From 369f3cb5f0b4284ab6b7d8faba2019137fb145d0 Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 29 Oct 2026 20:08:45 -0700 Subject: [PATCH 01/23] feat(cost-map): add baseten joey-1 decision model (#49123) Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/model_prices_and_context_window_backup.json | 9 +++++++++ model_prices_and_context_window.json | 9 +++++++++ 2 files changed, 18 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 1a1982fddd7..77ac4f49e8b 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -81059,5 +81059,14 @@ "supports_response_schema": true, "supports_tool_choice": false, "supports_vision": true + }, + "baseten/decision-models/baseten-joey-2b-latest": { + "input_cost_per_token": 1e-08, + "litellm_provider": "baseten", + "max_input_tokens": 16384, + "mode": "decisions", + "output_cost_per_token": 0.0, + "source": "https://inference.baseten.co/v1/models", + "supports_vision": true } } diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 1a1982fddd7..77ac4f49e8b 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -81059,5 +81059,14 @@ "supports_response_schema": true, "supports_tool_choice": false, "supports_vision": true + }, + "baseten/decision-models/baseten-joey-2b-latest": { + "input_cost_per_token": 1e-08, + "litellm_provider": "baseten", + "max_input_tokens": 16384, + "mode": "decisions", + "output_cost_per_token": 0.0, + "source": "https://inference.baseten.co/v1/models", + "supports_vision": true } } From 733cfeba935284056e77baf2dcc9517fdc96a2ba Mon Sep 17 00:00:00 2001 From: Kerry Lu Date: Thu, 29 Oct 2026 20:30:42 -0700 Subject: [PATCH 02/23] fix(cost): bill a provider reported usage.cost on every path unless the deployment is priced (#48953) * fix(cost): bill a provider reported usage.cost on every path unless the deployment is priced * test(integration): a reported cost is attributed to output in the gateway parity cases * fix(cost): skip a chat usage.cost that already holds litellm's stamped bill and Responses API objects * fix(cost): record litellm's bill on the Responses terminal event so a later pricing pass does not read it as the upstream's charge * fix(cost): mark usage.cost litellm wrote as its own bill and keep a breakdown for a reported cost * fix(cost): keep the token breakdown when a header reported cost repeats an already priced bill * fix(cost): read the stored breakdown total through the logging object's typed field * refactor(cost): name the providers billed from their own usage fields for what they are * fix(cost): keep the upstream's usage.cost beside litellm's bill and stamp a zero operator price on the stream frame --------- Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/cost_calculator.py | 21 +- litellm/litellm_core_utils/core_helpers.py | 37 ++- .../streaming_chunk_builder_utils.py | 23 +- litellm/responses/streaming_iterator.py | 3 +- .../cost_calculation/cases/openai.py | 225 ++++++++++++++++++ .../integration/cost_calculation/cost_map.py | 6 + .../litellm_core_utils/test_core_helpers.py | 78 +++--- .../test_streaming_handler.py | 6 +- .../llms/openai_like/test_kyma_provider.py | 8 +- tests/unit/test_main.py | 40 ++++ 10 files changed, 387 insertions(+), 60 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 6677b7327dc..ee102e82ca4 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -1544,6 +1544,12 @@ def _response_time_ms_for_cost( return litellm_logging_obj.get_response_ms() +def _breakdown_total(litellm_logging_obj: LitellmLoggingObject | None) -> float | None: + breakdown: Final = None if litellm_logging_obj is None else litellm_logging_obj.cost_breakdown + total: Final[object] = None if breakdown is None else breakdown.get("total_cost") + return float(total) if isinstance(total, (int, float)) and not isinstance(total, bool) else None + + def _adjusted_provider_billed_cost( reported_cost: float, custom_llm_provider: str | None, @@ -1561,9 +1567,9 @@ def _adjusted_provider_billed_cost( ) _store_cost_breakdown_in_logging_obj( litellm_logging_obj=litellm_logging_obj, - prompt_tokens_cost_usd_dollar=None, - completion_tokens_cost_usd_dollar=None, - cost_for_built_in_tools_cost_usd_dollar=None, + prompt_tokens_cost_usd_dollar=0.0, + completion_tokens_cost_usd_dollar=adjusted_cost, + cost_for_built_in_tools_cost_usd_dollar=0.0, total_cost_usd_dollar=adjusted_cost, original_cost=reported_cost, discount_percent=discount_percent, @@ -2376,6 +2382,15 @@ def response_cost_calculator( ) provider_response_cost: Final = get_response_cost_from_hidden_params(hidden_params) if provider_response_cost is not None and not custom_pricing: + if _breakdown_total(litellm_logging_obj) != provider_response_cost: + _store_cost_breakdown_in_logging_obj( + litellm_logging_obj=litellm_logging_obj, + prompt_tokens_cost_usd_dollar=0.0, + completion_tokens_cost_usd_dollar=provider_response_cost, + cost_for_built_in_tools_cost_usd_dollar=0.0, + total_cost_usd_dollar=provider_response_cost, + original_cost=provider_response_cost, + ) return provider_response_cost response_cost = completion_cost( diff --git a/litellm/litellm_core_utils/core_helpers.py b/litellm/litellm_core_utils/core_helpers.py index d940baf8b56..28f45d2bb2c 100644 --- a/litellm/litellm_core_utils/core_helpers.py +++ b/litellm/litellm_core_utils/core_helpers.py @@ -1017,13 +1017,44 @@ def chunk_usage(chunk: object) -> object: return getattr(chunk, "usage", None) -USAGE_COST_BILLED_PROVIDERS: Final = frozenset({"kyma"}) +PROVIDERS_BILLED_FROM_THEIR_OWN_USAGE_FIELDS: Final = frozenset({"xai", "perplexity", "opencode_go"}) + + +UPSTREAM_USAGE_COST: Final = "upstream_usage_cost" + + +def _numeric(value: object) -> float | None: + return float(value) if isinstance(value, (int, float)) and not isinstance(value, bool) else None + + +def stamp_litellm_bill_on_usage(response: object, usage: object, bill: float) -> None: + """Writes litellm's bill into ``usage.cost`` and the hidden ``response_cost`` cache, keeping the upstream's own + ``usage.cost`` under ``upstream_usage_cost`` so a later pricing pass still bills the upstream's charge instead of + reading litellm's bill back as if the upstream had reported it.""" + hidden: Final[object] = getattr(response, "_hidden_params", None) + if isinstance(hidden, dict): + hidden_params: Final = cast(dict[str, object], hidden) # cast-ok: hidden params ride on the response as a dict + upstream: Final = ( + hidden_params[UPSTREAM_USAGE_COST] + if UPSTREAM_USAGE_COST in hidden_params + else _numeric(getattr(usage, "cost", None)) + ) + hidden_params.update({"response_cost": bill, UPSTREAM_USAGE_COST: upstream}) + setattr(usage, "cost", bill) def billed_usage_cost(response: object, custom_llm_provider: str | None) -> float | None: - if custom_llm_provider not in USAGE_COST_BILLED_PROVIDERS: + """The charge an OpenAI-compatible upstream reports in ``usage.cost``, skipped for providers billed from their own + usage fields. Once litellm wrote its own bill into ``usage.cost`` (the stream builder and the Responses terminal + event do), the upstream's figure is read from ``upstream_usage_cost`` instead.""" + if custom_llm_provider in PROVIDERS_BILLED_FROM_THEIR_OWN_USAGE_FIELDS: return None - cost: Final = provider_reported_usage_cost(_usage_field(response, "usage"), custom_llm_provider) + hidden: Final = as_str_mapping(getattr(response, "_hidden_params", None)) + cost: Final = ( + _numeric(hidden[UPSTREAM_USAGE_COST]) + if hidden is not None and UPSTREAM_USAGE_COST in hidden + else provider_reported_usage_cost(_usage_field(response, "usage"), custom_llm_provider) + ) return cost if cost is not None and math.isfinite(cost) and cost >= 0 else None diff --git a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py index c69d9277d65..94351e72ab1 100644 --- a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py +++ b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py @@ -13,9 +13,10 @@ from typing_extensions import ReadOnly, Required import litellm from litellm._logging import verbose_logger from litellm.litellm_core_utils.core_helpers import ( - USAGE_COST_BILLED_PROVIDERS, + PROVIDERS_BILLED_FROM_THEIR_OWN_USAGE_FIELDS, chunk_usage, provider_reported_usage_cost, + stamp_litellm_bill_on_usage, ) from litellm.litellm_core_utils.hidden_params import HIDDEN_PARAMS_ATTR, set_hidden_params from litellm.litellm_core_utils.llm_cost_calc.guardrail_cost import BEDROCK_TRACE_FIELDS, object_items, object_mapping @@ -1380,9 +1381,6 @@ def stream_chunk_builder_text_completion( return TextCompletionResponse(**response) -_CALCULATOR_PRICED_REPORTED_COST_PROVIDERS: Final = frozenset({LlmProviders.XAI.value}) | USAGE_COST_BILLED_PROVIDERS - - def _logged_provider(logging_obj: Optional["Logging"]) -> object: if logging_obj is None: return None @@ -1390,10 +1388,6 @@ def _logged_provider(logging_obj: Optional["Logging"]) -> object: return provider -def _reported_cost_is_priced_by_calculator(logging_obj: Optional["Logging"]) -> bool: - return _logged_provider(logging_obj) in _CALCULATOR_PRICED_REPORTED_COST_PROVIDERS - - def _stream_builder_response_cost( response: ModelResponse, logging_obj: Optional["Logging"], chunks: Sequence[object] ) -> float | None: @@ -1404,9 +1398,6 @@ def _stream_builder_response_cost( for chunk in reversed(chunks) ) return next((cost for cost in reported if cost is not None), None) - usage_cost: Final = getattr(getattr(response, "usage", None), "cost", None) - if isinstance(usage_cost, (int, float)) and not _reported_cost_is_priced_by_calculator(logging_obj): - return float(usage_cost) if logging_obj is not None: return None provider_hint: Final = response._hidden_params.get( # pyright: ignore[reportPrivateUsage] # no public accessor @@ -1453,13 +1444,15 @@ def _set_stream_builder_response_cost( def _stamp_streaming_usage_cost(usage: Usage, response: ModelResponse, logging_obj: Optional["Logging"]) -> None: if logging_obj is None: return - if isinstance(getattr(usage, "cost", None), (int, float)): + if _logged_provider(logging_obj) in PROVIDERS_BILLED_FROM_THEIR_OWN_USAGE_FIELDS and isinstance( + getattr(usage, "cost", None), (int, float) + ): return computed_cost: Final = logging_obj.response_cost_calculator(result=response) - if not isinstance(computed_cost, (int, float)) or computed_cost <= 0: + if isinstance(computed_cost, bool) or not isinstance(computed_cost, (int, float)) or computed_cost < 0: return - setattr(usage, "cost", computed_cost) - _set_hidden_response_cost(response, float(computed_cost)) + if computed_cost > 0 or isinstance(getattr(usage, "cost", None), (int, float)): + stamp_litellm_bill_on_usage(response, usage, float(computed_cost)) def _field_of(value: object, name: str) -> object: diff --git a/litellm/responses/streaming_iterator.py b/litellm/responses/streaming_iterator.py index 6a82304b76a..f0bb2201162 100644 --- a/litellm/responses/streaming_iterator.py +++ b/litellm/responses/streaming_iterator.py @@ -43,6 +43,7 @@ from litellm.litellm_core_utils.core_helpers import ( process_response_headers, provider_reported_usage_cost, set_response_cost_in_hidden_params, + stamp_litellm_bill_on_usage, ) from litellm.litellm_core_utils.exception_mapping_utils import with_deployment_api_base from litellm.litellm_core_utils.internal_call_metadata import ( @@ -1869,7 +1870,7 @@ def stamp_responses_usage_cost( if isinstance(bill, bool) or not isinstance(bill, (int, float)): return if bill > 0 or isinstance(getattr(usage_obj, "cost", None), (int, float)): - setattr(usage_obj, "cost", float(bill)) + stamp_litellm_bill_on_usage(response_obj, usage_obj, float(bill)) def opening_item_payload(payload: Mapping[str, object]) -> Mapping[str, object]: diff --git a/tests/integration/cost_calculation/cases/openai.py b/tests/integration/cost_calculation/cases/openai.py index 53fd5c1cec2..1443cc5ab59 100644 --- a/tests/integration/cost_calculation/cases/openai.py +++ b/tests/integration/cost_calculation/cases/openai.py @@ -2,6 +2,7 @@ PARITY pairs the streamed case with its plain twin where both bill the same row.""" +from collections.abc import Mapping from typing import Final, cast from integration.cost_calculation.cost_tracking_case import ( @@ -11878,7 +11879,215 @@ GPT_5_6_TERRA_MESSAGES_BRIDGED_STREAM: Final = GPT_5_6_TERRA_MESSAGES_BRIDGED_JS } ) + +# An OpenAI-compatible gateway that reports what it charged under usage.cost (#47533). The reported total is the bill, +# split across input and output by the ratio of the cost map rates (100 * 1e-6 : 20 * 3e-6), on plain and streamed +# replies alike. A deployment the operator priced keeps the operator's rates on both. +_GATEWAY_REPORTED_USAGE: Final = {"prompt_tokens": 100, "completion_tokens": 20, "total_tokens": 120, "cost": 0.037} +_GATEWAY_REPORTED_EXPECTED: Final = ExactExpected( + spend=0.037, input_cost=0.0, output_cost=0.037, prompt_tokens=100, completion_tokens=20 +) +_GATEWAY_OPERATOR_PRICED: Final = Deployment(input_cost_per_token=2e-06, output_cost_per_token=6e-06) +_GATEWAY_OPERATOR_PRICED_EXPECTED: Final = ExactExpected( + spend=0.00032, input_cost=0.0002, output_cost=0.00012, prompt_tokens=100, completion_tokens=20 +) +_GATEWAY_CHUNK: Final = { + "id": "chatcmpl-$REQUEST_ID", + "object": "chat.completion.chunk", + "created": 1, + "model": "gateway-reported-cost-fixture", +} + +GATEWAY_REPORTED_COST: Final = CostTrackingTestCase( + name="gateway-reported-cost", + covers="quota_management.spend_tracking.cost_matrix.logs_cost", + model="gateway-reported-cost-fixture", + endpoint="/v1/chat/completions", + request={"model": "$MODEL", "messages": [{"role": "user", "content": "reported cost"}], "stream": False}, + response=JsonResponse( + content_type="application/json", + body={ + "id": "chatcmpl-$REQUEST_ID", + "object": "chat.completion", + "created": 1, + "model": "gateway-reported-cost-fixture", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": _GATEWAY_REPORTED_USAGE, + }, + ), + expected=_GATEWAY_REPORTED_EXPECTED, +) + +GATEWAY_REPORTED_COST_STREAM: Final = CostTrackingTestCase( + name="gateway-reported-cost-stream", + covers="quota_management.spend_tracking.cost_matrix.logs_cost", + model="gateway-reported-cost-fixture", + endpoint="/v1/chat/completions", + request={ + "model": "$MODEL", + "messages": [{"role": "user", "content": "reported cost"}], + "stream": True, + "stream_options": {"include_usage": True}, + }, + response=SseResponse( + content_type="text/event-stream", + frames=sse_frames( + {**_GATEWAY_CHUNK, "choices": [{"index": 0, "delta": {"role": "assistant", "content": "ok"}}]}, + {**_GATEWAY_CHUNK, "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]}, + {**_GATEWAY_CHUNK, "choices": [], "usage": _GATEWAY_REPORTED_USAGE}, + done=True, + ), + ), + expected=_GATEWAY_REPORTED_EXPECTED, +) + +GATEWAY_REPORTED_COST_OPERATOR_PRICED: Final = GATEWAY_REPORTED_COST.model_copy( + update={ + "name": "gateway-reported-cost-operator_priced", + "deployment": _GATEWAY_OPERATOR_PRICED, + "expected": _GATEWAY_OPERATOR_PRICED_EXPECTED, + } +) + +GATEWAY_REPORTED_COST_OPERATOR_PRICED_STREAM: Final = GATEWAY_REPORTED_COST_STREAM.model_copy( + update={ + "name": "gateway-reported-cost-operator_priced-stream", + "deployment": _GATEWAY_OPERATOR_PRICED, + "expected": _GATEWAY_OPERATOR_PRICED_EXPECTED, + } +) + +_GATEWAY_RESPONSES_USAGE: Final[Mapping[str, JsonValue]] = {"input_tokens": 100, "output_tokens": 20, "total_tokens": 120, "cost": 0.037} +_GATEWAY_RESPONSES_ITEM: Final[Mapping[str, JsonValue]] = { + "id": "msg_gateway", + "type": "message", + "role": "assistant", + "status": "completed", + "content": [{"type": "output_text", "text": "ok", "annotations": []}], +} +_GATEWAY_RESPONSES_BODY: Final[Mapping[str, JsonValue]] = { + "id": "resp_$REQUEST_ID", + "object": "response", + "created_at": 1, + "status": "completed", + "model": "gateway-reported-cost-fixture", + "output": [_GATEWAY_RESPONSES_ITEM], + "usage": _GATEWAY_RESPONSES_USAGE, +} + +GATEWAY_REPORTED_COST_RESPONSES: Final = CostTrackingTestCase( + name="gateway-reported-cost-responses", + covers="quota_management.spend_tracking.cost_matrix.logs_cost", + model="gateway-reported-cost-fixture", + endpoint="/v1/responses", + request={"model": "$MODEL", "input": "reported cost", "cache": {"no-cache": True}}, + response=JsonResponse(content_type="application/json", body=_GATEWAY_RESPONSES_BODY), + expected=_GATEWAY_REPORTED_EXPECTED, +) + +GATEWAY_REPORTED_COST_RESPONSES_STREAM: Final = GATEWAY_REPORTED_COST_RESPONSES.model_copy( + update={ + "name": "gateway-reported-cost-responses-stream", + "request": {**GATEWAY_REPORTED_COST_RESPONSES.request, "stream": True}, + "response": SseResponse( + content_type="text/event-stream", + frames=sse_frames( + { + "type": "response.created", + "sequence_number": 0, + "response": {**_GATEWAY_RESPONSES_BODY, "status": "in_progress", "output": [], "usage": None}, + }, + { + "type": "response.output_text.delta", + "sequence_number": 1, + "item_id": "msg_gateway", + "output_index": 0, + "content_index": 0, + "delta": "ok", + }, + {"type": "response.completed", "sequence_number": 2, "response": _GATEWAY_RESPONSES_BODY}, + ), + ), + } +) + +GATEWAY_REPORTED_COST_RESPONSES_OPERATOR_PRICED: Final = GATEWAY_REPORTED_COST_RESPONSES.model_copy( + update={ + "name": "gateway-reported-cost-responses-operator_priced", + "deployment": _GATEWAY_OPERATOR_PRICED, + "expected": _GATEWAY_OPERATOR_PRICED_EXPECTED, + } +) + +GATEWAY_REPORTED_COST_RESPONSES_OPERATOR_PRICED_STREAM: Final = GATEWAY_REPORTED_COST_RESPONSES_STREAM.model_copy( + update={ + "name": "gateway-reported-cost-responses-operator_priced-stream", + "deployment": _GATEWAY_OPERATOR_PRICED, + "expected": _GATEWAY_OPERATOR_PRICED_EXPECTED, + } +) + + +_GATEWAY_ZERO_USAGE: Final = {**_GATEWAY_REPORTED_USAGE, "cost": 0} + +GATEWAY_REPORTED_ZERO_OPERATOR_PRICED: Final = GATEWAY_REPORTED_COST_OPERATOR_PRICED.model_copy( + update={ + "name": "gateway-reported-zero-operator_priced", + "response": JsonResponse( + content_type="application/json", + body={**GATEWAY_REPORTED_COST.response.body, "usage": _GATEWAY_ZERO_USAGE}, + ), + } +) + +GATEWAY_REPORTED_ZERO_OPERATOR_PRICED_STREAM: Final = GATEWAY_REPORTED_COST_OPERATOR_PRICED_STREAM.model_copy( + update={ + "name": "gateway-reported-zero-operator_priced-stream", + "response": SseResponse( + content_type="text/event-stream", + frames=sse_frames( + {**_GATEWAY_CHUNK, "choices": [{"index": 0, "delta": {"role": "assistant", "content": "ok"}}]}, + {**_GATEWAY_CHUNK, "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]}, + {**_GATEWAY_CHUNK, "choices": [], "usage": _GATEWAY_ZERO_USAGE}, + done=True, + ), + ), + } +) + +_GATEWAY_OPENROUTER_OPERATOR_PRICED: Final = _GATEWAY_OPERATOR_PRICED.model_copy( + update={"model": "openrouter/vendor/gateway-reported-cost-fixture"} +) + +GATEWAY_REPORTED_COST_OPENROUTER_OPERATOR_PRICED: Final = GATEWAY_REPORTED_COST_OPERATOR_PRICED.model_copy( + update={ + "name": "gateway-reported-cost-openrouter-operator_priced", + "deployment": _GATEWAY_OPENROUTER_OPERATOR_PRICED, + } +) + +GATEWAY_REPORTED_COST_OPENROUTER_OPERATOR_PRICED_STREAM: Final = ( + GATEWAY_REPORTED_COST_OPERATOR_PRICED_STREAM.model_copy( + update={ + "name": "gateway-reported-cost-openrouter-operator_priced-stream", + "deployment": _GATEWAY_OPENROUTER_OPERATOR_PRICED, + } + ) +) + CASES: Final[tuple[CostTrackingTestCase, ...]] = ( + GATEWAY_REPORTED_COST, + GATEWAY_REPORTED_COST_STREAM, + GATEWAY_REPORTED_COST_OPERATOR_PRICED, + GATEWAY_REPORTED_COST_OPERATOR_PRICED_STREAM, + GATEWAY_REPORTED_COST_RESPONSES, + GATEWAY_REPORTED_COST_RESPONSES_STREAM, + GATEWAY_REPORTED_COST_RESPONSES_OPERATOR_PRICED, + GATEWAY_REPORTED_COST_RESPONSES_OPERATOR_PRICED_STREAM, + GATEWAY_REPORTED_ZERO_OPERATOR_PRICED, + GATEWAY_REPORTED_ZERO_OPERATOR_PRICED_STREAM, + GATEWAY_REPORTED_COST_OPENROUTER_OPERATOR_PRICED, + GATEWAY_REPORTED_COST_OPENROUTER_OPERATOR_PRICED_STREAM, GPT_5_6_SLASH_ALIAS, GPT_5_6_SLASH_ALIAS_STREAM, GPT_6_SOL_RESPONSES_DEPLOYMENT_JSON, @@ -12120,6 +12329,22 @@ CASES: Final[tuple[CostTrackingTestCase, ...]] = ( ) PARITY: Final[tuple[StreamParityTestCase, ...]] = ( + StreamParityTestCase(plain=GATEWAY_REPORTED_COST, streamed=GATEWAY_REPORTED_COST_STREAM), + StreamParityTestCase( + plain=GATEWAY_REPORTED_COST_OPERATOR_PRICED, streamed=GATEWAY_REPORTED_COST_OPERATOR_PRICED_STREAM + ), + StreamParityTestCase( + plain=GATEWAY_REPORTED_ZERO_OPERATOR_PRICED, streamed=GATEWAY_REPORTED_ZERO_OPERATOR_PRICED_STREAM + ), + StreamParityTestCase( + plain=GATEWAY_REPORTED_COST_OPENROUTER_OPERATOR_PRICED, + streamed=GATEWAY_REPORTED_COST_OPENROUTER_OPERATOR_PRICED_STREAM, + ), + StreamParityTestCase(plain=GATEWAY_REPORTED_COST_RESPONSES, streamed=GATEWAY_REPORTED_COST_RESPONSES_STREAM), + StreamParityTestCase( + plain=GATEWAY_REPORTED_COST_RESPONSES_OPERATOR_PRICED, + streamed=GATEWAY_REPORTED_COST_RESPONSES_OPERATOR_PRICED_STREAM, + ), StreamParityTestCase(plain=GPT_5_3_CODEX_INPUT_TEXT, streamed=GPT_5_3_CODEX_STREAM), StreamParityTestCase(plain=GPT_5_3_CODEX_INPUT_TEXT, streamed=GPT_5_3_CODEX_STREAM_INCOMPLETE), StreamParityTestCase(plain=GPT_5_3_CODEX_INPUT_TEXT, streamed=GPT_5_3_CODEX_STREAM_UNVALIDATED), diff --git a/tests/integration/cost_calculation/cost_map.py b/tests/integration/cost_calculation/cost_map.py index 32722b37c9b..da4c2b6e859 100644 --- a/tests/integration/cost_calculation/cost_map.py +++ b/tests/integration/cost_calculation/cost_map.py @@ -26,6 +26,12 @@ COST_MAP_ENTRIES: Final[dict[str, dict[str, JsonValue]]] = { "input_cost_per_token_above_128k_tokens": 2.2e-07, "output_cost_per_token_above_128k_tokens": 8.8e-07, }, + "gateway-reported-cost-fixture": { + "litellm_provider": "openai", + "mode": "chat", + "input_cost_per_token": 1e-06, + "output_cost_per_token": 3e-06, + }, "openrouter/anthropic/claude-sonnet-5": { "litellm_provider": "openrouter", "mode": "chat", diff --git a/tests/unit/litellm_core_utils/test_core_helpers.py b/tests/unit/litellm_core_utils/test_core_helpers.py index a4a3ae285b1..1c86c4e30d8 100644 --- a/tests/unit/litellm_core_utils/test_core_helpers.py +++ b/tests/unit/litellm_core_utils/test_core_helpers.py @@ -69,9 +69,7 @@ def test_preserve_upstream_non_openai_attributes_keeps_scalar_and_structured_val scalar_attribute="scalar", structured_attribute={"items": ["value"]}, ) - model_response: Final = ModelResponseStream( - choices=[StreamingChoices(index=0, delta=Delta(content="hello"))] - ) + model_response: Final = ModelResponseStream(choices=[StreamingChoices(index=0, delta=Delta(content="hello"))]) preserve_upstream_non_openai_attributes(model_response, original_chunk) @@ -262,9 +260,7 @@ class TestMapFinishReasonAnthropic: ("content_filtered", "content_filter"), ], ) - def test_anthropic_finish_reasons( - self, provider_reason: str, expected: str - ) -> None: + def test_anthropic_finish_reasons(self, provider_reason: str, expected: str) -> None: assert map_finish_reason(provider_reason) == expected def test_refusal(self): @@ -336,9 +332,7 @@ class TestMapFinishReasonZhipu: class TestMapFinishReasonOpenAIPassthrough: - @pytest.mark.parametrize( - "reason", ["stop", "length", "tool_calls", "function_call", "content_filter"] - ) + @pytest.mark.parametrize("reason", ["stop", "length", "tool_calls", "function_call", "content_filter"]) def test_openai_values_pass_through(self, reason): assert map_finish_reason(reason) == reason @@ -349,9 +343,7 @@ class TestMapFinishReasonGenericError: assert map_finish_reason("error") == "stop" def test_lowercase_error_does_not_warn(self, mocker): - warn = mocker.patch( - "litellm.litellm_core_utils.core_helpers.verbose_logger.warning" - ) + warn = mocker.patch("litellm.litellm_core_utils.core_helpers.verbose_logger.warning") assert map_finish_reason("error") == "stop" warn.assert_not_called() @@ -369,8 +361,7 @@ class TestFinishReasonMapOutputsAreValid: """Every value in _FINISH_REASON_MAP must be a valid OpenAI finish reason.""" for provider_reason, openai_reason in _FINISH_REASON_MAP.items(): assert openai_reason in VALID_OPENAI_FINISH_REASONS, ( - f"Mapped value '{openai_reason}' (from '{provider_reason}') " - f"is not a valid OpenAI finish reason" + f"Mapped value '{openai_reason}' (from '{provider_reason}') is not a valid OpenAI finish reason" ) @@ -380,28 +371,18 @@ class TestRedactNestedMatchAndRegexKeys: "assessments": [ { "sensitiveInformationPolicy": { - "piiEntities": [ - {"type": "NAME", "match": "secret-name", "action": "BLOCKED"} - ] - }, - "wordPolicy": { - "customWords": [{"match": "badword", "action": "BLOCKED"}] + "piiEntities": [{"type": "NAME", "match": "secret-name", "action": "BLOCKED"}] }, + "wordPolicy": {"customWords": [{"match": "badword", "action": "BLOCKED"}]}, } ], "regex": "should-redact-key-named-regex", } out = redact_nested_match_and_regex_keys(payload) - assert out["assessments"][0]["sensitiveInformationPolicy"]["piiEntities"][0][ - "match" - ] == "[REDACTED]" - assert out["assessments"][0]["wordPolicy"]["customWords"][0]["match"] == ( - "[REDACTED]" - ) + assert out["assessments"][0]["sensitiveInformationPolicy"]["piiEntities"][0]["match"] == "[REDACTED]" + assert out["assessments"][0]["wordPolicy"]["customWords"][0]["match"] == ("[REDACTED]") assert out["regex"] == "[REDACTED]" - assert payload["assessments"][0]["sensitiveInformationPolicy"]["piiEntities"][ - 0 - ]["match"] == "secret-name" + assert payload["assessments"][0]["sensitiveInformationPolicy"]["piiEntities"][0]["match"] == "secret-name" def test_passes_through_none_and_str(self): assert redact_nested_match_and_regex_keys(None) is None @@ -564,13 +545,17 @@ class TestIsExpectedClientError: assert is_expected_client_error(over_budget) is True litellm_limit = RateLimitError( - message="key over rpm", llm_provider="anthropic", model="claude-haiku-4-5", + message="key over rpm", + llm_provider="anthropic", + model="claude-haiku-4-5", category=RateLimitErrorCategory.LITELLM_RATE_LIMIT, ) assert is_expected_client_error(litellm_limit) is True vendor_limit = RateLimitError( - message="rate limited upstream", llm_provider="anthropic", model="claude-haiku-4-5", + message="rate limited upstream", + llm_provider="anthropic", + model="claude-haiku-4-5", category=RateLimitErrorCategory.VENDOR_RATE_LIMIT, ) assert is_expected_client_error(vendor_limit) is False @@ -803,3 +788,34 @@ class TestReceiveStampSurvivesASecondMetadataBucket: } assert get_litellm_metadata_from_kwargs(kwargs)["litellm_received_at"] is own + + +def test_billed_usage_cost_bills_any_chat_upstream_without_its_own_reader() -> None: + from types import SimpleNamespace + + from litellm.litellm_core_utils.core_helpers import billed_usage_cost + from litellm.types.utils import ModelResponse, Usage + + usage = Usage(prompt_tokens=100, completion_tokens=20, total_tokens=120, cost=0.037) + chat = ModelResponse(usage=usage) + assert billed_usage_cost(chat, "openai") == 0.037 + assert billed_usage_cost(chat, "openrouter") == 0.037 + assert billed_usage_cost(chat, "xai") is None + assert billed_usage_cost(SimpleNamespace(usage=usage), "openai") == 0.037 + + +def test_billed_usage_cost_keeps_the_upstream_charge_after_litellm_wrote_its_bill_into_usage_cost() -> None: + from litellm.litellm_core_utils.core_helpers import billed_usage_cost, stamp_litellm_bill_on_usage + from litellm.types.utils import ModelResponse, Usage + + reported = ModelResponse(usage=Usage(prompt_tokens=100, completion_tokens=20, total_tokens=120, cost=0.037)) + stamp_litellm_bill_on_usage(reported, reported.usage, 0.0407) + stamp_litellm_bill_on_usage(reported, reported.usage, 0.0407) + assert reported.usage.cost == 0.0407 + assert reported._hidden_params["response_cost"] == 0.0407 + assert billed_usage_cost(reported, "openai") == 0.037 + + unreported = ModelResponse(usage=Usage(prompt_tokens=100, completion_tokens=20, total_tokens=120)) + stamp_litellm_bill_on_usage(unreported, unreported.usage, 0.0015) + assert unreported.usage.cost == 0.0015 + assert billed_usage_cost(unreported, "openai") is None diff --git a/tests/unit/litellm_core_utils/test_streaming_handler.py b/tests/unit/litellm_core_utils/test_streaming_handler.py index f4502937fce..41e8c50fa85 100644 --- a/tests/unit/litellm_core_utils/test_streaming_handler.py +++ b/tests/unit/litellm_core_utils/test_streaming_handler.py @@ -2659,7 +2659,7 @@ def test_perplexity_streaming_dict_cost_bills_through_its_own_calculator(): ) == pytest.approx(0.00503) -def test_openai_compatible_streaming_cost_is_priced_from_the_cost_map(): +def test_openai_compatible_streaming_reported_cost_is_billed_instead_of_the_cost_map(): import litellm from litellm.cost_calculator import ( get_response_cost_from_hidden_params, @@ -2681,7 +2681,7 @@ def test_openai_compatible_streaming_cost_is_priced_from_the_cost_map(): id="chatcmpl-openai-compatible", model=model, choices=[], - usage=Usage(completion_tokens=5, prompt_tokens=10, total_tokens=15, cost=3_144_000), + usage=Usage(completion_tokens=5, prompt_tokens=10, total_tokens=15, cost=0.037), ) CustomStreamWrapper._propagate_usage_cost_to_hidden_params(complete_response, "openai") @@ -2693,7 +2693,7 @@ def test_openai_compatible_streaming_cost_is_priced_from_the_cost_map(): custom_llm_provider="openai", call_type="completion", optional_params={}, - ) == pytest.approx(2e-5) + ) == pytest.approx(0.037) def test_xai_streaming_reported_cost_still_takes_the_margin(monkeypatch): diff --git a/tests/unit/llms/openai_like/test_kyma_provider.py b/tests/unit/llms/openai_like/test_kyma_provider.py index 2b3740ca676..e834e0fc0e9 100644 --- a/tests/unit/llms/openai_like/test_kyma_provider.py +++ b/tests/unit/llms/openai_like/test_kyma_provider.py @@ -254,7 +254,7 @@ def test_interaction_on_kyma_is_billed_at_the_cost_kyma_reports(): @pytest.mark.parametrize("model", ("openrouter/openai/gpt-4o-mini", "deepseek/deepseek-chat")) -def test_interaction_on_another_provider_keeps_its_upstream_cost_out_of_the_body(model: str): +def test_interaction_on_another_openai_compatible_provider_is_billed_at_its_reported_cost(model: str): chat_usage: Final = {"prompt_tokens": 4, "completion_tokens": 3, "total_tokens": 7, "cost": 0.00721} responses_body: Final = { "id": "resp_upstream", @@ -278,7 +278,7 @@ def test_interaction_on_another_provider_keeps_its_upstream_cost_out_of_the_body upstream.route(path__regex=r".*/responses$").respond(200, json=responses_body) interaction: Final = litellm.interactions.create(model=model, input="ping", api_key="k") - assert "cost" not in interaction.model_dump(exclude_none=True)["usage"] + assert interaction.model_dump(exclude_none=True)["usage"]["cost"] == pytest.approx(0.00721) def test_sdk_completion_cost_prices_kyma_at_the_cost_kyma_reports(): @@ -410,8 +410,8 @@ def test_cost_headers_show_the_discount_and_margin_taken_from_the_cost_kyma_repo assert float(headers["x-litellm-response-cost-discount-amount"] or "nan") == pytest.approx(0.004) assert float(headers["x-litellm-response-cost-margin-amount"] or "nan") == pytest.approx(0.0004) assert float(headers["x-litellm-response-cost"] or "nan") == pytest.approx(0.0044) - assert headers["x-litellm-response-cost-input"] is None - assert headers["x-litellm-response-cost-output"] is None + assert float(headers["x-litellm-response-cost-input"] or "nan") == 0.0 + assert float(headers["x-litellm-response-cost-output"] or "nan") == pytest.approx(0.0044) @pytest.mark.parametrize("thinking", ({"type": "enabled", "budget_tokens": 1024}, {"type": "adaptive"})) diff --git a/tests/unit/test_main.py b/tests/unit/test_main.py index 571f890a4e6..b2ae6b43d26 100644 --- a/tests/unit/test_main.py +++ b/tests/unit/test_main.py @@ -5443,6 +5443,46 @@ def test_stream_chunk_builder_keeps_provider_reported_usage_cost(): assert response._hidden_params["response_cost"] == pytest.approx(0.5) +def test_completion_cost_on_a_stamped_stream_still_bills_the_reported_charge(): + usage_chunk: Final = _stream_builder_text_chunk("gpt-4o", "") + usage_chunk.usage = Usage(prompt_tokens=10, completion_tokens=5, total_tokens=15, cost=0.5) + chunks: Final = [ + _stream_builder_text_chunk("gpt-4o", "Hello ", finish_reason="stop"), + usage_chunk, + ] + + response: Final = litellm.stream_chunk_builder( + chunks=chunks, messages=[{"role": "user", "content": "hi"}], logging_obj=_stream_builder_logging_obj() + ) + + assert response is not None + assert litellm.completion_cost(completion_response=response) == pytest.approx(0.5) + + +def test_stream_chunk_builder_stamps_a_zero_operator_price_over_a_reported_cost(): + logging_obj: Final = _stream_builder_logging_obj() + logging_obj.litellm_params = { + **logging_obj.litellm_params, + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + logging_obj.custom_cost_per_token = {"input_cost_per_token": 0.0, "output_cost_per_token": 0.0} + usage_chunk: Final = _stream_builder_text_chunk("gpt-4o", "") + usage_chunk.usage = Usage(prompt_tokens=10, completion_tokens=5, total_tokens=15, cost=0.5) + chunks: Final = [ + _stream_builder_text_chunk("gpt-4o", "Hello ", finish_reason="stop"), + usage_chunk, + ] + + response: Final = litellm.stream_chunk_builder( + chunks=chunks, messages=[{"role": "user", "content": "hi"}], logging_obj=logging_obj + ) + + assert response is not None + assert getattr(response.usage, "cost", None) == 0.0 + assert response._hidden_params["response_cost"] == 0.0 + + def test_stream_chunk_builder_prices_alias_from_openai_sdk_usage_chunk(): from openai.types.completion_usage import CompletionUsage From 342ce6ae4a4549cacaf96e47aa80a096ae1035bd Mon Sep 17 00:00:00 2001 From: DewellNaa Date: Thu, 29 Oct 2026 20:30:50 -0700 Subject: [PATCH 03/23] feat(zai): support response_format and max_completion_tokens, add glm-5v-turbo pricing (#48335) * feat(zai): fall back to BIGMODEL/ZHIPUAI api keys and add GLM-5.3 aliases and glm-5v-turbo Resolve the zai api key from ZAI_API_KEY, BIGMODEL_API_KEY or ZHIPUAI_API_KEY so deployments predating the zai rename keep working. Accept response_format and max_completion_tokens as supported openai params. Register zai/GLM-5.3 and zai/GLM-5.3-Flash aliases plus the glm-5v-turbo vision model in both the main and backup cost maps * fix(zai): read only ZAI_API_KEY, dropping the BigModel key fallbacks * fix(zai): map max_completion_tokens to max_tokens, correct glm-5v-turbo pricing, drop redundant aliases * fix(zai): reject json_schema response_format, add glm-5v-turbo cache read price * fix(zai): forward every response_format type so allowed_openai_params keeps working * fix(zai): follow json_schema natively on glm-5.3, prompt other models, drop over-limit max_tokens under drop_params * fix(zai): keep BigModel's json_schema as sent and never drop the caller's max_tokens under drop_params * fix(zai): write the prompted JSON schema as JSON so models answer with valid JSON * test(zai): type the zai provider tests and mark their locals Final --------- Co-authored-by: DewellNaa <210961864+DewellNaa@users.noreply.github.com> Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../prompt_templates/factory.py | 34 ++- litellm/llms/bigmodel/chat/transformation.py | 4 + litellm/llms/zai/chat/transformation.py | 94 ++++++++- ...odel_prices_and_context_window_backup.json | 17 ++ model_prices_and_context_window.json | 17 ++ .../llms/bigmodel/test_bigmodel_provider.py | 23 ++ tests/unit/llms/zai/test_zai_provider.py | 197 ++++++++++++++++++ 7 files changed, 364 insertions(+), 22 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/factory.py b/litellm/litellm_core_utils/prompt_templates/factory.py index fad4caa59a0..834d26b4fe8 100644 --- a/litellm/litellm_core_utils/prompt_templates/factory.py +++ b/litellm/litellm_core_utils/prompt_templates/factory.py @@ -5878,38 +5878,36 @@ def function_call_prompt( ] -def response_schema_prompt(model: str, response_schema: dict, llm_provider: str) -> str: +def response_schema_prompt(model: str, response_schema: Mapping[str, object], llm_provider: str) -> str: """ Decides if a user-defined custom prompt or default needs to be used Returns the prompt str that's passed to the model as a user message """ - custom_prompt_details: dict | None = None - response_schema_as_message: Final = [{"role": "user", "content": f"{response_schema}"}] - if f"{model}/response_schema_prompt" in litellm.custom_prompt_dict: - custom_prompt_details = litellm.custom_prompt_dict[ - f"{model}/response_schema_prompt" - ] # allow user to define custom response schema prompt by model - elif "response_schema_prompt" in litellm.custom_prompt_dict: - custom_prompt_details = litellm.custom_prompt_dict["response_schema_prompt"] - - if custom_prompt_details is not None: - return render_custom_prompt_template( - template=custom_prompt_details, messages=response_schema_as_message, model=model, llm_provider=llm_provider - ) - else: + prompts: Final[Mapping[str, Mapping[str, object]]] = litellm.custom_prompt_dict + model_prompt_key: Final = f"{model}/response_schema_prompt" + custom_prompt_details: Final = prompts.get( + model_prompt_key if model_prompt_key in prompts else "response_schema_prompt" + ) + if custom_prompt_details is None: return default_response_schema_prompt(response_schema=response_schema) + return render_custom_prompt_template( + template=custom_prompt_details, + messages=[{"role": "user", "content": f"{response_schema}"}], + model=model, + llm_provider=llm_provider, + ) -def default_response_schema_prompt(response_schema: dict) -> str: +def default_response_schema_prompt(response_schema: Mapping[str, object]) -> str: """ Used if provider/model doesn't support 'response_schema' param. This is the default prompt. Allow user to override this with a custom_prompt. """ prompt_str: Final = f"""Use this JSON schema: - ```json - {response_schema} + ```json + {json.dumps(response_schema, default=str)} ```""" return prompt_str diff --git a/litellm/llms/bigmodel/chat/transformation.py b/litellm/llms/bigmodel/chat/transformation.py index 12d5e912c15..93e9e24fe7b 100644 --- a/litellm/llms/bigmodel/chat/transformation.py +++ b/litellm/llms/bigmodel/chat/transformation.py @@ -26,6 +26,10 @@ class BigModelChatConfig(ZAIChatConfig): def custom_llm_provider(self) -> str: return "bigmodel" + @property + def prompts_unsupported_json_schema(self) -> bool: + return False + def get_openai_compatible_provider_info(self, api_base: str | None, api_key: str | None) -> tuple[str, str | None]: return get_bigmodel_url(api_base, "chat"), get_bigmodel_api_key(api_key) diff --git a/litellm/llms/zai/chat/transformation.py b/litellm/llms/zai/chat/transformation.py index e0e42d772f8..9e2b97d2bed 100644 --- a/litellm/llms/zai/chat/transformation.py +++ b/litellm/llms/zai/chat/transformation.py @@ -1,11 +1,43 @@ +from collections.abc import Mapping from typing import Final +from typing_extensions import TypeIs # noqa: TID251 # TypeIs reaches typing only on 3.13 + +import litellm +from litellm.litellm_core_utils.prompt_templates.factory import response_schema_prompt from litellm.secret_managers.main import get_secret_str -from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam +from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam, ChatCompletionUserMessage +from litellm.utils import ModelInfoLookupMiss, lookup_model_info_for_registration from ...openai.chat.gpt_transformation import OpenAIGPTConfig ZAI_API_BASE: Final = "https://api.z.ai/api/paas/v4" +TOKEN_LIMIT_PARAMS: Final = frozenset({"max_tokens", "max_completion_tokens"}) + + +def _is_json_object( + value: object, +) -> TypeIs[Mapping[str, object]]: # guard-ok: cost-map rows and request params are JSON + return isinstance(value, Mapping) + + +def _json_field(value: object, key: str) -> object: + return value.get(key) if _is_json_object(value) else None + + +def _exceeds_output_limit(model: str, max_tokens: object) -> bool: + info: Final = lookup_model_info_for_registration(model=model, custom_llm_provider="zai") + limit: Final = None if isinstance(info, ModelInfoLookupMiss) else info.get("max_output_tokens") + return isinstance(max_tokens, int) and isinstance(limit, int) and max_tokens > limit + + +def _schema_without_native_support(model: str, response_format: object) -> Mapping[str, object] | None: + if _json_field(response_format, "type") != "json_schema": + return None + if litellm.supports_response_schema(model=model, custom_llm_provider="zai"): + return None + schema: Final = _json_field(_json_field(response_format, "json_schema"), "schema") + return schema if _is_json_object(schema) else None class ZAIChatConfig(OpenAIGPTConfig): @@ -13,6 +45,10 @@ class ZAIChatConfig(OpenAIGPTConfig): def custom_llm_provider(self) -> str | None: return "zai" + @property + def prompts_unsupported_json_schema(self) -> bool: + return True + def _get_openai_compatible_provider_info( self, api_base: str | None, api_key: str | None ) -> tuple[str | None, str | None]: @@ -40,9 +76,11 @@ class ZAIChatConfig(OpenAIGPTConfig): # GLM/ZAI supports cache_control, so return messages and tools unchanged return messages, tools - def get_supported_openai_params(self, model: str) -> list: + def get_supported_openai_params(self, model: str) -> list[str]: base_params: Final = [ "max_tokens", + "max_completion_tokens", + "response_format", "stream", "stream_options", "temperature", @@ -52,8 +90,6 @@ class ZAIChatConfig(OpenAIGPTConfig): "tool_choice", ] - import litellm - try: if litellm.supports_reasoning(model=model, custom_llm_provider=self.custom_llm_provider): base_params.append("thinking") @@ -61,3 +97,53 @@ class ZAIChatConfig(OpenAIGPTConfig): pass return base_params + + def map_openai_params( + self, + non_default_params: Mapping[str, object], + optional_params: Mapping[str, object], + model: str, + drop_params: bool, + ) -> dict[str, object]: # mutable-ok: request pipeline adds parameters to the returned body + supported_params: Final = self.get_supported_openai_params(model) + sent: Final = {param: value for param, value in non_default_params.items() if param in supported_params} + max_completion_tokens: Final = sent.get("max_completion_tokens") + falls_back_to_max_tokens: Final = max_completion_tokens is None or ( + (litellm.drop_params or drop_params) and _exceeds_output_limit(model, max_completion_tokens) + ) + max_tokens: Final = sent.get("max_tokens") if falls_back_to_max_tokens else max_completion_tokens + return { + **optional_params, + **{param: value for param, value in sent.items() if param not in TOKEN_LIMIT_PARAMS}, + **({} if max_tokens is None else {"max_tokens": max_tokens}), + } + + def transform_request( + self, + model: str, + messages: list[AllMessageValues], # mutable-ok: OpenAIGPTConfig.transform_request signature + optional_params: dict[str, object], # mutable-ok: OpenAIGPTConfig.transform_request signature + litellm_params: dict[str, object], # mutable-ok: OpenAIGPTConfig.transform_request signature + headers: dict[str, str], # mutable-ok: OpenAIGPTConfig.transform_request signature + ) -> dict[str, object]: # mutable-ok: inherited contract returns the request body + schema: Final = ( + _schema_without_native_support(model, optional_params.get("response_format")) + if self.prompts_unsupported_json_schema + else None + ) + schema_message: Final = ( + () + if schema is None + else ( + ChatCompletionUserMessage( + role="user", content=response_schema_prompt(model=model, response_schema=schema, llm_provider="zai") + ), + ) + ) + return super().transform_request( + model, + [*messages, *schema_message], + optional_params if schema is None else {**optional_params, "response_format": {"type": "json_object"}}, + litellm_params, + headers, + ) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 77ac4f49e8b..96a04c96f44 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -54512,6 +54512,7 @@ "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_response_schema": true, "supports_tool_choice": true }, "zai/glm-5.3-flash": { @@ -54526,10 +54527,26 @@ "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_response_schema": true, "supports_tool_choice": true, "source": "https://docs.z.ai/guides/overview/pricing", "supports_vision": true }, + "zai/glm-5v-turbo": { + "cache_read_input_token_cost": 2.4e-07, + "input_cost_per_token": 1.2e-06, + "litellm_provider": "zai", + "max_input_tokens": 200000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://vercel.com/ai-gateway/models/glm-5v-turbo", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": true + }, "zai/glm-5.1": { "cache_creation_input_token_cost": 0, "cache_read_input_token_cost": 2.6e-07, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 77ac4f49e8b..96a04c96f44 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -54512,6 +54512,7 @@ "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_response_schema": true, "supports_tool_choice": true }, "zai/glm-5.3-flash": { @@ -54526,10 +54527,26 @@ "supports_function_calling": true, "supports_prompt_caching": true, "supports_reasoning": true, + "supports_response_schema": true, "supports_tool_choice": true, "source": "https://docs.z.ai/guides/overview/pricing", "supports_vision": true }, + "zai/glm-5v-turbo": { + "cache_read_input_token_cost": 2.4e-07, + "input_cost_per_token": 1.2e-06, + "litellm_provider": "zai", + "max_input_tokens": 200000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4e-06, + "source": "https://vercel.com/ai-gateway/models/glm-5v-turbo", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_tool_choice": true, + "supports_vision": true + }, "zai/glm-5.1": { "cache_creation_input_token_cost": 0, "cache_read_input_token_cost": 2.6e-07, diff --git a/tests/unit/llms/bigmodel/test_bigmodel_provider.py b/tests/unit/llms/bigmodel/test_bigmodel_provider.py index cdf98c8f9a4..6abfcb060d0 100644 --- a/tests/unit/llms/bigmodel/test_bigmodel_provider.py +++ b/tests/unit/llms/bigmodel/test_bigmodel_provider.py @@ -197,6 +197,29 @@ def test_chat_does_not_send_the_openai_organization_to_bigmodel(monkeypatch: pyt assert "openai-organization" not in route.calls.last.request.headers +@respx.mock +def test_chat_sends_json_schema_response_format_as_given() -> None: + route: Final = respx.post(f"{BASE}/api/paas/v4/chat/completions").respond( + json={ + "id": "chat_test", + "object": "chat.completion", + "created": 1, + "model": "glm-4.6", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "{}"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + } + ) + messages: Final = [{"role": "user", "content": "Ann likes 7"}] + response_format: Final = { + "type": "json_schema", + "json_schema": {"name": "profile", "schema": {"type": "object", "properties": {"n": {"type": "integer"}}}}, + } + _ = litellm.completion(model="bigmodel/glm-4.6", messages=messages, response_format=response_format) + body: Final = json.loads(route.calls.last.request.content) + assert body["response_format"] == response_format + assert body["messages"] == messages + + @pytest.mark.parametrize("is_async", (False, True)) @pytest.mark.asyncio @respx.mock diff --git a/tests/unit/llms/zai/test_zai_provider.py b/tests/unit/llms/zai/test_zai_provider.py index 32849d5eef1..fb126bd8f41 100644 --- a/tests/unit/llms/zai/test_zai_provider.py +++ b/tests/unit/llms/zai/test_zai_provider.py @@ -2,13 +2,27 @@ Tests for Z.AI (Zhipu AI) provider - GLM models """ +import json import math +from collections.abc import Mapping +from typing import Final import pytest +import respx import litellm from litellm import completion from litellm.cost_calculator import cost_per_token +from litellm.llms.zai.chat.transformation import ZAIChatConfig +from litellm.types.utils import PromptTokensDetailsWrapper, Usage + +JSON_SCHEMA_RESPONSE_FORMAT: Final = { + "type": "json_schema", + "json_schema": { + "name": "profile", + "schema": {"type": "object", "properties": {"favorite_number": {"type": "integer"}}}, + }, +} @pytest.fixture @@ -99,3 +113,186 @@ def