + is refused at registration because the router was built with the license's ceiling.""" + from litellm.types.router import Deployment + + f = tmp_path / "c.yaml" + f.write_text(_TWO_HEURISTIC_V2_ROUTERS_YAML.replace(" - model_name: v2-b\n", " - model_name: v1-b\n", 1).replace( + "classifier_type: heuristic_v2\n tiers: {SIMPLE: gpt-4o-mini}\nrouter_settings", + "classifier_type: heuristic\n tiers: {SIMPLE: gpt-4o-mini}\nrouter_settings", + )) + monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None) + monkeypatch.setattr("litellm.proxy.proxy_server.store_model_in_db", False) + monkeypatch.delenv("LITELLM_CONFIG_BUCKET_NAME", raising=False) + monkeypatch.setattr("litellm.proxy.proxy_server._license_check.heuristic_v2_router_limit", lambda: 1) + + router, _model_list, _general_settings = await ProxyConfig().load_config(router=None, config_file_path=str(f)) + + assert router.heuristic_v2_router_limit is not None + assert router.heuristic_v2_router_limit() == 1 + assert sorted(router.complexity_routers) == ["v1-b", "v2-a"] + db_row = Deployment(**_heuristic_v2_row("v2-from-db"), model_info={"id": "db-id"}) + assert router.upsert_deployment(db_row) is None + assert sorted(router.complexity_routers) == ["v1-b", "v2-a"] + + def test_validate_deployment_max_agentic_loops_allows_a_deployment_without_the_key(): model = {"model_name": "gpt-4o", "litellm_params": {"model": "gpt-4o"}} diff --git a/tests/test_litellm/router_strategy/test_complexity_router.py b/tests/test_litellm/router_strategy/test_complexity_router.py index aa1b51afe10..da3791da39a 100644 --- a/tests/test_litellm/router_strategy/test_complexity_router.py +++ b/tests/test_litellm/router_strategy/test_complexity_router.py @@ -14,6 +14,7 @@ from pydantic import ValidationError import litellm from litellm import Router +from litellm.router_utils.auto_router_model_naming import count_heuristic_v2_routers from litellm._logging import verbose_router_logger from litellm.caching.dual_cache import DualCache from litellm.constants import RETURN_RAW_MODEL_NAME_METADATA_KEY, SESSION_ID_GENERATED_METADATA_KEY @@ -1085,6 +1086,165 @@ class TestRouterComplexityDeploymentMethods: router.init_complexity_router_deployment(deployment) assert "auto_router/complexity_router/test-router" in router.complexity_routers + @staticmethod + def _router_row(model_name: str, model_id: str, classifier_type: str) -> dict[str, object]: + return { + "model_name": model_name, + "litellm_params": { + "model": "auto_router/complexity_router", + "complexity_router_config": { + "classifier_type": classifier_type, + "tiers": {"SIMPLE": "gpt-4o-mini", "MEDIUM": "gpt-4o"}, + }, + }, + "model_info": {"id": model_id}, + } + + _POOL: dict[str, object] = { + "model_name": "gpt-4o-mini", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "k"}, + } + + def test_heuristic_v2_ceiling_keeps_the_first_router_and_drops_the_rest(self) -> None: + """The proxy runs with ignore_invalid_deployments, so the second heuristic_v2 router is dropped + at registration while a heuristic (v1) sibling and the first v2 router stay routable.""" + router = Router( + model_list=[ + self._POOL, + self._router_row("v2-a", "id-a", "heuristic_v2"), + self._router_row("v2-b", "id-b", "heuristic_v2"), + self._router_row("v1-c", "id-c", "heuristic"), + ], + heuristic_v2_router_limit=lambda: 1, + ignore_invalid_deployments=True, + ) + + assert sorted(router.complexity_routers) == ["v1-c", "v2-a"] + assert router.get_deployment(model_id="id-b") is None + + def test_heuristic_v2_ceiling_raises_without_ignore_invalid_deployments(self) -> None: + with pytest.raises(ValueError, match="At most 1 auto-router"): + Router( + model_list=[ + self._POOL, + self._router_row("v2-a", "id-a", "heuristic_v2"), + self._router_row("v2-b", "id-b", "heuristic_v2"), + ], + heuristic_v2_router_limit=lambda: 1, + ) + + def test_heuristic_v2_limit_is_resolved_on_every_registration(self) -> None: + """The Router never caches the limit: when the resolver's answer moves (the proxy re-verified + its license), the next registration and the next limit query see the new value.""" + limits = {"value": None} + router = Router( + model_list=[ + self._POOL, + self._router_row("v2-a", "id-a", "heuristic_v2"), + self._router_row("v2-b", "id-b", "heuristic_v2"), + ], + heuristic_v2_router_limit=lambda: limits["value"], + ignore_invalid_deployments=True, + ) + assert sorted(router.complexity_routers) == ["v2-a", "v2-b"] + assert router.heuristic_v2_router_limit_violation() is None + + limits["value"] = 1 + assert router.heuristic_v2_router_limit_violation() is not None + assert router.upsert_deployment(Deployment(**self._router_row("v2-c", "id-c", "heuristic_v2"))) is None + assert sorted(router.complexity_routers) == ["v2-a", "v2-b"] + + def test_heuristic_v2_ceiling_tightening_refuses_the_edit_and_keeps_the_live_router(self) -> None: + """Two heuristic_v2 routers registered under an unlimited ceiling, then the ceiling drops to one: + an edit to either must be refused before its live row is popped, or the failed re-add and + the failed restore would drop a serving router while the write reports success.""" + limits = {"value": None} + router = Router( + model_list=[ + self._POOL, + self._router_row("v2-a", "id-a", "heuristic_v2"), + self._router_row("v2-b", "id-b", "heuristic_v2"), + ], + heuristic_v2_router_limit=lambda: limits["value"], + ignore_invalid_deployments=True, + ) + limits["value"] = 1 + + assert router.upsert_deployment(Deployment(**self._router_row("v2-a-renamed", "id-a", "heuristic_v2"))) is None + assert sorted(router.complexity_routers) == ["v2-a", "v2-b"] + assert router.get_deployment(model_id="id-a") is not None + + assert router.upsert_deployment(Deployment(**self._router_row("v1-a", "id-a", "heuristic"))) is not None + assert sorted(router.complexity_routers) == ["v1-a", "v2-b"] + + def test_config_deployments_excludes_db_rows(self) -> None: + """The proxy counts config.yaml routers from here and DB rows from the database, so a DB-loaded + row (``model_info.db_model``) must not show up twice.""" + router = Router(model_list=[self._POOL, self._router_row("v2-a", "id-a", "heuristic_v2")]) + db_row = self._router_row("v2-db", "id-db", "heuristic_v2") + db_row["model_info"] = {"id": "id-db", "db_model": True} + assert router.upsert_deployment(Deployment(**db_row)) is not None + + assert sorted(str(row["model_name"]) for row in router.config_deployments()) == ["gpt-4o-mini", "v2-a"] + assert count_heuristic_v2_routers(router.config_deployments()) == 1 + + def test_failed_edit_of_a_live_v2_router_rolls_back_without_the_ceiling(self) -> None: + """A rollback after a failed upsert re-admits state that was already serving, so it must not be + judged by a ceiling that tightened since: converting one of two live heuristic_v2 routers to a + config whose registration fails must leave it serving its previous v2 configuration.""" + limits = {"value": None} + router = Router( + model_list=[ + self._POOL, + self._router_row("v2-a", "id-a", "heuristic_v2"), + self._router_row("v2-b", "id-b", "heuristic_v2"), + ], + heuristic_v2_router_limit=lambda: limits["value"], + ignore_invalid_deployments=True, + ) + limits["value"] = 1 + + broken = self._router_row("v1-a", "id-a", "heuristic") + broken["litellm_params"]["complexity_router_config"]["tiers"] = {} + assert router.upsert_deployment(Deployment(**broken)) is None + + assert sorted(router.complexity_routers) == ["v2-a", "v2-b"] + live = router.get_deployment(model_id="id-a") + assert live is not None and live.litellm_params.complexity_router_config["classifier_type"] == "heuristic_v2" + assert router.heuristic_v2_router_limit_violation() is not None + + def test_heuristic_v2_routers_are_unlimited_by_default(self) -> None: + router = Router( + model_list=[ + self._POOL, + self._router_row("v2-a", "id-a", "heuristic_v2"), + self._router_row("v2-b", "id-b", "heuristic_v2"), + ] + ) + + assert sorted(router.complexity_routers) == ["v2-a", "v2-b"] + assert router.heuristic_v2_router_limit_violation() is None + + def test_heuristic_v2_router_limit_violation_frees_the_slot_of_the_router_being_edited(self) -> None: + """A DB reload upserts the existing heuristic_v2 router again; that edit must keep its own slot + while a different deployment switching to heuristic_v2 is refused.""" + router = Router( + model_list=[self._POOL, self._router_row("v2-a", "id-a", "heuristic_v2")], + heuristic_v2_router_limit=lambda: 