diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 9ee62f94647..2e57d3e0708 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -9089,6 +9089,7 @@ class ProxyStartupEvent: prisma_client, pod_lock_manager=proxy_logging_obj.db_spend_update_writer.pod_lock_manager, alert=_alert_ptu_rollup_failure, + router=llm_router, ) scheduler.add_job( diff --git a/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py b/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py index f1f7248c064..a4eac992d89 100644 --- a/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py +++ b/litellm/proxy/spend_tracking/ptu_flat_cost_rollup.py @@ -14,7 +14,6 @@ and share the existing unique constraint. import asyncio import json -import sys from collections.abc import Awaitable, Callable, Mapping from dataclasses import dataclass from datetime import date, datetime, time, timedelta, timezone @@ -327,16 +326,6 @@ class _LoadedDeployments: config_sourced: bool -def _running_router() -> object | None: - """The proxy's router, or None outside a running proxy. - - Read out of ``sys.modules`` rather than imported, so a rollup driven from a test or a - script does not pull the whole proxy server in behind it. - """ - proxy_server: Final = sys.modules.get("litellm.proxy.proxy_server") - return getattr(proxy_server, "llm_router", None) if proxy_server is not None else None - - def _config_deployments(router: object | None, *, owned_by_db: frozenset[str]) -> tuple[_PTUDeployment, ...]: """Deployments the router holds that no ``LiteLLM_ProxyModelTable`` row owns. @@ -357,15 +346,17 @@ def _config_deployments(router: object | None, *, owned_by_db: frozenset[str]) - ) -async def _load_ptu_models(prisma_client: "PrismaClient") -> _LoadedDeployments: +async def _load_ptu_models(prisma_client: "PrismaClient", *, router: object | None) -> _LoadedDeployments: """Every deployment carrying valid manual PTU config, and every id the scan saw. Reserved capacity is billed by the provider whichever file declared it, so a deployment the proxy only knows from config.yaml accrues alongside the stored ones. + The router is handed in rather than read off the proxy module, so a run prices exactly + the deployments its caller declares and nothing a co-resident process left behind. """ rows: Final = await prisma_client.db.litellm_proxymodeltable.find_many() db_ids: Final = frozenset(model_id for row in rows if (model_id := str(getattr(row, "model_id", "") or ""))) - config_records: Final = _config_deployments(_running_router(), owned_by_db=db_ids) + config_records: Final = _config_deployments(router, owned_by_db=db_ids) models: Final = tuple( parsed for parsed in (_parse_ptu_model(row) for row in (*rows, *config_records)) if parsed is not None ) @@ -382,6 +373,7 @@ async def run_ptu_flat_cost_rollup( prisma_client: "PrismaClient", target_date: date | None = None, may_prune: bool = True, + router: object | None = None, ) -> RollupResult: """Rollup one UTC day of flat PTU cost across all PTU-configured model deployments. @@ -406,7 +398,7 @@ async def run_ptu_flat_cost_rollup( date_str: Final = day.isoformat() run_started: Final = datetime.now(timezone.utc) - loaded: Final = await _load_ptu_models(prisma_client) + loaded: Final = await _load_ptu_models(prisma_client, router=router) ptu_models: Final = loaded.models charges: Final = _aggregate_charges(ptu_models, day) @@ -527,6 +519,7 @@ async def _existing_sentinel_keys( async def run_ptu_flat_cost_backfill( prisma_client: "PrismaClient", today: date | None = None, + router: object | None = None, ) -> BackfillResult: """Price the elapsed days of every PTU window that carry no sentinel row yet. @@ -546,7 +539,7 @@ async def run_ptu_flat_cost_backfill( verbose_proxy_logger.warning("PTU backfill: prisma_client is None, skipping") return BackfillResult(start=end, end=end, days_scanned=0, rows_written=0) - ptu_models: Final = (await _load_ptu_models(prisma_client)).models + ptu_models: Final = (await _load_ptu_models(prisma_client, router=router)).models days: Final = _backfill_window(ptu_models, end) if not days: @@ -591,6 +584,7 @@ async def run_scheduled_ptu_rollup( pod_lock_manager: "PodLockManager | None" = None, target_date: date | None = None, alert: Callable[[str], Awaitable[None]] | None = None, + router: object | None = None, ) -> RollupResult | None: """Run the daily rollup under a cross-pod lock so only one proxy reconciles a day. @@ -615,7 +609,7 @@ async def run_scheduled_ptu_rollup( return None if pod_lock_manager is None or pod_lock_manager.redis_cache is None: - return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False) + return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False, router=router) if not await pod_lock_manager.acquire_lock(cronjob_id=PTU_ROLLUP_JOB_ID, ttl=PTU_ROLLUP_LOCK_TTL_SECONDS): if await _lock_is_held(pod_lock_manager): @@ -629,10 +623,10 @@ async def run_scheduled_ptu_rollup( "PTU rollup: could not take the rollup lock and no other pod holds it, " "running unguarded rather than skipping the day" ) - return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False) + return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=False, router=router) try: - return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=True) + return await _run_and_alert(prisma_client, target_date=target_date, alert=alert, may_prune=True, router=router) finally: await pod_lock_manager.release_lock(cronjob_id=PTU_ROLLUP_JOB_ID) @@ -657,6 +651,7 @@ async def _run_and_alert( target_date: date | None, alert: "Callable[[str], Awaitable[None]] | None", may_prune: bool = True, + router: object | None = None, ) -> RollupResult: """Reconcile the day, catch up any days left unpriced, and alert on charges that did not land. @@ -669,7 +664,9 @@ async def _run_and_alert( explicit date means reconcile exactly that day, so it stays a single-day operation. Its failure is contained: the day's own result is returned either way. """ - result: Final = await run_ptu_flat_cost_rollup(prisma_client, target_date=target_date, may_prune=may_prune) + result: Final = await run_ptu_flat_cost_rollup( + prisma_client, target_date=target_date, may_prune=may_prune, router=router + ) if result.rows_failed: await _deliver_alert( alert, @@ -686,7 +683,7 @@ async def _run_and_alert( "by the provider with nothing attributing it here. Extend the window, or retire the deployment.", ) if target_date is None: - await _backfill_and_alert(prisma_client, alert=alert) + await _backfill_and_alert(prisma_client, alert=alert, router=router) return result @@ -694,6 +691,7 @@ async def _backfill_and_alert( prisma_client: "PrismaClient", *, alert: "Callable[[str], Awaitable[None]] | None", + router: object | None = None, ) -> None: """Catch up unpriced PTU days, alerting on charges that did not land. @@ -701,7 +699,7 @@ async def _backfill_and_alert( caller whatever the catch-up pass does. """ try: - backfill: Final = await run_ptu_flat_cost_backfill(prisma_client) + backfill: Final = await run_ptu_flat_cost_backfill(prisma_client, router=router) except Exception as exc: # noqa: BLE001 # the catch-up pass must not fail the day's rollup verbose_proxy_logger.error("PTU backfill: catch-up pass failed, the day's rollup still stands: %s", exc) return diff --git a/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py b/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py index e039455607d..17a487ddd4b 100644 --- a/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py +++ b/tests/test_litellm/proxy/spend_tracking/test_ptu_flat_cost_rollup.py @@ -1767,16 +1767,20 @@ async def test_the_prune_cutoff_allows_for_clock_skew_between_hosts(): @pytest.mark.asyncio -async def test_a_run_pricing_config_cannot_prune_a_row_it_did_not_scan(monkeypatch): +async def test_a_run_pricing_config_cannot_prune_a_row_it_did_not_scan(): """Staleness alone stops being evidence once two hosts hold different configuration: a row this run never