test_zai_sync_completion(respx_mock, zai_response, monkeypatch): assert response.choices[0].message.content == "Hello! How can I help you today?" assert response.usage.total_tokens == 25 + + +def test_zai_sends_response_format_and_maps_max_completion_tokens( + respx_mock: respx.MockRouter, zai_response: Mapping[str, object], monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setenv("ZAI_API_KEY", "test-api-key") + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(json=zai_response) + + completion( + model="zai/glm-5.3", + messages=[{"role": "user", "content": "Reply in JSON"}], + response_format={"type": "json_object"}, + max_completion_tokens=64, + ) + + body: Final = json.loads(respx_mock.calls[0].request.content) + assert body["response_format"] == {"type": "json_object"} + assert body["max_tokens"] == 64 + assert "max_completion_tokens" not in body + + +@pytest.mark.parametrize( + "token_limits", + [{"max_tokens": 300, "max_completion_tokens": 15}, {"max_completion_tokens": 15, "max_tokens": 300}], +) +def test_zai_max_completion_tokens_wins_over_max_tokens(token_limits: dict[str, int]) -> None: + mapped: Final = ZAIChatConfig().map_openai_params(token_limits, {}, "glm-5.3", drop_params=False) + + assert mapped["max_tokens"] == 15 + assert "max_completion_tokens" not in mapped + + +@pytest.mark.parametrize( + ("over_limit_by", "drop_params", "forwarded"), + [(1, True, False), (1, False, True), (0, True, True)], +) +def test_zai_drop_params_drops_max_completion_tokens_above_model_limit( + local_model_cost_map: None, + respx_mock: respx.MockRouter, + zai_response: Mapping[str, object], + monkeypatch: pytest.MonkeyPatch, + over_limit_by: int, + drop_params: bool, + forwarded: bool, +) -> None: + monkeypatch.setenv("ZAI_API_KEY", "test-api-key") + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(json=zai_response) + requested: Final = litellm.model_cost["zai/glm-4.5-flash"]["max_output_tokens"] + over_limit_by + + completion( + model="zai/glm-4.5-flash", + messages=[{"role": "user", "content": "Count"}], + max_completion_tokens=requested, + drop_params=drop_params, + ) + + body: Final = json.loads(respx_mock.calls[0].request.content) + assert body.get("max_tokens") == (requested if forwarded else None) + + +@pytest.mark.parametrize("with_over_limit_max_completion_tokens", [False, True]) +def test_zai_drop_params_keeps_max_tokens_above_model_limit( + local_model_cost_map: None, + respx_mock: respx.MockRouter, + zai_response: Mapping[str, object], + monkeypatch: pytest.MonkeyPatch, + with_over_limit_max_completion_tokens: bool, +) -> None: + monkeypatch.setenv("ZAI_API_KEY", "test-api-key") + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(json=zai_response) + over_limit: Final = litellm.model_cost["zai/glm-4.5-flash"]["max_output_tokens"] + 1 + over_limit_max_completion_tokens: Final = ( + {"max_completion_tokens": over_limit + 1} if with_over_limit_max_completion_tokens else {} + ) + + completion( + model="zai/glm-4.5-flash", + messages=[{"role": "user", "content": "Count"}], + max_tokens=over_limit, + drop_params=True, + **over_limit_max_completion_tokens, + ) + + body: Final = json.loads(respx_mock.calls[0].request.content) + assert body["max_tokens"] == over_limit + + +@pytest.mark.parametrize("pass_through", [{}, {"allowed_openai_params": ["response_format"]}]) +def test_zai_forwards_json_schema_to_models_that_enforce_it( + local_model_cost_map: None, + respx_mock: respx.MockRouter, + zai_response: Mapping[str, object], + monkeypatch: pytest.MonkeyPatch, + pass_through: dict[str, list[str]], +) -> None: + monkeypatch.setenv("ZAI_API_KEY", "test-api-key") + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(json=zai_response) + messages: Final = [{"role": "user", "content": "Ann likes 7"}] + + completion(model="zai/glm-5.3", messages=messages, response_format=JSON_SCHEMA_RESPONSE_FORMAT, **pass_through) + + body: Final = json.loads(respx_mock.calls[0].request.content) + assert body["response_format"] == JSON_SCHEMA_RESPONSE_FORMAT + assert body["messages"] == messages + + +@pytest.mark.parametrize("pass_through", [{}, {"allowed_openai_params": ["response_format"]}]) +def test_zai_puts_json_schema_in_prompt_for_models_that_ignore_it( + local_model_cost_map: None, + respx_mock: respx.MockRouter, + zai_response: Mapping[str, object], + monkeypatch: pytest.MonkeyPatch, + pass_through: dict[str, list[str]], +) -> None: + monkeypatch.setenv("ZAI_API_KEY", "test-api-key") + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(json=zai_response) + + completion( + model="zai/glm-4.6", + messages=[{"role": "user", "content": "Ann likes 7"}], + response_format=JSON_SCHEMA_RESPONSE_FORMAT, + **pass_through, + ) + + body: Final = json.loads(respx_mock.calls[0].request.content) + assert body["response_format"] == {"type": "json_object"} + assert body["messages"][0] == {"role": "user", "content": "Ann likes 7"} + assert body["messages"][-1]["role"] == "user" + prompt: Final = body["messages"][-1]["content"] + rendered_schema: Final = prompt[prompt.index("{") : prompt.rindex("}") + 1] + assert json.loads(rendered_schema) == JSON_SCHEMA_RESPONSE_FORMAT["json_schema"]["schema"] + + +def test_zai_glm_5v_turbo_bills_cached_input_at_cache_read_price(local_model_cost_map: None) -> None: + usage: Final = Usage( + prompt_tokens=1000, + completion_tokens=0, + total_tokens=1000, + prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=800), + ) + + prompt_cost, _ = cost_per_token(model="glm-5v-turbo", custom_llm_provider="zai", usage_object=usage) + + entry: Final = litellm.model_cost["zai/glm-5v-turbo"] + assert math.isclose( + prompt_cost, 200 * entry["input_cost_per_token"] + 800 * entry["cache_read_input_token_cost"] + ) + + +@pytest.mark.parametrize("model", ["glm-5v-turbo", "GLM-5V-Turbo", "GLM-5.3", "GLM-5.3-Flash"]) +def test_zai_cost_resolves_case_insensitively(local_model_cost_map: None, model: str) -> None: + prompt_cost, completion_cost = cost_per_token( + model=model, custom_llm_provider="zai", prompt_tokens=1000, completion_tokens=1000 + ) + + entry: Final = litellm.model_cost[f"zai/{model.lower()}"] + assert math.isclose(prompt_cost, 1000 * entry["input_cost_per_token"]) + assert math.isclose(completion_cost, 1000 * entry["output_cost_per_token"]) + + +def test_zai_glm_5v_turbo_accepts_thinking( + local_model_cost_map: None, + respx_mock: respx.MockRouter, + zai_response: Mapping[str, object], + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setenv("ZAI_API_KEY", "test-api-key") + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(json=zai_response) + + completion( + model="zai/glm-5v-turbo", + messages=[{"role": "user", "content": "Hello"}], + thinking={"type": "disabled"}, + ) + + body: Final = json.loads(respx_mock.calls[0].request.content) + assert body["thinking"] == {"type": "disabled"} From b91fc92b6d0815df038e3496b669871fd917c056 Mon Sep 17 00:00:00 2001 From: paveln83 <349044532+paveln83@users.noreply.github.com> Date: Thu, 29 Oct 2026 20:32:04 -0700 Subject: [PATCH 04/23] fix(router): cap the default OTPM reservation and reconcile token counters after each call (#49054) * fix(router): cap the default OTPM reservation for requests without max_tokens * fix(router): reconcile ITPM and OTPM reservations after every successful call * fix(router): read the default OTPM reservation from the served model and make its fraction configurable The default reservation now looks up max output from litellm_params.model, as main does, so a pricing base_model never changes it. NO_MAX_TOKENS_TPM_RESERVATION_FRACTION sets the fraction for both the router check and the key and team TPM floor. A max_tokens that is not a number or string counts as no cap instead of raising inside the check. * fix(router): refund a failed attempt's io token reservation in make_call The proxy passes one logging object to every retry, and it runs failure callbacks only for the first attempt, so the last attempt's ITPM and OTPM reservation stayed held until the minute ended. Each attempt that raises now gives its reservation back before the retry or fallback * fix(router): count Anthropic cache writes once in ITPM reconcile Native /v1/messages usage reports input_tokens without cache reads and writes, so the reconcile now maps it through AnthropicConfig.calculate_usage before subtracting cached tokens. * fix(router): use a deployment max_tokens, charge late usage to its finish minute, and read the reservation fraction as a fraction * fix(router): reserve with the deployment cap only through the param the request's API sends * fix(router): never let a deployment's output cap raise the reservation above the default * test(router): move the reservation fraction env var test to the integration suite The unit test started a child interpreter, which the unit guide forbids. The integration test boots a real proxy with the env var set and checks admission against the held reservation. * test(router): retry the reservation fraction scenario when the minute rolls over The OTPM counter is keyed by minute, so a rollover between the held stream and the capped requests gave the capped request a fresh counter. The control request now leaves room for usage the first attempt left in the new minute. * test(router): give the rollover retry its own deployment --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/constants.py | 4 + litellm/litellm_core_utils/env_utils.py | 10 +- .../hooks/parallel_request_limiter_v3.py | 17 +- litellm/router.py | 30 +- .../io_token_rate_limit_check.py | 185 +++--- .../pre_call_checks/model_rate_limit_check.py | 37 +- .../test_otpm_default_reservation_fraction.py | 137 +++++ .../unit/litellm_core_utils/test_env_utils.py | 10 + tests/unit/test_constants.py | 9 +- .../test_router/test_io_token_rate_limits.py | 530 +++++++++++++++++- 10 files changed, 831 insertions(+), 138 deletions(-) create mode 100644 tests/integration/routing/test_otpm_default_reservation_fraction.py diff --git a/litellm/constants.py b/litellm/constants.py index e07d727973d..cfee777f36a 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -7,6 +7,7 @@ from typing import Final, Literal from litellm.litellm_core_utils.env_utils import ( get_env_float, + get_env_float_in_range, get_env_float_or_none, get_env_int, get_env_int_in_range, @@ -149,6 +150,9 @@ DEFAULT_FAILURE_THRESHOLD_PERCENT: Final = get_env_float( "DEFAULT_FAILURE_THRESHOLD_PERCENT", 0.5 ) # default cooldown a deployment if 50% of requests fail in a given minute DEFAULT_MAX_TOKENS: Final = get_env_int("DEFAULT_MAX_TOKENS", 4096) +NO_MAX_TOKENS_TPM_RESERVATION_FRACTION: Final = get_env_float_in_range( + "NO_MAX_TOKENS_TPM_RESERVATION_FRACTION", 0.25, 0.01, 1.0 +) DEFAULT_ALLOWED_FAILS: Final = get_env_int("DEFAULT_ALLOWED_FAILS", 3) DEFAULT_REDIS_SYNC_INTERVAL: Final = get_env_int("DEFAULT_REDIS_SYNC_INTERVAL", 1) DEFAULT_COOLDOWN_TIME_SECONDS: Final = get_env_int("DEFAULT_COOLDOWN_TIME_SECONDS", 5) diff --git a/litellm/litellm_core_utils/env_utils.py b/litellm/litellm_core_utils/env_utils.py index 6883af98bbd..462b0b3df47 100644 --- a/litellm/litellm_core_utils/env_utils.py +++ b/litellm/litellm_core_utils/env_utils.py @@ -71,7 +71,15 @@ def get_env_int_in_range(env_var: str, default: int, minimum: int, maximum: int) Values outside the range fall back to the default and warn, so a misconfigured knob can neither crash the caller nor silently change the meaning of what it computes. """ - value: Final = get_env_int(env_var, default) + return _within_range(env_var, get_env_int(env_var, default), default, minimum, maximum) + + +def get_env_float_in_range(env_var: str, default: float, minimum: float, maximum: float) -> float: + """Float counterpart of `get_env_int_in_range`.""" + return _within_range(env_var, get_env_float(env_var, default), default, minimum, maximum) + + +def _within_range(env_var: str, value: _NumberT, default: _NumberT, minimum: _NumberT, maximum: _NumberT) -> _NumberT: if minimum <= value <= maximum: return value logging.getLogger("LiteLLM").warning( diff --git a/litellm/proxy/hooks/parallel_request_limiter_v3.py b/litellm/proxy/hooks/parallel_request_limiter_v3.py index 404aff84994..14b08a09116 100644 --- a/litellm/proxy/hooks/parallel_request_limiter_v3.py +++ b/litellm/proxy/hooks/parallel_request_limiter_v3.py @@ -42,7 +42,11 @@ from litellm.caching.redis_batch import ( active_request_redis_batch, ) from litellm.caching.redis_cache import log_redis_failure -from litellm.constants import DYNAMIC_RATE_LIMIT_ERROR_THRESHOLD_PER_MINUTE, INTERNAL_CALL_ORIGIN_METADATA_KEY +from litellm.constants import ( + DYNAMIC_RATE_LIMIT_ERROR_THRESHOLD_PER_MINUTE, + INTERNAL_CALL_ORIGIN_METADATA_KEY, + NO_MAX_TOKENS_TPM_RESERVATION_FRACTION, +) from litellm.cost_calculator import get_usage_object from litellm.integrations.custom_logger import CustomLogger from litellm.litellm_core_utils.audio_utils.utils import estimate_audio_tokens @@ -508,11 +512,6 @@ REDIS_NODE_HASHTAG_NAME: Final = "all_keys" # *some* output budget; these define that fallback estimate. DEFAULT_MAX_TOKENS_ESTIMATE: Final = 4096 DEFAULT_CHARS_PER_TOKEN: Final = 4 -# Fraction of the available output budget reserved as the upfront floor when -# the request omits max_tokens. Applied to both DEFAULT_MAX_TOKENS_ESTIMATE -# (baseline floor) and to the smallest configured TPM limit (capped floor for -# small per-tenant TPM caps). -_TPM_FLOOR_FRACTION: Final = 4 # Both embeddings and the Responses API put their prompt in data["input"], # but only embeddings have no output tokens. Every "is this an embedding" # check on data["input"] must exclude these call types, or a Responses call @@ -1003,10 +1002,10 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): per-tenant cap can't be tripped by the floor alone. Returns the baseline floor when no limit is provided. """ - baseline: Final = DEFAULT_MAX_TOKENS_ESTIMATE // _TPM_FLOOR_FRACTION + baseline: Final = int(DEFAULT_MAX_TOKENS_ESTIMATE * NO_MAX_TOKENS_TPM_RESERVATION_FRACTION) if min_configured_tpm_limit is None: return baseline - return min(baseline, max(1, min_configured_tpm_limit // _TPM_FLOOR_FRACTION)) + return min(baseline, max(1, int(min_configured_tpm_limit * NO_MAX_TOKENS_TPM_RESERVATION_FRACTION))) @staticmethod def _is_embedding_request(data: object, call_type: str | None) -> bool: @@ -1138,7 +1137,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): if call_type in RESPONSES_API_CALL_TYPES else base_capped_floor ) - baseline_floor: Final = DEFAULT_MAX_TOKENS_ESTIMATE // _TPM_FLOOR_FRACTION + baseline_floor: Final = int(DEFAULT_MAX_TOKENS_ESTIMATE * NO_MAX_TOKENS_TPM_RESERVATION_FRACTION) is_embedding: Final = PROXY_MaxParallelRequestsHandler_v3._is_embedding_request(data, call_type) if ( capped_floor >= baseline_floor diff --git a/litellm/router.py b/litellm/router.py index 3a423b79e49..20af2e5649e 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -377,6 +377,7 @@ from litellm.router_utils.pre_call_checks.encrypted_content_affinity_check impor EncryptedContentAffinityCheck, ) from litellm.router_utils.pre_call_checks.io_token_rate_limit_check import ( + async_io_token_refund_failure, build_io_token_rate_limit_headers, deployment_has_io_token_limits, refund_stale_reservation_before_retry, @@ -9500,7 +9501,7 @@ class Router(DeprecatedRouterMembers): raise original_exception - async def make_call(self, original_function: Any, *args, **kwargs): + async def make_call(self, original_function: Any, *args: object, **kwargs: object): """ Handler for making a call to the .completion()/.embeddings()/etc. functions. """ @@ -9509,17 +9510,26 @@ class Router(DeprecatedRouterMembers): endpoint_function: Final = ( generic_function if callable(generic_function) else _OPAQUE_VALUE_ADAPTER.validate_python(original_function) ) - with endpoint_selection( - getattr(endpoint_function, "__name__", None), - getattr(endpoint_function, "__module__", None), - model=_OPAQUE_VALUE_ADAPTER.validate_python(model_group), - ): - response = original_function(*args, **kwargs) - if coroutine_checker.is_async_callable(response) or inspect.isawaitable(response): - response = await response + try: + with endpoint_selection( + getattr(endpoint_function, "__name__", None), + getattr(endpoint_function, "__module__", None), + model=_OPAQUE_VALUE_ADAPTER.validate_python(model_group), + ): + response = original_function(*args, **kwargs) + if coroutine_checker.is_async_callable(response) or inspect.isawaitable(response): + response = await response + except Exception: + with contextlib.suppress(Exception): + await async_io_token_refund_failure( + self.cache, kwargs, parent_otel_span=get_parent_otel_span_from_kwargs(kwargs) + ) + raise await self.increment_deployment_usage_for_response(response=response, request_kwargs=kwargs) ## PROCESS RESPONSE HEADERS - response = await self.set_response_headers(response=response, model_group=model_group, request_kwargs=kwargs) + response = await self.set_response_headers( + response=response, model_group=model_group if isinstance(model_group, str) else None, request_kwargs=kwargs + ) return response diff --git a/litellm/router_utils/pre_call_checks/io_token_rate_limit_check.py b/litellm/router_utils/pre_call_checks/io_token_rate_limit_check.py index 0b3d48c9e1d..7580c7dadd3 100644 --- a/litellm/router_utils/pre_call_checks/io_token_rate_limit_check.py +++ b/litellm/router_utils/pre_call_checks/io_token_rate_limit_check.py @@ -1,7 +1,8 @@ """ Separate ITPM/OTPM (input/output tokens per minute) deployment rate limits. -- Pre-call: atomically reserve estimated_input against ITPM and max_tokens against OTPM +- Pre-call: atomically reserve estimated_input against ITPM and max_tokens against OTPM; a request + without max_tokens reserves the model's max output, capped at a fraction of the OTPM limit - Post-call: reconcile ITPM to actual input tokens and OTPM to actual output tokens - Cached prompt-read tokens are excluded from ITPM post-call accounting @@ -22,6 +23,7 @@ from litellm import token_counter from litellm._internal_context import with_service_target from litellm._logging import verbose_router_logger from litellm.caching.dual_cache import DualCache +from litellm.constants import NO_MAX_TOKENS_TPM_RESERVATION_FRACTION from litellm.litellm_core_utils.token_counter import deployment_openai_model, offload_token_count from litellm.llms.azure.ptu_capacity import is_str_keyed_mapping from litellm.router_utils.routing_read_batch import ROUTER_USAGE_TARGET @@ -127,10 +129,15 @@ def _get_cache_keys(deployment: dict, current_minute: str) -> tuple[str, str] | return itpm_key, otpm_key -def _tokenizer_model(deployment: object) -> str: +def _served_model(deployment: object) -> str: params: Final = deployment.get("litellm_params") if is_str_keyed_mapping(deployment) else None served: Final = params.get("model") if is_str_keyed_mapping(params) else None - return deployment_openai_model(deployment) or (served if isinstance(served, str) else "") + return served if isinstance(served, str) else "" + + +def _tokenizer_and_served_models(deployment: object) -> tuple[str, str]: + served: Final = _served_model(deployment) + return deployment_openai_model(deployment) or served, served def _estimate_input_tokens(request_kwargs: dict[str, Any] | None, model: str = "") -> int: @@ -160,24 +167,61 @@ def _model_max_output_tokens(model_name: str) -> int | None: return None -def _resolve_max_tokens(request_kwargs: dict[str, Any] | None, deployment: dict) -> int: - if request_kwargs: - # An explicit max_tokens=0 must be honored, not treated as absent and - # replaced by the model default. - explicit = request_kwargs.get("max_tokens") - if explicit is None: - explicit = request_kwargs.get("max_completion_tokens") - if explicit is None: - explicit = request_kwargs.get("max_output_tokens") - if explicit is not None: - return max(0, int(explicit)) +_CHAT_OUTPUT_CAP_KEYS: Final = ("max_tokens", "max_completion_tokens") +_RESPONSES_OUTPUT_CAP_KEYS: Final = ("max_output_tokens",) +_OUTPUT_CAP_KEYS: Final = _CHAT_OUTPUT_CAP_KEYS + _RESPONSES_OUTPUT_CAP_KEYS - model_name: Final = (deployment.get("litellm_params") or {}).get("model") - if model_name: - model_max: Final = _model_max_output_tokens(model_name) - if model_max is not None: - return model_max - return 4096 + +def _explicit_max_tokens( + request_kwargs: Mapping[str, object] | None, cap_keys: tuple[str, ...] = _OUTPUT_CAP_KEYS +) -> int | None: + if not request_kwargs: + return None + explicit: Final = next( + (request_kwargs[key] for key in cap_keys if request_kwargs.get(key) is not None), + None, + ) + match explicit: + case int() | float() | str(): + return max(0, int(explicit)) + case _: + return None + + +def _sent_deployment_cap_keys(request_kwargs: Mapping[str, object] | None) -> tuple[str, ...]: + is_responses_request: Final = ( + request_kwargs is not None and "input" in request_kwargs and "messages" not in request_kwargs + ) + return _RESPONSES_OUTPUT_CAP_KEYS if is_responses_request else _CHAT_OUTPUT_CAP_KEYS + + +def _deployment_output_cap(request_kwargs: Mapping[str, object] | None, deployment: Mapping[str, object]) -> int | None: + params: Final = deployment.get("litellm_params") + return ( + _explicit_max_tokens(params, _sent_deployment_cap_keys(request_kwargs)) + if is_str_keyed_mapping(params) + else None + ) + + +def _default_otpm_reservation(model_name: str, otpm_limit: int) -> int: + model_max: Final = _model_max_output_tokens(model_name) if model_name else None + share_of_limit: Final = max(1, int(otpm_limit * NO_MAX_TOKENS_TPM_RESERVATION_FRACTION)) + return min(4096 if model_max is None else model_max, share_of_limit) + + +def _otpm_reservation( + request_kwargs: Mapping[str, object] | None, + deployment: Mapping[str, object], + model_name: str, + otpm_limit: int, +) -> int: + from_request: Final = _explicit_max_tokens(request_kwargs) + if from_request is not None: + return from_request + default: Final = _default_otpm_reservation(model_name, otpm_limit) + from_deployment: Final = _deployment_output_cap(request_kwargs, deployment) + return default if from_deployment is None else min(from_deployment, default) def _get_usage_tokens(usage: object) -> tuple[int, int, int]: @@ -210,6 +254,12 @@ def _extract_response_usage(response_obj: object) -> object: return getattr(response_obj, "usage", None) +def _openai_shaped_usage(usage: object) -> object: + if isinstance(usage, dict) and litellm.AnthropicConfig.is_anthropic_usage_object(usage): + return litellm.AnthropicConfig().calculate_usage(usage_object=usage, reasoning_content=None) + return usage + + def _usage_is_present(usage: object) -> bool: """ True only if usage carries an actual input/output breakdown. @@ -238,7 +288,7 @@ def _resolve_reconcile_usage_tokens( fields. When usage cannot be resolved, return ``usage_resolved=False`` so callers keep the pre-call reservation instead of refunding it as zero usage. """ - usage: Final = _extract_response_usage(response_obj) + usage: Final = _openai_shaped_usage(_extract_response_usage(response_obj)) if _usage_is_present(usage): prompt_tokens, completion_tokens, cached_tokens = _get_usage_tokens(usage) return max(0, prompt_tokens - cached_tokens), completion_tokens, True @@ -261,6 +311,21 @@ def _resolve_reconcile_usage_tokens( return 0, 0, False +def _key_for_this_minute(reserved_key: str) -> str: + counter, _, _ = reserved_key.rpartition(":") + return f"{counter}:{get_utc_datetime().strftime('%H-%M')}" + + +def _reconcile_increments( + reservation: tuple[int, int, str | None, str | None], billable_input: int, completion_tokens: int +) -> tuple[tuple[str, int], ...]: + itpm_reserved, otpm_reserved, itpm_key, otpm_key = reservation + deltas: Final = ((itpm_key, billable_input - itpm_reserved), (otpm_key, completion_tokens - otpm_reserved)) + return tuple( + (key if delta < 0 else _key_for_this_minute(key), delta) for key, delta in deltas if key is not None and delta + ) + + def _stash_reservation_in_metadata( request_kwargs: dict[str, object] | None, *, @@ -311,7 +376,9 @@ def _reservation_channels(kwargs: Mapping[str, object] | None) -> tuple[object, return () top_level: Final = (kwargs.get("metadata"), kwargs.get("litellm_metadata")) litellm_params: Final = _as_mutable_mapping(kwargs.get("litellm_params")) - from_params: Final = () if litellm_params is None else (litellm_params.get("metadata"),) + from_params: Final = ( + () if litellm_params is None else (litellm_params.get("metadata"), litellm_params.get("litellm_metadata")) + ) standard_logging_object: Final = _as_mutable_mapping(kwargs.get("standard_logging_object")) from_logging_object: Final = () if standard_logging_object is None else (standard_logging_object.get("metadata"),) return top_level + from_params + from_logging_object @@ -433,9 +500,11 @@ def io_token_pre_call_check( return deployment request_kwargs: Final = get_io_token_rate_limit_request_kwargs() - _model: Final = _tokenizer_model(deployment) + _model, served_model = _tokenizer_and_served_models(deployment) estimated_input: Final = _estimate_input_tokens(request_kwargs, model=_model) - max_tokens: Final = _resolve_max_tokens(request_kwargs, deployment) + otpm_to_reserve: Final = ( + _otpm_reservation(request_kwargs, deployment, served_model, otpm_limit) if otpm_limit is not None else 0 + ) dt: Final = get_utc_datetime() current_minute: Final = dt.strftime("%H-%M") @@ -458,7 +527,7 @@ def io_token_pre_call_check( ) if otpm_limit is not None: - otpm_reserved = 0 if max_tokens == 0 else _reservation_value(max_tokens, otpm_limit) + otpm_reserved = otpm_to_reserve try: _sync_increment_with_rollback( dual_cache, @@ -497,9 +566,11 @@ async def async_io_token_pre_call_check( return deployment request_kwargs: Final = get_io_token_rate_limit_request_kwargs() - _model: Final = _tokenizer_model(deployment) + _model, served_model = _tokenizer_and_served_models(deployment) estimated_input: Final = await offload_token_count(_estimate_input_tokens)(request_kwargs, model=_model) - max_tokens: Final = _resolve_max_tokens(request_kwargs, deployment) + otpm_to_reserve: Final = ( + _otpm_reservation(request_kwargs, deployment, served_model, otpm_limit) if otpm_limit is not None else 0 + ) dt: Final = get_utc_datetime() current_minute: Final = dt.strftime("%H-%M") @@ -523,7 +594,7 @@ async def async_io_token_pre_call_check( ) if otpm_limit is not None: - otpm_reserved = 0 if max_tokens == 0 else _reservation_value(max_tokens, otpm_limit) + otpm_reserved = otpm_to_reserve try: await _increment_with_rollback( dual_cache, @@ -564,7 +635,8 @@ def io_token_reconcile_success( ) -> None: request_kwargs: Final[Mapping[str, object] | None] = kwargs response: Final[object] = response_obj - itpm_reserved, otpm_reserved, itpm_key, otpm_key = _read_reservation_from_kwargs(request_kwargs) + reservation: Final = _read_reservation_from_kwargs(request_kwargs) + itpm_reserved, otpm_reserved, itpm_key, otpm_key = reservation if itpm_key is None and otpm_key is None: return @@ -572,23 +644,8 @@ def io_token_reconcile_success( try: if usage_resolved: - if itpm_key is not None: - itpm_delta: Final = billable_input - itpm_reserved - if itpm_delta != 0: - dual_cache.increment_cache( - key=itpm_key, - value=itpm_delta, - ttl=RoutingArgsTTL, - ) - - if otpm_key is not None: - otpm_delta: Final = completion_tokens - otpm_reserved - if otpm_delta != 0: - dual_cache.increment_cache( - key=otpm_key, - value=otpm_delta, - ttl=RoutingArgsTTL, - ) + for key, delta in _reconcile_increments(reservation, billable_input, completion_tokens): + dual_cache.increment_cache(key=key, value=delta, ttl=RoutingArgsTTL) else: verbose_router_logger.debug( "[IO TOKEN LIMIT] usage missing; keeping reservation (itpm_reserved=%s, otpm_reserved=%s)", @@ -618,39 +675,23 @@ async def async_io_token_reconcile_success( ) -> None: request_kwargs: Final[Mapping[str, object] | None] = kwargs response: Final[object] = response_obj - itpm_reserved, otpm_reserved, itpm_key, otpm_key = _read_reservation_from_kwargs(request_kwargs) + reservation: Final = _read_reservation_from_kwargs(request_kwargs) + itpm_reserved, otpm_reserved, itpm_key, otpm_key = reservation if itpm_key is None and otpm_key is None: return billable_input, completion_tokens, usage_resolved = _resolve_reconcile_usage_tokens(request_kwargs, response) - # Reconcile against the exact key that held the reservation (which encodes - # the reservation's minute), not a key recomputed at response time. This - # tracks actual usage even when the pre-call estimate was 0, and avoids a - # minute-boundary mismatch for calls that span into the next minute. Always - # clear the stash afterwards (even if an increment throws) so a retry or a - # duplicate success event can't re-process it. + # A refund goes back to the minute that held the reservation, while usage above the reservation is charged + # to the minute the response finished in, so a call that runs past its minute still counts against a live + # window. Always clear the stash afterwards (even if an increment throws) so a retry or a duplicate success + # event can't re-process it. try: if usage_resolved: - if itpm_key is not None: - itpm_delta: Final = billable_input - itpm_reserved - if itpm_delta != 0: - await dual_cache.async_increment_cache( - key=itpm_key, - value=itpm_delta, - ttl=RoutingArgsTTL, - parent_otel_span=parent_otel_span, - ) - - if otpm_key is not None: - otpm_delta: Final = completion_tokens - otpm_reserved - if otpm_delta != 0: - await dual_cache.async_increment_cache( - key=otpm_key, - value=otpm_delta, - ttl=RoutingArgsTTL, - parent_otel_span=parent_otel_span, - ) + for key, delta in _reconcile_increments(reservation, billable_input, completion_tokens): + await dual_cache.async_increment_cache( + key=key, value=delta, ttl=RoutingArgsTTL, parent_otel_span=parent_otel_span + ) else: verbose_router_logger.debug( "[IO TOKEN LIMIT] usage missing; keeping reservation (itpm_reserved=%s, otpm_reserved=%s)", diff --git a/litellm/router_utils/pre_call_checks/model_rate_limit_check.py b/litellm/router_utils/pre_call_checks/model_rate_limit_check.py index f4b41211c6e..f41d19cd62c 100644 --- a/litellm/router_utils/pre_call_checks/model_rate_limit_check.py +++ b/litellm/router_utils/pre_call_checks/model_rate_limit_check.py @@ -26,7 +26,6 @@ from litellm.caching.dual_cache import DualCache from litellm.caching.redis_cache import RedisCircuitBreakerOpenError from litellm.integrations.custom_logger import CustomLogger from litellm.router_utils.pre_call_checks.io_token_rate_limit_check import ( - ITPM_RESERVED_KEY, async_io_token_pre_call_check, async_io_token_reconcile_success, async_io_token_refund_failure, @@ -485,24 +484,12 @@ class ModelRateLimitingCheck(CustomLogger): try: standard_logging_object: Final[StandardLoggingPayload | None] = kwargs.get("standard_logging_object") - - # IO token reconciliation works purely from the cache keys stashed in - # kwargs/metadata, so it must run before the model_id guard below - # (which only the TPM-tracking path needs). Otherwise a request whose - # standard_logging_object lacks model_id would never return its - # reservation, leaving the counter elevated until the TTL expires. - slo_metadata: Final = (standard_logging_object.get("metadata") or {}) if standard_logging_object else {} - kwargs_metadata: Final = kwargs.get("metadata") or {} - if ITPM_RESERVED_KEY in slo_metadata or ITPM_RESERVED_KEY in kwargs_metadata: - await async_io_token_reconcile_success( - self.dual_cache, - kwargs, - response_obj, - parent_otel_span=get_parent_otel_span_from_kwargs(kwargs), - ) - # Fall through: a deployment can also configure tpm/rpm alongside - # itpm/otpm, and that path's pre-call check reads the tpm_key - # counter tracked below, so it must still be incremented here. + await async_io_token_reconcile_success( + self.dual_cache, + kwargs, + response_obj, + parent_otel_span=get_parent_otel_span_from_kwargs(kwargs), + ) if standard_logging_object is None: return @@ -557,17 +544,7 @@ class ModelRateLimitingCheck(CustomLogger): """ try: standard_logging_object: Final[StandardLoggingPayload | None] = kwargs.get("standard_logging_object") - slo_metadata: Final = (standard_logging_object.get("metadata") or {}) if standard_logging_object else {} - kwargs_metadata: Final = kwargs.get("metadata") or {} - if ITPM_RESERVED_KEY in slo_metadata or ITPM_RESERVED_KEY in kwargs_metadata: - io_token_reconcile_success( - self.dual_cache, - kwargs, - response_obj, - ) - # Fall through: a deployment can also configure tpm/rpm alongside - # itpm/otpm, and that path's pre-call check reads the tpm_key - # counter tracked below, so it must still be incremented here. + io_token_reconcile_success(self.dual_cache, kwargs, response_obj) if standard_logging_object is None: return diff --git a/tests/integration/routing/test_otpm_default_reservation_fraction.py b/tests/integration/routing/test_otpm_default_reservation_fraction.py new file mode 100644 index 00000000000..589a2535a03 --- /dev/null +++ b/tests/integration/routing/test_otpm_default_reservation_fraction.py @@ -0,0 +1,137 @@ +from __future__ import annotations + +import dataclasses +import threading +import uuid +from collections.abc import Callable, Iterator +from concurrent.futures import ThreadPoolExecutor +from dataclasses import dataclass +from datetime import UTC, datetime +from pathlib import Path +from queue import SimpleQueue +from typing import Final + +import httpx +import pytest +from integration._support.client import JSON_OBJECT, gateway_from_environment +from integration._support.openai_wire import chat_reply +from integration._support.process import OwnedProxy, graceful_stop_seconds, owned_proxy_process +from integration._support.wire import Reply, Request, Wire, wire_server +from integration.providers._cache_control_marks_support import owned_config +from integration.routing.test_payment_required_mapping import OPENAI_MODEL, is_model_info_probe, model_list_reply + +pytestmark: Final = pytest.mark.timeout(2 * graceful_stop_seconds() + 120) + +GROUPS: Final = ("otpm-fraction-first", "otpm-fraction-retry") +OTPM_LIMIT: Final = 1000 +FRACTION: Final = 0.5 +DEFAULT_RESERVATION: Final = int(OTPM_LIMIT * FRACTION) + + +@dataclass(frozen=True, slots=True) +class Rig: + owned: OwnedProxy + upstream: Wire + gates: SimpleQueue[threading.Event] + + +@dataclass(frozen=True, slots=True) +class Admissions: + held_forwarded_max_tokens: bool + statuses: tuple[int, int, int] + one_minute: bool + + +def _reply_for(gates: SimpleQueue[threading.Event]) -> Callable[[Request], Reply]: + def reply(request: Request) -> Reply: + if is_model_info_probe(request): + return model_list_reply() + if JSON_OBJECT.validate_json(request.body).get("stream") is True: + return dataclasses.replace( + chat_reply("chatcmpl-held", OPENAI_MODEL, "held", stream=True), gate_before_first=gates.get(timeout=60) + ) + return chat_reply("chatcmpl-capped", OPENAI_MODEL, "capped", stream=False) + + return reply + + +def _config(directory: Path, upstream: Wire) -> Path: + return owned_config( + directory, + [ + { + "model_name": group, + "litellm_params": { + "model": OPENAI_MODEL, + "api_base": upstream.url + "/v1", + "api_key": "synthetic-openai-key", + "otpm": OTPM_LIMIT, + }, + "model_info": {"id": group}, + } + for group in GROUPS + ], + litellm_settings={"cache": False}, + router_settings={"optional_pre_call_checks": ["enforce_model_rate_limits"], "num_retries": 0}, + ) + + +@pytest.fixture(scope="module") +def rig(tmp_path_factory: pytest.TempPathFactory) -> Iterator[Rig]: + directory: Final = tmp_path_factory.mktemp("otpm-fraction") + gates: Final[SimpleQueue[threading.Event]] = SimpleQueue() + with ( + gateway_from_environment() as gateway, + wire_server(_reply_for(gates)) as upstream, + owned_proxy_process( + gateway, + directory, + {"NO_MAX_TOKENS_TPM_RESERVATION_FRACTION": str(FRACTION)}, + config=_config(directory, upstream), + ) as owned, + ): + yield Rig(owned, upstream, gates) + + +def _chat(rig: Rig, group: str, **params: int | bool) -> httpx.Response: + return rig.owned.gateway.request( + "POST", + "/v1/chat/completions", + {"model": group, "messages": [{"role": "user", "content": f"fraction {uuid.uuid4().hex}"}], **params}, + ) + + +def _await_forwarded_chat(upstream: Wire) -> Request: + request: Final = upstream.received.get(timeout=60) + return _await_forwarded_chat(upstream) if is_model_info_probe(request) else request + + +def _utc_minute() -> str: + return datetime.now(UTC).strftime("%Y-%m-%d %H:%M") + + +def _admissions_beside_a_held_stream(rig: Rig, group: str) -> Admissions: + release: Final = threading.Event() + rig.gates.put(release) + started: Final = _utc_minute() + with ThreadPoolExecutor(max_workers=1) as pool: + held: Final = pool.submit(_chat, rig, group, stream=True) + try: + forwarded: Final = JSON_OBJECT.validate_json(_await_forwarded_chat(rig.upstream).body) + over: Final = _chat(rig, group, max_tokens=OTPM_LIMIT - DEFAULT_RESERVATION + 1) + fits: Final = _chat(rig, group, max_tokens=OTPM_LIMIT - DEFAULT_RESERVATION) + ended: Final = _utc_minute() + finally: + release.set() + return Admissions( + held_forwarded_max_tokens="max_tokens" in forwarded, + statuses=(over.status_code, fits.status_code, held.result(timeout=60).status_code), + one_minute=started == ended, + ) + + +def test_reservation_fraction_env_var_sets_what_a_request_without_max_tokens_holds(rig: Rig) -> None: + first: Final = _admissions_beside_a_held_stream(rig, GROUPS[0]) + admissions: Final = first if first.one_minute else _admissions_beside_a_held_stream(rig, GROUPS[1]) + + assert admissions == Admissions(held_forwarded_max_tokens=False, statuses=(429, 200, 200), one_minute=True) diff --git a/tests/unit/litellm_core_utils/test_env_utils.py b/tests/unit/litellm_core_utils/test_env_utils.py index 1ddd172e274..89b091e47c2 100644 --- a/tests/unit/litellm_core_utils/test_env_utils.py +++ b/tests/unit/litellm_core_utils/test_env_utils.py @@ -6,6 +6,7 @@ import pytest from litellm._logging import verbose_logger from litellm.litellm_core_utils.env_utils import ( get_env_float, + get_env_float_in_range, get_env_float_or_none, get_env_int, get_env_int_or_none, @@ -94,3 +95,12 @@ def test_unparsable_stdout_log_limit_does_not_break_logging( assert "first line" in messages assert "second line" in messages assert len([message for message in messages if raw in message]) == 1 + + +@pytest.mark.parametrize(("raw", "expected"), [("0.5", 0.5), ("1", 1.0), ("4", 0.25), ("0", 0.25), ("half", 0.25)]) +def test_float_in_range_keeps_values_inside_the_range_and_falls_back_otherwise( + monkeypatch: pytest.MonkeyPatch, raw: str, expected: float +) -> None: + monkeypatch.setenv("ENV_UTILS_TEST_FRACTION", raw) + + assert get_env_float_in_range("ENV_UTILS_TEST_FRACTION", 0.25, 0.01, 1.0) == expected diff --git a/tests/unit/test_constants.py b/tests/unit/test_constants.py index 99b049f1803..d9c34e95364 100644 --- a/tests/unit/test_constants.py +++ b/tests/unit/test_constants.py @@ -19,7 +19,14 @@ from unittest.mock import patch _ENV_HELPERS = frozenset( - {"get_env_int", "get_env_float", "get_env_int_in_range", "get_env_int_or_none", "get_env_float_or_none"} + { + "get_env_int", + "get_env_float", + "get_env_int_in_range", + "get_env_float_in_range", + "get_env_int_or_none", + "get_env_float_or_none", + } ) diff --git a/tests/unit/test_router/test_io_token_rate_limits.py b/tests/unit/test_router/test_io_token_rate_limits.py index 05921aa70b3..e5ff4531f84 100644 --- a/tests/unit/test_router/test_io_token_rate_limits.py +++ b/tests/unit/test_router/test_io_token_rate_limits.py @@ -3,6 +3,8 @@ Tests for separate ITPM/OTPM deployment rate limits (enforce_model_rate_limits). """ import asyncio +from collections.abc import Awaitable, Callable, Mapping +from datetime import datetime from typing import Final import pytest @@ -10,13 +12,16 @@ import pytest import litellm from litellm import Router from litellm.caching.dual_cache import DualCache +from litellm.constants import NO_MAX_TOKENS_TPM_RESERVATION_FRACTION +from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObject +from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from litellm.router_utils.pre_call_checks.io_token_rate_limit_check import ( ITPM_CACHE_KEY, ITPM_RESERVED_KEY, OTPM_CACHE_KEY, OTPM_RESERVED_KEY, + _explicit_max_tokens, _reservation_value, - _resolve_max_tokens, async_io_token_pre_call_check, async_io_token_reconcile_success, build_io_token_rate_limit_headers, @@ -30,7 +35,16 @@ from litellm.router_utils.pre_call_checks.io_token_rate_limit_check import ( from litellm.router_utils.pre_call_checks.model_rate_limit_check import ( ModelRateLimitingCheck, ) +from litellm.types.router import RouterCacheEnum from litellm.types.utils import ModelResponse, Usage +from litellm.utils import get_utc_datetime + +MANTLE_OPUS = "bedrock_mantle/anthropic.claude-opus-4-7" +MANTLE_OPUS_MAX_OUTPUT = litellm.get_model_info(model=MANTLE_OPUS)["max_output_tokens"] +UNMAPPED_MODEL = "openai/litellm-io-unmapped-model" +UNMAPPED_MODEL_DEFAULT_MAX_OUTPUT = 4096 +AZURE_BASE_MODEL = "azure/gpt-5" +AZURE_BASE_MODEL_MAX_OUTPUT = litellm.get_model_info(model=AZURE_BASE_MODEL)["max_output_tokens"] class TestIOTokenRateLimitHelpers: @@ -48,13 +62,23 @@ class TestIOTokenRateLimitHelpers: # A real non-zero estimate is reserved as-is. assert _reservation_value(42, 100) == 42 - def test_resolve_max_tokens_respects_explicit_zero(self): - deployment = {"litellm_params": {"model": "openai/gpt-4o-mini"}} - # An explicit max_tokens=0 is honored, not replaced by the model default. - assert _resolve_max_tokens({"max_tokens": 0}, deployment) == 0 - # max_completion_tokens is the fallback only when max_tokens is absent. - assert _resolve_max_tokens({"max_completion_tokens": 12}, deployment) == 12 - assert _resolve_max_tokens({"max_output_tokens": 9}, deployment) == 9 + @pytest.mark.parametrize( + "request_kwargs, expected", + [ + ({"max_tokens": 0}, 0), + ({"max_completion_tokens": 12}, 12), + ({"max_output_tokens": 9}, 9), + ({"max_tokens": 5, "max_completion_tokens": 12}, 5), + ({"max_tokens": None, "max_output_tokens": 9}, 9), + ({"max_tokens": [100]}, None), + ({"messages": []}, None), + (None, None), + ], + ) + def test_explicit_max_tokens_reads_every_output_cap_key( + self, request_kwargs: dict[str, object] | None, expected: int | None + ): + assert _explicit_max_tokens(request_kwargs) == expected def test_build_io_token_rate_limit_headers(self): headers = build_io_token_rate_limit_headers( @@ -386,6 +410,244 @@ class TestModelRateLimitingCheckIOTokens: normal_output_otpm_key = normal_output_kwargs["metadata"][OTPM_CACHE_KEY] assert await dual_cache.async_get_cache(key=normal_output_otpm_key) == 5 + @pytest.mark.asyncio + @pytest.mark.parametrize("use_sync_check", [False, True]) + @pytest.mark.parametrize( + "model, otpm_limit", + [ + (MANTLE_OPUS, MANTLE_OPUS_MAX_OUTPUT // 2), + (MANTLE_OPUS, MANTLE_OPUS_MAX_OUTPUT - 1), + (UNMAPPED_MODEL, 1000), + (UNMAPPED_MODEL, 3), + ], + ) + async def test_request_without_max_tokens_is_admitted_below_model_max_output( + self, model: str, otpm_limit: int, use_sync_check: bool + ): + dual_cache: Final = DualCache() + check: Final = ModelRateLimitingCheck(dual_cache=dual_cache) + deployment: Final = { + "litellm_params": {"model": model, "otpm": otpm_limit}, + "model_info": {"id": f"io-no-max-tokens-{otpm_limit}"}, + "model_name": "no-max-tokens", + } + request_kwargs: Final = {"messages": [{"role": "user", "content": "hello"}], "metadata": {}} + set_io_token_rate_limit_request_kwargs(request_kwargs) + + if use_sync_check: + check.pre_call_check(deployment) + else: + await check.async_pre_call_check(deployment) + + reserved: Final = request_kwargs["metadata"][OTPM_RESERVED_KEY] + assert isinstance(reserved, int) + assert 0 < reserved < otpm_limit + assert await dual_cache.async_get_cache(key=request_kwargs["metadata"][OTPM_CACHE_KEY]) == reserved + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "model, otpm_limit", + [(MANTLE_OPUS, MANTLE_OPUS_MAX_OUTPUT), (UNMAPPED_MODEL, UNMAPPED_MODEL_DEFAULT_MAX_OUTPUT)], + ) + async def test_concurrent_requests_without_max_tokens_share_otpm_without_overshoot( + self, model: str, otpm_limit: int + ): + dual_cache: Final = DualCache() + check: Final = ModelRateLimitingCheck(dual_cache=dual_cache) + deployment: Final = { + "litellm_params": {"model": model, "otpm": otpm_limit}, + "model_info": {"id": "io-no-max-tokens-race-id"}, + "model_name": "no-max-tokens", + } + + async def _attempt() -> dict[str, object] | None: + request_kwargs: Final = {"messages": [{"role": "user", "content": "hi"}], "metadata": {}} + set_io_token_rate_limit_request_kwargs(request_kwargs) + try: + await check.async_pre_call_check(deployment) + except litellm.RateLimitError: + return None + return request_kwargs["metadata"] + + admitted: Final = [m for m in await asyncio.gather(*[_attempt() for _ in range(8)]) if m is not None] + otpm_key: Final = admitted[0][OTPM_CACHE_KEY] + + assert len(admitted) == round(1 / NO_MAX_TOKENS_TPM_RESERVATION_FRACTION) + assert {m[OTPM_CACHE_KEY] for m in admitted} == {otpm_key} + assert await dual_cache.async_get_cache(key=otpm_key) == sum(m[OTPM_RESERVED_KEY] for m in admitted) + assert await dual_cache.async_get_cache(key=otpm_key) <= otpm_limit + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "model, model_max_output", + [(MANTLE_OPUS, MANTLE_OPUS_MAX_OUTPUT), (UNMAPPED_MODEL, UNMAPPED_MODEL_DEFAULT_MAX_OUTPUT)], + ) + async def test_request_without_max_tokens_reserves_model_max_output_when_otpm_allows( + self, model: str, model_max_output: int + ): + dual_cache: Final = DualCache() + check: Final = ModelRateLimitingCheck(dual_cache=dual_cache) + deployment: Final = { + "litellm_params": {"model": model, "otpm": model_max_output * 10}, + "model_info": {"id": "io-no-max-tokens-roomy-id"}, + "model_name": "roomy", + } + request_kwargs: Final = {"messages": [{"role": "user", "content": "hello"}], "metadata": {}} + set_io_token_rate_limit_request_kwargs(request_kwargs) + + await check.async_pre_call_check(deployment) + + assert request_kwargs["metadata"][OTPM_RESERVED_KEY] == model_max_output + + @pytest.mark.asyncio + @pytest.mark.parametrize("use_sync_check", [False, True]) + async def test_base_model_does_not_change_the_default_otpm_reservation(self, use_sync_check: bool): + check: Final = ModelRateLimitingCheck(dual_cache=DualCache()) + otpm_limit: Final = AZURE_BASE_MODEL_MAX_OUTPUT * 10 + + async def _reserved(model_info: dict[str, str]) -> int: + request_kwargs: Final = {"messages": [{"role": "user", "content": "hello"}], "metadata": {}} + set_io_token_rate_limit_request_kwargs(request_kwargs) + deployment: Final = { + "litellm_params": {"model": "azure/litellm-io-custom-deployment", "otpm": otpm_limit}, + "model_info": model_info, + "model_name": "custom-name", + } + if use_sync_check: + check.pre_call_check(deployment) + else: + await check.async_pre_call_check(deployment) + return request_kwargs["metadata"][OTPM_RESERVED_KEY] + + with_base_model: Final = await _reserved({"id": "io-with-base-model", "base_model": AZURE_BASE_MODEL}) + without_base_model: Final = await _reserved({"id": "io-without-base-model"}) + + assert with_base_model == without_base_model + + @pytest.mark.asyncio + @pytest.mark.parametrize("use_sync_check", [False, True]) + async def test_unparseable_max_tokens_leaves_itpm_uncharged(self, use_sync_check: bool): + dual_cache: Final = DualCache() + check: Final = ModelRateLimitingCheck(dual_cache=dual_cache) + deployment: Final = { + "litellm_params": {"model": MANTLE_OPUS, "itpm": 2000, "otpm": 2000}, + "model_info": {"id": "io-bad-max-tokens-id"}, + "model_name": "opus", + } + + async def _pre_call(request_kwargs: dict[str, object]) -> None: + set_io_token_rate_limit_request_kwargs(request_kwargs) + if use_sync_check: + check.pre_call_check(deployment) + else: + await check.async_pre_call_check(deployment) + + await _pre_call({"messages": [{"role": "user", "content": "hello there"}], "max_tokens": "abc", "metadata": {}}) + valid_kwargs: Final = { + "messages": [{"role": "user", "content": "hello there"}], + "max_tokens": 1, + "metadata": {}, + } + await _pre_call(valid_kwargs) + + itpm_counter: Final = await dual_cache.async_get_cache(key=valid_kwargs["metadata"][ITPM_CACHE_KEY]) + assert itpm_counter == valid_kwargs["metadata"][ITPM_RESERVED_KEY] + + @pytest.mark.asyncio + async def test_explicit_max_tokens_above_default_share_is_reserved_in_full(self): + dual_cache: Final = DualCache() + check: Final = ModelRateLimitingCheck(dual_cache=dual_cache) + deployment: Final = { + "litellm_params": {"model": MANTLE_OPUS, "otpm": 100}, + "model_info": {"id": "io-explicit-max-tokens-id"}, + "model_name": "opus", + } + request_kwargs: Final = {"messages": [{"role": "user", "content": "hello"}], "max_tokens": 100, "metadata": {}} + set_io_token_rate_limit_request_kwargs(request_kwargs) + + await check.async_pre_call_check(deployment) + + assert request_kwargs["metadata"][OTPM_RESERVED_KEY] == 100 + set_io_token_rate_limit_request_kwargs({"messages": [{"role": "user", "content": "hi"}], "metadata": {}}) + with pytest.raises(litellm.RateLimitError): + await check.async_pre_call_check(deployment) + + @pytest.mark.asyncio + @pytest.mark.parametrize("use_sync_check", [False, True]) + async def test_deployment_max_tokens_sets_the_reservation_unless_the_request_sets_one(self, use_sync_check: bool): + check: Final = ModelRateLimitingCheck(dual_cache=DualCache()) + deployment: Final = { + "litellm_params": {"model": MANTLE_OPUS, "otpm": 4000, "max_tokens": 100}, + "model_info": {"id": "io-deployment-max-tokens-id"}, + "model_name": "opus", + } + without_cap: Final = {"messages": [{"role": "user", "content": "hello"}], "metadata": {}} + with_cap: Final = {"messages": [{"role": "user", "content": "hello"}], "max_tokens": 300, "metadata": {}} + + for request_kwargs in (without_cap, with_cap): + set_io_token_rate_limit_request_kwargs(request_kwargs) + if use_sync_check: + check.pre_call_check(deployment) + else: + await check.async_pre_call_check(deployment) + + assert without_cap["metadata"][OTPM_RESERVED_KEY] == 100 + assert with_cap["metadata"][OTPM_RESERVED_KEY] == 300 + + @pytest.mark.asyncio + async def test_deployment_cap_sets_the_reservation_only_through_the_param_its_api_sends(self): + check: Final = ModelRateLimitingCheck(dual_cache=DualCache()) + chat: Final = {"messages": [{"role": "user", "content": "hello"}]} + responses: Final = {"input": "hello"} + + async def reserved(deployment_cap: Mapping[str, int], request_shape: Mapping[str, object]) -> int: + request_kwargs: Final = {**request_shape, "metadata": {}} + set_io_token_rate_limit_request_kwargs(request_kwargs) + await check.async_pre_call_check( + { + "litellm_params": {"model": MANTLE_OPUS, "otpm": 4000, **deployment_cap}, + "model_info": {"id": f"io-sent-cap-{sorted(deployment_cap)}-{sorted(request_shape)}"}, + "model_name": "opus", + } + ) + return request_kwargs["metadata"][OTPM_RESERVED_KEY] + + chat_default: Final = await reserved({}, chat) + responses_default: Final = await reserved({}, responses) + + assert chat_default == responses_default > 100 + assert await reserved({"max_tokens": 100}, chat) == 100 + assert await reserved({"max_completion_tokens": 100}, chat) == 100 + assert await reserved({"max_output_tokens": 100}, chat) == chat_default + assert await reserved({"max_output_tokens": 100}, responses) == 100 + assert await reserved({"max_tokens": 100}, responses) == responses_default + assert await reserved({"max_completion_tokens": 100}, responses) == responses_default + + @pytest.mark.asyncio + @pytest.mark.parametrize("use_sync_check", [False, True]) + async def test_deployment_cap_above_the_default_reserves_the_default(self, use_sync_check: bool): + check: Final = ModelRateLimitingCheck(dual_cache=DualCache()) + otpm_limit: Final = UNMAPPED_MODEL_DEFAULT_MAX_OUTPUT * 5 + + async def reserved(deployment_cap: Mapping[str, int]) -> int: + request_kwargs: Final = {"messages": [{"role": "user", "content": "hello"}], "metadata": {}} + set_io_token_rate_limit_request_kwargs(request_kwargs) + deployment: Final = { + "litellm_params": {"model": UNMAPPED_MODEL, "otpm": otpm_limit, **deployment_cap}, + "model_info": {"id": f"io-deployment-cap-above-default-{deployment_cap}"}, + "model_name": "unmapped", + } + if use_sync_check: + check.pre_call_check(deployment) + else: + await check.async_pre_call_check(deployment) + return request_kwargs["metadata"][OTPM_RESERVED_KEY] + + default: Final = await reserved({}) + + assert await reserved({"max_tokens": default + 1}) == default + assert await reserved({"max_tokens": otpm_limit * 2}) == default + def test_sync_io_pre_call_reserves_and_reconciles(self): from litellm.utils import get_utc_datetime @@ -415,9 +677,9 @@ class TestModelRateLimitingCheckIOTokens: "standard_logging_object": { "model_id": "io-sync-id", "hidden_params": {"litellm_model_name": "bedrock_mantle/anthropic.claude-opus-4-7"}, - "metadata": dict(request_kwargs["metadata"]), + "metadata": {}, }, - "metadata": request_kwargs["metadata"], + "litellm_params": {"metadata": request_kwargs["metadata"]}, } response = ModelResponse( choices=[{"message": {"role": "assistant", "content": "hi"}, "index": 0, "finish_reason": "stop"}], @@ -709,6 +971,32 @@ class TestModelRateLimitingCheckIOTokens: assert await dual_cache.async_get_cache(key=itpm_key) == 4 + @pytest.mark.asyncio + @pytest.mark.parametrize("use_sync_reconcile", [False, True]) + async def test_reconcile_charges_usage_above_the_reservation_to_the_minute_it_finished_in( + self, use_sync_reconcile: bool + ): + dual_cache: Final = DualCache() + counter: Final = "global_router:io-late-id:bedrock_mantle/test:otpm" + reserved_minute_key: Final = f"{counter}:99-99" + await dual_cache.async_increment_cache(key=reserved_minute_key, value=10, ttl=60) + kwargs: Final = {"metadata": {ITPM_RESERVED_KEY: 0, OTPM_RESERVED_KEY: 10, OTPM_CACHE_KEY: reserved_minute_key}} + response: Final = ModelResponse( + choices=[{"message": {"role": "assistant", "content": "hi"}, "index": 0, "finish_reason": "stop"}], + usage=Usage(prompt_tokens=4, completion_tokens=25, total_tokens=29), + ) + minute_before: Final = get_utc_datetime().strftime("%H-%M") + + if use_sync_reconcile: + io_token_reconcile_success(dual_cache, kwargs, response) + else: + await async_io_token_reconcile_success(dual_cache, kwargs, response) + + finished_minutes: Final = {minute_before, get_utc_datetime().strftime("%H-%M")} + charged: Final = [await dual_cache.async_get_cache(key=f"{counter}:{minute}") for minute in finished_minutes] + assert await dual_cache.async_get_cache(key=reserved_minute_key) == 10 + assert sum(value or 0 for value in charged) == 25 - 10 + @pytest.mark.asyncio async def test_reconcile_missing_usage_keeps_reservation(self): dual_cache = DualCache() @@ -799,16 +1087,16 @@ class TestModelRateLimitingCheckIOTokens: assert await dual_cache.async_get_cache(key=itpm_key) == 4 - def test_sync_reconcile_anthropic_dict_usage(self): + def test_sync_reconcile_anthropic_dict_usage_counts_uncached_input_once(self): dual_cache = DualCache() itpm_key = "global_router:io-anthropic:bedrock_mantle/test:itpm:00-00" otpm_key = "global_router:io-anthropic:bedrock_mantle/test:otpm:00-00" - dual_cache.set_cache(key=itpm_key, value=6, ttl=60) + dual_cache.set_cache(key=itpm_key, value=10, ttl=60) dual_cache.set_cache(key=otpm_key, value=4, ttl=60) kwargs = { "metadata": { - ITPM_RESERVED_KEY: 6, + ITPM_RESERVED_KEY: 10, OTPM_RESERVED_KEY: 4, ITPM_CACHE_KEY: itpm_key, OTPM_CACHE_KEY: otpm_key, @@ -816,12 +1104,17 @@ class TestModelRateLimitingCheckIOTokens: } response = { "type": "message", - "usage": {"input_tokens": 3, "output_tokens": 2, "cache_read_input_tokens": 1}, + "usage": { + "input_tokens": 3, + "output_tokens": 2, + "cache_read_input_tokens": 1, + "cache_creation_input_tokens": 4, + }, } io_token_reconcile_success(dual_cache, kwargs, response) - assert dual_cache.get_cache(key=itpm_key) == 2 + assert dual_cache.get_cache(key=itpm_key) == 3 + 4 assert dual_cache.get_cache(key=otpm_key) == 2 @pytest.mark.asyncio @@ -958,7 +1251,214 @@ class TestModelRateLimitingCheckIOTokens: assert current == 0 +ROUTER_IO_MODEL: Final = "anthropic/claude-haiku-5-5" +RouterCall = Callable[[Router, str | Exception], Awaitable[tuple[int, int]]] + + +def _router_with_io_limits() -> Router: + return Router( + model_list=[ + { + "model_name": "limited", + "litellm_params": {"model": ROUTER_IO_MODEL, "api_key": "sk-test", "itpm": 50000, "otpm": 50000}, + } + ], + optional_pre_call_checks=["enforce_model_rate_limits"], + num_retries=1, + ) + + +def _current_minute() -> str: + return get_utc_datetime().strftime("%H-%M") + + +def _counter(router: Router, template: RouterCacheEnum, minutes: frozenset[str]) -> int: + deployment: Final = router.model_list[0] + return sum( + router.cache.get_cache( + key=template.value.format( + id=deployment["model_info"]["id"], + model=deployment["litellm_params"]["model"], + current_minute=minute, + ) + ) + or 0 + for minute in minutes + ) + + +async def _io_counters_since(router: Router, first_minute: str) -> tuple[int, int]: + await asyncio.sleep(0) + await asyncio.wait_for(GLOBAL_LOGGING_WORKER.flush(), timeout=10) + minutes: Final = frozenset({first_minute, _current_minute()}) + return _counter(router, RouterCacheEnum.ITPM, minutes), _counter(router, RouterCacheEnum.OTPM, minutes) + + +async def _sdk_chat(router: Router, mock_response: str | Exception) -> tuple[int, int]: + response: Final = await router.acompletion( + model="limited", + messages=[{"role": "user", "content": "hello there"}], + max_tokens=100, + mock_response=mock_response, + ) + return response.usage.prompt_tokens, response.usage.completion_tokens + + +async def _proxy_chat(router: Router, mock_response: str | Exception) -> tuple[int, int]: + response: Final = await router.acompletion( + model="limited", + messages=[{"role": "user", "content": "hello there"}], + max_tokens=100, + metadata={"user_api_key_hash": "hashed-key"}, + mock_response=mock_response, + ) + return response.usage.prompt_tokens, response.usage.completion_tokens + + +async def _proxy_responses(router: Router, mock_response: str | Exception) -> tuple[int, int]: + response: Final = await router.aresponses( + model="limited", + input="hello there", + max_output_tokens=100, + litellm_metadata={"user_api_key_hash": "hashed-key"}, + mock_response=mock_response, + ) + return response.usage.input_tokens, response.usage.output_tokens + + +async def _proxy_messages(router: Router, mock_response: str | Exception) -> tuple[int, int]: + response: Final = await router.aanthropic_messages( + model="limited", + messages=[{"role": "user", "content": "hello there"}], + max_tokens=100, + litellm_metadata={"user_api_key_hash": "hashed-key"}, + mock_response=mock_response, + ) + return response["usage"]["input_tokens"], response["usage"]["output_tokens"] + + +ProxyCallSharingLogging = Callable[[Router, str | Exception, LiteLLMLoggingObject], Awaitable[object]] +PROXY_MESSAGES: Final = [{"role": "user", "content": "hello there"}] + + +def _logging_obj_shared_across_attempts(original_function: str) -> LiteLLMLoggingObject: + logging_obj, _ = litellm.utils.function_setup( + original_function, + litellm.utils.Rules(), + datetime.now(), + model="limited", + messages=PROXY_MESSAGES, + litellm_call_id="proxy-call", + ) + return logging_obj + + +async def _proxy_chat_sharing_logging( + router: Router, mock_error: str | Exception, logging_obj: LiteLLMLoggingObject +) -> object: + return await router.acompletion( + model="limited", + messages=PROXY_MESSAGES, + max_tokens=100, + metadata={"user_api_key_hash": "hashed-key"}, + mock_response=mock_error, + litellm_call_id="proxy-call", + litellm_logging_obj=logging_obj, + ) + + +async def _proxy_responses_sharing_logging( + router: Router, mock_error: str | Exception, logging_obj: LiteLLMLoggingObject +) -> object: + return await router.aresponses( + model="limited", + input="hello there", + max_output_tokens=100, + litellm_metadata={"user_api_key_hash": "hashed-key"}, + mock_response=mock_error, + litellm_call_id="proxy-call", + litellm_logging_obj=logging_obj, + ) + + +async def _proxy_messages_sharing_logging( + router: Router, mock_error: str | Exception, logging_obj: LiteLLMLoggingObject +) -> object: + return await router.aanthropic_messages( + model="limited", + messages=PROXY_MESSAGES, + max_tokens=100, + litellm_metadata={"user_api_key_hash": "hashed-key"}, + mock_response=mock_error, + litellm_call_id="proxy-call", + litellm_logging_obj=logging_obj, + ) + + class TestRouterIOTokenIntegration: + @pytest.mark.asyncio + @pytest.mark.parametrize( + "call", + [_sdk_chat, _proxy_chat, _proxy_responses, _proxy_messages], + ids=["sdk_chat", "proxy_chat_metadata", "proxy_responses_litellm_metadata", "proxy_messages_litellm_metadata"], + ) + async def test_success_reconciles_io_counters_to_real_usage(self, call: RouterCall): + router: Final = _router_with_io_limits() + first_minute: Final = _current_minute() + + usage: Final = await call(router, "hi") + + assert await _io_counters_since(router, first_minute) == usage + + @pytest.mark.asyncio + @pytest.mark.parametrize( + ("call", "mock_error"), + [ + ( + _proxy_responses, + litellm.InternalServerError(message="mock 500", llm_provider="anthropic", model=ROUTER_IO_MODEL), + ), + (_proxy_messages, "litellm.InternalServerError"), + ], + ids=["proxy_responses_litellm_metadata", "proxy_messages_litellm_metadata"], + ) + async def test_retried_failure_refunds_each_io_reservation_once( + self, call: RouterCall, mock_error: str | Exception + ): + router: Final = _router_with_io_limits() + first_minute: Final = _current_minute() + + with pytest.raises(litellm.InternalServerError): + await call(router, mock_error) + + assert await _io_counters_since(router, first_minute) == (0, 0) + + + @pytest.mark.asyncio + @pytest.mark.parametrize( + ("call", "original_function", "mock_error"), + [ + (_proxy_chat_sharing_logging, "acompletion", "litellm.InternalServerError"), + ( + _proxy_responses_sharing_logging, + "aresponses", + litellm.InternalServerError(message="mock 500", llm_provider="anthropic", model=ROUTER_IO_MODEL), + ), + (_proxy_messages_sharing_logging, "anthropic_messages", "litellm.InternalServerError"), + ], + ids=["proxy_chat", "proxy_responses", "proxy_messages"], + ) + async def test_retried_failure_sharing_one_logging_object_refunds_every_attempt( + self, call: ProxyCallSharingLogging, original_function: str, mock_error: str | Exception + ): + router: Final = _router_with_io_limits() + first_minute: Final = _current_minute() + + with pytest.raises(litellm.InternalServerError): + await call(router, mock_error, _logging_obj_shared_across_attempts(original_function)) + + assert await _io_counters_since(router, first_minute) == (0, 0) + @pytest.mark.asyncio async def test_model_group_info_aggregates_io_limits(self): router = Router( From ccc56d27adeeece2a1069bde587c76486ba9ab39 Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 29 Oct 2026 21:04:22 -0700 Subject: [PATCH 05/23] fix(azure): add batch prices to the unversioned o3, o4-mini, o3-mini, gpt-4o and gpt-4o-mini rows (#49182) Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../model_prices_and_context_window_backup.json | 15 +++++++++++++++ model_prices_and_context_window.json | 15 +++++++++++++++ 2 files changed, 30 insertions(+) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 96a04c96f44..0799a298942 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -6014,11 +6014,14 @@ "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 1.25e-06, "input_cost_per_token": 2.5e-06, + "input_cost_per_token_batches": 1.25e-06, "litellm_provider": "azure", "max_input_tokens": 128000, "max_output_tokens": 16384, "mode": "chat", "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -6313,16 +6316,19 @@ "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 7.5e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_batches": 7.5e-08, "litellm_provider": "azure", "max_input_tokens": 128000, "max_output_tokens": 16384, "mode": "chat", "output_cost_per_token": 6e-07, + "output_cost_per_token_batches": 3e-07, "search_context_cost_per_query": { "search_context_size_high": 0.014, "search_context_size_low": 0.014, "search_context_size_medium": 0.014 }, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -11303,16 +11309,19 @@ "deprecation_date": "2026-11-19", "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "azure", "max_input_tokens": 200000, "max_output_tokens": 100000, "mode": "chat", "output_cost_per_token": 8e-06, + "output_cost_per_token_batches": 4e-06, "search_context_cost_per_query": { "search_context_size_high": 0.014, "search_context_size_low": 0.014, "search_context_size_medium": 0.014 }, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -11411,11 +11420,14 @@ "deprecation_date": "2026-11-19", "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "azure", "max_input_tokens": 200000, "max_output_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -11513,16 +11525,19 @@ "deprecation_date": "2026-11-19", "cache_read_input_token_cost": 2.75e-07, "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "azure", "max_input_tokens": 200000, "max_output_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, "search_context_cost_per_query": { "search_context_size_high": 0.014, "search_context_size_low": 0.014, "search_context_size_medium": 0.014 }, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 96a04c96f44..0799a298942 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -6014,11 +6014,14 @@ "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 1.25e-06, "input_cost_per_token": 2.5e-06, + "input_cost_per_token_batches": 1.25e-06, "litellm_provider": "azure", "max_input_tokens": 128000, "max_output_tokens": 16384, "mode": "chat", "output_cost_per_token": 1e-05, + "output_cost_per_token_batches": 5e-06, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -6313,16 +6316,19 @@ "deprecation_date": "2027-04-14", "cache_read_input_token_cost": 7.5e-08, "input_cost_per_token": 1.5e-07, + "input_cost_per_token_batches": 7.5e-08, "litellm_provider": "azure", "max_input_tokens": 128000, "max_output_tokens": 16384, "mode": "chat", "output_cost_per_token": 6e-07, + "output_cost_per_token_batches": 3e-07, "search_context_cost_per_query": { "search_context_size_high": 0.014, "search_context_size_low": 0.014, "search_context_size_medium": 0.014 }, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supports_function_calling": true, "supports_parallel_function_calling": true, "supports_prompt_caching": true, @@ -11303,16 +11309,19 @@ "deprecation_date": "2026-11-19", "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 2e-06, + "input_cost_per_token_batches": 1e-06, "litellm_provider": "azure", "max_input_tokens": 200000, "max_output_tokens": 100000, "mode": "chat", "output_cost_per_token": 8e-06, + "output_cost_per_token_batches": 4e-06, "search_context_cost_per_query": { "search_context_size_high": 0.014, "search_context_size_low": 0.014, "search_context_size_medium": 0.014 }, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", @@ -11411,11 +11420,14 @@ "deprecation_date": "2026-11-19", "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "azure", "max_input_tokens": 200000, "max_output_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -11513,16 +11525,19 @@ "deprecation_date": "2026-11-19", "cache_read_input_token_cost": 2.75e-07, "input_cost_per_token": 1.1e-06, + "input_cost_per_token_batches": 5.5e-07, "litellm_provider": "azure", "max_input_tokens": 200000, "max_output_tokens": 100000, "mode": "chat", "output_cost_per_token": 4.4e-06, + "output_cost_per_token_batches": 2.2e-06, "search_context_cost_per_query": { "search_context_size_high": 0.014, "search_context_size_low": 0.014, "search_context_size_medium": 0.014 }, + "source": "https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/", "supported_endpoints": [ "/v1/chat/completions", "/v1/batch", From f32ac94ad1fc3bce79da08752c7bb6a17ebdbfa1 Mon Sep 17 00:00:00 2001 From: mika-lt <349045450+mika-lt@users.noreply.github.com> Date: Thu, 29 Oct 2026 23:30:02 -0700 Subject: [PATCH 06/23] fix(openai): keep the upstream error body on e.response for OpenAI and Azure errors (#49079) * fix(openai): keep the upstream error response on mapped exceptions so e.response.text has the wire body * fix: keep the upstream error body on 429 and 5xx errors RateLimitError, InternalServerError, BadGatewayError, and ServiceUnavailableError rebuilt their response and dropped the upstream body. They now keep the body and exactly the upstream headers, and an unread streaming response no longer raises. * fix: keep a gzip-encoded 429 or 5xx error readable The kept body is already decoded, so the rebuilt response drops content-encoding instead of decoding the body a second time and raising DecodingError. * fix: keep the upstream headers and drop credentials on every mapped provider error The copied error response now keeps every upstream header and carries a request with only the method and the URL, so no provider key, query secret, or prompt reaches e.request. Timeout and APIError get the response too, and azure/ chat, streaming, embedding, and image errors attach it * fix: keep the upstream error body on gateway timeouts and re-raised gateway errors * test(azure): cover the 404, 500, 502, and 504 error bodies and the missing-key placeholder * fix: keep no e.response on timeouts and errors litellm raises itself * fix: keep the upstream elapsed time on error responses and the upstream body on azure_text errors * fix: keep the upstream body without credentials on Azure content filter errors and message-matched OpenAI errors --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/exceptions.py | 93 +++--- .../exception_mapping_utils.py | 57 +++- litellm/llms/azure/azure.py | 12 +- litellm/llms/azure/completion/handler.py | 22 +- litellm/llms/azure/exception_mapping.py | 5 +- litellm/llms/openai/common_utils.py | 9 + litellm/llms/openai/completion/handler.py | 23 +- .../llms/openai/image_variations/handler.py | 16 +- litellm/llms/openai/openai.py | 24 +- .../test_exception_mapping_utils.py | 82 +++++- .../azure/test_azure_exception_mapping.py | 201 +++++++++++++ .../completion/test_completion_handler.py | 24 +- .../llms/openai/image_variations/__init__.py | 0 .../test_openai_image_variations_handler.py | 51 ++++ tests/unit/llms/openai/test_openai.py | 274 +++++++++++++++++- .../proxy/test_common_request_processing.py | 63 ++++ .../test_exception_header_preservation.py | 19 ++ 17 files changed, 895 insertions(+), 80 deletions(-) create mode 100644 tests/unit/llms/openai/image_variations/__init__.py create mode 100644 tests/unit/llms/openai/image_variations/test_openai_image_variations_handler.py diff --git a/litellm/exceptions.py b/litellm/exceptions.py index 28211b3dab2..12425c45b01 100644 --- a/litellm/exceptions.py +++ b/litellm/exceptions.py @@ -9,6 +9,7 @@ ## LiteLLM versions of the OpenAI Exception Types +import datetime import enum from collections.abc import Mapping, Sequence from typing import Any, Final @@ -115,6 +116,54 @@ def validate_rate_limit_type(value: object) -> str | None: return None +def _upstream_body(response: object) -> bytes: + if not isinstance(response, httpx.Response): + return b"" + try: + return response.content + except httpx.ResponseNotRead: + return b"" + + +def _request_without_credentials(upstream: object, fallback_url: str) -> httpx.Request: + if not isinstance(upstream, httpx.Response): + return httpx.Request(method="POST", url=fallback_url) + try: + upstream_request: Final = upstream.request + except RuntimeError: + return httpx.Request(method="POST", url=fallback_url) + return httpx.Request( + method=upstream_request.method, + url=upstream_request.url.copy_with(userinfo=b"", query=None, fragment=None), + ) + + +def _upstream_elapsed(upstream: object) -> datetime.timedelta | None: + if not isinstance(upstream, httpx.Response): + return None + try: + return upstream.elapsed + except RuntimeError: + return None + + +def error_response_without_credentials(status_code: int, upstream: object, fallback_url: str) -> httpx.Response: + response: Final = httpx.Response( + status_code=status_code, + stream=httpx.ByteStream(_upstream_body(upstream)), + request=_request_without_credentials(upstream, fallback_url), + ) + _ = response.read() + response.headers = httpx.Headers(upstream.headers if isinstance(upstream, httpx.Response) else None) + elapsed: Final = _upstream_elapsed(upstream) + if elapsed is not None: + response.elapsed = elapsed + return response + + +_NO_UPSTREAM_REQUEST_URL: Final = "https://cloud.google.com/vertex-ai/" + + _MINIMAL_ERROR_RESPONSE: httpx.Response | None = None @@ -519,20 +568,11 @@ class RateLimitError(openai.RateLimitError): # headers stay reachable on `e.response.headers` for callers that # explicitly want them; only the proxy-supplied `headers=` kwarg # makes it onto `self.headers`. - _response_headers: Final = getattr(response, "headers", None) if response is not None else None self.headers: dict[str, str] | None = {k: str(v) for k, v in headers.items()} if headers else None # Mirrors FastAPI HTTPException.detail so the same instance can be # serialized through both the ProxyException and HTTPException paths. self.detail = detail if detail is not None else self.message - self.response = httpx.Response( - status_code=429, - headers=_response_headers, - content=response.content if response is not None else None, - request=httpx.Request( - method="POST", - url=" https://cloud.google.com/vertex-ai/", - ), - ) + self.response = error_response_without_credentials(429, response, _NO_UPSTREAM_REQUEST_URL) super().__init__( self.message, response=self.response, # pyright: ignore[reportArgumentType] # SDK accepts HTTPX at runtime @@ -757,15 +797,7 @@ class ServiceUnavailableError(openai.APIStatusError): self.litellm_debug_info = litellm_debug_info self.max_retries = max_retries self.num_retries = num_retries - _response_headers: Final = getattr(response, "headers", None) if response is not None else None - self.response = httpx.Response( - status_code=self.status_code, - headers=_response_headers, - request=httpx.Request( - method="POST", - url=" https://cloud.google.com/vertex-ai/", - ), - ) + self.response = error_response_without_credentials(self.status_code, response, _NO_UPSTREAM_REQUEST_URL) super().__init__( self.message, response=self.response, body=None ) # Call the base class constructor with the parameters it needs @@ -808,15 +840,7 @@ class BadGatewayError(openai.APIStatusError): self.litellm_debug_info = litellm_debug_info self.max_retries = max_retries self.num_retries = num_retries - _response_headers: Final = getattr(response, "headers", None) if response is not None else None - self.response = httpx.Response( - status_code=self.status_code, - headers=_response_headers, - request=httpx.Request( - method="POST", - url=" https://cloud.google.com/vertex-ai/", - ), - ) + self.response = error_response_without_credentials(self.status_code, response, _NO_UPSTREAM_REQUEST_URL) super().__init__( self.message, response=self.response, # pyright: ignore[reportArgumentType] # SDK accepts HTTPX at runtime @@ -862,15 +886,7 @@ class InternalServerError(openai.InternalServerError): self.litellm_debug_info = litellm_debug_info self.max_retries = max_retries self.num_retries = num_retries - _response_headers: Final = getattr(response, "headers", None) if response is not None else None - self.response = httpx.Response( - status_code=self.status_code, - headers=_response_headers, - request=httpx.Request( - method="POST", - url=" https://cloud.google.com/vertex-ai/", - ), - ) + self.response = error_response_without_credentials(self.status_code, response, _NO_UPSTREAM_REQUEST_URL) super().__init__( self.message, response=self.response, # pyright: ignore[reportArgumentType] # SDK accepts HTTPX at runtime @@ -908,6 +924,7 @@ class APIError(openai.APIError): max_retries: int | None = None, num_retries: int | None = None, body: object | None = None, + response: httpx.Response | None = None, ): self.status_code = status_code self.message = f"litellm.APIError: {message}" @@ -916,6 +933,8 @@ class APIError(openai.APIError): self.litellm_debug_info = litellm_debug_info self.max_retries = max_retries self.num_retries = num_retries + if response is not None: + self.response = response if request is None: request = httpx.Request(method="POST", url="https://api.openai.com/v1") super().__init__(self.message, request=request, body=body) diff --git a/litellm/litellm_core_utils/exception_mapping_utils.py b/litellm/litellm_core_utils/exception_mapping_utils.py index 64a95eb5f28..5a27f5446ad 100644 --- a/litellm/litellm_core_utils/exception_mapping_utils.py +++ b/litellm/litellm_core_utils/exception_mapping_utils.py @@ -37,6 +37,7 @@ from ..exceptions import ( ServiceUnavailableError, Timeout, UnprocessableEntityError, + error_response_without_credentials, ) @@ -305,6 +306,16 @@ def _litellm_proxy_response( ) +def _without_credentials(response: httpx.Response | None) -> httpx.Response | None: + if not isinstance(response, httpx.Response): + return response + return error_response_without_credentials(response.status_code, response, "https://docs.litellm.ai/docs") + + +def _upstream_reply(response: httpx.Response | None) -> httpx.Response | None: + return response if response is not None and (response.headers or response.content) else None + + def _map_openai_exception( *, model: str, @@ -315,7 +326,7 @@ def _map_openai_exception( exception_provider: str, extra_information: str, ) -> None: - response: Final = _litellm_proxy_response(original_exception, custom_llm_provider) + response: Final = _without_credentials(_litellm_proxy_response(original_exception, custom_llm_provider)) if (cyber_policy_body := _openai_cyber_policy_body(original_exception)) is not None: raise ContentPolicyViolationError( message=f"ContentPolicyViolationError: {exception_provider} - {original_exception}", @@ -378,6 +389,7 @@ def _map_openai_exception( model=model, llm_provider=custom_llm_provider, litellm_debug_info=extra_information, + response=_upstream_reply(response), ) elif ( ("invalid_request_error" in error_str and "content_policy_violation" in error_str) @@ -449,6 +461,7 @@ def _map_openai_exception( model=model, llm_provider=custom_llm_provider, body=getattr(original_exception, "body", None), + response=_upstream_reply(response), ) elif "Request too large" in error_str: raise RateLimitError( @@ -479,6 +492,7 @@ def _map_openai_exception( model=model, request=_request, litellm_debug_info=extra_information, + response=_upstream_reply(response), ) elif hasattr(original_exception, "status_code"): if original_exception.status_code == 400: @@ -512,6 +526,7 @@ def _map_openai_exception( model=model, llm_provider=custom_llm_provider, litellm_debug_info=extra_information, + response=_upstream_reply(response), ) elif original_exception.status_code == 422: raise BadRequestError( @@ -563,6 +578,7 @@ def _map_openai_exception( llm_provider=custom_llm_provider, litellm_debug_info=extra_information, exception_status_code=original_exception.status_code, + response=_upstream_reply(response), ) else: raise APIError( @@ -572,6 +588,7 @@ def _map_openai_exception( model=model, request=getattr(original_exception, "request", None), litellm_debug_info=extra_information, + response=_upstream_reply(response), ) else: # if no status code then it is an APIConnectionError: https://github.com/openai/openai-python#handling-errors @@ -2019,6 +2036,7 @@ def _map_azure_exception( exception_provider: str, extra_information: str, ) -> None: + response: Final = _without_credentials(getattr(original_exception, "response", None)) message = get_error_message(error_obj=original_exception) if message is None: if hasattr(original_exception, "message"): @@ -2054,7 +2072,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif "DeploymentNotFound" in error_str: raise NotFoundError( @@ -2062,7 +2080,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif azure_error_code == "content_policy_violation" or ExceptionCheckers.is_azure_content_policy_violation_error( error_str @@ -2076,6 +2094,7 @@ def _map_azure_exception( model=model, extra_information=extra_information, original_exception=original_exception, + response=response, ) elif azure_error_code == "invalid_encrypted_content" or "could not be verified" in error_str: helpful_message: Final = ( @@ -2094,7 +2113,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, body=getattr(original_exception, "body", None), ) elif getattr(original_exception, "status_code", None) in (None, 400) and ( @@ -2106,7 +2125,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, body=getattr(original_exception, "body", None), ) elif "invalid_request_error" in error_str and getattr(original_exception, "status_code", None) in (None, 400): @@ -2115,7 +2134,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, body=getattr(original_exception, "body", None), ) elif "The api_key client option must be set either by passing api_key to the client or by setting" in error_str: @@ -2124,7 +2143,7 @@ def _map_azure_exception( llm_provider=custom_llm_provider, model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif "Connection error" in error_str: raise APIConnectionError( @@ -2140,7 +2159,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, body=getattr(original_exception, "body", None), ) elif original_exception.status_code == 401: @@ -2149,7 +2168,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif original_exception.status_code == 404: raise NotFoundError( @@ -2157,7 +2176,7 @@ def _map_azure_exception( llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif original_exception.status_code == 408: raise Timeout( @@ -2165,6 +2184,7 @@ def _map_azure_exception( model=model, litellm_debug_info=extra_information, llm_provider="azure", + response=_upstream_reply(response), ) elif original_exception.status_code == 422: raise BadRequestError( @@ -2172,7 +2192,7 @@ def _map_azure_exception( model=model, llm_provider="azure", litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif original_exception.status_code == 429: raise RateLimitError( @@ -2180,7 +2200,7 @@ def _map_azure_exception( model=model, llm_provider="azure", litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif original_exception.status_code == 502: raise BadGatewayError( @@ -2188,7 +2208,7 @@ def _map_azure_exception( model=model, llm_provider="azure", litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif original_exception.status_code == 503: raise ServiceUnavailableError( @@ -2196,7 +2216,7 @@ def _map_azure_exception( model=model, llm_provider="azure", litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, ) elif original_exception.status_code == 504: # gateway timeout error raise Timeout( @@ -2205,6 +2225,7 @@ def _map_azure_exception( litellm_debug_info=extra_information, llm_provider="azure", exception_status_code=original_exception.status_code, + response=_upstream_reply(response), ) else: raise APIError( @@ -2214,6 +2235,7 @@ def _map_azure_exception( litellm_debug_info=extra_information, model=model, request=httpx.Request(method="POST", url="https://openai.com/"), + response=_upstream_reply(response), ) else: # if no status code then it is an APIConnectionError: https://github.com/openai/openai-python#handling-errors @@ -2589,17 +2611,20 @@ def exception_type( model=model, llm_provider=custom_llm_provider, litellm_debug_info=extra_information, + response=_upstream_reply( + _without_credentials(_litellm_proxy_response(mappable_exception, custom_llm_provider)) + ), ) if ( custom_llm_provider == "litellm_proxy" ): # handle special case where calling litellm proxy + exception str contains error message extract_and_raise_litellm_exception( - response=_litellm_proxy_response(mappable_exception, custom_llm_provider), + response=_without_credentials(_litellm_proxy_response(mappable_exception, custom_llm_provider)), error_str=error_str, model=model, custom_llm_provider=custom_llm_provider, - body=getattr(original_exception, "body", None), + body=getattr(mappable_exception, "body", None), ) if ( custom_llm_provider == "openai" diff --git a/litellm/llms/azure/azure.py b/litellm/llms/azure/azure.py index a1ba8de6d7f..69ffac5d6f9 100644 --- a/litellm/llms/azure/azure.py +++ b/litellm/llms/azure/azure.py @@ -42,6 +42,7 @@ from ..openai.common_utils import ( build_output_token_limit_response, is_output_token_limit_error, model_dump_provider_response, + upstream_error_response, ) from .common_utils import ( AzureOpenAIError, @@ -399,6 +400,7 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM): raise AzureOpenAIError( status_code=status_code, message=str(e), + response=upstream_error_response(e), headers=error_headers, body=error_body, ) @@ -662,6 +664,7 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM): raise AzureOpenAIError( status_code=status_code, message=message, + response=upstream_error_response(e), headers=error_headers, body=error_body, ) @@ -855,7 +858,12 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM): if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) error_text = error_response.text - raise AzureOpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise AzureOpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) async def make_async_azure_httpx_request( self, @@ -1190,7 +1198,7 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM): except Exception as e: error_code: Final[object] = getattr(e, "status_code", None) if error_code is not None: - raise AzureOpenAIError(status_code=error_code, message=str(e)) + raise AzureOpenAIError(status_code=error_code, message=str(e), response=upstream_error_response(e)) else: raise AzureOpenAIError(status_code=500, message=str(e)) diff --git a/litellm/llms/azure/completion/handler.py b/litellm/llms/azure/completion/handler.py index 34e51eaf1ec..c2744e7f750 100644 --- a/litellm/llms/azure/completion/handler.py +++ b/litellm/llms/azure/completion/handler.py @@ -8,6 +8,7 @@ from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging from litellm.litellm_core_utils.prompt_templates.factory import prompt_factory from litellm.utils import CustomStreamWrapper, ModelResponse, TextCompletionResponse +from ...openai.common_utils import upstream_error_response from ...openai.completion.transformation import OpenAITextCompletionConfig from ..common_utils import AzureOpenAIError, BaseAzureLLM @@ -183,7 +184,12 @@ class AzureTextCompletion(BaseAzureLLM): error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise AzureOpenAIError(status_code=status_code, message=str(e), headers=error_headers) + raise AzureOpenAIError( + status_code=status_code, + message=str(e), + headers=error_headers, + response=upstream_error_response(e), + ) async def acompletion( self, @@ -244,7 +250,12 @@ class AzureTextCompletion(BaseAzureLLM): error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise AzureOpenAIError(status_code=status_code, message=str(e), headers=error_headers) + raise AzureOpenAIError( + status_code=status_code, + message=str(e), + headers=error_headers, + response=upstream_error_response(e), + ) def streaming( self, @@ -355,4 +366,9 @@ class AzureTextCompletion(BaseAzureLLM): error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise AzureOpenAIError(status_code=status_code, message=str(e), headers=error_headers) + raise AzureOpenAIError( + status_code=status_code, + message=str(e), + headers=error_headers, + response=upstream_error_response(e), + ) diff --git a/litellm/llms/azure/exception_mapping.py b/litellm/llms/azure/exception_mapping.py index f7146a36384..6eb5161f7e4 100644 --- a/litellm/llms/azure/exception_mapping.py +++ b/litellm/llms/azure/exception_mapping.py @@ -1,5 +1,7 @@ from typing import Final +import httpx + from litellm.exceptions import ContentPolicyViolationError @@ -14,6 +16,7 @@ class AzureOpenAIExceptionMapping: model: str, extra_information: str, original_exception: Exception, + response: httpx.Response | None, ) -> ContentPolicyViolationError: """ Create a content policy violation error @@ -39,7 +42,7 @@ class AzureOpenAIExceptionMapping: llm_provider="azure", model=model, litellm_debug_info=extra_information, - response=getattr(original_exception, "response", None), + response=response, provider_specific_fields={ # Preserve legacy key for backward compatibility. "innererror": inner_error, diff --git a/litellm/llms/openai/common_utils.py b/litellm/llms/openai/common_utils.py index 6c416124d1a..aa14a985afe 100644 --- a/litellm/llms/openai/common_utils.py +++ b/litellm/llms/openai/common_utils.py @@ -29,6 +29,7 @@ if TYPE_CHECKING: from aiohttp import ClientSession import litellm +from litellm.exceptions import error_response_without_credentials from litellm.litellm_core_utils.completion_timeout import CompletionTimeout from litellm.litellm_core_utils.core_helpers import as_str_mapping, decoded_json_object from litellm.litellm_core_utils.token_counter import token_counter, usage_tokenizer_model @@ -148,6 +149,14 @@ class OpenAIError(BaseLLMException): ) +def upstream_error_response(error: Exception) -> httpx.Response | None: + if not isinstance(error, (openai.APIStatusError, httpx.HTTPStatusError)): + return None + return error_response_without_credentials( + error.response.status_code, error.response, "https://docs.litellm.ai/docs" + ) + + _DECODED_JSON: Final = TypeAdapter(object) diff --git a/litellm/llms/openai/completion/handler.py b/litellm/llms/openai/completion/handler.py index 278427e8d77..8bfdd536f90 100644 --- a/litellm/llms/openai/completion/handler.py +++ b/litellm/llms/openai/completion/handler.py @@ -10,7 +10,7 @@ from litellm.types.llms.openai import AllMessageValues, OpenAITextCompletionUser from litellm.types.utils import LlmProviders, ModelResponse, TextCompletionResponse from litellm.utils import ProviderConfigManager -from ..common_utils import BaseOpenAILLM, OpenAIError +from ..common_utils import BaseOpenAILLM, OpenAIError, upstream_error_response from ..workload_identity import build_async_openai_client, build_openai_client from .transformation import OpenAITextCompletionConfig @@ -164,7 +164,12 @@ class OpenAITextCompletion(BaseLLM): error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise OpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise OpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) async def acompletion( self, @@ -220,7 +225,12 @@ class OpenAITextCompletion(BaseLLM): error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise OpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise OpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) def streaming( self, @@ -261,7 +271,12 @@ class OpenAITextCompletion(BaseLLM): error_response: object = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise OpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise OpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) streamwrapper: Final = CustomStreamWrapper( completion_stream=response, model=model, diff --git a/litellm/llms/openai/image_variations/handler.py b/litellm/llms/openai/image_variations/handler.py index 01f01d3de1d..504dfb50015 100644 --- a/litellm/llms/openai/image_variations/handler.py +++ b/litellm/llms/openai/image_variations/handler.py @@ -14,7 +14,7 @@ from litellm.utils import ProviderConfigManager from ...base_llm.image_variations.transformation import BaseImageVariationConfig from ...custom_httpx.llm_http_handler import LiteLLMLoggingObj -from ..common_utils import OpenAIError +from ..common_utils import OpenAIError, upstream_error_response from ..workload_identity import build_async_openai_client, build_openai_client @@ -129,7 +129,12 @@ class OpenAIImageVariationsHandler: error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise OpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise OpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) def image_variations( self, @@ -245,4 +250,9 @@ class OpenAIImageVariationsHandler: error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise OpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise OpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) diff --git a/litellm/llms/openai/openai.py b/litellm/llms/openai/openai.py index af2e85c5f20..e0a23d6c99f 100644 --- a/litellm/llms/openai/openai.py +++ b/litellm/llms/openai/openai.py @@ -64,6 +64,7 @@ from .common_utils import ( is_openai_backed_api_base, is_output_token_limit_error, model_dump_provider_response, + upstream_error_response, ) from .workload_identity import ( build_async_openai_client, @@ -983,6 +984,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): raise OpenAIError( status_code=status_code, message=error_text, + response=upstream_error_response(e), headers=error_headers, body=error_body, ) @@ -1110,6 +1112,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): raise OpenAIError( status_code=status_code, message=message, + response=upstream_error_response(e), headers=error_headers, body=exception_body, ) @@ -1282,6 +1285,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): raise OpenAIError( status_code=getattr(e, "status_code", 500), message=str(e), + response=upstream_error_response(e), headers=error_headers, body=exception_body, ) @@ -1409,7 +1413,12 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise OpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise OpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) def embedding( self, @@ -1499,7 +1508,12 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): error_response: Final[object] = getattr(e, "response", None) if error_headers is None and error_response: error_headers = getattr(error_response, "headers", None) - raise OpenAIError(status_code=status_code, message=error_text, headers=error_headers) + raise OpenAIError( + status_code=status_code, + message=error_text, + response=upstream_error_response(e), + headers=error_headers, + ) async def aimage_generation( self, @@ -1666,7 +1680,11 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): original_response=str(e), ) if hasattr(e, "status_code"): - raise OpenAIError(status_code=getattr(e, "status_code", 500), message=str(e)) + raise OpenAIError( + status_code=getattr(e, "status_code", 500), + message=str(e), + response=upstream_error_response(e), + ) else: raise OpenAIError(status_code=500, message=str(e)) diff --git a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py index 19e5387da9a..88501a23235 100644 --- a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py +++ b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py @@ -24,6 +24,7 @@ from litellm.llms.base_llm.chat.transformation import BaseLLMException from litellm.llms.bedrock.common_utils import BedrockError from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler from litellm.llms.openai.common_utils import OpenAIError +from litellm.llms.vertex_ai.common_utils import VertexAIError from litellm.proxy.common_utils.openai_error_payload import openai_error_type from litellm.types.utils import LlmProviders @@ -54,8 +55,8 @@ def test_openai_cyber_policy_uses_the_structured_error_code(shape: str, code: st assert caught.value.code == code assert caught.value.body == body assert caught.value.status_code == 400 - assert caught.value.response is response - assert caught.value.request is request + assert caught.value.response.status_code == 400 + assert caught.value.request.url == request.url assert caught.value.response.headers["x-request-id"] == "cyber-policy-request" @@ -2940,3 +2941,80 @@ def test_transport_timeout_maps_to_timeout( assert exc_info.value.status_code == 408 assert f"Request timed out. Error_str: {reason}" in str(exc_info.value) + + +def test_bedrock_server_error_keeps_the_aws_body_without_the_signed_request() -> None: + body: Final = '{"message":"Bedrock is unable to process your request."