1, + ignore_invalid_deployments=True, + ) + + assert router.heuristic_v2_router_limit_violation() is not None + + edited = self._router_row("v2-a-renamed", "id-a", "heuristic_v2") + assert router.upsert_deployment(Deployment(**edited)) is not None + assert sorted(router.complexity_routers) == ["v2-a-renamed"] + + assert router.upsert_deployment(Deployment(**self._router_row("v2-b", "id-b", "heuristic_v2"))) is None + assert sorted(router.complexity_routers) == ["v2-a-renamed"] + assert router.upsert_deployment(Deployment(**self._router_row("v1-c", "id-c", "heuristic"))) is not None + assert sorted(router.complexity_routers) == ["v1-c", "v2-a-renamed"] + def test_hybrid_initialization_waits_for_later_pool_deployments(self): router = Router( model_list=[ diff --git a/tests/test_litellm/router_utils/test_auto_router_model_naming.py b/tests/test_litellm/router_utils/test_auto_router_model_naming.py index 0007f09896a..238d0546518 100644 --- a/tests/test_litellm/router_utils/test_auto_router_model_naming.py +++ b/tests/test_litellm/router_utils/test_auto_router_model_naming.py @@ -1,8 +1,13 @@ +from collections.abc import Mapping + import pytest from litellm.router_utils.auto_router_model_naming import ( carries_complexity_router_settings, classify_strategy_router_model, + count_heuristic_v2_routers, + heuristic_v2_limit_violation, + is_heuristic_v2_router, strategy_router_dependencies, validate_complexity_router_config_placement, validate_complexity_router_config_write, @@ -369,3 +374,55 @@ def test_placement_is_scoped_to_complexity_router_deployments(model, present_fie flat param on an s3_vectors vector store, so an unscoped gate would reject a valid deployment. Either complexity field names one on its own, which is what the load itself requires.""" assert carries_complexity_router_settings(model, present_fields) is scoped + + +@pytest.mark.parametrize( + "litellm_params,expected", + [ + ({"model": "auto_router/complexity_router", "complexity_router_config": {"classifier_type": "heuristic_v2"}}, True), + ({"model": "auto_router/complexity_router-eu", "complexity_router_config": {"classifier_type": "heuristic_v2"}}, True), + ({"model": "auto_router/complexity_router", "complexity_router_config": {"classifier_type": "heuristic"}}, False), + ({"model": "auto_router/complexity_router", "complexity_router_config": {"tiers": {"SIMPLE": "a"}}}, False), + ({"model": "auto_router/complexity_router"}, False), + ({"model": "auto_router/quality_router", "complexity_router_config": {"classifier_type": "heuristic_v2"}}, False), + ({"model": "openai/gpt-4o", "complexity_router_config": {"classifier_type": "heuristic_v2"}}, False), + ({"model": "auto_router/complexity_router", "complexity_router_config": "heuristic_v2"}, False), + ({}, False), + ], +) +def test_is_heuristic_v2_router(litellm_params: Mapping[str, object], expected: bool) -> None: + """Only a complexity router whose config selects heuristic_v2 counts toward the license limit.""" + assert is_heuristic_v2_router(litellm_params) is expected + + +def test_count_heuristic_v2_routers_reads_model_list_rows_and_ignores_malformed_ones() -> None: + v2 = {"model": "auto_router/complexity_router", "complexity_router_config": {"classifier_type": "heuristic_v2"}} + rows: list[Mapping[str, object]] = [ + {"model_name": "a", "litellm_params": v2}, + {"model_name": "b", "litellm_params": {"model": "openai/gpt-4o"}}, + {"model_name": "c", "litellm_params": v2}, + {"model_name": "d"}, + {"model_name": "e", "litellm_params": "not a mapping"}, + ] + assert count_heuristic_v2_routers(rows) == 2 + assert count_heuristic_v2_routers(()) == 0 + + +@pytest.mark.parametrize( + "held,limit,violates", + [ + (1, 1, False), + (2, 1, True), + (0, 1, False), + (5, None, False), + (3, 3, False), + (4, 3, True), + ], +) +def test_heuristic_v2_limit_violation(held: int, limit: int | None, violates: bool) -> None: + violation = heuristic_v2_limit_violation(held=held, limit=limit) + assert (violation is not None) is violates + if violation is not None: + assert f"At most {limit} auto-router" in violation + assert f"would make {held}" in violation + assert "license" not in violation