considered belongs to a deployment another host is pricing from its own file, and sweeping it drops that charge.""" table = _FakeSentinelTable() table.seed("t", DAY, "dep-elsewhere", 480.0, updated_at=datetime(2020, 1, 1, tzinfo=timezone.utc)) entry = _router_entry(model_id="cfg-here", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - await run_scheduled_ptu_rollup(_prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY) + await run_scheduled_ptu_rollup( + _prisma_for([], table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=_router_holding(entry), + ) assert ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "dep-elsewhere") in table.rows assert ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "cfg-here") in table.rows @@ -1784,7 +1788,7 @@ async def test_a_run_pricing_config_cannot_prune_a_row_it_did_not_scan(monkeypat @pytest.mark.asyncio -async def test_a_deployment_deleted_from_the_table_keeps_the_day_it_was_charged(monkeypatch): +async def test_a_deployment_deleted_from_the_table_keeps_the_day_it_was_charged(): """The accepted cost of bounding the prune, driven through the sequence that produces it: charge the day while the deployment exists, remove it, run the day again. Nothing scans it now, so nothing may judge its row, and the amount it was billed stands.""" @@ -1793,18 +1797,19 @@ async def test_a_deployment_deleted_from_the_table_keeps_the_day_it_was_charged( live_row = _model_row(model_id="dep-live", model_info=ptu) doomed_row = _model_row(model_id="dep-doomed", model_info=ptu) charged_key = ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "dep-doomed") - monkeypatch.setattr( - ptu_rollup, "_running_router", lambda: _router_holding(_router_entry(model_id="cfg", model_info=dict(ptu))) - ) + router = _router_holding(_router_entry(model_id="cfg", model_info=dict(ptu))) await run_scheduled_ptu_rollup( - _prisma_for([live_row, doomed_row], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY + _prisma_for([live_row, doomed_row], table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=router, ) billed = table.rows[charged_key]["ptu_flat_cost"] table.rows[charged_key]["updated_at"] = datetime(2020, 1, 1, tzinfo=timezone.utc) await run_scheduled_ptu_rollup( - _prisma_for([live_row], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY + _prisma_for([live_row], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY, router=router ) assert table.rows[charged_key]["ptu_flat_cost"] == billed @@ -1842,7 +1847,7 @@ async def test_every_deployment_that_prices_is_inside_the_set_that_bounds_the_pr table, ) - loaded = await ptu_rollup._load_ptu_models(prisma) + loaded = await ptu_rollup._load_ptu_models(prisma, router=None) assert {model.model_id for model in loaded.models} <= loaded.scanned_ids assert loaded.scanned_ids == {"dep-a", "dep-b", "dep-unpriced"} @@ -1858,7 +1863,7 @@ async def test_a_priced_deployment_is_in_the_bound_even_with_an_id_the_scan_skip _FakeSentinelTable(), ) - loaded = await ptu_rollup._load_ptu_models(prisma) + loaded = await ptu_rollup._load_ptu_models(prisma, router=None) assert {model.model_id for model in loaded.models} <= loaded.scanned_ids @@ -1872,13 +1877,13 @@ async def test_the_prune_splits_the_id_set_across_statements(monkeypatch): table = _FakeSentinelTable() ptu = {"ptu_count": 5, "cost_per_ptu_per_hour": 2.0, "team_id": "t"} deployments = [_model_row(model_id=f"dep-{n}", model_info=ptu) for n in range(4)] - monkeypatch.setattr( - ptu_rollup, "_running_router", lambda: _router_holding(_router_entry(model_id="dep-4", model_info=dict(ptu))) - ) table.seed("t", DAY, "dep-3", 480.0, updated_at=datetime(2020, 1, 1, tzinfo=timezone.utc)) await run_scheduled_ptu_rollup( - _prisma_for(deployments, table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY + _prisma_for(deployments, table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=_router_holding(_router_entry(model_id="dep-4", model_info=dict(ptu))), ) chunks = [call["model"]["in"] for