}' + signed_request: Final = httpx.Request( + "POST", + "https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-opus-5-5/converse", + headers={"authorization": "AWS4-HMAC-SHA256 Credential=AKIDSIGNED/20261029/us-east-1/bedrock/aws4_request"}, + content=b'{"messages":[{"role":"user","content":[{"text":"hi"}]}]}', + ) + upstream: Final = httpx.Response( + 500, content=body, headers={"x-amzn-requestid": "req-aws"}, request=signed_request + ) + + with pytest.raises(litellm.ServiceUnavailableError) as raised: + exception_type( + model="bedrock/anthropic.claude-opus-5-5", + original_exception=BedrockError(status_code=500, message=body, response=upstream), + custom_llm_provider="bedrock", + ) + + assert raised.value.response.text == body + assert raised.value.response.headers["x-amzn-requestid"] == "req-aws" + assert "AKIDSIGNED" not in str(raised.value.request.headers) + assert raised.value.request.content == b"" + + +@pytest.mark.parametrize("provider", ("gemini", "vertex_ai")) +def test_gemini_server_error_response_carries_the_upstream_error_body(provider: str) -> None: + body: Final = '{"error":{"code":500,"message":"Internal error encountered.","status":"INTERNAL"}}' + + with pytest.raises(litellm.InternalServerError) as raised: + exception_type( + model="gemini-3.1-pro", + original_exception=VertexAIError(status_code=500, message=body, headers=None), + custom_llm_provider=provider, + ) + + assert raised.value.response.status_code == 500 + assert raised.value.response.json() == json.loads(body) + + +@pytest.mark.parametrize( + ("status_code", "message", "mapped_error"), + ( + (500, "The server had an error processing your request. Sorry about that!", litellm.InternalServerError), + (520, "Web server is returning an unknown error", litellm.InternalServerError), + (504, "A timeout occurred", litellm.Timeout), + (500, "Mistral API raised a streaming error", litellm.APIError), + ), +) +def test_openai_error_matched_by_its_message_keeps_the_upstream_body( + status_code: int, message: str, mapped_error: type[Exception] +) -> None: + body: Final = json.dumps({"error": {"message": message, "type": "server_error", "param": None, "code": None}}) + sent_request: Final = httpx.Request( + "POST", + "https://api.openai.com/v1/chat/completions", + headers={"authorization": "Bearer sk-openai-secret"}, + content=b'{"messages":[{"role":"user","content":"hi"}]}', + ) + upstream: Final = httpx.Response( + status_code, content=body, headers={"x-request-id": "req-openai"}, request=sent_request + ) + + with pytest.raises(mapped_error) as raised: + exception_type( + model="gpt-5.6", + original_exception=openai.APIStatusError(message, response=upstream, body=json.loads(body)), + custom_llm_provider="openai", + ) + + error_response: Final = getattr(raised.value, "response") + assert error_response.text == body + assert error_response.headers["x-request-id"] == "req-openai" + assert "sk-openai-secret" not in str(error_response.request.headers) + assert error_response.request.content == b"" diff --git a/tests/unit/llms/azure/test_azure_exception_mapping.py b/tests/unit/llms/azure/test_azure_exception_mapping.py index cbbf70c691b..018b499f401 100644 --- a/tests/unit/llms/azure/test_azure_exception_mapping.py +++ b/tests/unit/llms/azure/test_azure_exception_mapping.py @@ -1,3 +1,4 @@ +import datetime import json from collections.abc import Awaitable, Callable, Mapping from typing import Final @@ -12,6 +13,7 @@ import respx import litellm from litellm.exceptions import BadRequestError, ContentPolicyViolationError, ContextWindowExceededError from litellm.litellm_core_utils.exception_mapping_utils import exception_type +from litellm.llms.azure.common_utils import AzureOpenAIError @respx.mock @@ -632,3 +634,202 @@ def test_azure_content_policy_wins_over_context_length_code(monkeypatch: pytest. error: Final = _azure_chat_error(monkeypatch, _azure_error_response(400, policy_error)) assert type(error) is ContentPolicyViolationError + + +_AZURE_REFUSAL_BODY: Final = '{"error":{"message":"azure refused","type":"upstream_refusal","code":"refused"}}' +_AZURE_DEPLOYMENT_NOT_FOUND_BODY: Final = '{"error":{"code":"DeploymentNotFound","message":"no such deployment"}}' +_AZURE_INTERNAL_ERROR_BODY: Final = '{"error":{"code":"InternalServerError","message":"Internal server error"}}' +_AZURE_CONTENT_FILTER_BODY: Final = json.dumps( + { + "error": { + "message": "The response was filtered due to the prompt triggering Azure OpenAI's content management policy.", + "type": None, + "param": "prompt", + "code": "content_filter", + "status": 400, + "innererror": { + "code": "ResponsibleAIPolicyViolation", + "content_filter_result": {"violence": {"filtered": True, "severity": "high"}}, + }, + } + } +) +_AZURE_KEY: Final = "azure-provider-secret" +_AZURE_BASE: Final = "https://azure-upstream.openai.azure.com" +_AZURE_ARGS: Final = {"api_base": _AZURE_BASE, "api_version": "2024-10-21", "api_key": _AZURE_KEY, "max_retries": 0} + + +async def _azure_completion() -> None: + litellm.completion(model="azure/gpt-4.1-mini", messages=[{"role": "user", "content": "hi"}], **_AZURE_ARGS) + + +async def _azure_completion_stream() -> None: + list( + litellm.completion( + model="azure/gpt-4.1-mini", messages=[{"role": "user", "content": "hi"}], stream=True, **_AZURE_ARGS + ) + ) + + +async def _azure_acompletion() -> None: + await litellm.acompletion(model="azure/gpt-4.1-mini", messages=[{"role": "user", "content": "hi"}], **_AZURE_ARGS) + + +async def _azure_acompletion_stream() -> None: + stream: Final = await litellm.acompletion( + model="azure/gpt-4.1-mini", messages=[{"role": "user", "content": "hi"}], stream=True, **_AZURE_ARGS + ) + async for _ in stream: + pass + + +async def _azure_embedding() -> None: + litellm.embedding(model="azure/text-embedding-3-large", input=["hi"], **_AZURE_ARGS) + + +async def _azure_aembedding() -> None: + await litellm.aembedding(model="azure/text-embedding-3-large", input=["hi"], **_AZURE_ARGS) + + +async def _azure_image_generation() -> None: + litellm.image_generation(model="azure/gpt-image-2", prompt="a cat", **_AZURE_ARGS) + + +async def _azure_text_completion() -> None: + litellm.text_completion(model="azure_text/gpt-35-turbo-instruct", prompt="hi", **_AZURE_ARGS) + + +async def _azure_text_completion_stream() -> None: + list(litellm.text_completion(model="azure_text/gpt-35-turbo-instruct", prompt="hi", stream=True, **_AZURE_ARGS)) + + +async def _azure_atext_completion() -> None: + await litellm.atext_completion(model="azure_text/gpt-35-turbo-instruct", prompt="hi", **_AZURE_ARGS) + + +async def _azure_atext_completion_stream() -> None: + stream: Final = await litellm.atext_completion( + model="azure_text/gpt-35-turbo-instruct", prompt="hi", stream=True, **_AZURE_ARGS + ) + async for _ in stream: + pass + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "route", + ( + _azure_completion, + _azure_completion_stream, + _azure_acompletion, + _azure_acompletion_stream, + _azure_embedding, + _azure_aembedding, + _azure_image_generation, + _azure_text_completion, + _azure_text_completion_stream, + _azure_atext_completion, + _azure_atext_completion_stream, + ), + ids=lambda route: route.__name__.removeprefix("_azure_"), +) +@pytest.mark.parametrize( + ("status_code", "body", "mapped_error"), + ( + (400, _AZURE_REFUSAL_BODY, BadRequestError), + (400, _AZURE_CONTENT_FILTER_BODY, BadRequestError), + (403, _AZURE_REFUSAL_BODY, litellm.APIError), + (404, _AZURE_DEPLOYMENT_NOT_FOUND_BODY, litellm.NotFoundError), + (500, _AZURE_INTERNAL_ERROR_BODY, openai.APIError), + (502, _AZURE_REFUSAL_BODY, litellm.BadGatewayError), + (504, _AZURE_REFUSAL_BODY, litellm.Timeout), + ), +) +@respx.mock +async def test_azure_error_keeps_the_upstream_body_without_the_api_key( + route: Callable[[], Awaitable[None]], + status_code: int, + body: str, + mapped_error: type[Exception], + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx.post(url__regex=r"https://azure-upstream\.openai\.azure\.com/.*").mock( + return_value=httpx.Response(status_code, content=body, headers={"content-type": "application/json"}) + ) + + with pytest.raises(mapped_error) as raised: + await route() + + error_response: Final = getattr(raised.value, "response") + assert error_response.status_code == status_code + assert error_response.text == body + assert error_response.json() == json.loads(body) + for exposed in (error_response.request, getattr(raised.value, "request")): + assert _AZURE_KEY not in str(exposed.headers) + assert exposed.content == b"" + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "route", + ( + _azure_completion, + _azure_completion_stream, + _azure_acompletion, + _azure_acompletion_stream, + _azure_embedding, + _azure_aembedding, + _azure_text_completion, + _azure_atext_completion, + _azure_atext_completion_stream, + ), + ids=lambda route: route.__name__.removeprefix("_azure_"), +) +@respx.mock +async def test_azure_error_response_keeps_the_upstream_elapsed_time( + route: Callable[[], Awaitable[None]], monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + respx.post(url__regex=r"https://azure-upstream\.openai\.azure\.com/.*").mock( + return_value=httpx.Response(400, content=_AZURE_REFUSAL_BODY, headers={"content-type": "application/json"}) + ) + + with pytest.raises(BadRequestError) as raised: + await route() + + assert raised.value.response.elapsed > datetime.timedelta(0) + + +def test_azure_client_side_missing_key_error_keeps_the_empty_placeholder_response() -> None: + missing_key: Final = openai.OpenAIError( + "The api_key client option must be set either by passing api_key to the client or by setting the " + "OPENAI_API_KEY environment variable" + ) + + with pytest.raises(litellm.AuthenticationError) as raised: + exception_type(model="azure/gpt-4.1-mini", original_exception=missing_key, custom_llm_provider="azure") + + assert raised.value.response.status_code == 401 + assert raised.value.response.text == "" + + +@pytest.mark.parametrize( + ("status_code", "message", "mapped_error"), + ( + (500, "azure_client is not an instance of AzureOpenAI or OpenAI", litellm.APIError), + (408, "the deployment did not answer in time", litellm.Timeout), + (504, "the deployment did not answer in time", litellm.Timeout), + ), +) +def test_azure_error_litellm_raises_itself_has_no_upstream_response( + status_code: int, message: str, mapped_error: type[Exception] +) -> None: + with pytest.raises(mapped_error) as raised: + exception_type( + model="azure/gpt-4.1-mini", + original_exception=AzureOpenAIError(status_code=status_code, message=message), + custom_llm_provider="azure", + ) + + assert getattr(raised.value, "response", None) is None diff --git a/tests/unit/llms/openai/completion/test_completion_handler.py b/tests/unit/llms/openai/completion/test_completion_handler.py index 9e2754eef68..628c9bbd783 100644 --- a/tests/unit/llms/openai/completion/test_completion_handler.py +++ b/tests/unit/llms/openai/completion/test_completion_handler.py @@ -93,24 +93,34 @@ async def test_acompletion_forwards_client_headers_to_provider( assert request_headers["x-mycorp-llmcall-id"] == "abc-123" +_REJECTION_BODY: Final = '{"error":{"message":"bad prompt","type":"invalid_request_error"}}' + + @pytest.fixture def rejected_completions_endpoint(): return respx.post("https://api.openai.com/v1/completions").mock( return_value=Response( 400, - json={"error": {"message": "bad prompt", "type": "invalid_request_error"}}, - headers={"x-request-id": "req-1"}, + content=_REJECTION_BODY, + headers={"x-request-id": "req-1", "content-type": "application/json"}, ) ) +def _assert_carries_the_upstream_rejection(rejected: litellm.BadRequestError) -> None: + assert rejected.status_code == 400 + assert rejected.litellm_response_headers["x-request-id"] == "req-1" + assert rejected.response.text == _REJECTION_BODY + assert rejected.response.json() == json.loads(_REJECTION_BODY) + assert rejected.response.headers["x-request-id"] == "req-1" + + @respx.mock def test_completion_surfaces_the_status_and_headers_of_a_rejected_request(rejected_completions_endpoint): with pytest.raises(litellm.BadRequestError) as rejected: text_completion(model="gpt-3.5-turbo-instruct", prompt="hello") - assert rejected.value.status_code == 400 - assert rejected.value.litellm_response_headers["x-request-id"] == "req-1" + _assert_carries_the_upstream_rejection(rejected.value) @respx.mock @@ -118,8 +128,7 @@ def test_streaming_completion_surfaces_the_status_and_headers_of_a_rejected_requ with pytest.raises(litellm.BadRequestError) as rejected: list(text_completion(model="gpt-3.5-turbo-instruct", prompt="hello", stream=True)) - assert rejected.value.status_code == 400 - assert rejected.value.litellm_response_headers["x-request-id"] == "req-1" + _assert_carries_the_upstream_rejection(rejected.value) @respx.mock @@ -130,8 +139,7 @@ async def test_acompletion_surfaces_the_status_and_headers_of_a_rejected_request with pytest.raises(litellm.BadRequestError) as rejected: await atext_completion(model="gpt-3.5-turbo-instruct", prompt="hello") - assert rejected.value.status_code == 400 - assert rejected.value.litellm_response_headers["x-request-id"] == "req-1" + _assert_carries_the_upstream_rejection(rejected.value) @respx.mock diff --git a/tests/unit/llms/openai/image_variations/__init__.py b/tests/unit/llms/openai/image_variations/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/unit/llms/openai/image_variations/test_openai_image_variations_handler.py b/tests/unit/llms/openai/image_variations/test_openai_image_variations_handler.py new file mode 100644 index 00000000000..83ddbec6038 --- /dev/null +++ b/tests/unit/llms/openai/image_variations/test_openai_image_variations_handler.py @@ -0,0 +1,51 @@ +import json +from io import BytesIO +from typing import Final + +import httpx +import pytest + +import litellm +from litellm.llms.openai.common_utils import OpenAIError + +_UPSTREAM_ERROR_BODY: Final = '{"error":{"message":"image is not a valid PNG","type":"invalid_request_error"}}' + + +def _upstream_rejection(_request: httpx.Request) -> httpx.Response: + return httpx.Response( + 400, + content=_UPSTREAM_ERROR_BODY, + headers={"x-request-id": "req_variation_400", "content-type": "application/json"}, + ) + + +_VARIATION_KWARGS: Final = { + "custom_llm_provider": "openai", + "api_key": "transport-only", + "api_base": "https://upstream.example/v1", + "max_retries": 0, +} + + +async def _request_variation(is_async: bool) -> None: + if is_async: + await litellm.aimage_variation(image=BytesIO(b"image-bytes"), **_VARIATION_KWARGS) + return + litellm.image_variation(image=BytesIO(b"image-bytes"), **_VARIATION_KWARGS) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("is_async", (False, True), ids=("sync", "async")) +async def test_rejected_image_variation_carries_the_upstream_error_response( + is_async: bool, monkeypatch: pytest.MonkeyPatch +) -> None: + transport: Final = httpx.MockTransport(_upstream_rejection) + monkeypatch.setattr(litellm, "client_session", httpx.Client(transport=transport)) + monkeypatch.setattr(litellm, "aclient_session", httpx.AsyncClient(transport=transport)) + + with pytest.raises((litellm.BadRequestError, OpenAIError)) as raised: + await _request_variation(is_async) + + assert raised.value.response.text == _UPSTREAM_ERROR_BODY + assert raised.value.response.json() == json.loads(_UPSTREAM_ERROR_BODY) + assert raised.value.response.headers["x-request-id"] == "req_variation_400" diff --git a/tests/unit/llms/openai/test_openai.py b/tests/unit/llms/openai/test_openai.py index 5b27a446fb3..3e1cbad9308 100644 --- a/tests/unit/llms/openai/test_openai.py +++ b/tests/unit/llms/openai/test_openai.py @@ -1,8 +1,9 @@ import asyncio +import gzip import importlib import json import os -from collections.abc import Callable, Mapping +from collections.abc import Awaitable, Callable, Mapping from types import ModuleType from typing import Final, Literal, Protocol from unittest.mock import MagicMock, Mock, patch @@ -375,6 +376,277 @@ async def test_acompletion_preserves_public_errors_for_both_http_clients( assert not http_client.is_closed +_UPSTREAM_ERROR_BODY: Final = ( + '{"error":{"message":"upstream rejected the request","type":"invalid_request_error",' + '"param":"response_format.type","code":"invalid_value"}}' +) +_UPSTREAM_BASE: Final = "https://upstream.example/v1" +_UPSTREAM_TRANSPORT: Final = httpx.MockTransport( + lambda _request: httpx.Response( + 400, + content=_UPSTREAM_ERROR_BODY, + headers={"x-request-id": "req_upstream_400", "x-upstream-trace": "trace-7", "content-type": "application/json"}, + ) +) +_MESSAGES: Final = [{"role": "user", "content": "hi"}] + + +def _rejecting_client() -> OpenAI: + return OpenAI( + api_key="transport-only", + base_url=_UPSTREAM_BASE, + max_retries=0, + http_client=httpx.Client(transport=_UPSTREAM_TRANSPORT), + ) + + +def _rejecting_async_client() -> AsyncOpenAI: + return AsyncOpenAI( + api_key="transport-only", + base_url=_UPSTREAM_BASE, + max_retries=0, + http_client=httpx.AsyncClient(transport=_UPSTREAM_TRANSPORT), + ) + + +def _upstream(provider: str, model: str) -> dict[str, object]: + return {"model": f"{provider}/{model}", "api_base": _UPSTREAM_BASE, "api_key": "transport-only", "num_retries": 0} + + +async def _completion(provider: str) -> None: + litellm.completion(messages=_MESSAGES, client=_rejecting_client(), **_upstream(provider, "gpt-5.6")) + + +async def _completion_stream(provider: str) -> None: + stream: Final = litellm.completion( + messages=_MESSAGES, client=_rejecting_client(), stream=True, **_upstream(provider, "gpt-5.6") + ) + list(stream) + + +async def _acompletion(provider: str) -> None: + await litellm.acompletion(messages=_MESSAGES, client=_rejecting_async_client(), **_upstream(provider, "gpt-5.6")) + + +async def _acompletion_stream(provider: str) -> None: + stream: Final = await litellm.acompletion( + messages=_MESSAGES, client=_rejecting_async_client(), stream=True, **_upstream(provider, "gpt-5.6") + ) + async for _ in stream: + pass + + +async def _embedding(provider: str) -> None: + litellm.embedding(input=["hi"], client=_rejecting_client(), **_upstream(provider, "text-embedding-3-large")) + + +async def _aembedding(provider: str) -> None: + await litellm.aembedding( + input=["hi"], client=_rejecting_async_client(), **_upstream(provider, "text-embedding-3-large") + ) + + +async def _image_generation(provider: str) -> None: + litellm.image_generation(prompt="a cat", client=_rejecting_client(), **_upstream(provider, "gpt-image-2")) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("provider", ("openai", "litellm_proxy")) +@pytest.mark.parametrize( + "route", + (_completion, _completion_stream, _acompletion, _acompletion_stream, _embedding, _aembedding, _image_generation), + ids=lambda route: route.__name__.lstrip("_"), +) +async def test_mapped_error_carries_the_upstream_error_response( + provider: str, route: Callable[[str], Awaitable[None]] +) -> None: + with pytest.raises(litellm.BadRequestError) as raised: + await route(provider) + + assert raised.value.response.text == _UPSTREAM_ERROR_BODY + assert raised.value.response.json() == json.loads(_UPSTREAM_ERROR_BODY) + assert raised.value.response.headers["x-request-id"] == "req_upstream_400" + assert raised.value.response.headers["x-upstream-trace"] == "trace-7" + + +_UPSTREAM_SERVER_ERROR_BODY: Final = ( + '{"error":{"message":"The server had an error while processing your request.","type":"server_error",' + '"param":null,"code":null}}' +) +_CHUNKED_UPSTREAM_HEADERS: Final = { + "content-type": "application/json", + "transfer-encoding": "chunked", + "x-request-id": "req_upstream_5xx", +} + + +def _chunked_upstream_client(status_code: int, content_encoding: Literal["gzip"] | None) -> httpx.AsyncClient: + body: Final = _UPSTREAM_SERVER_ERROR_BODY.encode() + encoding_headers: Final = {"content-encoding": content_encoding} if content_encoding else {} + return httpx.AsyncClient( + transport=httpx.MockTransport( + lambda _request: httpx.Response( + status_code, + headers={**_CHUNKED_UPSTREAM_HEADERS, **encoding_headers}, + stream=httpx.ByteStream(gzip.compress(body) if content_encoding else body), + ) + ) + ) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("provider", ("openai", "litellm_proxy")) +@pytest.mark.parametrize("stream", (False, True)) +@pytest.mark.parametrize( + ("status_code", "mapped_error"), + ( + (429, litellm.RateLimitError), + (500, litellm.InternalServerError), + (502, litellm.BadGatewayError), + (503, litellm.ServiceUnavailableError), + ), +) +@pytest.mark.parametrize("content_encoding", (None, "gzip")) +async def test_retryable_error_keeps_the_decoded_upstream_body_and_its_headers( + provider: str, + stream: bool, + status_code: int, + mapped_error: type[Exception], + content_encoding: Literal["gzip"] | None, +) -> None: + client: Final = AsyncOpenAI( + api_key="transport-only", + base_url=_UPSTREAM_BASE, + max_retries=0, + http_client=_chunked_upstream_client(status_code, content_encoding), + ) + + with pytest.raises(mapped_error) as raised: + await litellm.acompletion(messages=_MESSAGES, client=client, stream=stream, **_upstream(provider, "gpt-5.6")) + + error_response: Final = getattr(raised.value, "response") + assert error_response.status_code == status_code + assert error_response.text == _UPSTREAM_SERVER_ERROR_BODY + assert error_response.json() == json.loads(_UPSTREAM_SERVER_ERROR_BODY) + assert dict(error_response.headers) == { + **_CHUNKED_UPSTREAM_HEADERS, + **({"content-encoding": content_encoding} if content_encoding else {}), + } + + +_REFUSAL_BODY: Final = '{"error":{"message":"upstream refused","type":"upstream_refusal","param":null,"code":null}}' +_GATEWAY_TIMEOUT_BODY: Final = ( + '{"error":{"message":"litellm.Timeout: APITimeoutError - Request timed out.","type":"invalid_request_error",' + '"param":null,"code":"408"}}' +) +_PROVIDER_KEY: Final = "sk-provider-secret" +_QUERY_SECRET: Final = "query-secret" + + +@pytest.mark.parametrize("provider", ("openai", "litellm_proxy")) +@pytest.mark.parametrize( + ("status_code", "mapped_error", "body"), + ( + (400, litellm.BadRequestError, _REFUSAL_BODY), + (403, litellm.APIError, _REFUSAL_BODY), + (408, litellm.Timeout, _REFUSAL_BODY), + (408, litellm.Timeout, _GATEWAY_TIMEOUT_BODY), + (409, litellm.APIError, _REFUSAL_BODY), + (413, litellm.APIError, _REFUSAL_BODY), + (429, litellm.RateLimitError, _REFUSAL_BODY), + (500, litellm.InternalServerError, _REFUSAL_BODY), + (504, litellm.Timeout, _REFUSAL_BODY), + ), +) +def test_mapped_error_keeps_the_upstream_body_without_the_provider_credentials( + provider: str, status_code: int, mapped_error: type[Exception], body: str +) -> None: + client: Final = OpenAI( + api_key=_PROVIDER_KEY, + base_url=_UPSTREAM_BASE, + default_query={"key": _QUERY_SECRET}, + max_retries=0, + http_client=httpx.Client( + transport=httpx.MockTransport( + lambda _request: httpx.Response( + status_code, content=body, headers={"content-type": "application/json"} + ) + ) + ), + ) + + with pytest.raises(mapped_error) as raised: + litellm.completion( + model=f"{provider}/gpt-5.6", + messages=_MESSAGES, + client=client, + api_base=_UPSTREAM_BASE, + api_key=_PROVIDER_KEY, + num_retries=0, + ) + + error_response: Final = getattr(raised.value, "response") + assert error_response.status_code == status_code + assert error_response.text == body + assert error_response.json() == json.loads(body) + assert str(error_response.request.url) == f"{_UPSTREAM_BASE}/chat/completions" + for exposed in (error_response.request, getattr(raised.value, "request")): + assert _QUERY_SECRET not in str(exposed.url) + assert _PROVIDER_KEY not in str(exposed.headers) + assert exposed.content == b"" + + +def _time_out(request: httpx.Request) -> httpx.Response: + raise httpx.ReadTimeout("no reply", request=request) + + +async def _sync_completion_timing_out(provider: str) -> None: + litellm.completion( + model=f"{provider}/gpt-5.6", + messages=_MESSAGES, + client=OpenAI( + api_key=_PROVIDER_KEY, + base_url=_UPSTREAM_BASE, + max_retries=0, + http_client=httpx.Client(transport=httpx.MockTransport(_time_out)), + ), + api_base=_UPSTREAM_BASE, + api_key=_PROVIDER_KEY, + num_retries=0, + ) + + +async def _async_stream_timing_out(provider: str) -> None: + stream: Final = await litellm.acompletion( + model=f"{provider}/gpt-5.6", + messages=_MESSAGES, + stream=True, + client=AsyncOpenAI( + api_key=_PROVIDER_KEY, + base_url=_UPSTREAM_BASE, + max_retries=0, + http_client=httpx.AsyncClient(transport=httpx.MockTransport(_time_out)), + ), + api_base=_UPSTREAM_BASE, + api_key=_PROVIDER_KEY, + num_retries=0, + ) + async for _ in stream: + pass + + +@pytest.mark.asyncio +@pytest.mark.parametrize("provider", ("openai", "litellm_proxy")) +@pytest.mark.parametrize("call", (_sync_completion_timing_out, _async_stream_timing_out)) +async def test_timeout_before_any_reply_has_no_upstream_response( + provider: str, call: Callable[[str], Awaitable[None]] +) -> None: + with pytest.raises(litellm.Timeout) as raised: + await call(provider) + + assert getattr(raised.value, "response", None) is None + + @pytest.mark.asyncio async def test_acompletion_streams_text_deltas_over_injected_transport( http_backend: ModuleType, sdk_provider: str, monkeypatch: pytest.MonkeyPatch diff --git a/tests/unit/proxy/test_common_request_processing.py b/tests/unit/proxy/test_common_request_processing.py index aeab7a3d348..c867db2fd02 100644 --- a/tests/unit/proxy/test_common_request_processing.py +++ b/tests/unit/proxy/test_common_request_processing.py @@ -2,6 +2,7 @@ import asyncio import contextvars import copy import datetime +import gzip import json from collections.abc import AsyncIterable, Mapping from types import MappingProxyType, SimpleNamespace @@ -13,6 +14,7 @@ from tests.unit.proxy.router_entry_point_stubs import route_entry_points_to_stub import anthropic import httpx +import openai import pytest import respx from fastapi import HTTPException, Request, Response, status @@ -11675,6 +11677,67 @@ async def test_handle_llm_api_exception_forwards_litellm_response_headers_when_r assert exc_info.value.headers["llm_provider-x-request-id"] == "req_openai_400" +_UPSTREAM_REJECTION_BODY: Final = b'{"error":{"message":"upstream said no","type":"upstream_rejection"}}' + + +def _upstream_rejection_wire(content_encoding: Literal["gzip"] | None) -> tuple[bytes, dict[str, str]]: + wire_body: Final = gzip.compress(_UPSTREAM_REJECTION_BODY) if content_encoding else _UPSTREAM_REJECTION_BODY + return wire_body, { + "x-request-id": "req_upstream", + "x-upstream-trace": "trace-7", + "content-type": "application/json", + "content-length": str(len(wire_body)), + **({"content-encoding": content_encoding} if content_encoding else {}), + } + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("status_code", "mapped_error"), + ( + (400, litellm.BadRequestError), + (429, litellm.RateLimitError), + (503, litellm.ServiceUnavailableError), + ), +) +@pytest.mark.parametrize("content_encoding", (None, "gzip")) +async def test_handle_llm_api_exception_forwards_exactly_the_upstream_headers_of_a_mapped_openai_error( + status_code: int, mapped_error: type[Exception], content_encoding: Literal["gzip"] | None +) -> None: + wire_body, upstream_headers = _upstream_rejection_wire(content_encoding) + transport: Final = httpx.MockTransport( + lambda _request: httpx.Response(status_code, content=wire_body, headers=upstream_headers) + ) + with pytest.raises(mapped_error) as mapped: + await litellm.acompletion( + model="openai/gpt-5.6", + messages=[{"role": "user", "content": "hi"}], + num_retries=0, + client=openai.AsyncOpenAI( + api_key="transport-only", + base_url="https://upstream.example/v1", + max_retries=0, + http_client=httpx.AsyncClient(transport=transport), + ), + ) + proxy_logging_obj: Final = MagicMock() + proxy_logging_obj.post_call_failure_hook = AsyncMock(return_value=None) + proxy_logging_obj.post_call_response_headers_hook = AsyncMock(return_value={}) + + with pytest.raises(ProxyException) as forwarded: + await ProxyBaseLLMRequestProcessing(data={}).handle_llm_api_exception( + e=mapped.value, + user_api_key_dict=ProxyUserAPIKeyAuth(api_key="sk-test"), + proxy_logging_obj=proxy_logging_obj, + ) + + assert forwarded.value.code == str(status_code) + assert "upstream said no" in forwarded.value.message + assert {name: value for name, value in forwarded.value.headers.items() if name.startswith("llm_provider-")} == { + f"llm_provider-{name}": value for name, value in upstream_headers.items() + } + + @pytest.mark.asyncio async def test_handle_llm_api_exception_logs_bug_report_for_unmapped_error( monkeypatch: pytest.MonkeyPatch, diff --git a/tests/unit/test_exception_header_preservation.py b/tests/unit/test_exception_header_preservation.py index 7acacf05226..c5a0868f36d 100644 --- a/tests/unit/test_exception_header_preservation.py +++ b/tests/unit/test_exception_header_preservation.py @@ -13,6 +13,7 @@ from typing import Final import httpx import pytest +import litellm from litellm.exceptions import ( BadRequestError, ContentPolicyViolationError, @@ -418,3 +419,21 @@ class TestBedrockErrorHeaders: headers = get_response_headers(dict(_response_headers)) assert headers.get("llm_provider-x-amzn-requestid") == "req-proxy-500" + + +@pytest.mark.parametrize( + "error_class", + (RateLimitError, ServiceUnavailableError, litellm.BadGatewayError, litellm.InternalServerError), +) +def test_error_built_from_an_unread_streaming_response_keeps_its_headers(error_class: type[Exception]) -> None: + unread: Final = httpx.Response( + 500, + headers={"x-request-id": "req-unread"}, + stream=httpx.ByteStream(b'{"error":"never read"}'), + ) + + error: Final = error_class(message="upstream failed", llm_provider="openai", model="gpt-5.6", response=unread) + + error_response: Final = getattr(error, "response") + assert error_response.text == "" + assert dict(error_response.headers) == {"x-request-id": "req-unread"} From dbcacc7ab7880d5115376cd6c22d22898c50d08f Mon Sep 17 00:00:00 2001 From: zoveki <349028329+zoveki@users.noreply.github.com> Date: Thu, 29 Oct 2026 23:34:57 -0700 Subject: [PATCH 07/23] fix(health): probe only models a partial wildcard serves (#49060) * fix(health): probe only models a partial wildcard serves A partial wildcard health check now probes the cheapest cost-map chat models the wildcard matches, keeps provider route prefixes such as Bedrock converse/ on the probe, and falls back only within the wildcard. A wildcard with no matching cost-map chat model reports an error naming model_info.health_check_model * fix(health): keep the provider-wide probe for an unmatched wildcard, skip retired models, and ignore Bedrock ARNs A partial wildcard with no served cost-map chat model falls back to the provider's cheapest chat models instead of failing, and fails with a message naming model_info.health_check_model only when the provider has none. The cheapest-model picker skips models past their cost-map deprecation_date, and a Bedrock ARN segment is never read as a route prefix. * fix(health): fall back to the provider's cheapest probe when a wildcard's own models fail A partial wildcard now probes up to three of its own cheapest chat models first and then the provider-wide cheapest ones, all from one cost-map scan, so stale or unpriced rows the wildcard matches cannot turn a healthy deployment red. Retired models are read by their resolved cost-map key, both deprecation checks use the UTC date, Bedrock only treats converse, invoke, and region names as route segments, and equal-price ties break by model id * fix(health): cap wildcard health probes at three and rank retired cost-map rows last A partial wildcard now probes two of its own cheapest cost-map chat models and then one provider-wide model, so a failing deployment makes no more calls than before. Rows past their cost-map deprecation date go last instead of being dropped, since some still serve, and the shared cheapest-model picker returns what it returned before. * fix(health): probe a wildcard's own models first, then the same three models main probes A partial wildcard's health check tries its two cheapest own cost-map models, then the provider's three cheapest models main already probes, so a row main shows healthy stays healthy. The deprecation check uses the server's local date, like the Baseten listing * fix(health): keep main's provider-wide probe order so equal-price ties match main --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../health_check_helpers.py | 80 +++- .../litellm_core_utils/llm_request_utils.py | 58 ++- .../test_health_check_helpers.py | 343 +++++++++++++++++- .../test_llm_request_utils.py | 84 ++++- .../proxy/test_health_check_max_tokens.py | 7 +- 5 files changed, 545 insertions(+), 27 deletions(-) diff --git a/litellm/litellm_core_utils/health_check_helpers.py b/litellm/litellm_core_utils/health_check_helpers.py index 940730eeb9f..e40299a218f 100644 --- a/litellm/litellm_core_utils/health_check_helpers.py +++ b/litellm/litellm_core_utils/health_check_helpers.py @@ -3,6 +3,8 @@ Helper functions for health check calls. """ import base64 +import itertools +import re from collections.abc import Awaitable, Callable, Mapping from types import MappingProxyType from typing import TYPE_CHECKING, Final, Literal @@ -26,6 +28,9 @@ IMAGE_EDIT_HEALTH_CHECK_PROMPT: Final = ( ) ANTHROPIC_MESSAGES_HEALTH_CHECK_MAX_TOKENS: Final = 16 +WILDCARD_HEALTH_CHECK_SERVED_MODEL_COUNT: Final = 2 +WILDCARD_HEALTH_CHECK_PROVIDER_MODEL_COUNT: Final = 3 +BEDROCK_ANY_MODEL_ROUTES: Final = frozenset({"converse", "invoke"}) EXTRACTION_HEALTH_CHECK_FIELDS: Final[Mapping[str, Mapping[str, str]]] = MappingProxyType( {"reachable": MappingProxyType({"type": "text", "instructions": "Any word from the text"})} ) @@ -90,6 +95,62 @@ def _ocr_health_check_document(model: str, custom_llm_provider: str) -> Document return native(model, custom_llm_provider) +def _provider_base_model(custom_llm_provider: str, model: str) -> str | None: + from litellm.types.utils import LlmProviders + from litellm.utils import ProviderConfigManager + + try: + model_info: Final = ProviderConfigManager.get_provider_model_info( + model=None, provider=LlmProviders(custom_llm_provider) + ) + return None if model_info is None else model_info.get_base_model(model) + except (ValueError, NotImplementedError): + return None + + +def _is_route_segment(custom_llm_provider: str, segment: str, model_after_segment: str) -> bool: + if custom_llm_provider == "bedrock": + from litellm.llms.bedrock.common_utils import get_all_bedrock_regions + + return segment in BEDROCK_ANY_MODEL_ROUTES or segment in get_all_bedrock_regions() + base_model: Final = _provider_base_model(custom_llm_provider, f"{segment}/{model_after_segment}") + return base_model is not None and not base_model.startswith(f"{segment}/") + + +def wildcard_health_check_models(wildcard: str, custom_llm_provider: str) -> tuple[str, ...]: + from litellm.litellm_core_utils.llm_request_utils import rank_chat_models_from_llm_provider + from litellm.router_utils.pattern_match_deployments import PatternMatchRouter + + provider_prefix: Final = f"{custom_llm_provider}/" + segments: Final = wildcard.removeprefix(provider_prefix).split("/") + route_segments: Final = tuple( + itertools.takewhile( + lambda segment_index: _is_route_segment( + custom_llm_provider, segments[segment_index], "/".join(segments[segment_index + 1 :]) + ), + range(len(segments) - 1), + ) + ) + route_prefix: Final = "".join(f"{segments[segment_index]}/" for segment_index in route_segments) + model_regex: Final = re.compile(PatternMatchRouter().pattern_to_regex("/".join(segments[len(route_segments) :]))) + ranked_model_ids: Final = tuple( + (ranked_model.model.removeprefix(provider_prefix), ranked_model) + for ranked_model in rank_chat_models_from_llm_provider(custom_llm_provider) + ) + served_model_ids: Final = sorted( + (ranked for ranked in ranked_model_ids if model_regex.fullmatch(ranked[0])), + key=lambda ranked: (ranked[1].is_past_deprecation_date, ranked[1].route_rank, ranked[1].cost, ranked[0]), + ) + served_probes: Final = itertools.islice( + dict.fromkeys(f"{provider_prefix}{route_prefix}{model_id}" for model_id, _ in served_model_ids), + WILDCARD_HEALTH_CHECK_SERVED_MODEL_COUNT, + ) + provider_probes: Final = ( + f"{provider_prefix}{model_id}" for model_id, _ in ranked_model_ids[:WILDCARD_HEALTH_CHECK_PROVIDER_MODEL_COUNT] + ) + return tuple(dict.fromkeys((*served_probes, *provider_probes))) + + class HealthCheckHelpers: @staticmethod async def ahealth_check_wildcard_models( @@ -99,23 +160,16 @@ class HealthCheckHelpers: litellm_logging_obj: "Logging", ) -> dict: from litellm import acompletion - from litellm.litellm_core_utils.llm_request_utils import ( - pick_cheapest_chat_models_from_llm_provider, - ) - # this is a wildcard model, we need to pick a random model from the provider - cheapest_models = pick_cheapest_chat_models_from_llm_provider(custom_llm_provider=custom_llm_provider, n=3) - if len(cheapest_models) == 0: + probe_models: Final = wildcard_health_check_models(wildcard=model, custom_llm_provider=custom_llm_provider) + if len(probe_models) == 0: raise Exception( - f"Unable to health check wildcard model for provider {custom_llm_provider}. Add a model on your config.yaml or contribute here - https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json" + f"Unable to health check wildcard model {model} for provider {custom_llm_provider}: the cost map has " + "no chat model for this provider. Set model_info.health_check_model to a model this wildcard serves" ) - if len(cheapest_models) > 1: - fallback_models = cheapest_models[1:] # Pick the last 2 models from the shuffled list - else: - fallback_models = None - model_params["model"] = cheapest_models[0] + model_params["model"] = probe_models[0] model_params["litellm_logging_obj"] = litellm_logging_obj - model_params["fallbacks"] = fallback_models + model_params["fallbacks"] = list(probe_models[1:]) or None model_params["max_tokens"] = model_params.get("max_tokens", 16) # GPT-5 models require max_output_tokens >= 16 await acompletion(**model_params) return {} diff --git a/litellm/litellm_core_utils/llm_request_utils.py b/litellm/litellm_core_utils/llm_request_utils.py index db881aa16c6..9c2ec0b8984 100644 --- a/litellm/litellm_core_utils/llm_request_utils.py +++ b/litellm/litellm_core_utils/llm_request_utils.py @@ -1,6 +1,11 @@ +import datetime from collections.abc import Iterable, Iterator, Mapping +from dataclasses import dataclass from typing import Final, cast +from pydantic import TypeAdapter +from typing_extensions import NotRequired, ReadOnly, TypedDict + import litellm from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH from litellm.litellm_core_utils.get_llm_provider_logic import inferred_provider @@ -151,19 +156,52 @@ def pick_cheapest_chat_models_from_llm_provider(custom_llm_provider: str, n=1): Returns: list[str]: A list of the n cheapest chat models. """ - if custom_llm_provider not in litellm.models_by_provider: - return [] + return [ranked_model.model for ranked_model in rank_chat_models_from_llm_provider(custom_llm_provider)[:n]] + +@dataclass(frozen=True, slots=True) +class RankedChatModel: + model: str + route_rank: int + cost: float + is_past_deprecation_date: bool + + +def rank_chat_models_from_llm_provider(custom_llm_provider: str) -> tuple[RankedChatModel, ...]: known_models: Final[Iterable[str]] = litellm.models_by_provider.get(custom_llm_provider, ()) - ranked_rows: Final = ( - (route_rank, cost, model) - for model, cost in _chat_row_costs(known_models, custom_llm_provider) + ranked_models: Final = ( + RankedChatModel( + model=model, route_rank=route_rank, cost=cost, is_past_deprecation_date=is_past_deprecation_date + ) + for model, cost, is_past_deprecation_date in _chat_row_costs(known_models, custom_llm_provider) if (route_rank := _route_rank(model, custom_llm_provider)) is not None ) - return [model for _, _, model in sorted(ranked_rows, key=lambda row: row[:2])[:n]] + return tuple(sorted(ranked_models, key=lambda ranked_model: (ranked_model.route_rank, ranked_model.cost))) -def _chat_row_costs(models: Iterable[str], custom_llm_provider: str) -> Iterator[tuple[str, float]]: +class _CostMapDeprecation(TypedDict): + deprecation_date: NotRequired[ReadOnly[object]] + + +_COST_MAP_DEPRECATION: Final = TypeAdapter(_CostMapDeprecation) + + +def _is_past_deprecation_date(cost_map_key: str) -> bool: + deprecation_date: Final = _COST_MAP_DEPRECATION.validate_python( + litellm.model_cost.get(cost_map_key) or {} # pyright: ignore[reportUnknownMemberType] # untyped cost map JSON + ).get("deprecation_date") + if not isinstance(deprecation_date, str): + return False + try: + return ( + datetime.date.fromisoformat(deprecation_date) + <= datetime.datetime.now(datetime.timezone.utc).astimezone().date() + ) + except ValueError: + return False + + +def _chat_row_costs(models: Iterable[str], custom_llm_provider: str) -> Iterator[tuple[str, float, bool]]: for model in models: try: model_info = litellm.get_model_info(model=model, custom_llm_provider=custom_llm_provider) @@ -171,7 +209,11 @@ def _chat_row_costs(models: Iterable[str], custom_llm_provider: str) -> Iterator continue if model_info.get("mode") != "chat": continue - yield model, (model_info.get("input_cost_per_token") or 0.0) + (model_info.get("output_cost_per_token") or 0.0) + yield ( + model, + (model_info.get("input_cost_per_token") or 0.0) + (model_info.get("output_cost_per_token") or 0.0), + _is_past_deprecation_date(model_info["key"]), + ) def _route_rank(model: str, custom_llm_provider: str) -> int | None: diff --git a/tests/unit/litellm_core_utils/test_health_check_helpers.py b/tests/unit/litellm_core_utils/test_health_check_helpers.py index 6979d72ff14..950a91beb81 100644 --- a/tests/unit/litellm_core_utils/test_health_check_helpers.py +++ b/tests/unit/litellm_core_utils/test_health_check_helpers.py @@ -8,6 +8,7 @@ from types import MappingProxyType from typing import Final, Literal from unittest.mock import AsyncMock, MagicMock, patch +import httpx import pytest import responses import respx @@ -19,6 +20,7 @@ from litellm.litellm_core_utils.health_check_helpers import ( HealthCheckHelpers, default_health_check_mode, native_health_check_mode, + wildcard_health_check_models, ) from litellm.main import ahealth_check from litellm.proxy._types import UserAPIKeyAuth @@ -1449,6 +1451,339 @@ async def test_groq_wildcard_health_check_uses_a_concrete_model( assert response == {} +_WILDCARD_CATALOG: Final = MappingProxyType( + { + "databricks": ( + "databricks/databricks-claude-fixture-d", + "databricks/databricks-claude-fixture-c", + "databricks/databricks-claude-fixture-b", + "databricks/databricks-claude-fixture-a", + "databricks/databricks-cheap-chat", + "databricks/databricks-claude-fixture-retired", + ), + "bedrock": ("meta.llama-fixture", "anthropic.claude-fixture"), + "cloudflare": ("cloudflare/@cf/meta/fixture-chat",), + "vertex_ai": ("claude-fixture@default",), + } +) +_WILDCARD_CATALOG_COST: Final = MappingProxyType( + { + "databricks/databricks-claude-fixture-retired": 1e-10, + "databricks/databricks-cheap-chat": 1e-9, + "databricks/databricks-claude-fixture-a": 2e-9, + "databricks/databricks-claude-fixture-b": 3e-9, + "databricks/databricks-claude-fixture-c": 4e-9, + "databricks/databricks-claude-fixture-d": 5e-9, + "meta.llama-fixture": 1e-9, + "anthropic.claude-fixture": 2e-9, + "cloudflare/@cf/meta/fixture-chat": 1e-9, + "claude-fixture@default": 1e-9, + } +) +_DATABRICKS_FIXTURE_API_BASE: Final = "https://fixture.cloud.databricks.com/serving-endpoints" + + +@pytest.fixture +def wildcard_catalog(monkeypatch: pytest.MonkeyPatch) -> None: + providers: Final = {model: provider for provider, models in _WILDCARD_CATALOG.items() for model in models} + litellm.register_model( + { + model: { + "litellm_provider": providers[model], + "mode": "chat", + "max_tokens": 1024, + "input_cost_per_token": cost, + "output_cost_per_token": cost, + **({"deprecation_date": "2000-01-01"} if model.endswith("-retired") else {}), + } + for model, cost in _WILDCARD_CATALOG_COST.items() + } + ) + for provider, models in _WILDCARD_CATALOG.items(): + monkeypatch.setitem(litellm.models_by_provider, provider, list(models)) + + +_DATABRICKS_PROVIDER_PROBES: Final = ( + "databricks/databricks-claude-fixture-retired", + "databricks/databricks-cheap-chat", + "databricks/databricks-claude-fixture-a", +) +_BEDROCK_PROVIDER_PROBES: Final = ("bedrock/meta.llama-fixture", "bedrock/anthropic.claude-fixture") +_DATABRICKS_SERVED_CLAUDE_PROBES: Final = ( + "databricks/databricks-claude-fixture-a", + "databricks/databricks-claude-fixture-b", +) + + +@pytest.mark.parametrize( + ("custom_llm_provider", "wildcard", "expected"), + [ + ("databricks", "databricks-claude-*", (*_DATABRICKS_SERVED_CLAUDE_PROBES, *_DATABRICKS_PROVIDER_PROBES[:2])), + ( + "databricks", + "databricks/databricks-claude-*", + (*_DATABRICKS_SERVED_CLAUDE_PROBES, *_DATABRICKS_PROVIDER_PROBES[:2]), + ), + ("databricks", "*", (*_DATABRICKS_PROVIDER_PROBES[1:], _DATABRICKS_PROVIDER_PROBES[0])), + ("databricks", "databricks-claude-fixture-r*", _DATABRICKS_PROVIDER_PROBES), + ("databricks", "databricks-llama-*", _DATABRICKS_PROVIDER_PROBES), + ("databricks", "*-fixture", _DATABRICKS_PROVIDER_PROBES), + ("databricks", "team-endpoints/*", _DATABRICKS_PROVIDER_PROBES), + ("bedrock", "converse/anthropic.*", ("bedrock/converse/anthropic.claude-fixture", *_BEDROCK_PROVIDER_PROBES)), + ( + "bedrock", + "converse/*", + ( + "bedrock/converse/meta.llama-fixture", + "bedrock/converse/anthropic.claude-fixture", + *_BEDROCK_PROVIDER_PROBES, + ), + ), + ("bedrock", "invoke/anthropic.*", ("bedrock/invoke/anthropic.claude-fixture", *_BEDROCK_PROVIDER_PROBES)), + ("bedrock", "us-east-1/anthropic.*", ("bedrock/us-east-1/anthropic.claude-fixture", *_BEDROCK_PROVIDER_PROBES)), + ("bedrock", "agent/anthropic.*", _BEDROCK_PROVIDER_PROBES), + ("bedrock", "nova/*", _BEDROCK_PROVIDER_PROBES), + ("bedrock", "nova-2/*", _BEDROCK_PROVIDER_PROBES), + ("bedrock", "mantle/*", _BEDROCK_PROVIDER_PROBES), + ("bedrock", "openai/*", _BEDROCK_PROVIDER_PROBES), + ("bedrock", "chat_completions/*", _BEDROCK_PROVIDER_PROBES), + ("bedrock", "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/*", _BEDROCK_PROVIDER_PROBES), + ( + "bedrock", + "converse/arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/*", + _BEDROCK_PROVIDER_PROBES, + ), + ("cloudflare", "@cf/*", ("cloudflare/@cf/meta/fixture-chat",)), + ("vertex_ai", "claude-*", ("vertex_ai/claude-fixture@default",)), + ("vertex_ai", "meta/*", ("vertex_ai/claude-fixture@default",)), + ("fixture-gateway", "team/*", ()), + ], +) +@pytest.mark.usefixtures("wildcard_catalog") +def test_wildcard_health_check_probes_the_cheapest_served_chat_models_before_the_provider_cheapest( + custom_llm_provider: str, wildcard: str, expected: tuple[str, ...] +) -> None: + assert wildcard_health_check_models(wildcard=wildcard, custom_llm_provider=custom_llm_provider) == expected + + +def _fail_first_databricks_claude_probe(request: httpx.Request) -> httpx.Response: + model: Final = json.loads(request.content)["model"] + if model == "databricks-claude-fixture-a": + return httpx.Response(500, json={"error_code": "INTERNAL_ERROR", "message": "fixture outage"}) + return httpx.Response( + 200, + json={ + "id": "chatcmpl-health", + "object": "chat.completion", + "created": 1, + "model": model, + "choices": [{"index": 0, "message": {"role": "assistant", "content": "2"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + }, + ) + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("wildcard_catalog") +async def test_partial_wildcard_health_check_falls_back_within_the_wildcard_before_any_other_model( + monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + litellm.in_memory_llm_clients_cache.flush_cache() + upstream: Final = respx_mock.post(f"{_DATABRICKS_FIXTURE_API_BASE}/chat/completions").mock( + side_effect=_fail_first_databricks_claude_probe + ) + + response: Final = await ahealth_check( + { + "model": "databricks/databricks-claude-*", + "api_key": "fake-key", + "api_base": _DATABRICKS_FIXTURE_API_BASE, + "messages": [{"role": "user", "content": "What is 1 + 1?"}], + } + ) + + probed_models: Final = [json.loads(call.request.content)["model"] for call in upstream.calls] + assert probed_models[0] == "databricks-claude-fixture-a" + assert probed_models[-1] == "databricks-claude-fixture-b" + assert all(model.startswith("databricks-claude-") for model in probed_models) + assert "databricks-claude-fixture-a" not in probed_models[1:] + assert response == {} + + +def _answer_only_the_retired_databricks_claude_model(request: httpx.Request) -> httpx.Response: + model: Final = json.loads(request.content)["model"] + if model != "databricks-claude-fixture-retired": + return httpx.Response(404, json={"error_code": "ENDPOINT_NOT_FOUND", "message": f"{model} does not exist"}) + return _fail_first_databricks_claude_probe(request) + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("wildcard_catalog") +async def test_partial_wildcard_health_check_stays_healthy_while_any_model_the_provider_probes_answers( + monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + litellm.in_memory_llm_clients_cache.flush_cache() + upstream: Final = respx_mock.post(f"{_DATABRICKS_FIXTURE_API_BASE}/chat/completions").mock( + side_effect=_answer_only_the_retired_databricks_claude_model + ) + + response: Final = await ahealth_check( + { + "model": "databricks/databricks-claude-*", + "api_key": "fake-key", + "api_base": _DATABRICKS_FIXTURE_API_BASE, + "messages": [{"role": "user", "content": "What is 1 + 1?"}], + } + ) + + assert [json.loads(call.request.content)["model"] for call in upstream.calls] == [ + "databricks-claude-fixture-a", + "databricks-claude-fixture-b", + "databricks-claude-fixture-retired", + ] + assert response == {} + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("wildcard_catalog") +async def test_partial_wildcard_health_check_that_fails_on_every_probe_tries_its_own_models_then_the_provider_probes( + monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + litellm.in_memory_llm_clients_cache.flush_cache() + upstream: Final = respx_mock.post(f"{_DATABRICKS_FIXTURE_API_BASE}/chat/completions").respond( + 401, json={"error_code": "UNAUTHENTICATED", "message": "fixture bad token"} + ) + + response: Final = await ahealth_check( + { + "model": "databricks/databricks-claude-*", + "api_key": "fake-key", + "api_base": _DATABRICKS_FIXTURE_API_BASE, + "messages": [{"role": "user", "content": "What is 1 + 1?"}], + } + ) + + assert [json.loads(call.request.content)["model"] for call in upstream.calls] == [ + "databricks-claude-fixture-a", + "databricks-claude-fixture-b", + "databricks-claude-fixture-retired", + "databricks-cheap-chat", + ] + assert "error" in response + + +_DATABRICKS_EQUAL_PRICE_MODELS: Final = ( + "databricks/databricks-free-d", + "databricks/databricks-free-c", + "databricks/databricks-free-b", + "databricks/databricks-free-a", +) + + +def _register_equal_price_databricks_models(monkeypatch: pytest.MonkeyPatch, listed_order: tuple[str, ...]) -> None: + litellm.register_model( + { + model: { + "litellm_provider": "databricks", + "mode": "chat", + "max_tokens": 1024, + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + for model in listed_order + } + ) + monkeypatch.setitem(litellm.models_by_provider, "databricks", list(listed_order)) + + +@pytest.mark.parametrize("listed_order", [_DATABRICKS_EQUAL_PRICE_MODELS, _DATABRICKS_EQUAL_PRICE_MODELS[::-1]]) +def test_partial_wildcard_health_check_orders_its_own_equal_price_models_by_id_whatever_the_listing_order( + monkeypatch: pytest.MonkeyPatch, listed_order: tuple[str, ...] +) -> None: + _register_equal_price_databricks_models(monkeypatch, listed_order) + + assert wildcard_health_check_models(wildcard="databricks/databricks-free-*", custom_llm_provider="databricks")[ + :2 + ] == ("databricks/databricks-free-a", "databricks/databricks-free-b") + + +def _answer_only_the_first_listed_databricks_free_model(request: httpx.Request) -> httpx.Response: + model: Final = json.loads(request.content)["model"] + if model != "databricks-free-d": + return httpx.Response(404, json={"error_code": "ENDPOINT_NOT_FOUND", "message": f"{model} does not exist"}) + return _fail_first_databricks_claude_probe(request) + + +@pytest.mark.asyncio +async def test_wildcard_health_check_falls_back_to_equal_price_provider_models_in_the_provider_listing_order( + monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter +) -> None: + _register_equal_price_databricks_models(monkeypatch, _DATABRICKS_EQUAL_PRICE_MODELS) + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + litellm.in_memory_llm_clients_cache.flush_cache() + upstream: Final = respx_mock.post(f"{_DATABRICKS_FIXTURE_API_BASE}/chat/completions").mock( + side_effect=_answer_only_the_first_listed_databricks_free_model + ) + + response: Final = await ahealth_check( + { + "model": "databricks/databricks-claude-*", + "api_key": "fake-key", + "api_base": _DATABRICKS_FIXTURE_API_BASE, + "messages": [{"role": "user", "content": "What is 1 + 1?"}], + } + ) + + assert [json.loads(call.request.content)["model"] for call in upstream.calls] == ["databricks-free-d"] + assert response == {} + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("wildcard_catalog") +async def test_partial_wildcard_health_check_without_a_served_cost_map_model_probes_the_provider_cheapest_chat_model( + monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + litellm.in_memory_llm_clients_cache.flush_cache() + upstream: Final = respx_mock.post(f"{_DATABRICKS_FIXTURE_API_BASE}/chat/completions").mock( + side_effect=_fail_first_databricks_claude_probe + ) + + response: Final = await ahealth_check( + { + "model": "databricks/databricks-llama-*", + "api_key": "fake-key", + "api_base": _DATABRICKS_FIXTURE_API_BASE, + "messages": [{"role": "user", "content": "What is 1 + 1?"}], + } + ) + + assert json.loads(upstream.calls[0].request.content)["model"] == "databricks-claude-fixture-retired" + assert response == {} + + +@pytest.mark.asyncio +async def test_partial_wildcard_health_check_on_a_provider_without_cost_map_chat_models_names_health_check_model( + monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter +) -> None: + monkeypatch.setitem(litellm.models_by_provider, "databricks", []) + + response: Final = await ahealth_check( + { + "model": "databricks/databricks-llama-*", + "api_key": "fake-key", + "api_base": _DATABRICKS_FIXTURE_API_BASE, + "messages": [{"role": "user", "content": "What is 1 + 1?"}], + } + ) + + assert "model_info.health_check_model" in response["error"], response["error"] + assert respx_mock.calls.call_count == 0 + + @pytest.mark.asyncio async def test_cohere_rerank_health_check_returns_a_successful_response( monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter @@ -1632,7 +1967,12 @@ async def test_wildcard_health_check_with_pinned_provider_calls_a_model_without_ monkeypatch.setitem( litellm.model_cost, "fake-together-chat", - {"input_cost_per_token": 1e-9, "output_cost_per_token": 1e-9, "mode": "chat", "litellm_provider": "together_ai"}, + { + "input_cost_per_token": 1e-9, + "output_cost_per_token": 1e-9, + "mode": "chat", + "litellm_provider": "together_ai", + }, ) monkeypatch.setitem(litellm.models_by_provider, "together_ai", ("fake-together-chat",)) litellm.in_memory_llm_clients_cache.flush_cache() @@ -1647,4 +1987,3 @@ async def test_wildcard_health_check_with_pinned_provider_calls_a_model_without_ assert "error" not in result, result assert json.loads(together_route.calls.last.request.content)["model"] == "fake-together-chat" assert not other_hosts.called - diff --git a/tests/unit/litellm_core_utils/test_llm_request_utils.py b/tests/unit/litellm_core_utils/test_llm_request_utils.py index 0a2bd5cec1b..e7350fae395 100644 --- a/tests/unit/litellm_core_utils/test_llm_request_utils.py +++ b/tests/unit/litellm_core_utils/test_llm_request_utils.py @@ -8,6 +8,7 @@ from litellm.litellm_core_utils.llm_request_utils import ( ensure_extra_body_is_safe, flatten_form_field_values, pick_cheapest_chat_models_from_llm_provider, + rank_chat_models_from_llm_provider, serialize_multipart_form_fields, ) @@ -149,6 +150,24 @@ def deepseek_rows_whose_cheapest_ids_route_elsewhere(monkeypatch: pytest.MonkeyP "deepseek/fake-cheapest-embedding": {"input_cost_per_token": 1e-13, "mode": "embedding"}, "openai/fake-free-chat": {"input_cost_per_token": 0.0, "output_cost_per_token": 0.0, "mode": "chat"}, "deepseek/fake-cheap-chat": {"input_cost_per_token": 1e-12, "output_cost_per_token": 1e-12, "mode": "chat"}, + "deepseek/fake-retired-chat": { + "input_cost_per_token": 1e-14, + "output_cost_per_token": 1e-14, + "mode": "chat", + "deprecation_date": "2000-01-01", + }, + "deepseek/fake-retiring-chat": { + "input_cost_per_token": 2e-12, + "output_cost_per_token": 2e-12, + "mode": "chat", + "deprecation_date": "2999-01-01", + }, + "deepseek/fake-misdated-chat": { + "input_cost_per_token": 2.5e-12, + "output_cost_per_token": 2.5e-12, + "mode": "chat", + "deprecation_date": "next spring", + }, "fake-free-chat-without-provider": {"input_cost_per_token": 0.0, "output_cost_per_token": 0.0, "mode": "chat"}, } for name, row in rows.items(): @@ -156,12 +175,73 @@ def deepseek_rows_whose_cheapest_ids_route_elsewhere(monkeypatch: pytest.MonkeyP monkeypatch.setitem(litellm.models_by_provider, PICKER_PROVIDER, tuple(rows)) -def test_pick_cheapest_chat_models_prefers_provider_routed_ids_and_skips_ids_routed_elsewhere( +def test_pick_cheapest_chat_models_prefers_provider_routed_ids_skips_ids_routed_elsewhere_and_keeps_retired_ones( deepseek_rows_whose_cheapest_ids_route_elsewhere: None, ) -> None: assert pick_cheapest_chat_models_from_llm_provider(PICKER_PROVIDER, n=10) == [ + "deepseek/fake-retired-chat", "deepseek/fake-cheap-chat", + "deepseek/fake-retiring-chat", + "deepseek/fake-misdated-chat", "deepseek/fake-pricey-chat", "fake-free-chat-without-provider", ] - assert pick_cheapest_chat_models_from_llm_provider(PICKER_PROVIDER, n=1) == ["deepseek/fake-cheap-chat"] + assert pick_cheapest_chat_models_from_llm_provider(PICKER_PROVIDER, n=1) == ["deepseek/fake-retired-chat"] + + +def test_rank_chat_models_flags_only_the_rows_whose_cost_map_deprecation_date_has_passed( + deepseek_rows_whose_cheapest_ids_route_elsewhere: None, +) -> None: + assert { + ranked_model.model: ranked_model.is_past_deprecation_date + for ranked_model in rank_chat_models_from_llm_provider(PICKER_PROVIDER) + } == { + "deepseek/fake-retired-chat": True, + "deepseek/fake-cheap-chat": False, + "deepseek/fake-retiring-chat": False, + "deepseek/fake-misdated-chat": False, + "deepseek/fake-pricey-chat": False, + "fake-free-chat-without-provider": False, + } + + +@pytest.mark.parametrize( + "listed_order", + [("deepseek/fake-tie-b", "deepseek/fake-tie-a"), ("deepseek/fake-tie-a", "deepseek/fake-tie-b")], +) +def test_pick_cheapest_chat_models_keeps_the_provider_listing_order_for_equal_price_models( + monkeypatch: pytest.MonkeyPatch, listed_order: tuple[str, str] +) -> None: + for name in listed_order: + monkeypatch.setitem( + litellm.model_cost, + name, + { + "input_cost_per_token": 1e-12, + "output_cost_per_token": 1e-12, + "mode": "chat", + "litellm_provider": PICKER_PROVIDER, + }, + ) + monkeypatch.setitem(litellm.models_by_provider, PICKER_PROVIDER, listed_order) + + assert pick_cheapest_chat_models_from_llm_provider(PICKER_PROVIDER, n=2) == list(listed_order) + + +def test_rank_chat_models_flags_a_retired_model_the_provider_lists_without_its_cost_map_prefix( + monkeypatch: pytest.MonkeyPatch, +) -> None: + rows: Final = { + "vertex_ai/fake-retired-vertex-chat": {"input_cost_per_token": 1e-14, "deprecation_date": "2000-01-01"}, + "vertex_ai/fake-live-vertex-chat": {"input_cost_per_token": 1e-12, "deprecation_date": "2999-01-01"}, + } + for name, row in rows.items(): + monkeypatch.setitem(litellm.model_cost, name, {**row, "mode": "chat", "litellm_provider": "vertex_ai"}) + monkeypatch.setitem( + litellm.models_by_provider, "vertex_ai", tuple(name.removeprefix("vertex_ai/") for name in rows) + ) + + assert [ + (ranked_model.model, ranked_model.is_past_deprecation_date) + for ranked_model in rank_chat_models_from_llm_provider("vertex_ai") + ] == [("fake-retired-vertex-chat", True), ("fake-live-vertex-chat", False)] diff --git a/tests/unit/proxy/test_health_check_max_tokens.py b/tests/unit/proxy/test_health_check_max_tokens.py index e48bf4edc09..df0429316ed 100644 --- a/tests/unit/proxy/test_health_check_max_tokens.py +++ b/tests/unit/proxy/test_health_check_max_tokens.py @@ -11,6 +11,7 @@ import respx import litellm from litellm.litellm_core_utils.health_check_helpers import HealthCheckHelpers +from litellm.litellm_core_utils.llm_request_utils import RankedChatModel from litellm.proxy import health_check as hc_module from litellm.proxy.health_check import ( _is_strategy_router_deployment, @@ -85,8 +86,10 @@ async def test_ahealth_check_wildcard_models_respects_max_tokens(): """ with ( patch( - "litellm.litellm_core_utils.llm_request_utils.pick_cheapest_chat_models_from_llm_provider", - return_value=["gpt-4o-mini"], + "litellm.litellm_core_utils.llm_request_utils.rank_chat_models_from_llm_provider", + return_value=( + RankedChatModel(model="gpt-4o-mini", route_rank=0, cost=0.0, is_past_deprecation_date=False), + ), ), patch("litellm.acompletion", new_callable=AsyncMock), ): From 4a5c5665faec4efd3bd2dc31436c9f6f8b25cf2e Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Thu, 29 Oct 2026 23:49:55 -0700 Subject: [PATCH 08/23] fix(azure_ai)!: keep OpenAI prompt cache breakpoints on the Foundry chat completions path (#45299) * fix(azure_ai): keep OpenAI prompt cache breakpoints on the Foundry chat completions path * fix(azure_ai): keep only OpenAI content part types when preserving prompt cache breakpoints * fix(azure_ai): share the breakpoint check with the cache-mark step and keep only text parts A deployment's model_info opt-in is registered under its provider-prefixed cost-map key without a litellm_provider, so that prefix now names the provider the flag speaks for. * refactor(azure_ai): read text parts through a walrus and mark new test values Final * fix(azure_ai): read a router-registered breakpoint opt-in only for azure_ai and keep a message mark off dropped file parts * fix(azure_ai): read the breakpoint opt-in from each deployment's own model_info and base_model A router deployment without its own supports_prompt_cache_breakpoint flag now reads the cost map as loaded, so a sibling's opt-in on the same model string no longer turns its requests into 400s. The shared hook no longer reads router registered keys for azure_ai. * fix(azure_ai): send breakpoints in Foundry's shape under drop_params and fall back to the model name * test(integration): expect the injected breakpoint on the Foundry GPT-6 chat wire --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../anthropic_cache_control_hook.py | 33 +- litellm/llms/azure_ai/chat/transformation.py | 111 +++++- ...ect_prompt_cache_breakpoint_owned_proxy.py | 17 +- .../chat/test_azure_ai_transformation.py | 356 ++++++++++++++++++ 4 files changed, 499 insertions(+), 18 deletions(-) diff --git a/litellm/integrations/anthropic_cache_control_hook.py b/litellm/integrations/anthropic_cache_control_hook.py index fc1f1830129..45c82589d2d 100644 --- a/litellm/integrations/anthropic_cache_control_hook.py +++ b/litellm/integrations/anthropic_cache_control_hook.py @@ -23,6 +23,7 @@ from litellm._logging import verbose_logger from litellm.integrations.custom_logger import CustomLogger from litellm.integrations.custom_prompt_management import CustomPromptManagement from litellm.integrations.prompt_management_base import PromptManagementClient +from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap from litellm.litellm_core_utils.prompt_templates.common_utils import ( cache_control_as_prompt_cache_breakpoint, with_prompt_cache_breakpoint, @@ -175,7 +176,10 @@ def _hosted_entry_flag(entry: Mapping[str, object], resolve_provider: Callable[[ def _hosted_openai_dialect_flag( - model: str, custom_llm_provider: str | None, resolve_provider: Callable[[str], str | None] + model: str, + custom_llm_provider: str | None, + resolve_provider: Callable[[str], str | None], + model_cost: Mapping[str, Mapping[str, object]] | None = None, ) -> bool | None: """ Explicit opt-in for an OpenAI-shaped deployment served by another provider. @@ -194,14 +198,19 @@ def _hosted_openai_dialect_flag( """ import litellm - exact_entry: Final = litellm.model_cost.get(model) - exact_provider: Final = exact_entry.get("litellm_provider") if isinstance(exact_entry, dict) else None + cost_map: Final = ( + cast(Mapping[str, Mapping[str, object]], litellm.model_cost) # cast-ok: the live map holds catalog entries + if model_cost is None + else model_cost + ) + exact_entry: Final = cost_map.get(model) + exact_provider: Final = exact_entry.get("litellm_provider") if isinstance(exact_entry, Mapping) else None if custom_llm_provider is None and (exact_provider == "openai" or ("/" not in model and exact_provider is None)): return None provider: Final = custom_llm_provider or resolve_provider(model) if provider is None or provider == "openai": return None - if isinstance(exact_entry, dict) and exact_provider == provider: + if isinstance(exact_entry, Mapping) and exact_provider == provider: return _hosted_entry_flag(exact_entry, lambda: provider) from litellm.utils import get_potential_model_names @@ -214,8 +223,8 @@ def _hosted_openai_dialect_flag( names["stripped_model_name"], names["provider_prefixed_model_name"], ) - entries: Final = (litellm.model_cost.get(candidate) for candidate in candidates) - flags: Final = (_hosted_entry_flag(entry, lambda: provider) for entry in entries if isinstance(entry, dict)) + entries: Final = (cost_map.get(candidate) for candidate in candidates) + flags: Final = (_hosted_entry_flag(entry, lambda: provider) for entry in entries if isinstance(entry, Mapping)) return next((flag for flag in flags if flag is not None), None) @@ -537,6 +546,18 @@ class AnthropicCacheControlHook(CustomPromptManagement): return False return prompt_cache_options is not None or targets_openai_api(api_base) + @staticmethod + def catalog_targets_openai_prompt_cache_breakpoint(model: str, custom_llm_provider: str) -> bool: + return ( + _hosted_openai_dialect_flag( + model, + custom_llm_provider, + AnthropicCacheControlHook._resolve_provider, + GetModelCostMap.loaded_model_cost_map(), + ) + is True + ) + @staticmethod def _resolve_provider(model: str) -> str | None: return _served_provider_for_model(model) diff --git a/litellm/llms/azure_ai/chat/transformation.py b/litellm/llms/azure_ai/chat/transformation.py index d382d653a53..5381447f28f 100644 --- a/litellm/llms/azure_ai/chat/transformation.py +++ b/litellm/llms/azure_ai/chat/transformation.py @@ -6,15 +6,19 @@ from typing import TYPE_CHECKING, Final, cast import httpx from httpx import Response +from pydantic import TypeAdapter, ValidationError import litellm from litellm._logging import verbose_logger +from litellm.integrations.anthropic_cache_control_hook import AnthropicCacheControlHook from litellm.litellm_core_utils.prompt_templates.common_utils import ( content_parts, convert_content_list_to_str, drop_tool_reference_parts_from_tool_messages, filter_value_from_dict, is_pdf_file_part, + prompt_cache_breakpoint_for_wire, + with_prompt_cache_breakpoint, ) from litellm.llms.azure.common_utils import BaseAzureLLM from litellm.llms.azure_ai.common_utils import ( @@ -27,7 +31,7 @@ from litellm.llms.openai.common_utils import drop_params_from_unprocessable_enti from litellm.llms.openai.openai import OpenAIChatCompletionResponseIterator, OpenAIConfig from litellm.llms.xai.chat.transformation import XAIChatConfig from litellm.secret_managers.main import get_secret_str -from litellm.types.llms.openai import AllMessageValues +from litellm.types.llms.openai import AllMessageValues, PromptCacheBreakpoint from litellm.types.router import GenericLiteLLMParams from litellm.types.utils import ModelResponse, ModelResponseStream, ProviderField from litellm.utils import add_path_to_api_base, supports_tool_choice @@ -62,6 +66,72 @@ def _has_list_only_content_part(message: Mapping[str, object]) -> bool: return any(_is_list_only_content_part(part) for part in content_parts(message)) +_CONTENT_PARTS_ADAPTER: Final = TypeAdapter(list[dict[str, object]]) + + +def _text_parts(message: Mapping[str, object]) -> tuple[Mapping[str, object], ...]: + content: Final = message.get("content") + if not isinstance(content, list): + return () + try: + parts: Final = _CONTENT_PARTS_ADAPTER.validate_python(content) + except ValidationError: + return () + return tuple(part for part in parts if part.get("type") == "text") + + +def _with_only_text_parts(message: AllMessageValues) -> AllMessageValues: + if _has_list_only_content_part(message) or not (text_parts := _text_parts(message)): + return message + text_only: Final = {**message, "content": list(text_parts)} + return cast(AllMessageValues, text_only) # cast-ok: the same message, text parts only + + +_FIELDS: Final = TypeAdapter(Mapping[str, object]) + + +def _foundry_breakpoint(marker: object) -> PromptCacheBreakpoint | None: + if prompt_cache_breakpoint_for_wire(marker, drop_params=True) is None: + return None + return PromptCacheBreakpoint(mode="explicit") + + +def _part_with_foundry_breakpoint(part: object) -> object: + if not isinstance(part, Mapping): + return part + fields: Final = _FIELDS.validate_python(part) + if "prompt_cache_breakpoint" not in fields: + return fields + unmarked: Final = {key: value for key, value in fields.items() if key != "prompt_cache_breakpoint"} + return with_prompt_cache_breakpoint(unmarked, _foundry_breakpoint(fields["prompt_cache_breakpoint"])) + + +def _with_breakpoints_for_foundry(message: AllMessageValues, drop_params: bool) -> AllMessageValues: + renamed: Final = AnthropicCacheControlHook.cache_control_as_prompt_cache_breakpoints(_with_only_text_parts(message)) + parts: Final = content_parts(renamed) if drop_params else () + if not parts: + return renamed + foundry_message: Final = {**renamed, "content": [_part_with_foundry_breakpoint(part) for part in parts]} + return cast(AllMessageValues, foundry_message) # cast-ok: the same message, its breakpoints in Foundry's shape + + +def _deployment_reads_breakpoints(model: str, litellm_params: Mapping[str, object]) -> bool: + base_model: Final = litellm_params.get("base_model") + named_models: Final = (base_model, model) if isinstance(base_model, str) and base_model else (model,) + model_info: Final = litellm_params.get("model_info") + if not isinstance(model_info, Mapping): + return any( + AnthropicCacheControlHook.targets_openai_prompt_cache_breakpoint(name, "azure_ai") for name in named_models + ) + deployment_flag: Final = _FIELDS.validate_python(model_info).get("supports_prompt_cache_breakpoint") + if isinstance(deployment_flag, bool): + return deployment_flag + return any( + AnthropicCacheControlHook.catalog_targets_openai_prompt_cache_breakpoint(name, "azure_ai") + for name in named_models + ) + + class AzureAIChatCompletionResponseIterator(OpenAIChatCompletionResponseIterator): def chunk_parser(self, chunk: dict[str, object]) -> ModelResponseStream: # mutable-ok: mirrors override contract usage: Final = chunk.get("usage") @@ -251,6 +321,8 @@ class AzureAIStudioConfig(OpenAIConfig): self, messages: list[AllMessageValues], model: str, + reads_breakpoints: bool | None = None, + drop_params: bool = False, ) -> list: """ - Azure AI Studio doesn't support content as a list. This handles: @@ -260,13 +332,31 @@ class AzureAIStudioConfig(OpenAIConfig): Azure AI Foundry backends set additionalProperties=false and reject these with "Extra inputs are not permitted", which breaks multi-turn Anthropic-format clients that echo thinking blocks back as history. - 2. Transforms list content to a string. + 2. Transforms list content to a string, except on deployments that read OpenAI explicit + prompt cache breakpoints (the deployment's own model_info opt-in, else the cost-map entry + of its base_model or name, as loaded and without a sibling deployment's opt-in): only a + content part can carry the breakpoint (a message-level one is silently ignored), so those + keep their text parts, with each Anthropic `cache_control` mark renamed to a breakpoint the + way the openai provider renames it. Every other part is dropped, exactly as the string + conversion drops it, before the marks are renamed, so a message-level mark lands on the + last text part. Under drop_params each breakpoint goes out as `{"mode": "explicit"}`, the only + shape Foundry takes, and a malformed one is dropped. 3. If message contains an image, audio, or PDF file part, send as is (user-intended) Operates on a deep copy so the caller's messages keep their thinking blocks and provider metadata, which a fallback to another provider still needs. """ - stripped_messages: Final = copy.deepcopy(drop_tool_reference_parts_from_tool_messages(messages)) + deployment_reads_breakpoints: Final = ( + AnthropicCacheControlHook.targets_openai_prompt_cache_breakpoint(model, "azure_ai") + if reads_breakpoints is None + else reads_breakpoints + ) + copied_messages: Final = copy.deepcopy(drop_tool_reference_parts_from_tool_messages(messages)) + stripped_messages: Final = ( + [_with_breakpoints_for_foundry(message, drop_params) for message in copied_messages] + if deployment_reads_breakpoints + else copied_messages + ) for message in stripped_messages: message_dict = cast(dict, message) # cast-ok: TypedDict is a runtime dict stripped on our copy for field in NON_OPENAI_SPEC_MESSAGE_FIELDS: @@ -275,6 +365,10 @@ class AzureAIStudioConfig(OpenAIConfig): if _has_list_only_content_part(message): continue + if deployment_reads_breakpoints and (text_parts := _text_parts(message)): + message_dict["content"] = list(text_parts) + continue + texts = convert_content_list_to_str(message=message) if texts: message["content"] = texts @@ -329,7 +423,16 @@ class AzureAIStudioConfig(OpenAIConfig): if extra_body and isinstance(extra_body, dict): optional_params.update(extra_body) optional_params.pop("max_retries", None) - return super().transform_request(model, messages, optional_params, litellm_params, headers) + return { + "model": model, + "messages": self._transform_messages( + messages, + model, + _deployment_reads_breakpoints(model, _FIELDS.validate_python(litellm_params)), + bool(litellm_params.get("drop_params") or litellm.drop_params), + ), + **optional_params, + } def transform_response( self, diff --git a/tests/integration/providers/test_openai_dialect_prompt_cache_breakpoint_owned_proxy.py b/tests/integration/providers/test_openai_dialect_prompt_cache_breakpoint_owned_proxy.py index 3df73d3b767..d470605b1e5 100644 --- a/tests/integration/providers/test_openai_dialect_prompt_cache_breakpoint_owned_proxy.py +++ b/tests/integration/providers/test_openai_dialect_prompt_cache_breakpoint_owned_proxy.py @@ -153,15 +153,16 @@ def _foundry_deployment(rig: _Rig, scenario: Scenario, *, bare: bool) -> str: return name -def _assert_foundry_chat_wire(received: Request, prompt: str, *, options: JsonValue | None) -> None: +def _assert_foundry_chat_wire(received: Request, user_content: JsonValue, *, options: JsonValue | None) -> None: body: Final = body_of(received) + marked_system: Final = [{"type": "text", "text": SYSTEM, "prompt_cache_breakpoint": EXPLICIT}] assert urlsplit(received.target).netloc == _FOUNDRY_HOST, received.target assert urlsplit(received.target).path == _FOUNDRY_CHAT_PATH, received.target - assert "prompt_cache_breakpoint" not in received.body.decode(), received.body assert body["model"] == _FOUNDRY_MODEL, received.body - assert body["messages"] == [{"role": "system", "content": SYSTEM}, {"role": "user", "content": prompt}], ( - received.body - ) + assert body["messages"] == [ + {"role": "system", "content": marked_system}, + {"role": "user", "content": user_content}, + ], received.body assert body.get("prompt_cache_options") == options, received.body @@ -185,7 +186,7 @@ def test_b9_b12_foundry_gpt6_responses_reach_the_openai_dialect_on_the_foundry_h @pytest.mark.timeout(_CELL_TIMEOUT) -def test_b10_foundry_gpt6_messages_bridge_to_chat_without_a_breakpoint(rig: _Rig) -> None: +def test_b10_foundry_gpt6_messages_bridge_to_chat_keeps_the_injected_breakpoint(rig: _Rig) -> None: marker: Final = fresh_marker() with rig.gateway.scenario() as scenario: name: Final = _foundry_deployment(rig, scenario, bare=False) @@ -193,12 +194,12 @@ def test_b10_foundry_gpt6_messages_bridge_to_chat_without_a_breakpoint(rig: _Rig rig.gateway, rig.wire, "messages", request_body("messages", name, prompt_text(marker)) ) assert_answered(outcome, marker) - _assert_foundry_chat_wire(received, prompt_text(marker), options=IMPLICIT) + _assert_foundry_chat_wire(received, [{"type": "text", "text": prompt_text(marker)}], options=IMPLICIT) success_row(name, marker) @pytest.mark.timeout(_CELL_TIMEOUT) -def test_b11_foundry_gpt6_chat_stays_on_the_plain_chat_wire(rig: _Rig) -> None: +def test_b11_foundry_gpt6_chat_stays_on_the_chat_wire_with_the_injected_breakpoint(rig: _Rig) -> None: marker: Final = fresh_marker() with rig.gateway.scenario() as scenario: name: Final = _foundry_deployment(rig, scenario, bare=False) diff --git a/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py b/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py index 6269fd367fe..145d45317c3 100644 --- a/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py +++ b/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py @@ -1,6 +1,7 @@ import json import traceback from collections.abc import Callable +from types import MappingProxyType from typing import Final from unittest.mock import MagicMock, patch @@ -9,6 +10,7 @@ import respx from httpx import Response import litellm +from litellm import Router from litellm.llms.custom_httpx.http_handler import HTTPHandler from litellm.litellm_core_utils.get_model_cost_map import get_model_cost_map from litellm.llms.azure_ai.azure_model_router.transformation import ( @@ -880,6 +882,360 @@ def test_azure_ai_stripping_does_not_mutate_caller_messages(): } +def test_azure_ai_keeps_prompt_cache_breakpoint_for_a_model_that_supports_it(_local_model_cost_map): + """ + Foundry GPT-6 deployments take OpenAI explicit prompt caching, and the breakpoint only + counts on a content part, so a system message carrying one must reach the request body as + the content-part list it arrived in while the Anthropic-only fields are still stripped. + """ + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", + messages=[ + { + "role": "system", + "content": [ + { + "type": "text", + "text": "stable instructions", + "prompt_cache_breakpoint": {"mode": "explicit"}, + "cache_control": {"type": "ephemeral"}, + } + ], + }, + {"role": "user", "content": [{"type": "text", "text": "hi"}]}, + ], + optional_params={"prompt_cache_options": {"mode": "explicit"}}, + litellm_params={}, + headers={}, + ) + + assert request["messages"][0]["content"] == [ + {"type": "text", "text": "stable instructions", "prompt_cache_breakpoint": {"mode": "explicit"}} + ] + assert request["messages"][1]["content"] == [{"type": "text", "text": "hi"}] + assert request["prompt_cache_options"] == {"mode": "explicit"} + + +def test_azure_ai_still_flattens_list_content_for_a_model_without_the_breakpoint_flag(_local_model_cost_map): + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-4o", + messages=[ + { + "role": "system", + "content": [ + {"type": "text", "text": "stable ", "prompt_cache_breakpoint": {"mode": "explicit"}}, + {"type": "text", "text": "instructions"}, + ], + }, + {"role": "user", "content": [{"type": "text", "text": "hi"}]}, + ], + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert request["messages"][0]["content"] == "stable instructions" + assert request["messages"][1]["content"] == "hi" + assert not _find_key_anywhere(request["messages"], "prompt_cache_breakpoint") + + +def test_azure_ai_drops_thinking_parts_from_history_while_keeping_the_breakpoint(_local_model_cost_map): + """ + A multi-turn client can echo an assistant turn back with thinking parts inside its content + list. Foundry rejects those parts, and the string conversion always dropped them, so a model + that keeps content parts must still drop them while the text parts and the breakpoint stay. + """ + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", + messages=[ + { + "role": "system", + "content": [ + {"type": "text", "text": "stable instructions", "prompt_cache_breakpoint": {"mode": "explicit"}} + ], + }, + {"role": "user", "content": "read the file"}, + { + "role": "assistant", + "content": [ + {"type": "thinking", "thinking": "The user wants me to read a file.", "signature": "sig"}, + {"type": "redacted_thinking", "data": "opaque"}, + {"type": "text", "text": "Reading it now."}, + ], + }, + {"role": "user", "content": [{"type": "text", "text": "thanks"}]}, + ], + optional_params={"prompt_cache_options": {"mode": "explicit"}}, + litellm_params={}, + headers={}, + ) + + assert request["messages"][0]["content"] == [ + {"type": "text", "text": "stable instructions", "prompt_cache_breakpoint": {"mode": "explicit"}} + ] + assert request["messages"][2]["content"] == [{"type": "text", "text": "Reading it now."}] + assert request["messages"][3]["content"] == [{"type": "text", "text": "thanks"}] + assert not _find_key_anywhere(request["messages"], "thinking") + assert not _find_key_anywhere(request["messages"], "data") + + + +def test_azure_ai_renames_a_cache_control_mark_to_a_breakpoint_for_a_model_that_reads_them(_local_model_cost_map): + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", + messages=[ + { + "role": "system", + "content": [{"type": "text", "text": "stable instructions", "cache_control": {"type": "ephemeral"}}], + }, + {"role": "user", "content": "hi", "cache_control": {"type": "ephemeral"}}, + ], + optional_params={"prompt_cache_options": {"mode": "explicit"}}, + litellm_params={}, + headers={}, + ) + + assert request["messages"][0]["content"] == [ + {"type": "text", "text": "stable instructions", "prompt_cache_breakpoint": {"mode": "explicit"}} + ] + assert request["messages"][1]["content"] == [ + {"type": "text", "text": "hi", "prompt_cache_breakpoint": {"mode": "explicit"}} + ] + assert not _find_key_anywhere(request["messages"], "cache_control") + + +def test_azure_ai_drops_a_file_part_next_to_the_breakpoint_like_the_string_conversion(_local_model_cost_map): + messages: Final = [ + { + "role": "user", + "content": [ + {"type": "text", "text": "summarize", "prompt_cache_breakpoint": {"mode": "explicit"}}, + {"type": "file", "file": {"file_id": "file-held-by-another-provider"}}, + ], + } + ] + + kept: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", messages=messages, optional_params={}, litellm_params={}, headers={} + ) + flattened: Final = AzureAIStudioConfig().transform_request( + model="gpt-4o", messages=messages, optional_params={}, litellm_params={}, headers={} + ) + + assert kept["messages"][0]["content"] == [ + {"type": "text", "text": "summarize", "prompt_cache_breakpoint": {"mode": "explicit"}} + ] + assert flattened["messages"][0]["content"] == "summarize" + + +def test_azure_ai_sends_a_pdf_part_next_to_the_breakpoint_as_is(_local_model_cost_map): + content: Final = [ + {"type": "file", "file": {"filename": "document.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjQK"}}, + {"type": "text", "text": "What is the codeword?", "prompt_cache_breakpoint": {"mode": "explicit"}}, + ] + + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", + messages=[{"role": "user", "content": content}], + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert request["messages"] == [{"role": "user", "content": content}] + + +def test_azure_ai_puts_a_message_level_mark_on_the_last_text_part_when_a_file_part_follows(_local_model_cost_map): + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", + messages=[ + { + "role": "user", + "cache_control": {"type": "ephemeral"}, + "content": [ + {"type": "text", "text": "summarize"}, + {"type": "file", "file": {"file_id": "file-held-by-another-provider"}}, + ], + } + ], + optional_params={"prompt_cache_options": {"mode": "explicit"}}, + litellm_params={}, + headers={}, + ) + + assert request["messages"][0]["content"] == [ + {"type": "text", "text": "summarize", "prompt_cache_breakpoint": {"mode": "explicit"}} + ] + + +@pytest.mark.parametrize( + ("registered", "keeps_parts"), + [ + ({"azure_ai/qa-foundry-alias": {"litellm_provider": "azure_ai", "supports_prompt_cache_breakpoint": True}}, True), + ({"qa-foundry-alias": {"litellm_provider": "openai", "supports_prompt_cache_breakpoint": True}}, False), + ], + ids=["azure_ai-entry-opts-in", "openai-entry-does-not-speak-for-azure_ai"], +) +def test_azure_ai_keeps_parts_only_when_the_azure_ai_entry_reads_breakpoints( + _local_model_cost_map, registered: dict[str, dict[str, object]], keeps_parts: bool +): + litellm.register_model(model_cost=registered, persist_across_reloads=False, add_to_provider_model_sets=False) + breakpoint_part: Final = {"type": "text", "text": "stable", "prompt_cache_breakpoint": {"mode": "explicit"}} + + request: Final = AzureAIStudioConfig().transform_request( + model="qa-foundry-alias", + messages=[{"role": "system", "content": [breakpoint_part]}], + optional_params={}, + litellm_params={"custom_llm_provider": "azure_ai", "drop_params": True}, + headers={}, + ) + + assert request["messages"][0]["content"] == ([breakpoint_part] if keeps_parts else "stable") + + +@pytest.mark.parametrize( + "registration_order", [("opted-in", "opted-out", "unflagged"), ("unflagged", "opted-out", "opted-in")] +) +def test_azure_ai_deployment_reads_only_its_own_breakpoint_flag_not_a_siblings_sharing_its_model_string( + _local_model_cost_map, respx_mock: respx.MockRouter, registration_order: tuple[str, ...] +) -> None: + breakpoint_flags: Final = MappingProxyType( + { + "opted-in": {"supports_prompt_cache_breakpoint": True}, + "opted-out": {"supports_prompt_cache_breakpoint": False}, + } + ) + def deployment(host: str, model_info: dict[str, object]) -> dict[str, object]: + return { + "model_name": host, + "litellm_params": { + "model": "azure_ai/qa-shared-deployment", + "api_base": f"https://{host}.services.ai.azure.com/models", + "api_key": "azure-ai-test-key", + }, + "model_info": model_info, + } + + routes: Final = { + host: respx_mock.post(host=f"{host}.services.ai.azure.com", path="/models/chat/completions").mock( + return_value=Response( + 200, + json={ + "id": "chatcmpl-azure-ai", + "object": "chat.completion", + "created": 1700000000, + "model": "qa-shared-deployment", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 5, "completion_tokens": 1, "total_tokens": 6}, + }, + ) + ) + for host in registration_order + } + router: Final = Router( + model_list=[ + deployment(host, {"id": f"{host}-id", **breakpoint_flags.get(host, {})}) for host in registration_order + ] + ) + marked_system: Final = { + "role": "system", + "content": [{"type": "text", "text": "stable", "cache_control": {"type": "ephemeral"}}], + } + + for host in registration_order: + router.completion(model=host, messages=[marked_system, {"role": "user", "content": "hi"}]) + + def sent_system_content(host: str) -> object: + return json.loads(routes[host].calls.last.request.content)["messages"][0]["content"] + + assert sent_system_content("opted-in") == [ + {"type": "text", "text": "stable", "prompt_cache_breakpoint": {"mode": "explicit"}} + ] + assert sent_system_content("opted-out") == "stable" + assert sent_system_content("unflagged") == "stable" + + +@pytest.mark.parametrize( + "litellm_params", + [ + {"base_model": "azure_ai/gpt-6-astra"}, + {"base_model": "azure_ai/gpt-6-astra", "model_info": {"id": "my-astra-deployment-id"}}, + ], +) +def test_azure_ai_keeps_parts_for_a_custom_deployment_name_whose_base_model_reads_breakpoints( + _local_model_cost_map, litellm_params: dict[str, object] +): + breakpoint_part: Final = {"type": "text", "text": "stable", "prompt_cache_breakpoint": {"mode": "explicit"}} + + request: Final = AzureAIStudioConfig().transform_request( + model="my-astra-deployment", + messages=[{"role": "system", "content": [breakpoint_part]}], + optional_params={}, + litellm_params=litellm_params, + headers={}, + ) + + assert request["messages"][0]["content"] == [breakpoint_part] + + +@pytest.mark.parametrize( + "litellm_params", + [ + {"base_model": "azure/cost-tracking-row"}, + {"base_model": "azure/cost-tracking-row", "model_info": {"id": "astra-deployment-id"}}, + ], +) +def test_azure_ai_falls_back_to_the_model_name_when_the_base_model_row_does_not_read_breakpoints( + _local_model_cost_map, litellm_params: dict[str, object] +): + breakpoint_part: Final = {"type": "text", "text": "stable", "prompt_cache_breakpoint": {"mode": "explicit"}} + + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", + messages=[{"role": "system", "content": [breakpoint_part]}], + optional_params={}, + litellm_params=litellm_params, + headers={}, + ) + + assert request["messages"][0]["content"] == [breakpoint_part] + + +@pytest.mark.parametrize( + "sent_marker, litellm_params, wire_part", + [ + ( + {"mode": "explicit", "ttl": "30m"}, + {"drop_params": True}, + {"type": "text", "text": "stable", "prompt_cache_breakpoint": {"mode": "explicit"}}, + ), + ({"type": "ephemeral"}, {"drop_params": True}, {"type": "text", "text": "stable"}), + ( + {"mode": "explicit", "ttl": "30m"}, + {}, + {"type": "text", "text": "stable", "prompt_cache_breakpoint": {"mode": "explicit", "ttl": "30m"}}, + ), + ], +) +def test_azure_ai_reshapes_breakpoints_for_foundry_only_under_drop_params( + _local_model_cost_map, + sent_marker: dict[str, object], + litellm_params: dict[str, object], + wire_part: dict[str, object], +): + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-6-astra", + messages=[ + {"role": "system", "content": [{"type": "text", "text": "stable", "prompt_cache_breakpoint": sent_marker}]} + ], + optional_params={}, + litellm_params=litellm_params, + headers={}, + ) + + assert request["messages"][0]["content"] == [wire_part] + + def test_azure_ai_reference_only_tool_result_is_sent_as_empty_text(): messages: Final = [ {"role": "user", "content": "load the WebFetch tool"}, From 953750ed41c69ea909fa4064a42317b04fa8a39b Mon Sep 17 00:00:00 2001 From: Devon Krisman <844724+dkrisman@users.noreply.github.com> Date: Fri, 30 Oct 2026 03:06:57 -0400 Subject: [PATCH 09/23] fix(websearch-interception)!: stream native server-tool blocks to clients (#40967) * fix(websearch-interception): stream native server-tool blocks to clients Streaming native clients (Anthropic SDK, Claude Desktop, Claude Code) never received the server_tool_use and web_search_tool_result blocks the interception layer builds for them, so citation panels stayed empty and result counts read zero. Two gaps in the same flow: FakeAnthropicMessagesStreamIterator._create_content_block_chunks only knew text, thinking, redacted_thinking and tool_use; any other block type emitted a bare content_block_stop with no start and no payload, silently dropping the block from the fake stream. server_tool_use now streams as content_block_start plus input_json_delta and any other server-generated block rides whole in content_block_start, matching Anthropic's native streaming of these types. try_short_circuit_search only recognized a native client by finding a web_search_* tool in the tools list, but the pre-request and deployment hooks convert that tool to the LiteLLM standard shape before the short-circuit runs, recording the conversion in WEBSEARCH_EMIT_NATIVE_BLOCKS_KEY. The flag now counts as native too, so the synthetic response carries the native blocks on the converted path instead of only when hook ordering happens to leave the tool untouched. Fixes #35333 * fix(anthropic): accept server_tool_use and web_search_tool_result blocks when costing a /v1/messages response The short-circuit now returns native web search blocks, and the cost path validated the served body against a content union that only allowed text, tool_use and thinking blocks, so non-streaming requests answered 500 * fix(anthropic): keep every server_tool_use field when re-streaming a /v1/messages reply * fix(anthropic): keep tool_use fields, redacted thinking data and text citations when re-streaming a /v1/messages reply * test(websearch-interception): build the cost test's logging object with a datetime start time * fix(websearch-interception): send search snippets once in the direct search answer * fix(guardrails): mask web search queries, titles and urls on /v1/messages and deliver stream rewrites past server_tool_use blocks Output guardrails now scan the native web search blocks the short-circuit emits, ahead of the reply text so guardrails that read the last text still scan the reply. A guardrail reply of a different length falls back to the reply-only write-back. Ended-stream tool call rewrites land on server_tool_use blocks too, the ended-stream scan key covers the result strings, and legacy-hook pipeline steps pass through the result strings their response view lacks. The wire test reads the streamed server_tool_use input from its input_json_delta. * fix(guardrails): deliver a legacy post-call hook's tool-call rewrites on streams instead of discarding the stream rewrite A legacy hook that masked a tool call's arguments on a buffered stream made the whole rewrite undeliverable, text included, so the original stream went out. With web search blocks in the stream that now hit any hook masking the search query. The adapter hands same-count tool-call rewrites to the translation's write-back, as a unified guardrail's already are, and still discards a rewrite that adds or drops a tool call. * fix(guardrails): show a legacy post-call hook the web search result blocks of a /v1/messages stream The legacy hook's view of a buffered Anthropic stream is rebuilt from the assembled ModelResponse, which keeps web_search_tool_result blocks only in provider-specific fields, so a hook masking result titles on a stream never saw them. The hook's view now carries those blocks, its rewrite of them goes back through the ended-stream write-back, and the adapter no longer needs to pass texts outside the hook's view through untouched. * test(guardrails): type the legacy hook stream test helpers * fix(guardrails): scan a streamed web search query as text like the non-streaming reply * test(guardrails): type the masking guardrail test