call in table.delete_many_calls] @@ -1911,140 +1916,124 @@ def _router_holding(*entries): @pytest.mark.asyncio -async def test_a_config_declared_deployment_is_priced(monkeypatch): +async def test_a_config_declared_deployment_is_priced(): """The whole point. A PTU deployment the proxy only knows from config.yaml is not in LiteLLM_ProxyModelTable, so a DB-only scan bills the provider's reservation to nobody.""" entry = _router_entry(model_id="cfg-1", model_name="gpt-4o-ptu", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=_router_holding(entry)) assert [(m.model_id, m.model_name, m.team_id) for m in loaded.models] == [("cfg-1", "gpt-4o-ptu", "t")] assert "cfg-1" in loaded.scanned_ids @pytest.mark.asyncio -async def test_a_database_backed_router_entry_is_not_counted_twice(monkeypatch): +async def test_a_database_backed_router_entry_is_not_counted_twice(): """Every deployment loaded from the table is also in the router, flagged db_model. Pricing both copies would write two charges for one reservation.""" row = _model_row(model_id="db-1", model_info=dict(_VALID_PTU)) mirrored = _router_entry(model_id="db-1", model_info={**_VALID_PTU, "db_model": True}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(mirrored)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([row], _FakeSentinelTable())) - - assert [m.model_id for m in loaded.models] == ["db-1"] - - -@pytest.mark.asyncio -async def test_a_router_entry_sharing_an_id_with_the_table_is_priced_once(monkeypatch): - """db_model is data the router carries rather than something this module controls, so the - id anti-join is what actually maps onto the failure: two charges under one id.""" - row = _model_row(model_id="both-1", model_info=dict(_VALID_PTU)) - unflagged = _router_entry(model_id="both-1", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(unflagged)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([row], _FakeSentinelTable())) - - assert [m.model_id for m in loaded.models] == ["both-1"] - - -@pytest.mark.asyncio -async def test_a_client_credential_clone_is_not_priced(monkeypatch): - """Supplying an api_key on a request mints a clone of the deployment under a fresh id, - carrying the source's PTU config. Pricing it bills one reservation per distinct caller key.""" - source = _router_entry(model_id="cfg-1", model_info=dict(_VALID_PTU)) - clone = _router_entry(model_id="cfg-1-clone", model_info={**_VALID_PTU, "original_model_id": "cfg-1"}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(source, clone)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) - - assert [m.model_id for m in loaded.models] == ["cfg-1"] - - -@pytest.mark.asyncio -async def test_a_config_deployment_without_ptu_config_is_scanned_but_not_priced(monkeypatch): - """It has to stay in the scanned set or its leftover sentinel rows become unprunable.""" - entry = _router_entry(model_id="cfg-plain", model_info={"team_id": "t"}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) - - assert loaded.models == () - assert "cfg-plain" in loaded.scanned_ids - - -@pytest.mark.asyncio -async def test_no_router_in_the_process_prices_the_database_alone(monkeypatch): - """The rollup is importable and callable outside a running proxy.""" - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: None) loaded = await ptu_rollup._load_ptu_models( - _prisma_for([_model_row(model_id="db-1", model_info=dict(_VALID_PTU))], _FakeSentinelTable()) + _prisma_for([row], _FakeSentinelTable()), router=_router_holding(mirrored) ) assert [m.model_id for m in loaded.models] == ["db-1"] @pytest.mark.asyncio -async def test_a_config_deployment_is_charged_end_to_end(monkeypatch): +async def test_a_router_entry_sharing_an_id_with_the_table_is_priced_once(): + """db_model is data the router carries rather than something this module controls, so the + id anti-join is what actually maps onto the failure: two charges under one id.""" + row = _model_row(model_id="both-1", model_info=dict(_VALID_PTU)) + unflagged = _router_entry(model_id="both-1", model_info=dict(_VALID_PTU)) + + loaded = await ptu_rollup._load_ptu_models( + _prisma_for([row], _FakeSentinelTable()), router=_router_holding(unflagged) + ) + + assert [m.model_id for m in loaded.models] == ["both-1"] + + +@pytest.mark.asyncio +async def test_a_client_credential_clone_is_not_priced(): + """Supplying an api_key on a request mints a clone of the deployment under a fresh id, + carrying the source's PTU config. Pricing it bills one reservation per distinct caller key.""" + source = _router_entry(model_id="cfg-1", model_info=dict(_VALID_PTU)) + clone = _router_entry(model_id="cfg-1-clone", model_info={**_VALID_PTU, "original_model_id": "cfg-1"}) + + loaded = await ptu_rollup._load_ptu_models( + _prisma_for([], _FakeSentinelTable()), router=_router_holding(source, clone) + ) + + assert [m.model_id for m in loaded.models] == ["cfg-1"] + + +@pytest.mark.asyncio +async def test_a_config_deployment_without_ptu_config_is_scanned_but_not_priced(): + """It has to stay in the scanned set or its leftover sentinel rows become unprunable.""" + entry = _router_entry(model_id="cfg-plain", model_info={"team_id": "t"}) + + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=_router_holding(entry)) + + assert loaded.models == () + assert "cfg-plain" in loaded.scanned_ids + + +@pytest.mark.asyncio +async def test_no_router_in_the_process_prices_the_database_alone(): + """The rollup is importable and callable outside a running proxy.""" + loaded = await ptu_rollup._load_ptu_models( + _prisma_for([_model_row(model_id="db-1", model_info=dict(_VALID_PTU))], _FakeSentinelTable()), router=None + ) + + assert [m.model_id for m in loaded.models] == ["db-1"] + + +@pytest.mark.asyncio +async def test_a_config_deployment_is_charged_end_to_end(): """Through the scheduled entry point, so the charge lands in a sentinel row rather than stopping at the loader.""" table = _FakeSentinelTable() entry = _router_entry(model_id="cfg-1", model_name="gpt-4o-ptu", model_info=dict(_VALID_PTU)) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - await run_scheduled_ptu_rollup(_prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True), target_date=DAY) + await run_scheduled_ptu_rollup( + _prisma_for([], table), + pod_lock_manager=_pod_lock(acquired=True), + target_date=DAY, + router=_router_holding(entry), + ) assert ("t", DAY.isoformat(), PTU_SENTINEL_API_KEY, "cfg-1") in table.rows @pytest.mark.asyncio -async def test_a_stale_database_backed_router_entry_is_not_treated_as_config(monkeypatch): +async def test_a_stale_database_backed_router_entry_is_not_treated_as_config(): """The reconcile can leave a deployment on the router after its row is gone. The id anti-join cannot see that one, so the flag is what keeps it from being priced as though config.yaml had declared it.""" stale = _router_entry(model_id="db-gone", model_info={**_VALID_PTU, "db_model": True}) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(stale)) - loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable())) + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=_router_holding(stale)) assert loaded.models == () -def test_the_router_lookup_reads_the_proxys_own_global(): - """Every other config test replaces this helper, so without one test driving the real - body a typo in the module path or the attribute name leaves the whole feature dead in - production with the suite still green.""" - import sys - import types as _types +@pytest.mark.asyncio +async def test_a_router_left_on_the_proxy_module_is_not_scanned(monkeypatch): + """A run scans the router its caller hands it and nothing else. Reading the proxy module's + global instead made every run depend on whatever else in the process had set one, which + is what a caller passing no router is asking not to happen.""" + import litellm.proxy.proxy_server as proxy_server - assert ptu_rollup._running_router() is None or "litellm.proxy.proxy_server" in