helpers * fix(guardrails): keep the block id when scanning a streamed web search query --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../websearch_interception/handler.py | 15 +- .../websearch_interception/transformation.py | 10 +- .../chat/guardrail_translation/handler.py | 335 ++++++++++++++++-- .../messages/native/fake_stream_iterator.py | 56 +-- .../proxy/policy_engine/pipeline_executor.py | 36 +- litellm/types/llms/anthropic.py | 19 + .../test_websearch_interception_wire.py | 27 +- .../test_websearch_short_circuit.py | 122 +++++++ .../test_anthropic_guardrail_handler.py | 307 +++++++++++++++- .../native/test_fake_stream_iterator.py | 145 +++++++- .../policy_engine/test_pipeline_executor.py | 238 +++++++++++-- 11 files changed, 1191 insertions(+), 119 deletions(-) diff --git a/litellm/integrations/websearch_interception/handler.py b/litellm/integrations/websearch_interception/handler.py index 4739b3ee561..2bb89248be4 100644 --- a/litellm/integrations/websearch_interception/handler.py +++ b/litellm/integrations/websearch_interception/handler.py @@ -398,14 +398,17 @@ class WebSearchInterceptionLogger(CustomLogger): (t for t in tools if is_anthropic_native_web_search_tool(t)), None, ) + emit_native_blocks: Final = native_tool is not None or ( + kwargs is not None and bool(kwargs.get(WEBSEARCH_EMIT_NATIVE_BLOCKS_KEY)) + ) outcome: Final = await self._short_circuit_search_outcome(query, kwargs=kwargs) search_result_text: Final = WebSearchTransformation.search_outcome_text(outcome) content: Final[list[dict[str, object]]] = [] - if native_tool is not None: + if emit_native_blocks: tool_use_id: Final = f"srvtoolu_{uuid.uuid4().hex}" - tool_name: Final = native_tool.get("name") or "web_search" + tool_name: Final = (native_tool.get("name") if native_tool is not None else None) or "web_search" content.append( { "type": "server_tool_use", @@ -415,10 +418,10 @@ class WebSearchInterceptionLogger(CustomLogger): } ) content.append( - WebSearchTransformation.build_web_search_outcome_block(tool_use_id=tool_use_id, outcome=outcome) + WebSearchTransformation.build_web_search_outcome_block( + tool_use_id=tool_use_id, outcome=outcome, with_snippets=False + ) ) - # Keep the text block so non-native short-circuit callers (Claude Code, - # github_copilot, etc.) see the same payload they always have. content.append({"type": "text", "text": search_result_text}) response: Final[dict[str, object]] = { @@ -435,7 +438,7 @@ class WebSearchInterceptionLogger(CustomLogger): verbose_logger.debug( "WebSearchInterception: Short-circuit search completed, returning synthetic response (%s chars, native_blocks=%s)", len(search_result_text), - native_tool is not None, + emit_native_blocks, ) return response diff --git a/litellm/integrations/websearch_interception/transformation.py b/litellm/integrations/websearch_interception/transformation.py index 58ae8bd6e48..55564dd1134 100644 --- a/litellm/integrations/websearch_interception/transformation.py +++ b/litellm/integrations/websearch_interception/transformation.py @@ -412,6 +412,7 @@ class WebSearchTransformation: def build_web_search_tool_result_block( tool_use_id: str, search_response: SearchResponse | None, + with_snippets: bool = True, ) -> dict[str, object]: """ Build an Anthropic-native ``web_search_tool_result`` content block. @@ -445,6 +446,8 @@ class WebSearchTransformation: ``litellm.asearch()``. If None or empty, the block is still emitted with an empty result list (signals "search ran, no results" rather than "search did not run"). + with_snippets: False when a text block in the same reply already + carries the snippets. """ items: Final[list[dict[str, object]]] = [] if search_response is not None: @@ -460,7 +463,7 @@ class WebSearchTransformation: "title": title, "page_age": page_age, "encrypted_content": "", - "snippet": getattr(r, "snippet", "") or "", + **({"snippet": getattr(r, "snippet", "") or ""} if with_snippets else {}), } ) return { @@ -481,12 +484,15 @@ class WebSearchTransformation: } @staticmethod - def build_web_search_outcome_block(tool_use_id: str, outcome: SearchOutcome) -> dict[str, object]: + def build_web_search_outcome_block( + tool_use_id: str, outcome: SearchOutcome, with_snippets: bool = True + ) -> dict[str, object]: match outcome: case SearchSucceeded(response=response): return WebSearchTransformation.build_web_search_tool_result_block( tool_use_id=tool_use_id, search_response=response, + with_snippets=with_snippets, ) case SearchFailed(error_code=error_code): return WebSearchTransformation.build_web_search_tool_result_error_block( diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py index e351542da01..7f688ae733b 100644 --- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py +++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py @@ -24,7 +24,7 @@ from pydantic import ConfigDict, TypeAdapter, ValidationError, with_config from typing_extensions import ReadOnly, TypedDict, assert_never from litellm._logging import verbose_proxy_logger -from litellm.litellm_core_utils.core_helpers import as_str_mapping +from litellm.litellm_core_utils.core_helpers import as_str_mapping, iter_items from litellm.llms.anthropic.chat.transformation import AnthropicConfig from litellm.llms.anthropic.messages.adapters.transformation import ( OPENAI_MAX_TOOL_NAME_LENGTH, @@ -61,6 +61,7 @@ from litellm.llms.base_llm.guardrail_translation.utils import ( unappliable_request_rewrite, ) from litellm.types.llms.anthropic import ( + ANTHROPIC_SERVER_TOOL_USE_ID_PREFIX, AllAnthropicToolsValues, AnthropicMessagesRequest, ) @@ -153,6 +154,7 @@ class _TextAttrBlock(Protocol): text: str +@runtime_checkable class _WritableMessage(Protocol): @overload def get(self, key: str, /) -> object | None: ... @@ -195,6 +197,140 @@ class _ToolCallShape: arguments: str +@dataclass(frozen=True, slots=True) +class _WebSearchString: + block_idx: int + result_idx: int | None + field: str + value: str + + +def _streamed_web_search_result_blocks(response: ModelResponse) -> tuple[Mapping[str, object], ...]: + choice: Final = response.choices[0] if response.choices else None + fields: Final = as_str_mapping(choice.message.provider_specific_fields) if isinstance(choice, Choices) else None + blocks: Final = ( + (as_str_mapping(item) for item in iter_items(fields.get("web_search_results"))) + if fields is not None and isinstance(fields.get("web_search_results"), list) + else () + ) + return tuple( + deepcopy(block) for block in blocks if block is not None and block.get("type") == "web_search_tool_result" + ) + + +def _web_search_results(block: Mapping[str, object]) -> Iterable[object]: + return iter_items(block.get("content")) if isinstance(block.get("content"), list) else () + + +def _web_search_result_strings(block_idx: int, result_idx: int, result: object) -> tuple[_WebSearchString, ...]: + fields: Final = as_str_mapping(result) + if fields is None or fields.get("type") != "web_search_result": + return () + values: Final = ((field, fields.get(field)) for field in ("title", "url")) + return tuple( + _WebSearchString(block_idx, result_idx, field, value) + for field, value in values + if isinstance(value, str) and value + ) + + +def _is_server_tool_use(block: Mapping[str, object]) -> bool: + block_id: Final = block.get("id") + match block.get("type"): + case "server_tool_use": + return True + case "tool_use": + return isinstance(block_id, str) and block_id.startswith(ANTHROPIC_SERVER_TOOL_USE_ID_PREFIX) + case _: + return False + + +def _web_search_block_strings( + block_idx: int, block: Mapping[str, object], *, with_query: bool +) -> tuple[_WebSearchString, ...]: + """The strings a client reads off a native web search block: the ``server_tool_use`` query and each + ``web_search_tool_result`` result's title and url.""" + match block.get("type"): + case "server_tool_use" | "tool_use" if with_query and _is_server_tool_use(block): + tool_input: Final = as_str_mapping(block.get("input")) + query: Final = None if tool_input is None else tool_input.get("query") + return (_WebSearchString(block_idx, None, "query", query),) if isinstance(query, str) and query else () + case "web_search_tool_result": + return tuple( + chain.from_iterable( + _web_search_result_strings(block_idx, result_idx, result) + for result_idx, result in enumerate(_web_search_results(block)) + ) + ) + case _: + return () + + +def _rewritten_web_search_result( + result_idx: int, result: object, rewrites: Mapping[tuple[int | None, str], str] +) -> object: + fields: Final = as_str_mapping(result) + if fields is None: + return result + return { + **fields, + **{field: rewrites[(result_idx, field)] for field in ("title", "url") if (result_idx, field) in rewrites}, + } + + +def _rewritten_web_search_block( + block: Mapping[str, object], rewrites: Sequence[tuple[_WebSearchString, str]] +) -> Mapping[str, object]: + by_location: Final = MappingProxyType({(target.result_idx, target.field): value for target, value in rewrites}) + if _is_server_tool_use(block): + query: Final = by_location.get((None, "query")) + tool_input: Final = as_str_mapping(block.get("input")) or MappingProxyType({}) + return block if query is None else {**block, "input": {**tool_input, "query": query}} + if not isinstance(block.get("content"), list): + return block + return { + **block, + "content": [ + _rewritten_web_search_result(result_idx, result, by_location) + for result_idx, result in enumerate(_web_search_results(block)) + ], + } + + +def _write_back_web_search_block(content_block: object, rewrites: Sequence[tuple[_WebSearchString, str]]) -> None: + block: Final = as_str_mapping(content_block) + if block is None or not isinstance(content_block, _WritableMessage): + return + rewritten: Final = _rewritten_web_search_block(block, rewrites) + writable: Final = _as_writable(content_block) + for key in ("input", "content"): + if key in rewritten: + writable[key] = rewritten[key] + + +def _split_web_search_texts( + guardrailed: Sequence[str], web_search_count: int, reply_count: int +) -> tuple[Sequence[str], Sequence[str]]: + split_at: Final = web_search_count if len(guardrailed) == web_search_count + reply_count else 0 + return guardrailed[:split_at], guardrailed[split_at:] + + +def _web_search_rewrites_by_block( + strings: Sequence[_WebSearchString], guardrailed: Sequence[object] +) -> Mapping[int, tuple[tuple[_WebSearchString, str], ...]]: + rewrites: Final = tuple( + (target, value) + for target, value in zip(strings, guardrailed) + if isinstance(value, str) and value != target.value + ) + return MappingProxyType( + { + block_idx: tuple(rewrite for rewrite in rewrites if rewrite[0].block_idx == block_idx) + for block_idx in {target.block_idx for target, _ in rewrites} + } + ) + + def _is_client_tool_use(block: Mapping[str, object]) -> bool: return ( block.get("type") == "tool_use" @@ -289,6 +425,7 @@ def _write_back_message_text(message: _WritableMessage, target: MessageTextTarge _TOOL_USE_INPUT_ADAPTER: Final = TypeAdapter(dict[str, object]) _DECODED_SSE_DATA: Final = TypeAdapter(object) +_REBUILT_TOOL_CALL_BLOCK_TYPES: Final = frozenset({"tool_use", "server_tool_use"}) _RELEASED_TOOL_USE_STOP: Final = ( b"event: message_delta\n" b'data: {"type": "message_delta", "delta": {"stop_reason": "tool_use", "stop_sequence": null}, ' @@ -315,6 +452,51 @@ def _write_back_tool_use( block["name"] = shape.name +def _streamed_query_strings( + tool_call_blocks: Sequence[tuple[int, Mapping[str, object]]], tool_calls: Sequence[_ToolCallShape] +) -> tuple[_WebSearchString, ...]: + if len(tool_call_blocks) != len(tool_calls): + return () + return tuple( + chain.from_iterable( + _web_search_block_strings( + index, {**block, "input": _rewritten_tool_use_input(tool_call.arguments)}, with_query=True + ) + for (index, block), tool_call in zip(tool_call_blocks, tool_calls) + ) + ) + + +def _query_rewrites(strings: Sequence[_WebSearchString], guardrailed: Sequence[object]) -> Mapping[int, str]: + return MappingProxyType( + { + target.block_idx: value + for target, value in zip(strings, guardrailed) + if isinstance(value, str) and value != target.value + } + ) + + +def _with_query(tool_call: _ToolCallShape, query: str) -> _ToolCallShape: + arguments: Final = _rewritten_tool_use_input(tool_call.arguments) + if arguments is None: + return tool_call + return _ToolCallShape(name=tool_call.name, arguments=json.dumps({**arguments, "query": query})) + + +def _with_query_rewrites( + tool_calls: tuple[_ToolCallShape, ...], + tool_call_blocks: Sequence[tuple[int, Mapping[str, object]]], + query_rewrites: Mapping[int, str], +) -> tuple[_ToolCallShape, ...]: + if not query_rewrites or len(tool_call_blocks) != len(tool_calls): + return tool_calls + return tuple( + _with_query(tool_call, query_rewrites[index]) if index in query_rewrites else tool_call + for (index, _), tool_call in zip(tool_call_blocks, tool_calls) + ) + + @dataclass(frozen=True, slots=True) class _SSEFieldRewrite: """One field of one nested section of a buffered SSE event, rewritten.""" @@ -402,7 +584,12 @@ class AnthropicMessagesHandler(BaseTranslation): def post_call_hook_response(self, response: object) -> object: if not isinstance(response, ModelResponse): return response - return self.adapter.translate_openai_response_to_anthropic(response) + translated: Final = self.adapter.translate_openai_response_to_anthropic(response) + result_blocks: Final = _streamed_web_search_result_blocks(response) + if not result_blocks: + return translated + content: Final = translated.get("content") or () + return {**translated, "content": [*result_blocks, *content]} @staticmethod def _build_streaming_usage_response( @@ -1273,9 +1460,17 @@ class AnthropicMessagesHandler(BaseTranslation): task_mappings, tool_calls_to_check, ) + block_fields: Final = tuple(map(self._output_block_fields, response_content)) + web_search_strings: Final = tuple( + chain.from_iterable( + _web_search_block_strings(content_idx, fields[1], with_query=True) + for content_idx, fields in enumerate(block_fields) + if fields is not None + ) + ) # Step 2: Apply guardrail to all texts in batch - if texts_to_check or tool_calls_to_check: + if texts_to_check or tool_calls_to_check or web_search_strings: request_data = self._prepare_request_data( request_data, response, @@ -1284,7 +1479,7 @@ class AnthropicMessagesHandler(BaseTranslation): ) inputs: Final = self._build_guardrail_inputs( - texts_to_check, + [*(target.value for target in web_search_strings), *texts_to_check], images_to_check, tool_calls_to_check, response, @@ -1297,14 +1492,19 @@ class AnthropicMessagesHandler(BaseTranslation): logging_obj=litellm_logging_obj, ) - guardrailed_texts: Final = guardrailed_inputs.get("texts", []) + web_search_rewrites, reply_rewrites = _split_web_search_texts( + guardrailed_inputs.get("texts", []), len(web_search_strings), len(texts_to_check) + ) # Step 3: Map guardrail responses back to original response structure await self._apply_guardrail_responses_to_output( response=response, - responses=guardrailed_texts, + responses=list(reply_rewrites), task_mappings=task_mappings, ) + self._write_back_web_search_strings( + response_content, _web_search_rewrites_by_block(web_search_strings, web_search_rewrites) + ) verbose_proxy_logger.debug("Anthropic Messages: Processed output response: %s", response) @@ -1350,8 +1550,14 @@ class AnthropicMessagesHandler(BaseTranslation): string_so_far = first_choice.message.content pre_guardrail_tool_calls: Final = _tool_call_shapes(tool_calls_list or ()) guardrail_inputs: Final = GenericGuardrailAPIInputs() - if string_so_far: - guardrail_inputs["texts"] = [string_so_far] + leading_texts: Final = (string_so_far,) if isinstance(string_so_far, str) and string_so_far else () + tool_call_blocks: Final = self._streamed_tool_call_blocks(responses_so_far) + streamed_query_strings: Final = _streamed_query_strings(tool_call_blocks, pre_guardrail_tool_calls) + streamed_result_strings: Final = self._streamed_web_search_result_strings(responses_so_far) + streamed_web_search_strings: Final = (*streamed_result_strings, *streamed_query_strings) + texts_to_check: Final = [*(target.value for target in streamed_web_search_strings), *leading_texts] + if texts_to_check: + guardrail_inputs["texts"] = texts_to_check if tool_calls_list: guardrail_inputs["tool_calls"] = tool_calls_list @@ -1374,29 +1580,42 @@ class AnthropicMessagesHandler(BaseTranslation): responses_so_far, request_data ) raise - guardrailed_texts: Final = _guardrailed_inputs.get("texts") + web_search_rewrites, reply_rewrites = _split_web_search_texts( + _guardrailed_inputs.get("texts") or [], len(streamed_web_search_strings), len(leading_texts) + ) if ( deliver_ended_stream_rewrites and isinstance(string_so_far, str) and string_so_far - and guardrailed_texts - and guardrailed_texts[0] != string_so_far + and reply_rewrites + and reply_rewrites[0] != string_so_far ): self._write_ended_stream_text_rewrite( responses_so_far, - guardrailed_texts[0], + reply_rewrites[0], guardrail_name=guardrail_to_apply.guardrail_name or "unknown", ) + if deliver_ended_stream_rewrites: + self._write_ended_stream_web_search_rewrites( + responses_so_far, + _web_search_rewrites_by_block(streamed_result_strings, web_search_rewrites), + ) if deliver_ended_stream_rewrites: returned_tool_calls: Final = _guardrailed_inputs.get("tool_calls") self._write_ended_stream_tool_call_rewrites( responses_so_far, pre_guardrail_tool_calls=pre_guardrail_tool_calls, - post_guardrail_tool_calls=_tool_call_shapes( - returned_tool_calls - if isinstance(returned_tool_calls, list) - and len(returned_tool_calls) == len(pre_guardrail_tool_calls) - else tool_calls_list or () + post_guardrail_tool_calls=_with_query_rewrites( + _tool_call_shapes( + returned_tool_calls + if isinstance(returned_tool_calls, list) + and len(returned_tool_calls) == len(pre_guardrail_tool_calls) + else tool_calls_list or () + ), + tool_call_blocks, + _query_rewrites( + streamed_query_strings, web_search_rewrites[len(streamed_result_strings) :] + ), ), guardrail_name=guardrail_to_apply.guardrail_name or "unknown", ) @@ -1451,6 +1670,14 @@ class AnthropicMessagesHandler(BaseTranslation): request_data["litellm_metadata"] = user_metadata return request_data + @staticmethod + def _write_back_web_search_strings( + response_content: Sequence[object], + rewrites_by_block: Mapping[int, tuple[tuple[_WebSearchString, str], ...]], + ) -> None: + for block_idx, rewrites in rewrites_by_block.items(): + _write_back_web_search_block(_content_block_at(response_content, block_idx), rewrites) + @staticmethod def _get_response_content(response: object) -> Sequence[object]: """Extract content list from a dict or object response.""" @@ -1573,30 +1800,22 @@ class AnthropicMessagesHandler(BaseTranslation): ) -> None: """Deliver ended-stream guardrail tool-call rewrites by rewriting the buffered chunks in place: the rebuilt response lists tool calls in the - order of the stream's ``tool_use`` blocks, so the nth rewritten call lands - on the nth block, its first ``input_json_delta`` carrying the full rewritten + order of the stream's ``tool_use`` and ``server_tool_use`` blocks, so the nth + rewritten call lands on the nth block, its first ``input_json_delta`` carrying the full rewritten arguments, every later one blanked, and ``content_block_start`` carrying the rewritten name. Blocks that do not line up with the rebuilt tool calls make the rewrite undeliverable, so the pipeline executor discards it and releases the original chunks.""" if post_guardrail_tool_calls == pre_guardrail_tool_calls: return - block_indices: Final = tuple( - index - for item in responses_so_far - for event in cls._iter_sse_events(item) - if event.get("type") == "content_block_start" - and isinstance(block := event.get("content_block"), Mapping) - and block.get("type") == "tool_use" - and isinstance(index := event.get("index"), int) - ) + block_indices: Final = tuple(index for index, _ in cls._streamed_tool_call_blocks(responses_so_far)) if len(block_indices) != len(post_guardrail_tool_calls): from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite raise UndeliverableStreamRewrite( guardrail_name, f"the guardrail returned {len(post_guardrail_tool_calls)} tool calls for a stream that carried " - f"{len(block_indices)} tool_use blocks", + f"{len(block_indices)} tool_use or server_tool_use blocks", ) rewrites_by_block: Final = MappingProxyType( { @@ -1636,6 +1855,60 @@ class AnthropicMessagesHandler(BaseTranslation): ), ) + @classmethod + def _streamed_web_search_result_strings(cls, responses_so_far: Sequence[object]) -> tuple[_WebSearchString, ...]: + events: Final = chain.from_iterable(cls._iter_sse_events(item) for item in responses_so_far) + result_blocks: Final = ( + (index, block) + for event in events + if event.get("type") == "content_block_start" + and isinstance(index := event.get("index"), int) + and (block := as_str_mapping(event.get("content_block"))) is not None + ) + return tuple( + chain.from_iterable( + _web_search_block_strings(index, block, with_query=False) for index, block in result_blocks + ) + ) + + @classmethod + def _write_ended_stream_web_search_rewrites( + cls, + responses_so_far: MutableSequence[object], # mutable-ok: rewrites the caller's buffered chunks in place + rewrites_by_block: Mapping[int, tuple[tuple[_WebSearchString, str], ...]], + ) -> None: + """Deliver ended-stream guardrail rewrites of web search result titles and urls into the + ``content_block_start`` that carries each whole ``web_search_tool_result`` block.""" + if not rewrites_by_block: + return + + def rewrite_result_block(event: Mapping[str, object]) -> _SSEFieldRewrite | None: + index: Final = event.get("index") + block: Final = as_str_mapping(event.get("content_block")) + rewrites: Final = rewrites_by_block.get(index) if isinstance(index, int) else None + if event.get("type") != "content_block_start" or block is None or rewrites is None: + return None + return _SSEFieldRewrite( + "content_block", "content", _rewritten_web_search_block(block, rewrites).get("content") + ) + + cls._rewrite_ended_stream_events(responses_so_far, rewrite_result_block) + + @classmethod + def _streamed_tool_call_blocks( + cls, responses_so_far: Sequence[object] + ) -> tuple[tuple[int, Mapping[str, object]], ...]: + events: Final = chain.from_iterable(cls._iter_sse_events(item) for item in responses_so_far) + return tuple( + (index, block) + for event in events + if event.get("type") == "content_block_start" + and (block := as_str_mapping(event.get("content_block"))) is not None + and isinstance(block_type := block.get("type"), str) + and block_type in _REBUILT_TOOL_CALL_BLOCK_TYPES + and isinstance(index := event.get("index"), int) + ) + @classmethod def _streamed_tool_use_ids(cls, responses_so_far: Sequence[object]) -> Mapping[int, str]: """The id of each buffered ``tool_use`` block, by its content block index.""" @@ -1738,8 +2011,10 @@ class AnthropicMessagesHandler(BaseTranslation): def get_streaming_scan_key(self, responses_so_far: Sequence[object]) -> StreamingScanKey | None: stream_ended: Final = self._check_streaming_has_ended(responses_so_far) tool_use_fingerprints: Final = self._streamed_tool_use_fingerprints(responses_so_far) + streamed_text: Final = self.get_streaming_string_so_far(responses_so_far) + result_strings: Final = self._streamed_web_search_result_strings(responses_so_far) if stream_ended else () return StreamingScanKey( - texts=(self.get_streaming_string_so_far(responses_so_far),), + texts=(*(target.value for target in result_strings), streamed_text), tool_calls=tool_use_fingerprints if stream_ended else (), stream_ended=stream_ended, tool_calls_in_flight=bool(tool_use_fingerprints) and not stream_ended, diff --git a/litellm/llms/anthropic/messages/native/fake_stream_iterator.py b/litellm/llms/anthropic/messages/native/fake_stream_iterator.py index 39bdfbbc53a..3a26737a8ac 100644 --- a/litellm/llms/anthropic/messages/native/fake_stream_iterator.py +++ b/litellm/llms/anthropic/messages/native/fake_stream_iterator.py @@ -12,11 +12,20 @@ import json from collections.abc import Mapping, Sequence from typing import Final, cast +from pydantic import TypeAdapter + from litellm.llms.anthropic.messages.native.utils import anthropic_message_delta from litellm.types.llms.anthropic_messages.anthropic_response import ( AnthropicMessagesResponse, ) +_CITATIONS: Final = TypeAdapter(tuple[object, ...]) + + +def _citations_of(block: Mapping[str, object]) -> tuple[object, ...]: + citations: Final = block.get("citations") + return _CITATIONS.validate_python(citations) if isinstance(citations, list) else () + class FakeAnthropicMessagesStreamIterator: """ @@ -47,16 +56,30 @@ class FakeAnthropicMessagesStreamIterator: def _create_content_block_chunks(self, block_dict: Mapping[str, object], index: int) -> list[bytes]: """Build SSE chunks for a single content block.""" - chunks: Final = [] + chunks: Final = list[bytes]() block_type: Final = block_dict.get("type") if block_type == "text": + citations: Final = _citations_of(block_dict) content_block_start = { "type": "content_block_start", "index": index, - "content_block": {"type": "text", "text": ""}, + "content_block": {"type": "text", "text": "", "citations": ()} + if citations + else {"type": "text", "text": ""}, } chunks.append(f"event: content_block_start\ndata: {json.dumps(content_block_start)}\n\n".encode()) + citation_deltas: Final = ( + { + "type": "content_block_delta", + "index": index, + "delta": {"type": "citations_delta", "citation": citation}, + } + for citation in citations + ) + chunks.extend( + f"event: content_block_delta\ndata: {json.dumps(delta)}\n\n".encode() for delta in citation_deltas + ) text: Final = block_dict.get("text", "") content_block_delta = { "type": "content_block_delta", @@ -89,36 +112,23 @@ class FakeAnthropicMessagesStreamIterator: } chunks.append(f"event: content_block_delta\ndata: {json.dumps(signature_delta)}\n\n".encode()) - elif block_type == "redacted_thinking": - content_block_start = { + elif block_type in ("tool_use", "server_tool_use"): + empty_input: Final[Mapping[str, object]] = {} + tool_use_start: Final = { "type": "content_block_start", "index": index, - "content_block": {"type": "redacted_thinking"}, + "content_block": {**block_dict, "input": empty_input}, } - chunks.append(f"event: content_block_start\ndata: {json.dumps(content_block_start)}\n\n".encode()) - - elif block_type == "tool_use": - content_block_start = { - "type": "content_block_start", - "index": index, - "content_block": { - "type": "tool_use", - "id": block_dict.get("id"), - "name": block_dict.get("name"), - "input": {}, - }, - } - chunks.append(f"event: content_block_start\ndata: {json.dumps(content_block_start)}\n\n".encode()) - input_data: Final = block_dict.get("input", {}) - content_block_delta = { + chunks.append(f"event: content_block_start\ndata: {json.dumps(tool_use_start)}\n\n".encode()) + tool_use_delta: Final = { "type": "content_block_delta", "index": index, "delta": { "type": "input_json_delta", - "partial_json": json.dumps(input_data), + "partial_json": json.dumps(block_dict.get("input", {})), }, } - chunks.append(f"event: content_block_delta\ndata: {json.dumps(content_block_delta)}\n\n".encode()) + chunks.append(f"event: content_block_delta\ndata: {json.dumps(tool_use_delta)}\n\n".encode()) else: passthrough_start: Final = { diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py index 4ad29bd3478..47572dd7ceb 100644 --- a/litellm/proxy/policy_engine/pipeline_executor.py +++ b/litellm/proxy/policy_engine/pipeline_executor.py @@ -98,15 +98,11 @@ def _count_change(sent: tuple[object, ...] | None, returned: tuple[object, ...] return (len(sent), len(returned)) -def _tool_call_mismatch_reason( - sent: tuple[tuple[object, object], ...] | None, returned: tuple[tuple[object, object], ...] | None -) -> str | None: - if sent == returned: - return None +def _tool_call_count_mismatch_reason(sent: Sequence[object] | None, returned: Sequence[object] | None) -> str | None: sent_count: Final = len(sent or ()) returned_count: Final = len(returned or ()) if sent_count == returned_count: - return "the legacy hook changed a tool call's name or arguments, which this path cannot write back" + return None return f"the legacy hook returned {returned_count} tool calls for a stream that carried {sent_count}" @@ -205,11 +201,10 @@ class _LegacyHookStreamAdapter(CustomGuardrail): non-streaming hooks, an exception it raises ends the stream through the executor's fail/error classification, and the response it hands back, or the one it changed in place and returned ``None`` for, is re-scanned by the same translation so its texts reach the - client through the translation's ended-stream write-back. A - replacement whose scanned texts do not line up with the originals, or whose tool calls - differ from them, is undeliverable, so the executor releases the original chunks. A stream - that carried no text to scan, such as a tool-only Anthropic message, stays deliverable as - long as the hook left the tool calls alone.""" + client through the translation's ended-stream write-back. Tool calls the hook rewrote go to the translation's + tool-call write-back like a unified guardrail's. A replacement whose scanned texts or tool + calls do not line up with the originals by count is undeliverable, so the executor releases + the original chunks.""" def __init__( self, @@ -258,14 +253,19 @@ class _LegacyHookStreamAdapter(CustomGuardrail): guardrail_name, f"the legacy hook returned {returned_text_count} texts for a stream that carried {sent_text_count}", ) - tool_call_mismatch: Final = _tool_call_mismatch_reason( - _tool_call_shapes(inputs.get("tool_calls")), _tool_call_shapes(rescanned.get("tool_calls")) + sent_tool_calls: Final = inputs.get("tool_calls") + rewritten_tool_calls: Final = rescanned.get("tool_calls") + tool_call_count_mismatch: Final = _tool_call_count_mismatch_reason(sent_tool_calls, rewritten_tool_calls) + if tool_call_count_mismatch is not None: + raise UndeliverableStreamRewrite(guardrail_name, tool_call_count_mismatch) + texts_update: Final[GenericGuardrailAPIInputs] = {"texts": rewritten} if rewritten else {} + tool_calls_update: Final[GenericGuardrailAPIInputs] = ( + {"tool_calls": rewritten_tool_calls} + if rewritten_tool_calls is not None + and _rewrote(_tool_call_shapes(sent_tool_calls), _tool_call_shapes(rewritten_tool_calls)) + else {} ) - if tool_call_mismatch is not None: - raise UndeliverableStreamRewrite(guardrail_name, tool_call_mismatch) - if not rewritten: - return inputs - rewritten_inputs: Final[GenericGuardrailAPIInputs] = {**inputs, "texts": rewritten} + rewritten_inputs: Final[GenericGuardrailAPIInputs] = {**inputs, **texts_update, **tool_calls_update} return rewritten_inputs async def _rescan( diff --git a/litellm/types/llms/anthropic.py b/litellm/types/llms/anthropic.py index 50c4a2340a3..a648376c542 100644 --- a/litellm/types/llms/anthropic.py +++ b/litellm/types/llms/anthropic.py @@ -713,6 +713,23 @@ class AnthropicResponseContentBlockRedactedThinking(LiteLLMBaseModel): data: str +class AnthropicResponseContentBlockServerToolUse(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True) + + type: Literal["server_tool_use"] + id: str + name: str + input: Mapping[str, object] + + +class AnthropicResponseContentBlockWebSearchToolResult(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True) + + type: Literal["web_search_tool_result"] + tool_use_id: str + content: Sequence[Mapping[str, object]] | Mapping[str, object] + + class AnthropicResponseUsageBlock(LiteLLMBaseModel): model_config = ConfigDict(extra="allow") @@ -744,6 +761,8 @@ class AnthropicResponse(LiteLLMBaseModel): | AnthropicResponseContentBlockToolUse | AnthropicResponseContentBlockThinking | AnthropicResponseContentBlockRedactedThinking + | AnthropicResponseContentBlockServerToolUse + | AnthropicResponseContentBlockWebSearchToolResult ] """Content generated by the model.""" diff --git a/tests/integration/messages_endpoint/providers/anthropic/test_websearch_interception_wire.py b/tests/integration/messages_endpoint/providers/anthropic/test_websearch_interception_wire.py index 5ec860910e5..1bd9f18253e 100644 --- a/tests/integration/messages_endpoint/providers/anthropic/test_websearch_interception_wire.py +++ b/tests/integration/messages_endpoint/providers/anthropic/test_websearch_interception_wire.py @@ -29,6 +29,24 @@ def sse_events(text: str) -> tuple[tuple[str, dict[str, object]], ...]: ) +def streamed_blocks(events: tuple[tuple[str, dict[str, object]], ...]) -> tuple[dict[str, object], ...]: + input_json: Final = tuple( + (event["index"], event["delta"]["partial_json"]) + for name, event in events + if name == "content_block_delta" and event["delta"]["type"] == "input_json_delta" + ) + return tuple( + { + **event["content_block"], + "input": json.loads("".join(part for index, part in input_json if index == event["index"])), + } + if any(index == event["index"] for index, _ in input_json) + else event["content_block"] + for name, event in events + if name == "content_block_start" + ) + + @pytest.mark.covers("other.provider_wire.bedrock.websearch_interception_streamed_capped_turn_ends_with_native_results") def test_streamed_web_search_turn_capped_by_max_agentic_loops_ends_turn_with_snippets_and_ordered_blocks( gateway: Gateway, tmp_path: Path @@ -136,7 +154,7 @@ def test_streamed_web_search_turn_capped_by_max_agentic_loops_ends_turn_with_sni for earlier, earlier_event in events[:position] if earlier == "content_block_start" }, response.text - started: Final = tuple(event["content_block"] for name, event in events if name == "content_block_start") + started: Final = streamed_blocks(events) search_ids: Final = tuple(block["id"] for block in started if block["type"] == "server_tool_use") assert search_ids and all(search_id.startswith("srvtoolu_") for search_id in search_ids), response.text assert started[-1] == {"type": "text", "text": ""}, response.text @@ -162,7 +180,12 @@ def test_streamed_web_search_turn_capped_by_max_agentic_loops_ends_turn_with_sni ) ), response.text assert ( - "".join(event["delta"]["text"] for name, event in events if name == "content_block_delta") == "turn-1" + "".join( + event["delta"]["text"] + for name, event in events + if name == "content_block_delta" and event["delta"]["type"] == "text_delta" + ) + == "turn-1" ), response.text assert [event["delta"]["stop_reason"] for name, event in events if name == "message_delta"] == [ "end_turn" diff --git a/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py b/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py index ef050d96dc5..10cb6ff12e9 100644 --- a/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py +++ b/tests/unit/integrations/websearch_interception/test_websearch_short_circuit.py @@ -5,7 +5,9 @@ Tests the short-circuit path that detects web-search-only /v1/messages requests and executes the search directly without routing through the backend LLM. """ +import datetime import json +from collections.abc import Iterator, Mapping, Sequence from typing import Final from unittest.mock import AsyncMock, patch @@ -16,6 +18,7 @@ import litellm from litellm.integrations.websearch_interception.handler import ( WebSearchInterceptionLogger, ) +from litellm.messages.dispatch import MessagesResult # --------------------------------------------------------------------------- # Detection tests @@ -73,6 +76,42 @@ class TestTryShortCircuitSearch: assert "Result" in text_block["text"] mock_search.assert_called_once_with("Search for Claude Code releases") + @pytest.mark.asyncio + async def test_emits_native_blocks_when_tools_already_converted(self, perplexity_search: respx.Route): + """Pre-request/deployment hooks convert the native tool to the LiteLLM + standard shape before the short-circuit runs and record that in + WEBSEARCH_EMIT_NATIVE_BLOCKS_KEY; the flag alone must be enough for + native blocks to be emitted.""" + from litellm.integrations.websearch_interception.handler import ( + WEBSEARCH_EMIT_NATIVE_BLOCKS_KEY, + ) + + logger: Final = WebSearchInterceptionLogger(enabled_providers=["hosted_vllm"]) + + result: Final = await logger.try_short_circuit_search( + model="hosted_vllm/some-model", + messages=[{"role": "user", "content": "Search for something"}], + tools=[ + { + "name": "litellm_web_search", + "description": "Search the web for information.", + "input_schema": { + "type": "object", + "properties": {"query": {"type": "string"}}, + "required": ["query"], + }, + } + ], + custom_llm_provider="hosted_vllm", + kwargs={WEBSEARCH_EMIT_NATIVE_BLOCKS_KEY: True}, + ) + + assert perplexity_search.call_count == 1 + assert result is not None + assert [b["type"] for b in result["content"]] == ["server_tool_use", "web_search_tool_result", "text"] + assert result["content"][0]["name"] == "web_search" + assert [r["url"] for r in result["content"][1]["content"]] == ["https://example.com/releases"] + @pytest.mark.asyncio async def test_does_not_short_circuit_mixed_tools(self): """Mix of web_search and other tools → NOT short-circuited""" @@ -468,6 +507,89 @@ class TestShortCircuitEntryPoint: assert text_block["text"] == "results" +class TestShortCircuitThroughMessagesEntryPoint: + """The pre-request hook renames the native tool before the short-circuit + runs, so only a request through anthropic_messages shows whether native + clients still get their server_tool_use and web_search_tool_result blocks.""" + + NATIVE_TOOL: Final = {"type": "web_search_20250305", "name": "web_search"} + + @pytest.fixture(autouse=True) + def _intercept_web_search(self, perplexity_search: respx.Route) -> Iterator[respx.Route]: + with patch("litellm.callbacks", [WebSearchInterceptionLogger(enabled_providers=["bedrock"])]): + yield perplexity_search + + @staticmethod + async def _send(*, tools: Sequence[Mapping[str, object]], stream: bool, **kwargs: object) -> MessagesResult: + return await litellm.anthropic_messages( + model="bedrock/converse/us.anthropic.claude-sonnet-5-5", + messages=[{"role": "user", "content": "latest python release"}], + tools=tools, + max_tokens=64, + stream=stream, + **kwargs, + ) + + @pytest.mark.asyncio + async def test_native_tool_gets_native_blocks(self, perplexity_search: respx.Route): + response: Final = await self._send(tools=[self.NATIVE_TOOL], stream=False) + + assert perplexity_search.call_count == 1 + assert [block["type"] for block in response["content"]] == [ + "server_tool_use", + "web_search_tool_result", + "text", + ] + server_tool_use, tool_result, text = response["content"] + assert server_tool_use["name"] == "web_search" + assert server_tool_use["input"] == {"query": "latest python release"} + assert tool_result["tool_use_id"] == server_tool_use["id"] + assert [result["url"] for result in tool_result["content"]] == ["https://example.com/releases"] + assert "Release notes" in text["text"] + assert "Release notes" not in json.dumps(tool_result) + + @pytest.mark.asyncio + async def test_native_blocks_reply_can_be_costed(self): + from litellm.litellm_core_utils.litellm_logging import Logging + + logging_obj: Final = Logging( + model="bedrock/converse/us.anthropic.claude-sonnet-5-5", + messages=[{"role": "user", "content": "latest python release"}], + stream=False, + call_type="anthropic_messages", + start_time=datetime.datetime(2026, 1, 1), + litellm_call_id="short-circuit-cost", + function_id="short-circuit-cost", + ) + logging_obj.update_environment_variables(litellm_params={}, optional_params={}) + response: Final = await self._send(tools=[self.NATIVE_TOOL], stream=False, litellm_logging_obj=logging_obj) + + assert logging_obj.response_cost_calculator(result=response) == 0.0 + + @pytest.mark.asyncio + async def test_native_tool_streams_native_blocks(self): + stream: Final = await self._send(tools=[self.NATIVE_TOOL], stream=True) + body: Final = b"".join([chunk async for chunk in stream]) + + starts: Final = [ + json.loads(line.removeprefix(b"data: "))["content_block"] + for line in body.splitlines() + if line.startswith(b"data: ") and b'"content_block_start"' in line + ] + assert [block["type"] for block in starts] == ["server_tool_use", "web_search_tool_result", "text"] + assert starts[1]["tool_use_id"] == starts[0]["id"] + + @pytest.mark.asyncio + async def test_litellm_standard_tool_stays_text_only(self): + from litellm.integrations.websearch_interception.tools import ( + get_litellm_web_search_tool, + ) + + response: Final = await self._send(tools=[get_litellm_web_search_tool()], stream=False) + + assert [block["type"] for block in response["content"]] == ["text"] + + class TestShortCircuitSafeguardResults: """A short-circuited reply never reaches the backend, so the proxy answers Claude Code's safeguards itself.""" diff --git a/tests/unit/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py b/tests/unit/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py index 141ad8a0481..d815e1eedd3 100644 --- a/tests/unit/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py +++ b/tests/unit/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py @@ -6,6 +6,7 @@ with guardrail transformations, specifically testing edge cases with empty choic """ import json +import re from typing import Any, Literal, Optional, Final from unittest.mock import MagicMock, patch @@ -552,13 +553,25 @@ class TestAnthropicMessagesHandlerStreamingOutputProcessing: assert chunks == original @pytest.mark.asyncio - async def test_deliver_ended_stream_tool_use_rewrite_with_server_tool_use_block_fails_closed(self): - from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite - + async def test_deliver_ended_stream_tool_use_rewrite_lands_on_server_tool_use_and_tool_use_blocks(self): handler = AnthropicMessagesHandler() server_tool_use = [ - ("content_block_start", {"type": "content_block_start", "index": 0, "content_block": {"type": "server_tool_use", "id": "srvtoolu_1", "name": "web_search", "input": {}}}), - ("content_block_delta", {"type": "content_block_delta", "index": 0, "delta": {"type": "input_json_delta", "partial_json": '{"query": "fruit"}'}}), + ( + "content_block_start", + { + "type": "content_block_start", + "index": 0, + "content_block": {"type": "server_tool_use", "id": "srvtoolu_1", "name": "web_search", "input": {}}, + }, + ), + ( + "content_block_delta", + { + "type": "content_block_delta", + "index": 0, + "delta": {"type": "input_json_delta", "partial_json": '{"query": "fruit"}'}, + }, + ), ("content_block_stop", {"type": "content_block_stop", "index": 0}), ] tool_use = self._ended_tool_use_sse_chunks() @@ -568,6 +581,26 @@ class TestAnthropicMessagesHandlerStreamingOutputProcessing: + [chunk.replace(b'"index": 0', b'"index": 1') for chunk in tool_use[1:]] ) + await handler.process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=self._argument_masking_guardrail(), + litellm_logging_obj=MagicMock(), + deliver_ended_stream_rewrites=True, + ) + + assert self._partial_jsons(chunks) == ['{"fruit": "[MASKED]"}', '{"fruit": "[MASKED]"}', "", ""] + raw = b"".join(chunks).decode() + assert '"id": "srvtoolu_1"' in raw and '"id": "toolu_1"' in raw + assert "persim" not in raw and '"query"' not in raw + + @pytest.mark.asyncio + async def test_deliver_ended_stream_tool_use_rewrite_on_an_unindexed_tool_block_fails_closed(self): + from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite + + handler = AnthropicMessagesHandler() + tool_use = self._ended_tool_use_sse_chunks() + chunks = tool_use[:1] + [tool_use[1].replace(b'"index": 0', b'"index": "0"')] + tool_use[2:] + with pytest.raises(UndeliverableStreamRewrite): await handler.process_output_streaming_response( responses_so_far=chunks, @@ -576,6 +609,270 @@ class TestAnthropicMessagesHandlerStreamingOutputProcessing: deliver_ended_stream_rewrites=True, ) + @staticmethod + def _python_masking_guardrail() -> CustomGuardrail: + class MaskPython(CustomGuardrail): + async def apply_guardrail( + self, + inputs: GenericGuardrailAPIInputs, + request_data: dict[str, object], + input_type: Literal["request", "response"], + logging_obj: Optional[LiteLLMLoggingObj] = None, + ) -> GenericGuardrailAPIInputs: + for tool_call in inputs.get("tool_calls", []): + if isinstance(tool_call, ChatCompletionMessageToolCall): + tool_call.function.arguments = re.sub( + "python", "[MASKED]", tool_call.function.arguments, flags=re.I + ) + return { + **inputs, + "texts": [re.sub("python", "[MASKED]", text, flags=re.I) for text in inputs.get("texts", [])], + } + + return MaskPython(guardrail_name="test") + + @staticmethod + def _python_texts_masking_guardrail() -> CustomGuardrail: + class MaskPythonTexts(CustomGuardrail): + async def apply_guardrail( + self, + inputs: GenericGuardrailAPIInputs, + request_data: dict[str, object], + input_type: Literal["request", "response"], + logging_obj: Optional[LiteLLMLoggingObj] = None, + ) -> GenericGuardrailAPIInputs: + return { + **inputs, + "texts": [re.sub("python", "[MASKED]", text, flags=re.I) for text in inputs.get("texts", [])], + } + + return MaskPythonTexts(guardrail_name="test") + + @staticmethod + def _web_search_result_block() -> dict[str, object]: + return { + "type": "web_search_tool_result", + "tool_use_id": "srvtoolu_1", + "content": [ + { + "type": "web_search_result", + "url": "https://docs.python.org/3.14/", + "title": "Python 3.14 docs", + "encrypted_content": "", + "page_age": None, + }, + { + "type": "web_search_result", + "url": "https://peps.example/745", + "title": "PEP 745", + "encrypted_content": "", + "page_age": None, + }, + ], + } + + @classmethod + def _web_search_stream_chunks(cls, server_tool_block_type: str = "server_tool_use") -> list[bytes]: + events = [ + ( + "message_start", + { + "type": "message_start", + "message": { + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "claude-sonnet-5-5", + "content": [], + "stop_reason": None, + "usage": {"input_tokens": 1, "output_tokens": 0}, + }, + }, + ), + ( + "content_block_start", + { + "type": "content_block_start", + "index": 0, + "content_block": { + "type": server_tool_block_type, + "id": "srvtoolu_1", + "name": "web_search", + "input": {}, + }, + }, + ), + ( + "content_block_delta", + { + "type": "content_block_delta", + "index": 0, + "delta": {"type": "input_json_delta", "partial_json": '{"query": "python 3.14 release date"}'}, + }, + ), + ("content_block_stop", {"type": "content_block_stop", "index": 0}), + ( + "content_block_start", + {"type": "content_block_start", "index": 1, "content_block": cls._web_search_result_block()}, + ), + ("content_block_stop", {"type": "content_block_stop", "index": 1}), + ( + "content_block_start", + {"type": "content_block_start", "index": 2, "content_block": {"type": "text", "text": ""}}, + ), + ( + "content_block_delta", + { + "type": "content_block_delta", + "index": 2, + "delta": {"type": "text_delta", "text": "Python 3.14 shipped"}, + }, + ), + ("content_block_stop", {"type": "content_block_stop", "index": 2}), + ( + "message_delta", + { + "type": "message_delta", + "delta": {"stop_reason": "end_turn", "stop_sequence": None}, + "usage": {"output_tokens": 2}, + }, + ), + ("message_stop", {"type": "message_stop"}), + ] + return [f"event: {name}\ndata: {json.dumps(payload)}\n\n".encode() for name, payload in events] + + @pytest.mark.parametrize("server_tool_block_type", ["server_tool_use", "tool_use"]) + @pytest.mark.parametrize( + "guardrail_factory", + [_python_masking_guardrail, _python_texts_masking_guardrail], + ids=["texts_and_tool_calls", "texts_only"], + ) + @pytest.mark.asyncio + async def test_deliver_ended_stream_rewrites_mask_the_web_search_query_results_and_text( + self, guardrail_factory, server_tool_block_type + ): + chunks = self._web_search_stream_chunks(server_tool_block_type) + + await AnthropicMessagesHandler().process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=guardrail_factory(), + litellm_logging_obj=MagicMock(), + deliver_ended_stream_rewrites=True, + ) + + raw = b"".join(chunks).decode() + assert "python" not in raw.lower() + result_block = next( + json.loads(line[len("data:") :])["content_block"] + for line in raw.split("\n") + if line.startswith("data:") and '"web_search_tool_result"' in line + ) + assert [(result["title"], result["url"]) for result in result_block["content"]] == [ + ("[MASKED] 3.14 docs", "https://docs.[MASKED].org/3.14/"), + ("PEP 745", "https://peps.example/745"), + ] + assert '"id": "srvtoolu_1"' in raw and '"tool_use_id": "srvtoolu_1"' in raw + + @staticmethod + def _last_text_redacting_guardrail() -> CustomGuardrail: + class RedactLastText(CustomGuardrail): + async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None): + return {**inputs, "texts": ["[REDACTED] " + inputs["texts"][-1]]} + + return RedactLastText(guardrail_name="test") + + @pytest.mark.asyncio + async def test_output_response_keeps_the_reply_as_the_last_text_a_guardrail_scans(self): + response = { + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "claude-sonnet-5-5", + "content": [ + { + "type": "server_tool_use", + "id": "srvtoolu_1", + "name": "web_search", + "input": {"query": "python 3.14 release date"}, + }, + self._web_search_result_block(), + {"type": "text", "text": "Python 3.14 shipped"}, + ], + "stop_reason": "end_turn", + } + + await AnthropicMessagesHandler().process_output_response( + response=response, + guardrail_to_apply=self._last_text_redacting_guardrail(), + litellm_logging_obj=MagicMock(), + ) + + assert response["content"][0]["input"] == {"query": "python 3.14 release date"} + assert response["content"][1] == self._web_search_result_block() + assert response["content"][2] == {"type": "text", "text": "[REDACTED] Python 3.14 shipped"} + + @pytest.mark.asyncio + async def test_deliver_ended_stream_keeps_the_reply_as_the_last_text_a_guardrail_scans(self): + chunks = self._web_search_stream_chunks() + + await AnthropicMessagesHandler().process_output_streaming_response( + responses_so_far=chunks, + guardrail_to_apply=self._last_text_redacting_guardrail(), + litellm_logging_obj=MagicMock(), + deliver_ended_stream_rewrites=True, + ) + + raw = b"".join(chunks).decode() + assert "[REDACTED] Python 3.14 shipped" in raw + assert raw.count("[REDACTED]") == 1 + assert json.dumps(self._web_search_result_block()) in raw + + def test_ended_stream_scan_key_differs_from_a_mid_stream_key_by_the_web_search_results(self): + handler = AnthropicMessagesHandler() + chunks = self._web_search_stream_chunks() + + mid_stream_key = handler.get_streaming_scan_key(chunks[:-2]) + ended_key = handler.get_streaming_scan_key(chunks) + + assert mid_stream_key.texts == ("Python 3.14 shipped",) + assert ended_key != mid_stream_key + assert ended_key.texts[-1] == "Python 3.14 shipped" + assert {"Python 3.14 docs", "https://docs.python.org/3.14/"} <= set(ended_key.texts) + + @pytest.mark.asyncio + async def test_output_response_masks_the_web_search_query_results_and_text(self): + response = { + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "claude-sonnet-5-5", + "content": [ + { + "type": "server_tool_use", + "id": "srvtoolu_1", + "name": "web_search", + "input": {"query": "python 3.14 release date"}, + }, + self._web_search_result_block(), + {"type": "text", "text": "Python 3.14 shipped"}, + ], + "stop_reason": "end_turn", + } + + await AnthropicMessagesHandler().process_output_response( + response=response, + guardrail_to_apply=self._python_masking_guardrail(), + litellm_logging_obj=MagicMock(), + ) + + assert "python" not in json.dumps(response).lower() + assert response["content"][0]["input"] == {"query": "[MASKED] 3.14 release date"} + assert [(result["title"], result["url"]) for result in response["content"][1]["content"]] == [ + ("[MASKED] 3.14 docs", "https://docs.[MASKED].org/3.14/"), + ("PEP 745", "https://peps.example/745"), + ] + assert response["content"][2] == {"type": "text", "text": "[MASKED] 3.14 shipped"} + @pytest.mark.asyncio async def test_ended_stream_rewrite_leaves_chunks_untouched_by_default(self): handler = AnthropicMessagesHandler() diff --git a/tests/unit/llms/anthropic/messages/native/test_fake_stream_iterator.py b/tests/unit/llms/anthropic/messages/native/test_fake_stream_iterator.py index 233145f2706..ece5bb39225 100644 --- a/tests/unit/llms/anthropic/messages/native/test_fake_stream_iterator.py +++ b/tests/unit/llms/anthropic/messages/native/test_fake_stream_iterator.py @@ -1,6 +1,10 @@ import json -from collections.abc import Iterable, Mapping -from typing import Final +from collections.abc import Iterable, Mapping, Sequence +from typing import Final, cast + +import anthropic +import httpx +import pytest from litellm.llms.anthropic.messages.native.fake_stream_iterator import ( FakeAnthropicMessagesStreamIterator, @@ -91,3 +95,140 @@ def test_message_delta_omits_absent_optional_fields(): "usage": {"input_tokens": 10, "output_tokens": 2}, } assert _message_delta(response) == {"stop_reason": "end_turn", "stop_sequence": None} + + +_WEB_SEARCH_RESULT_BLOCK: Final = { + "type": "web_search_tool_result", + "tool_use_id": "srvtoolu_01", + "content": [ + { + "type": "web_search_result", + "url": "https://example.com/a", + "title": "Example A", + "encrypted_content": "abc123", + } + ], +} + + +def _content_events(content: Sequence[Mapping[str, object]]) -> tuple[tuple[str, dict[str, object]], ...]: + response: Final = { + "id": "msg_ws", + "type": "message", + "role": "assistant", + "model": "claude-sonnet-4-6", + "content": list(content), + "stop_reason": "end_turn", + "usage": {"input_tokens": 10, "output_tokens": 5}, + } + return tuple( + event for event in _decoded_events(FakeAnthropicMessagesStreamIterator(response=response)) + if event[0].startswith("content_block_") + ) + + +def test_server_tool_use_streams_empty_start_then_input_json_delta(): + """A real Anthropic stream opens server_tool_use with an empty input and + sends the query as an input_json_delta, which SDK accumulators read.""" + caller: Final = {"type": "code_execution_20260120", "tool_id": "srvtoolu_00"} + events: Final = _content_events( + [ + { + "type": "server_tool_use", + "id": "srvtoolu_01", + "name": "web_search", + "input": {"query": "example search"}, + "caller": caller, + } + ] + ) + assert [event_type for event_type, _ in events] == [ + "content_block_start", + "content_block_delta", + "content_block_stop", + ] + assert events[0][1]["content_block"] == { + "type": "server_tool_use", + "id": "srvtoolu_01", + "name": "web_search", + "input": {}, + "caller": caller, + } + delta: Final = events[1][1]["delta"] + assert isinstance(delta, dict) + assert delta["type"] == "input_json_delta" + assert json.loads(delta["partial_json"]) == {"query": "example search"} + + +def test_web_search_tool_result_rides_whole_in_content_block_start(): + events: Final = _content_events([_WEB_SEARCH_RESULT_BLOCK]) + assert [event_type for event_type, _ in events] == ["content_block_start", "content_block_stop"] + assert events[0][1]["content_block"] == _WEB_SEARCH_RESULT_BLOCK + + +def test_native_search_blocks_keep_indices_aligned_with_text(): + events: Final = _content_events( + [ + {"type": "server_tool_use", "id": "srvtoolu_01", "name": "web_search", "input": {"query": "q"}}, + _WEB_SEARCH_RESULT_BLOCK, + {"type": "text", "text": "Here is what I found."}, + ] + ) + starts: Final = tuple(payload for event_type, payload in events if event_type == "content_block_start") + stops: Final = tuple(payload["index"] for event_type, payload in events if event_type == "content_block_stop") + assert [(start["index"], cast(dict[str, object], start["content_block"])["type"]) for start in starts] == [ + (0, "server_tool_use"), + (1, "web_search_tool_result"), + (2, "text"), + ] + assert stops == (0, 1, 2) + + +_CALLER: Final = {"type": "code_execution_20260120", "tool_id": "srvtoolu_00"} + +_REPLAYED_BLOCKS: Final = ( + {"type": "thinking", "thinking": "Look up the release page.", "signature": "sig_01"}, + {"type": "redacted_thinking", "data": "EncryptedThinking=="}, + {"type": "tool_use", "id": "toolu_01", "name": "get_weather", "input": {"city": "Paris"}, "caller": _CALLER}, + {"type": "server_tool_use", "id": "srvtoolu_01", "name": "web_fetch", "input": {"url": "https://example.com/a"}, "caller": _CALLER}, + _WEB_SEARCH_RESULT_BLOCK, + { + "type": "text", + "text": "Example A is the newest release.", + "citations": [ + { + "type": "web_search_result_location", + "cited_text": "Example A was released today.", + "url": "https://example.com/a", + "title": "Example A", + "encrypted_index": "idx_01", + } + ], + }, +) + + +def _sdk_rebuilt_content(block: Mapping[str, object]) -> list[dict[str, object]]: + response: Final = { + "id": "msg_replay", + "type": "message", + "role": "assistant", + "model": "claude-sonnet-5-5", + "content": [block], + "stop_reason": "end_turn", + "usage": {"input_tokens": 10, "output_tokens": 5}, + } + body: Final = b"".join(FakeAnthropicMessagesStreamIterator(response=response)) + transport: Final = httpx.MockTransport( + lambda _request: httpx.Response(200, content=body, headers={"content-type": "text/event-stream"}) + ) + client: Final = anthropic.Anthropic(api_key="sk-ant-test", http_client=httpx.Client(transport=transport)) + with client.messages.stream( + model="claude-sonnet-5-5", max_tokens=64, messages=[{"role": "user", "content": "hi"}] + ) as stream: + return [content_block.model_dump(exclude_none=True) for content_block in stream.get_final_message().content] + + +@pytest.mark.parametrize("block", _REPLAYED_BLOCKS, ids=lambda block: str(block["type"])) +def test_anthropic_sdk_rebuilds_every_block_field_from_the_stream(block: Mapping[str, object]): + assert _sdk_rebuilt_content(block) == [block] diff --git a/tests/unit/proxy/policy_engine/test_pipeline_executor.py b/tests/unit/proxy/policy_engine/test_pipeline_executor.py index c42878bd4f7..da855753b38 100644 --- a/tests/unit/proxy/policy_engine/test_pipeline_executor.py +++ b/tests/unit/proxy/policy_engine/test_pipeline_executor.py @@ -5,9 +5,12 @@ Uses mock guardrails to validate pipeline execution without external services. """ import copy +import json import logging import pickle -from typing import Literal +from collections.abc import Callable, Iterable, Mapping, Sequence +from itertools import chain +from typing import Final, Literal from unittest.mock import MagicMock import pytest @@ -15,6 +18,7 @@ import pytest import litellm from litellm.caching.dual_cache import DualCache from litellm.integrations.custom_guardrail import CustomGuardrail +from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicMessagesHandler from litellm.llms.base_llm.guardrail_translation.utils import UnappliableRequestRewrite from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.guardrails.guardrail_hooks.custom_code.custom_code_guardrail import ( @@ -1726,21 +1730,6 @@ async def test_streaming_step_discards_legacy_rewrite_whose_texts_do_not_line_up assert chunks == [_chunk()] -@pytest.mark.asyncio -async def test_streaming_step_discards_legacy_rewrite_that_changes_a_tool_call(monkeypatch, caplog): - masked_tool_call = {"function": {"name": "lookup", "arguments": '{"ssn": "[MASKED]"}'}} - guardrail = _LegacyHookGuardrail( - replacement=_legacy_replacement("[REWRITTEN] hello world", tool_calls=[masked_tool_call]) - ) - chunks = [_chunk()] - - with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): - result = await _run_legacy_streaming_step(monkeypatch, guardrail, chunks) - - _assert_passed_with_discard_warning(result, caplog, "the legacy hook changed a tool call's name or arguments") - assert chunks == [_chunk()] - - @pytest.mark.asyncio async def test_streaming_step_discards_legacy_rewrite_that_drops_the_tool_calls(monkeypatch, caplog): guardrail = _LegacyHookGuardrail(replacement=_legacy_replacement("[REWRITTEN] hello world", tool_calls=[])) @@ -1772,21 +1761,6 @@ async def test_streaming_step_passes_a_tool_only_stream_the_legacy_hook_left_alo assert not any("discarded" in record.getMessage() for record in caplog.records) -@pytest.mark.asyncio -async def test_streaming_step_discards_a_legacy_tool_call_rewrite_on_a_tool_only_stream(monkeypatch, caplog): - masked_tool_call = {"function": {"name": "lookup", "arguments": '{"ssn": "[MASKED]"}'}} - guardrail = _LegacyHookGuardrail(replacement=_legacy_replacement(tool_calls=[masked_tool_call])) - chunks = [_tool_only_chunk()] - - with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): - result = await _run_legacy_streaming_step( - monkeypatch, guardrail, chunks, translation=_ToolOnlyLegacyScanningTranslation() - ) - - _assert_passed_with_discard_warning(result, caplog, "the legacy hook changed a tool call's name or arguments") - assert chunks == [_tool_only_chunk()] - - class _NoHooksGuardrail(CustomGuardrail): pass @@ -1825,6 +1799,208 @@ async def test_streaming_step_discards_a_legacy_rewrite_the_translation_cannot_r assert chunks == [_chunk()] +_AnthropicHookResponse = dict[str, list[dict[str, object]]] + + +class _AnthropicTextRewritingLegacyGuardrail(_LegacyHookGuardrail): + async def async_post_call_success_hook( + self, data: dict[str, object], user_api_key_dict: UserAPIKeyAuth, response: _AnthropicHookResponse + ) -> None: + self.calls.append({"response": copy.deepcopy(response)}) + for block in response["content"]: + if block["type"] == "text" and isinstance(text := block.get("text"), str): + block["text"] = "[REWRITTEN] " + text + + +def _anthropic_web_search_stream() -> list[bytes]: + result_block: Final = { + "type": "web_search_tool_result", + "tool_use_id": "srvtoolu_1", + "content": [ + {"type": "web_search_result", "url": "https://example.test/a", "title": "A", "encrypted_content": ""} + ], + } + events: Final = ( + { + "type": "message_start", + "message": { + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "m", + "content": [], + "stop_reason": None, + "usage": {"input_tokens": 1, "output_tokens": 0}, + }, + }, + { + "type": "content_block_start", + "index": 0, + "content_block": {"type": "server_tool_use", "id": "srvtoolu_1", "name": "web_search", "input": {}}, + }, + { + "type": "content_block_delta", + "index": 0, + "delta": {"type": "input_json_delta", "partial_json": '{"query": "a"}'}, + }, + {"type": "content_block_stop", "index": 0}, + {"type": "content_block_start", "index": 1, "content_block": result_block}, + {"type": "content_block_stop", "index": 1}, + {"type": "content_block_start", "index": 2, "content_block": {"type": "text", "text": ""}}, + {"type": "content_block_delta", "index": 2, "delta": {"type": "text_delta", "text": "hello world"}}, + {"type": "content_block_stop", "index": 2}, + {"type": "message_delta", "delta": {"stop_reason": "end_turn"}, "usage": {"output_tokens": 2}}, + {"type": "message_stop"}, + ) + return _anthropic_sse(events) + + +def _anthropic_sse(events: Iterable[Mapping[str, object]]) -> list[bytes]: + return [f"event: {event['type']}\ndata: {json.dumps(event)}\n\n".encode() for event in events] + + +def _anthropic_tool_only_stream() -> list[bytes]: + return _anthropic_sse( + ( + { + "type": "message_start", + "message": { + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "m", + "content": [], + "stop_reason": None, + "usage": {"input_tokens": 1, "output_tokens": 0}, + }, + }, + { + "type": "content_block_start", + "index": 0, + "content_block": {"type": "tool_use", "id": "toolu_1", "name": "lookup", "input": {}}, + }, + { + "type": "content_block_delta", + "index": 0, + "delta": {"type": "input_json_delta", "partial_json": '{"ssn"'}, + }, + { + "type": "content_block_delta", + "index": 0, + "delta": {"type": "input_json_delta", "partial_json": ': "123"}'}, + }, + {"type": "content_block_stop", "index": 0}, + {"type": "message_delta", "delta": {"stop_reason": "tool_use"}, "usage": {"output_tokens": 2}}, + {"type": "message_stop"}, + ) + ) + + +def _streamed_events(chunks: Sequence[bytes]) -> list[dict[str, object]]: + return [ + json.loads(line[len("data: ") :]) + for line in b"".join(chunks).decode().splitlines() + if line.startswith("data: ") + ] + + +def _streamed_block_deltas(chunks: Sequence[bytes], index: int, field: str) -> str: + deltas: Final = ( + event["delta"] + for event in _streamed_events(chunks) + if event.get("index") == index and event["type"] == "content_block_delta" + ) + return "".join(str(delta[field]) for delta in deltas if isinstance(delta, dict)) + + +def _streamed_web_search_results(chunks: Sequence[bytes]) -> list[dict[str, str]]: + blocks: Final = ( + event["content_block"] for event in _streamed_events(chunks) if event["type"] == "content_block_start" + ) + results: Final = ( + block["content"] for block in blocks if isinstance(block, dict) and block["type"] == "web_search_tool_result" + ) + return list(chain.from_iterable(results)) + + +@pytest.mark.asyncio +async def test_streaming_step_delivers_a_legacy_hook_mask_of_web_search_result_titles( + monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture +) -> None: + chunks: Final = _anthropic_web_search_stream() + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + result: Final = await _run_legacy_streaming_step( + monkeypatch, _AnthropicMaskingLegacyGuardrail(), chunks, translation=AnthropicMessagesHandler() + ) + + assert [step.outcome for step in result.step_results] == ["pass"] + assert [(item["title"], item["url"]) for item in _streamed_web_search_results(chunks)] == [ + ("[MASKED]", "https://example.test/a") + ] + assert _streamed_block_deltas(chunks, 2, "text") == "[MASKED] world" + assert not any("discarded" in record.getMessage() for record in caplog.records) + + +class _AnthropicMaskingLegacyGuardrail(_LegacyHookGuardrail): + async def async_post_call_success_hook( + self, data: dict[str, object], user_api_key_dict: UserAPIKeyAuth, response: _AnthropicHookResponse + ) -> None: + for block in response["content"]: + if block["type"] == "text" and isinstance(text := block.get("text"), str): + block["text"] = text.replace("hello", "[MASKED]") + if block["type"] == "tool_use" and isinstance(tool_input := block.get("input"), dict): + block["input"] = {key: "[MASKED]" for key in tool_input} + if block["type"] == "web_search_tool_result" and isinstance(results := block.get("content"), list): + block["content"] = [{**result, "title": "[MASKED]"} for result in results] + + +@pytest.mark.parametrize( + "stream, masked_key", + [(_anthropic_web_search_stream, "query"), (_anthropic_tool_only_stream, "ssn")], + ids=["web_search_query", "tool_only"], +) +@pytest.mark.asyncio +async def test_streaming_step_delivers_a_legacy_hook_mask_of_a_streamed_tool_input( + monkeypatch: pytest.MonkeyPatch, + caplog: pytest.LogCaptureFixture, + stream: Callable[[], list[bytes]], + masked_key: str, +) -> None: + chunks: Final = stream() + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + result: Final = await _run_legacy_streaming_step( + monkeypatch, _AnthropicMaskingLegacyGuardrail(), chunks, translation=AnthropicMessagesHandler() + ) + + assert result.terminal_action == "allow" + assert [step.outcome for step in result.step_results] == ["pass"] + assert json.loads(_streamed_block_deltas(chunks, 0, "partial_json")) == {masked_key: "[MASKED]"} + assert "hello" not in b"".join(chunks).decode() + assert not any("discarded" in record.getMessage() for record in caplog.records) + + +@pytest.mark.asyncio +async def test_streaming_step_delivers_a_legacy_text_rewrite_on_a_web_search_stream( + monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture +) -> None: + guardrail: Final = _AnthropicTextRewritingLegacyGuardrail() + chunks: Final = _anthropic_web_search_stream() + + with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"): + result: Final = await _run_legacy_streaming_step( + monkeypatch, guardrail, chunks, translation=AnthropicMessagesHandler() + ) + + assert result.terminal_action == "allow" + assert [step.outcome for step in result.step_results] == ["pass"] + raw: Final = b"".join(chunks).decode() + assert "[REWRITTEN] hello world" in raw + assert '"title": "A"' in raw and '"url": "https://example.test/a"' in raw + assert not any("discarded" in record.getMessage() for record in caplog.records) + + @pytest.mark.asyncio async def test_later_legacy_step_sees_the_stream_as_the_earlier_step_left_it(monkeypatch): masker = _LegacyHookGuardrail(replacement=_legacy_replacement("[REWRITTEN] hello world")) From 56388fd13de12f13e7f49a43e0050ad590baa94d Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 30 Oct 2026 00:11:55 -0700 Subject: [PATCH 10/23] fix(count_tokens): count a history that carries a compaction block (#44784) * fix(count_tokens): count a compaction block instead of answering 500 A message history that replays Anthropic's signed compaction block made /v1/messages/count_tokens answer HTTP 500: neither the Anthropic nor the Vertex AI partner count-tokens request carried the compact-2026-09-04 beta, the provider rejected the block with a 400, the proxy fell back to the local tokenizer, and that raised on the unknown block type. Both provider count paths now add the compaction beta whenever the history carries a compaction block, and the local tokenizer counts the block's summary text like a text block, skipping the signature. * fix(token_counter): skip the tokenizer for empty thinking and compaction text * fix(count_tokens): forward context_management on Anthropic counts and count chat compaction blocks locally Vertex AI compaction support for inference and counting moves to its own PR, so the Vertex count handler is back to main here * fix(anthropic): map list-form context_management to edits for count_tokens and inference betas Count requests forwarded the OpenAI list form of context_management unchanged, while inference maps it to Anthropic edits, and the inference beta helper read the list as no edits. Both now map it once. A replayed compaction block with an empty summary no longer reaches the tokenizer. * fix(anthropic): forward count context_management only with edits, map Vertex list-form betas, and return provider context_management * fix(anthropic): drop an unmappable context_management list under drop_params --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/litellm_core_utils/token_counter.py | 43 ++++-- litellm/llms/anthropic/common_utils.py | 74 ++++++++-- .../llms/anthropic/count_tokens/handler.py | 4 +- .../anthropic/count_tokens/token_counter.py | 1 + .../anthropic/count_tokens/transformation.py | 31 ++++- .../messages/native/transformation.py | 26 +--- .../anthropic/messages/transformation.py | 25 +--- .../proxy/anthropic_endpoints/endpoints.py | 18 ++- .../litellm_core_utils/test_token_counter.py | 91 ++++++++++++ .../test_anthropic_messages_transformation.py | 62 ++++++++- ...t_anthropic_count_tokens_transformation.py | 131 +++++++++++++++++- .../llms/anthropic/test_count_tokens_oauth.py | 2 +- .../vertex_ai/count_tokens/test_context.py | 3 +- ...artner_models_anthropic_messages_config.py | 23 +++ .../anthropic_endpoints/test_endpoints.py | 68 ++++++++- 15 files changed, 520 insertions(+), 82 deletions(-) diff --git a/litellm/litellm_core_utils/token_counter.py b/litellm/litellm_core_utils/token_counter.py index ea3e645ffdf..800a0939cec 100644 --- a/litellm/litellm_core_utils/token_counter.py +++ b/litellm/litellm_core_utils/token_counter.py @@ -17,6 +17,7 @@ import anyio import anyio.lowlevel import httpx import tiktoken +from pydantic import TypeAdapter from tiktoken.model import MODEL_TO_ENCODING from typing_extensions import ParamSpec, TypeVar @@ -69,6 +70,7 @@ from litellm.types.llms.anthropic import ( AnthropicMessagesToolResultParam, AnthropicMessagesToolUseParam, AnthropicSearchResultParam, + CompactionBlock, ) from litellm.types.llms.openai import ( AllMessageValues, @@ -937,6 +939,8 @@ def _count_messages( default_token_count, params.pdf_pricing, ) + elif key == "provider_specific_fields": + num_tokens += sum(map(params.count_function, _compaction_summaries(value))) elif key == "search_results" and isinstance(value, list): from litellm.litellm_core_utils.prompt_templates.common_utils import ( extract_search_results_text, @@ -951,6 +955,25 @@ def _count_messages( return num_tokens +_OBJECT_LIST: Final = TypeAdapter(list[object]) + + +def _compaction_summaries(provider_specific_fields: object) -> tuple[str, ...]: + fields: Final = as_str_mapping(provider_specific_fields) + blocks: Final = fields.get("compaction_blocks") if fields is not None else None + block_fields: Final = ( + tuple(map(as_str_mapping, _OBJECT_LIST.validate_python(blocks))) if isinstance(blocks, list) else () + ) + return tuple( + content + for block in block_fields + if block is not None + and block.get("type") == "compaction" + and isinstance(content := block.get("content"), str) + and content + ) + + def _count_extra( count_function: TokenCounterFunction, tools: list[ChatCompletionToolParam] | None, @@ -1648,6 +1671,11 @@ def _count_anthropic_content( return tokens +_TEXT_FIELD_BY_BLOCK_TYPE: Final = MappingProxyType( + {"text": "text", "thinking": "thinking", "redacted_thinking": "thinking", "compaction": "content"} +) + + def _count_content_list( count_function: TokenCounterFunction, content_list: str @@ -1659,6 +1687,7 @@ def _count_content_list( | AnthropicMessagesImageParam | AnthropicMessagesDocumentParam | AnthropicSearchResultParam + | CompactionBlock ], use_default_image_token_count: bool, default_token_count: int | None, @@ -1670,8 +1699,10 @@ def _count_content_list( for c in content_list: if isinstance(c, str): num_tokens += count_function(c) - elif c["type"] == "text": - num_tokens += count_function(str(c.get("text", ""))) + elif c["type"] in _TEXT_FIELD_BY_BLOCK_TYPE: + block_text = str(c.get(_TEXT_FIELD_BY_BLOCK_TYPE[c["type"]]) or "") + if block_text: + num_tokens += count_function(block_text) elif c["type"] == "image_url": image_url = c.get("image_url") num_tokens += _count_image_tokens(image_url, use_default_image_token_count) @@ -1704,12 +1735,6 @@ def _count_content_list( default_token_count, pdf_pricing, ) - elif c["type"] in ("thinking", "redacted_thinking"): - # Claude extended thinking content block - # Count the thinking text and skip the opaque blobs (signature, redacted data) - thinking_text = str(c.get("thinking", "")) - if thinking_text: - num_tokens += count_function(thinking_text) elif c["type"] == "tool_reference": # Anthropic tool-search reference block: a lightweight pointer to # a deferred tool, e.g. {"type": "tool_reference", "tool_name": ...}. @@ -1727,7 +1752,7 @@ def _count_content_list( f"Invalid content item type: {content_type}. " f"Expected str or dict with 'type' field " f"(text, image_url, image, document, file, tool_use, tool_result, thinking, redacted_thinking, " - f"tool_reference, search_result)." + f"compaction, tool_reference, search_result)." ) return num_tokens except Exception as e: diff --git a/litellm/llms/anthropic/common_utils.py b/litellm/llms/anthropic/common_utils.py index 0bc519187f4..82977f97854 100644 --- a/litellm/llms/anthropic/common_utils.py +++ b/litellm/llms/anthropic/common_utils.py @@ -8,12 +8,13 @@ import re from collections.abc import Container, Mapping, MutableMapping, Sequence from dataclasses import dataclass from datetime import datetime, timezone +from itertools import chain from types import MappingProxyType from typing import Annotated, Any, ClassVar, Final, Literal, TypeVar from urllib.parse import quote, urlparse import httpx -from pydantic import BeforeValidator, ConfigDict, Field, StrictBool, TypeAdapter, ValidationError +from pydantic import BeforeValidator, ConfigDict, Field, JsonValue, StrictBool, TypeAdapter, ValidationError from typing_extensions import ReadOnly, TypedDict import litellm @@ -24,6 +25,7 @@ from litellm.constants import ( DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET, ) from litellm.exceptions import UnsupportedParamsError +from litellm.litellm_core_utils.core_helpers import as_str_mapping from litellm.litellm_core_utils.prompt_templates.common_utils import ( get_file_ids_from_messages, is_encrypted_reasoning_block, @@ -46,6 +48,7 @@ from litellm.llms.anthropic.wif import ( from litellm.llms.base_llm.base_utils import BaseLLMModelInfo, BaseTokenCounter from litellm.llms.base_llm.chat.transformation import BaseLLMException from litellm.types.llms.anthropic import ( + ANTHROPIC_BETA_HEADER_VALUES, ANTHROPIC_HOSTED_TOOLS, ANTHROPIC_MID_CONVERSATION_OUTPUT_CONFIG_BETA_HEADER, ANTHROPIC_MID_CONVERSATION_TOOL_CHANGES_BETA_HEADER, @@ -118,6 +121,7 @@ _CLAUDE_CODE_BILLING_HEADER_PREFIX: Final = "x-anthropic-billing-header:" OBJECT_MAPPING_ADAPTER: Final = TypeAdapter(dict[object, object]) OBJECT_LIST_ADAPTER: Final = TypeAdapter(list[object]) _DECODED_JSON: Final = TypeAdapter(object) +_JSON_VALUE: Final = TypeAdapter(JsonValue) _CLAUDE_CODE_USER_AGENT_PREFIXES: Final = ("claude-cli/", "claude-code/") @@ -177,24 +181,68 @@ def is_anthropic_messages_url(url: str) -> bool: return parsed_url.hostname == "api.anthropic.com" or parsed_url.path.removesuffix("/").endswith("/v1/messages") +def _message_blocks(message: object) -> Sequence[object]: + fields: Final = as_str_mapping(message) + blocks: Final = validated_object_list(fields.get("content")) if fields is not None else None + return blocks or () + + +def _is_signed_compaction_block(block: object) -> bool: + fields: Final = as_str_mapping(block) + if fields is None or fields.get("type") != "compaction": + return False + signature: Final = fields.get("signature") + return isinstance(signature, str) and bool(signature) + + +def carries_signed_compaction_block(messages: Sequence[object]) -> bool: + return any(map(_is_signed_compaction_block, chain.from_iterable(map(_message_blocks, messages)))) + + def requires_native_compaction_beta( custom_llm_provider: str, optional_params: Mapping[str, object], messages: Sequence[object], ) -> bool: return custom_llm_provider == "anthropic" and ( - optional_params.get("compaction") is not None - or any( - isinstance(block, Mapping) - and block.get("type") == "compaction" - and isinstance(block.get("signature"), str) - and bool(block.get("signature")) - for message in messages - if isinstance(message, Mapping) - for content in (message.get("content"),) - if isinstance(content, (list, tuple)) - for block in content + optional_params.get("compaction") is not None or carries_signed_compaction_block(messages) + ) + + +def anthropic_context_management(context_management: JsonValue) -> JsonValue: + if not isinstance(context_management, list): + return context_management + from litellm.llms.anthropic.chat.transformation import AnthropicConfig + + edits: Final = AnthropicConfig.map_openai_context_management_to_anthropic( + [entry for entry in context_management if isinstance(entry, dict)] + ) + return context_management if edits is None else _JSON_VALUE.validate_python(edits) + + +def _context_management_edits(spec: JsonValue) -> Sequence[object]: + mapping: Final = as_str_mapping(spec) + return (validated_object_list(mapping.get("edits")) if mapping is not None else None) or () + + +def context_management_with_edits(context_management: JsonValue) -> JsonValue: + spec: Final = anthropic_context_management(context_management) + return spec if _context_management_edits(spec) else None + + +def context_management_betas(context_management: JsonValue) -> tuple[str, ...]: + edits: Final = _context_management_edits(anthropic_context_management(context_management)) + edit_types: Final = tuple(edit.get("type") for edit in map(as_str_mapping, edits) if edit is not None) + return tuple( + beta + for beta, used in ( + (ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value, "compact_20260112" in edit_types), + ( + ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value, + any(edit_type != "compact_20260112" for edit_type in edit_types), + ), ) + if used ) @@ -1367,8 +1415,6 @@ class AnthropicModelInfo(BaseLLMModelInfo): if is_vertex_request is True: # Vertex AI requires web search beta header for web search to work if web_search_tool_used: - from litellm.types.llms.anthropic import ANTHROPIC_BETA_HEADER_VALUES - headers["anthropic-beta"] = ANTHROPIC_BETA_HEADER_VALUES.WEB_SEARCH_2025_03_05.value elif len(all_betas) > 0: headers["anthropic-beta"] = ",".join(all_betas) diff --git a/litellm/llms/anthropic/count_tokens/handler.py b/litellm/llms/anthropic/count_tokens/handler.py index f8bf7232d9c..7a7d1d776d3 100644 --- a/litellm/llms/anthropic/count_tokens/handler.py +++ b/litellm/llms/anthropic/count_tokens/handler.py @@ -39,6 +39,7 @@ class AnthropicCountTokensHandler(AnthropicCountTokensConfig): tools: list[dict[str, JsonValue]] | None = None, system: JsonValue = None, optional_params: Mapping[str, JsonValue] | None = None, + context_management: JsonValue = None, drop_params: bool = False, ) -> dict[str, JsonValue]: """ @@ -70,6 +71,7 @@ class AnthropicCountTokensHandler(AnthropicCountTokensConfig): tools=tools, system=system, optional_params=optional_params, + context_management=context_management, drop_params=drop_params, ) @@ -81,7 +83,7 @@ class AnthropicCountTokensHandler(AnthropicCountTokensConfig): verbose_logger.debug("Making request to: %s", endpoint_url) # Get required headers - headers: Final = self.get_count_tokens_headers(auth_header) + headers: Final = self.get_count_tokens_headers(auth_header, messages, context_management) # Use LiteLLM's async httpx client async_client: Final = get_async_httpx_client(llm_provider=litellm.LlmProviders.ANTHROPIC) diff --git a/litellm/llms/anthropic/count_tokens/token_counter.py b/litellm/llms/anthropic/count_tokens/token_counter.py index 91fc77773d1..99b1cd7f241 100644 --- a/litellm/llms/anthropic/count_tokens/token_counter.py +++ b/litellm/llms/anthropic/count_tokens/token_counter.py @@ -132,6 +132,7 @@ class AnthropicTokenCounter(BaseTokenCounter): tools=tools, system=system, optional_params=count_options, + context_management=context_management.spec if context_management is not None else None, drop_params=drops_unsupported_params(litellm_params.get("drop_params")), ) diff --git a/litellm/llms/anthropic/count_tokens/transformation.py b/litellm/llms/anthropic/count_tokens/transformation.py index e60ffe67c8d..1c8f0c206cd 100644 --- a/litellm/llms/anthropic/count_tokens/transformation.py +++ b/litellm/llms/anthropic/count_tokens/transformation.py @@ -27,11 +27,15 @@ from litellm.litellm_core_utils.prompt_templates.mid_conversation_system import ) from litellm.litellm_core_utils.url_utils import join_api_base from litellm.llms.anthropic.common_utils import ( + carries_signed_compaction_block, + context_management_betas, + context_management_with_edits, is_claude_model_target, merge_anthropic_beta_headers, tools_without_custom_field, ) from litellm.llms.anthropic.wif import resolve_anthropic_base +from litellm.types.llms.anthropic import ANTHROPIC_BETA_HEADER_VALUES from litellm.types.llms.openai import AllMessageValues, ChatCompletionImageObject, ValidUserMessageContentTypes from litellm.types.utils import TokenCountMessagesFormat, TokenCountResponse @@ -376,6 +380,7 @@ class AnthropicCountTokensConfig: tools: list[dict[str, JsonValue]] | None = None, system: JsonValue = None, optional_params: Mapping[str, JsonValue] | None = None, + context_management: JsonValue = None, drop_params: bool = False, ) -> dict[str, JsonValue]: # mutable-ok: provider transport requires JSON dictionaries """ @@ -408,6 +413,7 @@ class AnthropicCountTokensConfig: drop_custom=drop_params, ), ), + ("context_management", context_management_with_edits(context_management)), ) if value is not None } @@ -419,18 +425,33 @@ class AnthropicCountTokensConfig: ) ) - def get_count_tokens_headers(self, auth_header: Mapping[str, str]) -> dict[str, str]: + def get_count_tokens_headers( + self, + auth_header: Mapping[str, str], + messages: Sequence[Mapping[str, JsonValue]], + context_management: JsonValue = None, + ) -> dict[str, str]: """The count-tokens headers around a resolved Anthropic auth header (``AnthropicModelInfo.get_auth_header``): x-api-key for a static key, an Authorization bearer for ``ANTHROPIC_AUTH_TOKEN`` and for sk-ant-oat tokens, whose mandatory oauth beta - merges with the token-counting beta instead of replacing it.""" + merges with the token-counting beta instead of replacing it. A history that carries a + signed compaction block also needs the compaction beta, or Anthropic rejects the block + as an unknown content type, and ``context_management`` edits need the betas inference sends.""" + compaction_betas: Final = ( + (ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_09_04.value,) + if carries_signed_compaction_block(messages) + else () + ) + request_betas: Final = ( + ANTHROPIC_TOKEN_COUNTING_BETA_VERSION, + *compaction_betas, + *context_management_betas(context_management), + ) return { "Content-Type": "application/json", "anthropic-version": "2023-06-01", **auth_header, - "anthropic-beta": merge_anthropic_beta_headers( - auth_header.get("anthropic-beta"), ANTHROPIC_TOKEN_COUNTING_BETA_VERSION - ), + "anthropic-beta": merge_anthropic_beta_headers(auth_header.get("anthropic-beta"), request_betas), } def validate_request( diff --git a/litellm/llms/anthropic/messages/native/transformation.py b/litellm/llms/anthropic/messages/native/transformation.py index c086618e829..ef855eff180 100644 --- a/litellm/llms/anthropic/messages/native/transformation.py +++ b/litellm/llms/anthropic/messages/native/transformation.py @@ -43,6 +43,7 @@ from ...common_utils import ( OBJECT_LIST_ADAPTER, AnthropicError, AnthropicModelInfo, + context_management_betas, is_claude_model_target, merge_anthropic_beta_headers, optionally_handle_anthropic_oauth, @@ -745,6 +746,8 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig): ) if transformed_context_management is not None: anthropic_messages_optional_request_params["context_management"] = transformed_context_management + elif drops_unsupported_params(litellm_params.drop_params): + anthropic_messages_optional_request_params.pop("context_management") ####### get required params for all anthropic messages requests ###### # Lazy %s: the f-string previously stringified the entire messages @@ -873,28 +876,7 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig): if requires_native_compaction_beta(custom_llm_provider, optional_params, messages): beta_values.add(ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_09_04.value) - # Check for context management - context_management_param: Final = optional_params.get("context_management") - if context_management_param is not None: - # Check edits array for compact_20260112 type - edits: Final = context_management_param.get("edits", ()) - has_compact = False - has_other = False - - for edit in edits: - edit_type = edit.get("type", "") - if edit_type == "compact_20260112": - has_compact = True - else: - has_other = True - - # Add compact header if any compact edits exist - if has_compact: - beta_values.add(ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value) - - # Add context management header if any other edits exist - if has_other: - beta_values.add(ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value) + beta_values.update(context_management_betas(optional_params.get("context_management"))) # Check for structured outputs. Anthropic's newer request shape nests # the schema under output_config.format; the older top-level diff --git a/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/messages/transformation.py b/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/messages/transformation.py index bfa94396ea7..da82766c2b9 100644 --- a/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/messages/transformation.py +++ b/litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/messages/transformation.py @@ -3,7 +3,7 @@ from typing import Final from litellm.exceptions import UnsupportedParamsError from litellm.litellm_core_utils.core_helpers import drops_unsupported_params -from litellm.llms.anthropic.common_utils import AnthropicModelInfo +from litellm.llms.anthropic.common_utils import AnthropicModelInfo, context_management_betas from litellm.llms.anthropic.messages.native.transformation import ( AnthropicMessagesConfig, messages_carry_output_config, @@ -171,28 +171,7 @@ class VertexAIPartnerModelsAnthropicMessagesConfig(AnthropicMessagesConfig, Vert if existing_beta: beta_values.update(b.strip() for b in existing_beta.split(",")) - # Check for context management - context_management_param: Final = optional_params.get("context_management") - if context_management_param is not None: - # Check edits array for compact_20260112 type - edits: Final = context_management_param.get("edits", []) - has_compact = False - has_other = False - - for edit in edits: - edit_type = edit.get("type", "") - if edit_type == "compact_20260112": - has_compact = True - else: - has_other = True - - # Add compact header if any compact edits exist - if has_compact: - beta_values.add(ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value) - - # Add context management header if any other edits exist - if has_other: - beta_values.add(ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value) + beta_values.update(context_management_betas(optional_params.get("context_management"))) # Check for web search tool for tool in tools: diff --git a/litellm/proxy/anthropic_endpoints/endpoints.py b/litellm/proxy/anthropic_endpoints/endpoints.py index 1a2e610d5c4..0e204cf9211 100644 --- a/litellm/proxy/anthropic_endpoints/endpoints.py +++ b/litellm/proxy/anthropic_endpoints/endpoints.py @@ -8,7 +8,7 @@ from typing import Final from fastapi import APIRouter, Depends, HTTPException, Request, Response, status from fastapi.responses import JSONResponse -from pydantic import JsonValue, ValidationError +from pydantic import JsonValue, TypeAdapter, ValidationError from typing_extensions import ReadOnly, TypedDict import litellm @@ -54,6 +54,7 @@ from litellm.proxy.litellm_pre_call_utils import ( should_auto_drop_params_for_agentic_cli, ) from litellm.router import Router +from litellm.types.utils import TokenCountResponse router: Final = APIRouter() @@ -334,6 +335,15 @@ def _count_tokens_body_options(data: Mapping[str, JsonValue], auto_drop_params: } +_PROVIDER_COUNT_REPLY: Final[TypeAdapter[Mapping[str, object] | None]] = TypeAdapter(Mapping[str, object] | None) + + +def _provider_context_management(response: TokenCountResponse) -> Mapping[str, object]: + original: Final = _PROVIDER_COUNT_REPLY.validate_python(response.original_response) + context_management: Final = original.get("context_management") if original is not None else None + return {} if context_management is None else {"context_management": context_management} + + @router.post( "/v1/messages/count_tokens", tags=["[beta] Anthropic Messages Token Counting"], @@ -413,7 +423,11 @@ async def count_tokens( log_llm_api_exception(e, litellm_call_id) return _anthropic_error_json_response(proxy_exception_from_route_error(e), request) else: - return {"input_tokens": token_response.total_tokens, **count_source_fields(token_response)} + return { + "input_tokens": token_response.total_tokens, + **_provider_context_management(token_response), + **count_source_fields(token_response), + } @router.post( diff --git a/tests/unit/litellm_core_utils/test_token_counter.py b/tests/unit/litellm_core_utils/test_token_counter.py index c2f8e12a726..697ac76b4e5 100644 --- a/tests/unit/litellm_core_utils/test_token_counter.py +++ b/tests/unit/litellm_core_utils/test_token_counter.py @@ -1962,6 +1962,97 @@ def test_token_counter_with_redacted_thinking_content(): assert token_counter(model=model, messages=with_block) == token_counter(model=model, messages=without_block) + +def test_token_counter_with_compaction_block(): + """ + A replayed on-demand compaction block (Anthropic's signed summary of the earlier turns) counts its + summary text like a text block and nothing for its signature. It used to raise, which turned every + /v1/messages/count_tokens call on a compacted history into a 500 once the provider count fell back here. + """ + model = "anthropic/claude-sonnet-5-5" + summary = "The user is building a recipe app in Python and asked for one-sentence class descriptions." + follow_up = {"role": "user", "content": [{"type": "text", "text": "Now do the same for Ingredient."}]} + signed_block = {"type": "compaction", "content": summary, "signature": "EqQBCkYIBRgCKkBjZ2xhc3M" * 40} + + as_compaction = [{"role": "assistant", "content": [signed_block]}, follow_up] + unsigned = [{"role": "assistant", "content": [{"type": "compaction", "content": summary}]}, follow_up] + as_text = [{"role": "assistant", "content": [{"type": "text", "text": summary}]}, follow_up] + + assert token_counter(model=model, messages=as_compaction) == token_counter(model=model, messages=as_text) + assert token_counter(model=model, messages=unsigned) == token_counter(model=model, messages=as_text) + + +def test_token_counter_counts_chat_compaction_blocks_like_their_summary_text(): + """ + Chat clients replay a compaction block in the assistant message's provider_specific_fields, and inference + sends it to Anthropic as a content block, so the local count includes its summary text and skips its signature. + """ + model = "anthropic/claude-sonnet-5-5" + summary = "The user is building a recipe app in Python and asked for one-sentence class descriptions." + follow_up = {"role": "user", "content": "Now do the same for Ingredient."} + signed_block = {"type": "compaction", "content": summary, "signature": "EqQBCkYIBRgCKkBjZ2xhc3M" * 40} + + replayed = [ + {"role": "assistant", "content": None, "provider_specific_fields": {"compaction_blocks": [signed_block]}}, + follow_up, + ] + as_text = [{"role": "assistant", "content": summary}, follow_up] + + assert token_counter(model=model, messages=replayed) == token_counter(model=model, messages=as_text) + + +class _BosEncoder: + def encode_batch_fast(self, texts: list[str]) -> list[list[int]]: + return [[0] * (len(text.split()) + 1) for text in texts] + + +def test_token_counter_charges_nothing_for_a_chat_compaction_block_with_no_summary(): + """ + A tokenizer that prepends a BOS token charges one token for the empty string, so a replayed + compaction block with an empty summary must not reach it, like an empty thinking block. + """ + model = "anthropic/claude-sonnet-5-5" + tokenizer = {"type": "huggingface_tokenizer", "tokenizer": _BosEncoder()} + follow_up = {"role": "user", "content": "Now do the same for Ingredient."} + replayed = [ + { + "role": "assistant", + "content": None, + "provider_specific_fields": {"compaction_blocks": [{"type": "compaction", "content": ""}]}, + }, + follow_up, + ] + bare = [{"role": "assistant", "content": None}, follow_up] + + assert token_counter(model=model, messages=replayed, custom_tokenizer=tokenizer) == token_counter( + model=model, messages=bare, custom_tokenizer=tokenizer + ) + + +def test_count_content_list_skips_the_tokenizer_for_empty_block_text(): + """ + An empty thinking block, a redacted_thinking block (no thinking text at all) and a compaction + block with no summary carry no tokens, even under a tokenizer that charges for the empty string + (the llama-2 tokenizer prepends its BOS token to every encode call). Counting them through the + tokenizer added one token per block on such models. + """ + from litellm.litellm_core_utils.token_counter import _count_content_list + + def charges_for_empty_text(text: str) -> int: + return len(text.split()) + 1 + + empty_blocks = [ + {"type": "thinking", "thinking": "", "signature": "EqcLCkYICxgCKkCrqu6lP..."