sys.modules + ambient = _router_holding(_router_entry(model_id="ambient-1", model_info=dict(_VALID_PTU))) + monkeypatch.setattr(proxy_server, "llm_router", ambient, raising=False) - sentinel = object() - stub = _types.SimpleNamespace(llm_router=sentinel) - real = sys.modules.get("litellm.proxy.proxy_server") - sys.modules["litellm.proxy.proxy_server"] = stub - try: - assert ptu_rollup._running_router() is sentinel - del stub.llm_router - assert ptu_rollup._running_router() is None - finally: - if real is None: - del sys.modules["litellm.proxy.proxy_server"] - else: - sys.modules["litellm.proxy.proxy_server"] = real + loaded = await ptu_rollup._load_ptu_models(_prisma_for([], _FakeSentinelTable()), router=None) - -def test_the_router_lookup_returns_none_outside_a_proxy(): - import sys - - real = sys.modules.pop("litellm.proxy.proxy_server", None) - try: - assert ptu_rollup._running_router() is None - finally: - if real is not None: - sys.modules["litellm.proxy.proxy_server"] = real + assert loaded.models == () + assert loaded.scanned_ids == frozenset() + assert loaded.config_sourced is False @pytest.mark.parametrize("chunk", [None, ("dep-a", "dep-b")], ids=["unbounded", "bounded"]) @@ -2063,7 +2052,7 @@ def test_the_prune_filter_is_a_plain_dict(chunk): @pytest.mark.asyncio -async def test_the_catch_up_pass_reaches_a_config_declared_deployment(monkeypatch): +async def test_the_catch_up_pass_reaches_a_config_declared_deployment(): """The catch-up shares the loader, so config deployments join it without being wired in. That is what prices the elapsed days of a reservation declared before today.""" table = _FakeSentinelTable() @@ -2073,9 +2062,10 @@ async def test_the_catch_up_pass_reaches_a_config_declared_deployment(monkeypatc model_id="cfg-back", model_info={"ptu_count": 100, "cost_per_ptu_per_hour": 0.02, "team_id": "t", "ptu_effective_from": started}, ) - monkeypatch.setattr(ptu_rollup, "_running_router", lambda: _router_holding(entry)) - await run_scheduled_ptu_rollup(_prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True)) + await run_scheduled_ptu_rollup( + _prisma_for([], table), pod_lock_manager=_pod_lock(acquired=True), router=_router_holding(entry) + ) charged = sorted(day for (_, day, _, model) in table.rows if model == "cfg-back") yesterday = (now.date() - timedelta(days=1)).isoformat() diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 83e9095c8ec..71fb184eb4b 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -11048,6 +11048,35 @@ async def test_ptu_rollup_job_registered_at_startup(monkeypatch): assert scheduler.get_job(PTU_ROLLUP_JOB_ID) is not None +@pytest.mark.asyncio +async def test_ptu_rollup_job_hands_the_rollup_the_proxys_router(monkeypatch): + """The rollup prices PTU deployments declared in config.yaml, which only the router + knows about. It takes the router as an argument, so nothing but this call site puts the + proxy's own router in front of it: without it that half of the feature is dead.""" + monkeypatch.delenv("STORE_MODEL_IN_DB", raising=False) + from litellm.proxy.spend_tracking import ptu_flat_cost_rollup + from litellm.proxy.spend_tracking.ptu_feature_flag import PTU_COST_ATTRIBUTION_ENV_VAR + from litellm.proxy.spend_tracking.ptu_flat_cost_rollup import PTU_ROLLUP_JOB_ID + + monkeypatch.setenv(PTU_COST_ATTRIBUTION_ENV_VAR, "true") + calls = [] + monkeypatch.setattr( + ptu_flat_cost_rollup, + "run_scheduled_ptu_rollup", + AsyncMock(side_effect=lambda *args, **kwargs: calls.append(kwargs)), + ) + + scheduler = await _run_scheduled_background_jobs() + + import litellm.proxy.proxy_server as ps + + router = MagicMock() + monkeypatch.setattr(ps, "llm_router", router) + await scheduler.get_job(PTU_ROLLUP_JOB_ID).func() + + assert [call["router"] for call in calls] == [router] + + @pytest.mark.asyncio async def test_ptu_rollup_job_not_registered_without_opt_in(monkeypatch): """Without LITELLM_ENABLE_PTU_COST_ATTRIBUTION the rollup never runs, so no sentinel row