}, + {"type": "thinking", "thinking": None, "signature": "EqcLCkYICxgCKkCrqu6lP..."}, + {"type": "redacted_thinking", "data": "EmwKAhgBEgy3va3pzix/LafPsn4aDFIT2Xlxh0L5L8rLVyIw"}, + {"type": "compaction", "content": ""}, + {"type": "text", "text": ""}, + ] + + assert _count_content_list(charges_for_empty_text, empty_blocks, False, None) == 0 + assert _count_content_list(charges_for_empty_text, [{"type": "thinking", "thinking": "two words"}], False, None) == 3 + + def test_token_counter_with_tool_reference_block(): """ Regression test: a message containing an Anthropic tool-search diff --git a/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_transformation.py b/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_transformation.py index 1c004d8c36e..8fa9d534a8c 100644 --- a/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_transformation.py +++ b/tests/unit/llms/anthropic/messages/native/test_anthropic_messages_transformation.py @@ -24,7 +24,7 @@ from litellm.llms.tencent.messages.transformation import TencentAnthropicMessage from litellm.llms.vertex_ai.vertex_ai_partner_models.anthropic.experimental_pass_through.transformation import ( VertexAIPartnerModelsAnthropicMessagesConfig, ) -from litellm.types.llms.anthropic import AnthropicFinishReason +from litellm.types.llms.anthropic import ANTHROPIC_BETA_HEADER_VALUES, AnthropicFinishReason from litellm.types.router import GenericLiteLLMParams from tests.unit.llms.anthropic.test_anthropic_wif import ( ScriptedPoster, @@ -321,6 +321,66 @@ def _derived_beta_header(config: AnthropicMessagesConfig) -> dict: ) +_COMPACT_EDIT_BETA: Final = ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value +_CLEARING_EDIT_BETA: Final = ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value + + +@pytest.mark.parametrize( + ("context_management", "betas"), + ( + pytest.param({"edits": [{"type": "compact_20260112"}]}, {_COMPACT_EDIT_BETA}, id="compact"), + pytest.param({"edits": [{"type": "clear_tool_uses_20250919"}]}, {_CLEARING_EDIT_BETA}, id="clearing"), + pytest.param( + {"edits": [{"type": "compact_20260112"}, {"type": "clear_thinking_20251015"}]}, + {_COMPACT_EDIT_BETA, _CLEARING_EDIT_BETA}, + id="both", + ), + pytest.param( + [{"type": "compaction", "compact_threshold": 150000}], {_COMPACT_EDIT_BETA}, id="openai_list_form" + ), + ), +) +def test_context_management_edits_add_the_beta_each_edit_needs( + context_management: JsonValue, betas: set[str], local_beta_headers_config: None +) -> None: + headers: Final = AnthropicMessagesConfig()._update_headers_with_anthropic_beta( + headers={}, optional_params={"context_management": context_management}, messages=[] + ) + + assert set(headers["anthropic-beta"].split(",")) == betas + + +_UNMAPPABLE_LIST_FORM: Final = [{"type": "clear_tool_uses_20250919"}] + + +@pytest.mark.parametrize("config_class", [AnthropicMessagesConfig, VertexAIPartnerModelsAnthropicMessagesConfig]) +@pytest.mark.parametrize( + ("context_management", "drop_params", "forwarded"), + ( + pytest.param(_UNMAPPABLE_LIST_FORM, True, None, id="unmappable_dropped_under_drop_params"), + pytest.param(_UNMAPPABLE_LIST_FORM, False, _UNMAPPABLE_LIST_FORM, id="unmappable_kept_for_the_provider_400"), + pytest.param( + [{"type": "compaction"}], + True, + {"edits": [{"type": "compact_20260112"}]}, + id="mappable_still_mapped_under_drop_params", + ), + ), +) +def test_context_management_the_list_form_cannot_map_is_dropped_only_under_drop_params( + config_class: type[AnthropicMessagesConfig], context_management: JsonValue, drop_params: bool, forwarded: JsonValue +) -> None: + request: Final = config_class().transform_anthropic_messages_request( + model="claude-haiku-5-5", + messages=[{"role": "user", "content": "Hi"}], + anthropic_messages_optional_request_params={"max_tokens": 50, "context_management": context_management}, + litellm_params=GenericLiteLLMParams(drop_params=drop_params), + headers={}, + ) + + assert request.get("context_management") == forwarded + + @pytest.mark.parametrize("config", _classifier_configs(), ids=type) def test_routes_whose_platform_runs_the_classifier_keep_safeguards_and_the_beta( config: AnthropicMessagesConfig, local_beta_headers_config: None diff --git a/tests/unit/llms/anthropic/test_anthropic_count_tokens_transformation.py b/tests/unit/llms/anthropic/test_anthropic_count_tokens_transformation.py index a3e723eaca2..aa7fa2727b4 100644 --- a/tests/unit/llms/anthropic/test_anthropic_count_tokens_transformation.py +++ b/tests/unit/llms/anthropic/test_anthropic_count_tokens_transformation.py @@ -26,7 +26,24 @@ from litellm.llms.azure_ai.anthropic.count_tokens.token_counter import AzureAIAn from litellm.llms.azure_ai.anthropic.count_tokens.transformation import ( AzureAIAnthropicCountTokensConfig, ) -from litellm.types.utils import TokenCountMessagesFormat, TokenCountResponse +from litellm.types.utils import CountContextManagement, TokenCountMessagesFormat, TokenCountResponse +from litellm.constants import ANTHROPIC_TOKEN_COUNTING_BETA_VERSION +from litellm.types.llms.anthropic import ANTHROPIC_BETA_HEADER_VALUES + +COMPACTION_BETA = ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_09_04.value +COMPACTED_HISTORY = [ + { + "role": "assistant", + "content": [ + { + "type": "compaction", + "content": "The user is building a recipe app and asked for one-sentence class descriptions.", + "signature": "EqQBCkYIBRgCKkBjZ2xhc3M" * 40, + } + ], + }, + {"role": "user", "content": "Now do the same for Ingredient."}, +] @pytest.mark.parametrize( @@ -890,3 +907,115 @@ async def test_counter_sends_the_structured_output_tool_choice_and_thinking_chat inference["tool_choice"], inference["tools"], ) + + +def test_headers_add_the_compaction_beta_only_for_a_compacted_history(): + config = AnthropicCountTokensConfig() + auth_header = {"x-api-key": "sk-ant-api03-test-key"} + + with_block = config.get_count_tokens_headers(auth_header, COMPACTED_HISTORY)["anthropic-beta"].split(",") + plain = config.get_count_tokens_headers(auth_header, [{"role": "user", "content": "hi"}])["anthropic-beta"] + + assert COMPACTION_BETA in with_block + assert ANTHROPIC_TOKEN_COUNTING_BETA_VERSION in with_block + assert plain == ANTHROPIC_TOKEN_COUNTING_BETA_VERSION + + +@pytest.mark.asyncio +async def test_handler_sends_the_compaction_beta_with_a_compacted_history(httpx_transport_clients): + with respx.mock: + route = respx.post("https://gateway.example/v1/messages/count_tokens").mock( + return_value=httpx.Response(200, json={"input_tokens": 404}) + ) + result = await AnthropicCountTokensHandler().handle_count_tokens_request( + model="claude-sonnet-5-5", + messages=COMPACTED_HISTORY, + auth_header={"x-api-key": "sk-ant-api03-test-key"}, + api_base="https://gateway.example", + ) + + assert result == {"input_tokens": 404} + betas = route.calls.last.request.headers["anthropic-beta"].split(",") + assert COMPACTION_BETA in betas + assert ANTHROPIC_TOKEN_COUNTING_BETA_VERSION in betas + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("httpx_transport_clients") +@pytest.mark.parametrize( + ("context_management", "forwarded", "sent_beta", "unsent_beta"), + ( + pytest.param( + {"edits": [{"type": "compact_20260112"}]}, + {"edits": [{"type": "compact_20260112"}]}, + ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value, + ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value, + id="compact", + ), + pytest.param( + {"edits": [{"type": "clear_tool_uses_20250919"}]}, + {"edits": [{"type": "clear_tool_uses_20250919"}]}, + ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value, + ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value, + id="clearing", + ), + pytest.param( + [{"type": "compaction", "compact_threshold": 150000}], + {"edits": [{"type": "compact_20260112", "trigger": {"type": "input_tokens", "value": 150000}}]}, + ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value, + ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value, + id="openai_list_form", + ), + ), +) +async def test_counter_forwards_context_management_with_the_beta_inference_sends( + context_management: JsonValue, forwarded: JsonValue, sent_beta: str, unsent_beta: str +) -> None: + """Anthropic counts an unsigned compaction block only under the compact_20260112 edit that inference sends, + and a clearing edit changes the count, so the count carries the request's context_management as inference + sends it, the OpenAI list form mapped to edits, and its beta.""" + with respx.mock: + route: Final = respx.post("https://gateway.example/v1/messages/count_tokens").mock( + return_value=httpx.Response(200, json={"input_tokens": 259}) + ) + result: Final = await AnthropicTokenCounter().count_tokens( + model_to_use="claude-opus-5-5", + messages=[{"role": "assistant", "content": [{"type": "compaction", "content": "Recipe app summary."}]}], + contents=None, + deployment={"litellm_params": {"api_key": "sk-ant-api03-test-key", "api_base": "https://gateway.example"}}, + context_management=CountContextManagement(spec=context_management, drop_invalid_edits=False), + ) + + sent: Final = route.calls.last.request + betas: Final = sent.headers["anthropic-beta"].split(",") + assert result is not None and result.total_tokens == 259 + assert TypeAdapter(dict[str, JsonValue]).validate_json(sent.content)["context_management"] == forwarded + assert (sent_beta in betas, unsent_beta in betas) == (True, False) + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("httpx_transport_clients") +@pytest.mark.parametrize("context_management", ({}, {"edits": []}, []), ids=("empty_object", "empty_edits", "empty_list")) +async def test_counter_leaves_out_context_management_without_edits(context_management: JsonValue) -> None: + """Anthropic's count endpoint rejects a context_management with no edits, which inference accepts, so the count + leaves it out and adds no context-management beta.""" + with respx.mock: + route: Final = respx.post("https://gateway.example/v1/messages/count_tokens").mock( + return_value=httpx.Response(200, json={"input_tokens": 11}) + ) + result: Final = await AnthropicTokenCounter().count_tokens( + model_to_use="claude-opus-5-5", + messages=[{"role": "user", "content": "Hello"}], + contents=None, + deployment={"litellm_params": {"api_key": "sk-ant-api03-test-key", "api_base": "https://gateway.example"}}, + context_management=CountContextManagement(spec=context_management, drop_invalid_edits=False), + ) + + sent: Final = route.calls.last.request + betas: Final = frozenset(sent.headers.get("anthropic-beta", "").split(",")) + assert result is not None and result.total_tokens == 11 + assert "context_management" not in TypeAdapter(dict[str, JsonValue]).validate_json(sent.content) + assert not betas & { + ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value, + ANTHROPIC_BETA_HEADER_VALUES.CONTEXT_MANAGEMENT_2025_06_27.value, + } diff --git a/tests/unit/llms/anthropic/test_count_tokens_oauth.py b/tests/unit/llms/anthropic/test_count_tokens_oauth.py index b8a7c001b37..3e3b408d825 100644 --- a/tests/unit/llms/anthropic/test_count_tokens_oauth.py +++ b/tests/unit/llms/anthropic/test_count_tokens_oauth.py @@ -48,7 +48,7 @@ FEDERATED_DEPLOYMENT = { def count_tokens_headers_for(api_key: str) -> dict[str, str]: auth_header = AnthropicModelInfo.get_auth_header(api_key=api_key) assert auth_header is not None - return AnthropicCountTokensConfig().get_count_tokens_headers(auth_header) + return AnthropicCountTokensConfig().get_count_tokens_headers(auth_header, [{"role": "user", "content": "hi"}]) @pytest.fixture diff --git a/tests/unit/llms/vertex_ai/count_tokens/test_context.py b/tests/unit/llms/vertex_ai/count_tokens/test_context.py index 38eb87debfd..67733884e12 100644 --- a/tests/unit/llms/vertex_ai/count_tokens/test_context.py +++ b/tests/unit/llms/vertex_ai/count_tokens/test_context.py @@ -666,7 +666,8 @@ async def test_anthropic_view_clears_only_the_turns_after_a_compaction_block_whe *_agent_turns(2, first_round=3), ) - view: Final = await inference_request_view(turns, None, "anthropic", _edits(clear_before_compact)) + with patch.object(litellm, "token_counter", side_effect=RuntimeError("tokenizer unavailable")): + view: Final = await inference_request_view(turns, None, "anthropic", _edits(clear_before_compact)) assert isinstance(view, InferenceRequestView) assert _tool_outputs_kept(view) == (True, True) diff --git a/tests/unit/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_messages_config.py b/tests/unit/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_messages_config.py index 1e7c3d52814..7f16957ddc4 100644 --- a/tests/unit/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_messages_config.py +++ b/tests/unit/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_messages_config.py @@ -526,6 +526,29 @@ def test_both_compact_and_context_management_headers_added(): ) +def test_openai_list_form_context_management_sends_the_compact_beta(): + config: Final = VertexAIPartnerModelsAnthropicMessagesConfig() + + with ( + patch.object(config, "_ensure_access_token", return_value=("token", "test-project")), + patch.object(config, "get_complete_vertex_url", return_value="https://mock-url"), + ): + updated_headers, _ = config.validate_anthropic_messages_environment( + headers={}, + model="claude-vertex-ai-opus-4-6", + messages=[], + optional_params={"context_management": [{"type": "compaction", "compact_threshold": 150000}]}, + litellm_params={ + "vertex_ai_project": "test-project", + "vertex_ai_location": "us-central1", + "vertex_credentials": "{}", + }, + api_base=None, + ) + + assert updated_headers["anthropic-beta"].split(",") == [ANTHROPIC_BETA_HEADER_VALUES.COMPACT_2026_01_12.value] + + def test_validate_environment_always_refreshes_token_ignoring_stale_bearer(): """Regression: stale Authorization in shared deployment extra_headers must not skip token refresh on /v1/messages — _ensure_access_token is always called.""" diff --git a/tests/unit/proxy/anthropic_endpoints/test_endpoints.py b/tests/unit/proxy/anthropic_endpoints/test_endpoints.py index f0317af915f..604fc84a096 100644 --- a/tests/unit/proxy/anthropic_endpoints/test_endpoints.py +++ b/tests/unit/proxy/anthropic_endpoints/test_endpoints.py @@ -6,6 +6,7 @@ import json import logging import unittest from collections.abc import AsyncGenerator, AsyncIterator, Callable, Mapping +from types import MappingProxyType from typing import Final from unittest.mock import AsyncMock, MagicMock, patch @@ -587,6 +588,48 @@ class TestErrorLogCarriesCallId: assert call_id in record.getMessage() + +class TestCountTokensCompactedHistory: + """A history carrying Anthropic's signed compaction block used to fail /v1/messages/count_tokens once the count + ran on the local tokenizer, which rejected the block type.""" + + @pytest.mark.asyncio + async def test_count_tokens_counts_a_compaction_block_like_its_summary_text(self): + import litellm.proxy.anthropic_endpoints.endpoints as ep + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy._types import UserAPIKeyAuth + + summary = "The user is building a recipe app and asked for one-sentence class descriptions." + follow_up = {"role": "user", "content": "Now do the same for Ingredient."} + compacted = { + "model": "claude-sonnet-5-5", + "messages": [ + { + "role": "assistant", + "content": [{"type": "compaction", "content": summary, "signature": "EqQBCkYIBRgCKkBjZ2xhc3M" * 40}], + }, + follow_up, + ], + } + as_text = { + "model": "claude-sonnet-5-5", + "messages": [{"role": "assistant", "content": [{"type": "text", "text": summary}]}, follow_up], + } + request = MagicMock() + request.headers = {} + + async def count(body): + with ( + patch.object(ep, "read_request_body", new=AsyncMock(return_value=body)), # test-quality-ok: endpoint reads the body via a module function; no injection seam + patch.object(proxy_server, "llm_router", None), # test-quality-ok: module global; with no router the count runs on the local tokenizer, the path that raised + ): + return await ep.count_tokens(request=request, user_api_key_dict=UserAPIKeyAuth()) + + counted = await count(compacted) + + assert counted["input_tokens"] > 0 + assert counted == await count(as_text) + def _claude_router() -> Router: return Router( model_list=[ @@ -759,7 +802,10 @@ async def test_count_tokens_drops_malformed_tools_the_counted_deployment_drops_e def _count_on_claude( - monkeypatch: pytest.MonkeyPatch, peer: Callable[[httpx.Request], httpx.Response], headers: Mapping[str, str] + monkeypatch: pytest.MonkeyPatch, + peer: Callable[[httpx.Request], httpx.Response], + headers: Mapping[str, str], + extra_body: Mapping[str, object] = MappingProxyType({}), ) -> httpx.Response: from fastapi import FastAPI @@ -799,7 +845,7 @@ def _count_on_claude( app.dependency_overrides[user_api_key_auth] = lambda: UserAPIKeyAuth() return TestClient(app).post( "/v1/messages/count_tokens", - json={"model": "count-log-claude", "messages": [{"role": "user", "content": "Hello"}]}, + json={"model": "count-log-claude", "messages": [{"role": "user", "content": "Hello"}], **extra_body}, headers=dict(headers), ) @@ -827,6 +873,24 @@ def test_count_tokens_logs_one_line_with_the_callers_call_id_model_and_count( ) +def test_count_tokens_returns_the_context_management_the_provider_applied(monkeypatch: pytest.MonkeyPatch) -> None: + applied: Final = {"original_input_tokens": 640} + response: Final = _count_on_claude( + monkeypatch, + lambda request: httpx.Response( + 200, json={"input_tokens": 576, "context_management": applied}, request=request + ), + {}, + {"context_management": {"edits": [{"type": "clear_tool_uses_20250919"}]}}, + ) + + assert response.json() == { + "input_tokens": 576, + "context_management": applied, + "litellm_count_source": "provider", + }, response.text + + def test_count_tokens_provider_rejection_logs_its_status_and_answers_the_same_error( monkeypatch: pytest.MonkeyPatch, token_count_log_lines: Callable[[], tuple[dict[str, object], ...]] ) -> None: From cfb77ecc3ad72d9f3602dbda22d78d7e84739289 Mon Sep 17 00:00:00 2001 From: mika-lt <349045450+mika-lt@users.noreply.github.com> Date: Fri, 30 Oct 2026 00:14:49 -0700 Subject: [PATCH 11/23] fix(proxy): list only routable models for a wildcard alias over a provider org (#49152) * fix(proxy): list only routable models for a wildcard alias over a provider org A wildcard alias over an org-scoped deployment (together_ai/* over together_ai/nosuchorg/*) listed the provider's whole catalog on /v1/models and /model/info, and every listed id routed to a model the provider 404s. A skipped deployment path segment now counts as a routing prefix only when the routed model resolves to a catalog id through the get_model_info lookup order, so responses/, Bedrock converse/, and region paths list as before. Together's model list now also carries the cost map's together_ai rows, skipping the together-ai-* pricing tiers, so an alias over a real org lists the models Together serves today. * fix(proxy): keep flat catalog ids under routes the cost map can't parse, drop retired Together rows * fix(proxy): check skipped path segments only on Together so Hugging Face provider routes keep listing --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/__init__.py | 5 ++ litellm/constants.py | 2 +- litellm/proxy/auth/model_checks.py | 34 ++++++++- litellm/responses/main.py | 8 +- litellm/utils.py | 46 ++++++------ tests/unit/proxy/auth/test_model_checks.py | 87 +++++++++++++++++++++- 6 files changed, 150 insertions(+), 32 deletions(-) diff --git a/litellm/__init__.py b/litellm/__init__.py index d59edcaff89..5f14c605cd0 100644 --- a/litellm/__init__.py +++ b/litellm/__init__.py @@ -872,6 +872,11 @@ def _populate_provider_model_sets(model_cost_map: Mapping[str, Mapping[str, obje # ignore the 'up-to', '-to-' model names -> not real models. just for cost tracking based on model params. if "-to-" not in key and "fireworks-ai-default" not in key: fireworks_ai_models.add(key) + elif value.get("litellm_provider") == "together_ai": + if _is_past_deprecation_date(value): + together_ai_models.discard(key) + elif not key.startswith("together-ai-"): + together_ai_models.add(key) elif value.get("litellm_provider") == "baseten": if _is_past_deprecation_date(value): baseten_models.discard(key) diff --git a/litellm/constants.py b/litellm/constants.py index cfee777f36a..e7c9c5dcbfb 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1274,7 +1274,7 @@ empower_models: Final = set( ] ) -together_ai_models: Final[set] = set( +together_ai_models: Final[set[str]] = set( [ # llama llms - chat "togethercomputer/llama-2-70b-chat", diff --git a/litellm/proxy/auth/model_checks.py b/litellm/proxy/auth/model_checks.py index c40ee190e12..b21315529eb 100644 --- a/litellm/proxy/auth/model_checks.py +++ b/litellm/proxy/auth/model_checks.py @@ -14,14 +14,16 @@ from litellm.litellm_core_utils.credential_accessor import CredentialAccessor from litellm.llms.azure.ptu_capacity import is_str_keyed_mapping from litellm.proxy._types import SpecialModelNames, UserAPIKeyAuth from litellm.repositories.object_permission_repository import ObjectPermissionRepository +from litellm.responses.main import strip_responses_routing_prefix from litellm.router import Router from litellm.router_utils.fallback_event_handlers import get_fallback_model_group from litellm.router_utils.pattern_match_deployments import PatternMatchRouter from litellm.types.router import CredentialLiteLLMParams, LiteLLM_Params from litellm.types.utils import LlmProviders -from litellm.utils import ProviderConfigManager, get_valid_models +from litellm.utils import ProviderConfigManager, get_valid_models, potential_model_names_in_lookup_order LLM_PROVIDER_PREFIXES: Final = frozenset(llm_provider.value for llm_provider in LlmProviders) +_PROVIDERS_SENDING_THE_DEPLOYMENT_PATH_AS_THE_MODEL_ID: Final = frozenset({LlmProviders.TOGETHER_AI.value}) _CREDENTIAL_LITELLM_PARAM_FIELDS = set(CredentialLiteLLMParams.model_fields) @@ -331,6 +333,25 @@ def _routes_back_to( ) +def _resolves_to_catalog_model(model: str, provider: str, provider_model_ids: frozenset[str]) -> bool: + provider_id_prefix: Final = f"{provider}/" + resolved_model: Final = ( + f"{provider_id_prefix}{strip_responses_routing_prefix(model.removeprefix(provider_id_prefix))}" + ) + return any( + name.removeprefix(provider_id_prefix) in provider_model_ids + for name in potential_model_names_in_lookup_order(resolved_model, provider) + ) + + +def _routes_to_catalog_model( + routes_to_model: Callable[[str, str], bool], provider: str, provider_model_ids: frozenset[str] +) -> Callable[[str, str], bool]: + return lambda listed_id, model: ( + routes_to_model(listed_id, model) and _resolves_to_catalog_model(model, provider, provider_model_ids) + ) + + def _ids_routing_to_catalog_models( name_pattern: str, deployment_pattern: str, @@ -344,8 +365,15 @@ def _ids_routing_to_catalog_models( return () provider_id_prefix: Final = f"{provider}/" provider_model_ids: Final = tuple(catalog_id.removeprefix(provider_id_prefix) for catalog_id in catalog_ids) + deployment_model_pattern_without_provider: Final = deployment_pattern.removeprefix(provider_id_prefix) catalog_pattern: Final = _without_routing_segments( - deployment_pattern.removeprefix(provider_id_prefix), provider, name_pattern, provider_model_ids + deployment_model_pattern_without_provider, provider, name_pattern, provider_model_ids + ) + routing_prefix: Final = deployment_model_pattern_without_provider.removesuffix(catalog_pattern) + reaches_catalog_model: Final = ( + _routes_to_catalog_model(routes_to_model, provider, frozenset(provider_model_ids)) + if routing_prefix and provider in _PROVIDERS_SENDING_THE_DEPLOYMENT_PATH_AS_THE_MODEL_ID + else routes_to_model ) deployment_model_pattern: Final = f"{provider_id_prefix}{catalog_pattern}" deployment_regex: Final = re.compile(PatternMatchRouter().pattern_to_regex(deployment_model_pattern)) @@ -369,7 +397,7 @@ def _ids_routing_to_catalog_models( for listed_id, catalog_id in listed_ids if listed_id and _routes_back_to( - listed_id, name_regex, (deployment_model_pattern, catalog_id), deployment_pattern, routes_to_model + listed_id, name_regex, (deployment_model_pattern, catalog_id), deployment_pattern, reaches_catalog_model ) ) ) diff --git a/litellm/responses/main.py b/litellm/responses/main.py index 98356e9a5d4..063ef9d91ea 100644 --- a/litellm/responses/main.py +++ b/litellm/responses/main.py @@ -759,7 +759,7 @@ def _dispatched_model_name( from litellm.litellm_core_utils.get_llm_provider_logic import resolve_model_and_provider_for_metadata provider_model, _ = resolve_model_and_provider_for_metadata(model=model, custom_llm_provider=custom_llm_provider) - return _strip_responses_routing_prefix(provider_model) + return strip_responses_routing_prefix(provider_model) def _will_bridge_to_chat_completions( @@ -1245,7 +1245,7 @@ def _pop_use_chat_completions_api_kw(kwargs: dict[str, object]) -> bool: _RESPONSES_ROUTING_PREFIX: Final = "responses/" -def _strip_responses_routing_prefix(model: str) -> str: +def strip_responses_routing_prefix(model: str) -> str: if not model.startswith(_RESPONSES_ROUTING_PREFIX): return model return model[len(_RESPONSES_ROUTING_PREFIX) :] @@ -1273,7 +1273,7 @@ def _resolve_model_provider_for_responses( litellm_params.api_key = dynamic_api_key if dynamic_api_base is not None: litellm_params.api_base = dynamic_api_base - return _strip_responses_routing_prefix(provider_model), resolved_provider + return strip_responses_routing_prefix(provider_model), resolved_provider def _apply_managed_file_id_mapping( @@ -2893,7 +2893,7 @@ async def _aresponses_websocket( api_key=api_key, litellm_params=litellm_params, ) - resolved_model: Final = _strip_responses_routing_prefix(provider_model) + resolved_model: Final = strip_responses_routing_prefix(provider_model) litellm_params_dict["data_residency"] = infer_openai_data_residency( _custom_llm_provider, diff --git a/litellm/utils.py b/litellm/utils.py index 41a3fddf3a3..e85e089289a 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -6708,6 +6708,26 @@ def _first_registered_match( return next(matches, (None, None)) +def _model_cost_lookup_order( + model: str, potential_model_names: PotentialModelNamesAndCustomLLMProvider +) -> tuple[str, ...]: + return ( + potential_model_names["combined_model_name"], + model, + potential_model_names["region_free_combined_model_name"], + potential_model_names["split_model"], + potential_model_names["combined_stripped_model_name"], + potential_model_names["stripped_model_name"], + potential_model_names["provider_prefixed_model_name"], + ) + + +def potential_model_names_in_lookup_order(model: str, custom_llm_provider: str) -> tuple[str, ...]: + return _model_cost_lookup_order( + model, get_potential_model_names(model=model, custom_llm_provider=custom_llm_provider) + ) + + def _get_model_info_from_generalization( model: str, potential_model_names: PotentialModelNamesAndCustomLLMProvider, @@ -6726,15 +6746,7 @@ def _get_model_info_from_generalization( still had a priced exact name to try. O(number of rules); only call after the exact lookups have missed. """ - candidates: Final = ( - potential_model_names["combined_model_name"], - model, - potential_model_names["region_free_combined_model_name"], - potential_model_names["split_model"], - potential_model_names["combined_stripped_model_name"], - potential_model_names["stripped_model_name"], - potential_model_names["provider_prefixed_model_name"], - ) + candidates: Final = _model_cost_lookup_order(model, potential_model_names) if any(get_model_cost_key(candidate) is not None for candidate in candidates): return None for candidate in candidates: @@ -6973,11 +6985,6 @@ def get_model_info_helper( model=model, custom_llm_provider=custom_llm_provider or declared_authenticating_provider(model) ) - combined_model_name: Final = potential_model_names["combined_model_name"] - stripped_model_name: Final = potential_model_names["stripped_model_name"] - combined_stripped_model_name: Final = potential_model_names["combined_stripped_model_name"] - provider_prefixed_model_name: Final = potential_model_names["provider_prefixed_model_name"] - split_model: Final = potential_model_names["split_model"] custom_llm_provider = potential_model_names["custom_llm_provider"] model_cost_custom_llm_provider: Final = custom_llm_provider ######################### @@ -7044,16 +7051,9 @@ def get_model_info_helper( and strip it. Tried last so no model that already resolves through 1-6 can change. """ - lookup_order: Final = ( - combined_model_name, - model, - potential_model_names["region_free_combined_model_name"], - split_model, - combined_stripped_model_name, - stripped_model_name, - provider_prefixed_model_name, + lookup: Final = _first_registered_match( + _model_cost_lookup_order(model, potential_model_names), model_cost_custom_llm_provider ) - lookup: Final = _first_registered_match(lookup_order, model_cost_custom_llm_provider) key: str | None = lookup[0] _model_info: dict[str, Any] | None = lookup[1] diff --git a/tests/unit/proxy/auth/test_model_checks.py b/tests/unit/proxy/auth/test_model_checks.py index 3f0f97e63b0..c67610c24a1 100644 --- a/tests/unit/proxy/auth/test_model_checks.py +++ b/tests/unit/proxy/auth/test_model_checks.py @@ -1057,7 +1057,9 @@ def test_wildcard_lists_each_model_once_under_its_full_owner_id(provider: str): import litellm from litellm.proxy.auth.model_checks import get_known_models_from_wildcard - full_ids: Final = {f"{provider}/{model.removeprefix(f'{provider}/')}" for model in litellm.models_by_provider[provider]} + full_ids: Final = { + f"{provider}/{model.removeprefix(f'{provider}/')}" for model in litellm.models_by_provider[provider] + } assert any(model.count("/") > 1 for model in full_ids) assert sorted(get_known_models_from_wildcard(f"{provider}/*")) == sorted(full_ids) @@ -1372,12 +1374,57 @@ def test_partial_wildcard_lists_catalog_ids_stored_with_the_provider_prefix( assert sorted(listed) == expected +@pytest.mark.parametrize( + ("provider", "routing_segment"), [("bedrock", "converse"), ("bedrock", "us-east-1"), ("openai", "responses")] +) +def test_routing_segment_lists_the_same_catalog_ids_as_the_bare_provider_wildcard( + provider: str, routing_segment: str +) -> None: + from litellm.proxy.auth.model_checks import get_known_models_from_wildcard + from litellm.types.router import LiteLLM_Params + + bare_listing: Final = get_known_models_from_wildcard(f"{provider}/*", LiteLLM_Params(model=f"{provider}/*")) + routed_listing: Final = get_known_models_from_wildcard( + f"{provider}/*", LiteLLM_Params(model=f"{provider}/{routing_segment}/*") + ) + + assert bare_listing + assert sorted(routed_listing) == sorted(bare_listing) + + +def test_provider_wildcard_lists_cost_map_models_but_not_pricing_tiers() -> None: + from litellm.proxy.auth.model_checks import get_known_models_from_wildcard + + model_id: Final = "together_ai/partial-wildcard-org/model-t" + pricing_tier: Final = "together-ai-partial-wildcard-tier" + with _registered_chat_models({model_id: "together_ai", pricing_tier: "together_ai"}): + listed: Final = get_known_models_from_wildcard("together_ai/*") + + assert model_id in listed + assert [listed_id for listed_id in listed if pricing_tier in listed_id] == [] + + +def test_provider_wildcard_stops_listing_a_cost_map_model_once_it_is_retired() -> None: + import litellm + from litellm.proxy.auth.model_checks import get_known_models_from_wildcard + + model_id: Final = "together_ai/partial-wildcard-org/model-retired" + with _registered_chat_models({model_id: "together_ai"}): + listed_while_served: Final = get_known_models_from_wildcard("together_ai/*") + litellm.add_known_models({model_id: {"litellm_provider": "together_ai", "deprecation_date": "2000-01-01"}}) + listed_once_retired: Final = get_known_models_from_wildcard("together_ai/*") + + assert model_id in listed_while_served + assert model_id not in listed_once_retired + + _ROUTE_CF_ID: Final = "cloudflare/@cf/partial-route-org/model-a-instruct" _ROUTE_CF_SUFFIXED_ID: Final = "cloudflare/@cf/partial-route-org/model-b-instruct-v2" _ROUTE_HF_ID: Final = "cloudflare/@hf/partial-route-org/model-c" _ROUTE_GEMINI_ID: Final = "gemini-partial-route-9-flash" _ROUTE_OPENAI_ID: Final = "gpt-partial-route-9-mini" _ROUTE_BEDROCK_ID: Final = "anthropic.partial-route-9-v1:0" +_ROUTE_TOGETHER_ID: Final = "together_ai/partial-route-org/model-t" _ROUTE_CF_IDS: Final = {_ROUTE_CF_ID, _ROUTE_CF_SUFFIXED_ID} @@ -1456,6 +1503,29 @@ _ROUTE_CF_IDS: Final = {_ROUTE_CF_ID, _ROUTE_CF_SUFFIXED_ID} set(), id="same-name-namespace-the-catalog-lacks", ), + pytest.param( + "together/*", "together_ai/partial-route-missing-org/*", None, set(), id="alias-over-org-the-catalog-lacks" + ), + pytest.param( + "openai/*", + "openai/chat_completions/*", + None, + {f"openai/chat_completions/{_ROUTE_OPENAI_ID}"}, + id="chat-completions-bridge-route", + ), + pytest.param( + "bedrock/*", "bedrock/llama/*", None, {f"bedrock/llama/{_ROUTE_BEDROCK_ID}"}, id="invoke-format-route" + ), + pytest.param( + "together/*", + "together_ai/partial-route-org/*", + None, + {_ROUTE_TOGETHER_ID}, + id="alias-over-org-a-cost-map-catalog-carries", + ), + pytest.param( + "bedrock/*", "bedrock/us-east-1/*", None, {f"bedrock/us-east-1/{_ROUTE_BEDROCK_ID}"}, id="region-route" + ), pytest.param( "*", "openai/*", "openai/gpt-partial-route-9*", set(), id="provider-prefixed-key-filter-over-catch-all" ), @@ -1501,6 +1571,7 @@ def test_every_listed_partial_wildcard_id_routes_to_the_catalog_model_it_names( _ROUTE_GEMINI_ID: "gemini", _ROUTE_OPENAI_ID: "openai", _ROUTE_BEDROCK_ID: "bedrock", + _ROUTE_TOGETHER_ID: "together_ai", } ): v1_models: Final = get_complete_model_list( @@ -1535,6 +1606,20 @@ def test_every_listed_partial_wildcard_id_routes_to_the_catalog_model_it_names( assert listed_model_info_names == sorted(reachable) +def test_inference_provider_route_lists_every_id_the_plain_provider_wildcard_lists() -> None: + from litellm.proxy.auth.model_checks import get_known_models_from_wildcard + from litellm.types.router import LiteLLM_Params + + plain: Final = get_known_models_from_wildcard("huggingface/*", LiteLLM_Params(model="huggingface/*")) + routed: Final = get_known_models_from_wildcard( + "huggingface/*", LiteLLM_Params(model="huggingface/featherless-ai/*") + ) + + assert plain + assert [model for model in plain if "/" not in model.removeprefix("huggingface/")] == [] + assert sorted(routed) == sorted(plain) + + def test_catch_all_lists_no_id_a_more_specific_sibling_wildcard_routes_elsewhere() -> None: from litellm import Router from litellm.proxy.auth.model_checks import expand_wildcard_deployments_for_model_info, get_complete_model_list From e81dd348a425c180c7a632c1eecea7a37682ad96 Mon Sep 17 00:00:00 2001 From: Filip Sinitsin Date: Fri, 30 Oct 2026 07:18:15 +0000 Subject: [PATCH 12/23] fix(proxy)!: log and bill Responses-style bodies sent to /cursor/chat/completions (#30277) * fix(proxy): restore spend logging for /cursor/chat/completions (#30126) The /cursor/chat/completions route processes requests through the Responses API pipeline, but _get_metadata_variable_name did not treat it as a litellm_metadata route. Proxy auth metadata therefore landed in the user-facing metadata param, so the spend tracking callback saw user_api_key=None and silently skipped the SpendLogs write. Add the route to LITELLM_METADATA_ROUTES and cover it with regression tests. * test(proxy): pin provider metadata and key billing on /cursor/chat/completions A Responses-format body on the Cursor route now sends OpenAI only the metadata the client sent and writes a spend row that bills the key. The chat-body branch, which this fix also moves onto litellm_metadata, still bills the key with its team and request tags. The scripted chat reply gets a unique id per test, since a shared id let only the first spend row per database land * fix(proxy): pick the Cursor route's metadata field from the body shape A Responses-style body on /cursor/chat/completions keeps proxy metadata in litellm_metadata, the same as /v1/responses, while a Chat Completions body keeps metadata, the same as /v1/chat/completions. Adding the whole route to the internal-metadata list let a client's chat-body tags replace the key's tags in spend logs and skip tag budgets. * fix(proxy): keep get_metadata_variable_name callable with only the request * fix(proxy): check client tags against tag budgets and read agent limits from litellm_metadata on Responses bodies A Responses-style body on /cursor/chat/completions now keeps its proxy fields in litellm_metadata, so the pre-call tag budget check and the v3 agent rate limiter stopped seeing the client's metadata.tags and the x-litellm-agent-id value. The pre-auth tag check now merges the caller's metadata.tags into the selected metadata field, sharing one helper with request setup, and the limiter reads agent_id from either field * fix(proxy): check client tags sent in a JSON-string metadata against tag budgets on Responses bodies * fix(proxy): answer 422, not 500, when a budget rejects a Responses body whose metadata is a JSON string * fix(proxy): keep key guardrails, metadata type errors and failure-row client details on Responses-style bodies --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/integrations/custom_guardrail.py | 32 ++-- litellm/proxy/auth/auth_checks.py | 10 +- litellm/proxy/common_request_processing.py | 16 +- .../hooks/parallel_request_limiter_v3.py | 3 +- .../proxy/hooks/proxy_track_cost_callback.py | 37 +++- litellm/proxy/litellm_pre_call_utils.py | 88 ++++++++-- .../proxy/response_api_endpoints/endpoints.py | 12 +- litellm/proxy/utils.py | 10 +- .../providers/test_cursor_responses_wire.py | 163 +++++++++++++++--- .../integrations/test_custom_guardrail.py | 14 ++ .../hooks/test_parallel_request_limiter_v3.py | 5 +- .../hooks/test_proxy_track_cost_callback.py | 50 ++++++ .../response_api_endpoints/test_endpoints.py | 2 +- .../proxy/test_common_request_processing.py | 34 ++++ .../unit/proxy/test_litellm_pre_call_utils.py | 68 ++++++++ .../test_post_call_failure_hook.py | 29 ++++ 16 files changed, 482 insertions(+), 91 deletions(-) diff --git a/litellm/integrations/custom_guardrail.py b/litellm/integrations/custom_guardrail.py index 9e2dc902d55..a9fb5c29bad 100644 --- a/litellm/integrations/custom_guardrail.py +++ b/litellm/integrations/custom_guardrail.py @@ -9,6 +9,7 @@ from types import MappingProxyType from typing import TYPE_CHECKING, Any, ClassVar, Final, Literal, Optional, cast, get_args import httpx +from pydantic import TypeAdapter from litellm._internal_context import with_service_target from litellm._logging import verbose_logger @@ -218,6 +219,15 @@ def _request_is_streaming(data: object, event_type: GuardrailEventHooks | None = ) +_REQUEST_FIELDS: Final = TypeAdapter(Mapping[str, object]) + + +def _guardrails_listed_in(metadata: object) -> list[str | dict[str, DynamicGuardrailParams]]: + if not isinstance(metadata, dict): + return [] + return metadata.get("guardrails") or [] + + class CustomGuardrail(CustomLogger): # If True, during_call runs async_moderation_hook instead of the unified apply_guardrail path. use_native_during_call_hook: ClassVar[bool] = False @@ -791,28 +801,22 @@ class CustomGuardrail(CustomLogger): return True raise - def get_guardrail_from_metadata(self, data: dict) -> list[str] | list[dict[str, DynamicGuardrailParams]]: + def get_guardrail_from_metadata(self, data: dict) -> list[str | dict[str, DynamicGuardrailParams]]: """ - Returns the guardrail(s) to be run from the metadata or root + Returns the guardrail(s) to be run from the root, else from both metadata fields together """ if "guardrails" in data: return data["guardrails"] - # Check both metadata locations. For regular endpoints move_guardrails_to_metadata - # writes to "metadata"; for thread/assistant endpoints it writes to - # "litellm_metadata". We check the one that actually contains the "guardrails" - # key so that a non-empty litellm_metadata without guardrails does not shadow - # the merged list stored in metadata (which would cause team guardrails to be - # silently skipped while default_on=True policy guardrails still fire). - for meta_key in ("metadata", "litellm_metadata"): - meta = data.get(meta_key) or {} - if isinstance(meta, dict) and "guardrails" in meta: - return meta.get("guardrails") or [] - return [] + request_fields: Final = _REQUEST_FIELDS.validate_python(data) + return [ + *_guardrails_listed_in(request_fields.get("metadata")), + *_guardrails_listed_in(request_fields.get("litellm_metadata")), + ] def _guardrail_is_in_requested_guardrails( self, - requested_guardrails: list[str] | list[dict[str, DynamicGuardrailParams]], + requested_guardrails: Sequence[str | dict[str, DynamicGuardrailParams]], ) -> bool: for _guardrail in requested_guardrails: if isinstance(_guardrail, dict): diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py index 1ef5b1072ed..b8d9530d2eb 100644 --- a/litellm/proxy/auth/auth_checks.py +++ b/litellm/proxy/auth/auth_checks.py @@ -446,7 +446,7 @@ def _request_has_trace_id(request_body: Mapping[str, object], request: Request, from litellm.proxy.litellm_pre_call_utils import ( _trace_id_from_traceparent, # pyright: ignore[reportPrivateUsage] # reuse W3C traceparent parser get_chain_id_from_headers, - metadata_variable_name_for_route, + metadata_variable_name_for_request, ) if get_chain_id_from_headers(dict(request.headers)): @@ -460,7 +460,7 @@ def _request_has_trace_id(request_body: Mapping[str, object], request: Request, traceparent: Final[str | None] = request.headers.get("traceparent") if isinstance(traceparent, str) and _trace_id_from_traceparent(traceparent): return True - metadata_variable_name: Final = metadata_variable_name_for_route(get_request_route(request)) + metadata_variable_name: Final = metadata_variable_name_for_request(get_request_route(request), request_body) selected_metadata: Final[object] = request_body.get(metadata_variable_name) if isinstance(selected_metadata, dict): selected_metadata_dict: Final = _TRACE_ID_METADATA_ADAPTER.validate_python(selected_metadata) @@ -1367,11 +1367,15 @@ async def common_checks( if not skip_all_budget_checks: # Key metadata.tags are injected into request_body here so the tag budget # check can read them; this mutation must run before the gathered checks. - from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup + from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup, metadata_variable_name_for_request if valid_token is not None: for tag_target in LiteLLMProxyRequestSetup.pre_auth_tag_merge_targets(request, request_body): LiteLLMProxyRequestSetup.pre_seed_litellm_metadata_for_route(request_data=tag_target, route=route) + LiteLLMProxyRequestSetup.merge_caller_body_tags( + request_data=tag_target, + metadata_variable_name=metadata_variable_name_for_request(route, tag_target), + ) LiteLLMProxyRequestSetup.apply_key_tags_pre_auth(request_data=tag_target, user_api_key_dict=valid_token) LiteLLMProxyRequestSetup.apply_authenticated_baggage_tags(request, request_body) diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index 7cd752c2297..a79d8fe29db 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -2401,9 +2401,10 @@ class ProxyBaseLLMRequestProcessing: version=version, proxy_config=proxy_config, ) + request_fields: Final = _REQUEST_FIELDS.validate_python(self.data) + metadata_variable_name: Final = get_metadata_variable_name(request, request_fields) if logging_auth is not None: - metadata_variable_name: Final = get_metadata_variable_name(request) - request_metadata: Final = self.data.get(metadata_variable_name) + request_metadata: Final = request_fields.get(metadata_variable_name) if isinstance(request_metadata, dict): self.data[metadata_variable_name] = { **TypeAdapter(dict[str, object]).validate_python(request_metadata), @@ -2431,12 +2432,11 @@ class ProxyBaseLLMRequestProcessing: # Store queue time in metadata after add_litellm_data_to_request to ensure it's preserved if queue_time_seconds is not None: - _metadata_variable_name: Final = get_metadata_variable_name(request) - if _metadata_variable_name not in self.data: - self.data[_metadata_variable_name] = {} - if not isinstance(self.data[_metadata_variable_name], dict): - self.data[_metadata_variable_name] = {} - self.data[_metadata_variable_name]["queue_time_seconds"] = queue_time_seconds + if metadata_variable_name not in self.data: + self.data[metadata_variable_name] = {} + if not isinstance(self.data[metadata_variable_name], dict): + self.data[metadata_variable_name] = {} + self.data[metadata_variable_name]["queue_time_seconds"] = queue_time_seconds if isinstance(model, str): reject_url_valued_destination("model", model) diff --git a/litellm/proxy/hooks/parallel_request_limiter_v3.py b/litellm/proxy/hooks/parallel_request_limiter_v3.py index 14b08a09116..15282679622 100644 --- a/litellm/proxy/hooks/parallel_request_limiter_v3.py +++ b/litellm/proxy/hooks/parallel_request_limiter_v3.py @@ -3312,7 +3312,8 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): if key_agent_id: return key_agent_id metadata: Final = data.get("metadata") or {} - return metadata.get("agent_id") + litellm_metadata: Final = data.get("litellm_metadata") or {} + return metadata.get("agent_id") or litellm_metadata.get("agent_id") def _get_session_id_from_data(self, data: dict) -> str | None: """Extract session_id from request metadata or litellm_session_id.""" diff --git a/litellm/proxy/hooks/proxy_track_cost_callback.py b/litellm/proxy/hooks/proxy_track_cost_callback.py index 7d8b2e8329c..9c4a4bc775b 100644 --- a/litellm/proxy/hooks/proxy_track_cost_callback.py +++ b/litellm/proxy/hooks/proxy_track_cost_callback.py @@ -37,7 +37,7 @@ from litellm.proxy.db.db_spend_update_writer import ( get_llm_router, ) from litellm.proxy.hooks.search_free_tier import apply_search_free_tier -from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup, metadata_variable_name_for_route +from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup, metadata_variable_name_for_request from litellm.proxy.route_llm_request import ProxyModelNotFoundError from litellm.proxy.spend_tracking.spend_counter_batch import post_call_counter_keys, spend_counter_batch_scope from litellm.proxy.spend_tracking.spend_event import ( @@ -137,19 +137,34 @@ _CAPTURED_IDENTITY_CALL_TYPES: Final[frozenset[str]] = frozenset( _OBJECT_MAPPING: Final = TypeAdapter(ObjectMapping) -def _proxy_stamped_used_client_oauth_token( - request_data: Mapping[str, object], request_route: str | None -) -> bool | None: +def _proxy_stamped_metadata(request_data: Mapping[str, object], request_route: str | None) -> Mapping[str, object]: proxy_bucket: Final = ( get_metadata_variable_name_from_kwargs(request_data) if request_route is None - else metadata_variable_name_for_route(request_route) + else metadata_variable_name_for_request(request_route, request_data) ) - proxy_metadata: Final = request_data.get(proxy_bucket) - stamped: Final = proxy_metadata.get("used_client_oauth_token") if isinstance(proxy_metadata, dict) else None + return _str_keyed_fields(request_data.get(proxy_bucket)) + + +def _proxy_stamped_used_client_oauth_token( + request_data: Mapping[str, object], request_route: str | None +) -> bool | None: + stamped: Final = _proxy_stamped_metadata(request_data, request_route).get("used_client_oauth_token") return stamped if isinstance(stamped, bool) else None +_PROXY_STAMPED_REQUEST_FIELDS: Final = ("requester_ip_address", "user_agent") + + +def _proxy_stamped_request_fields( + request_data: Mapping[str, object], request_route: str | None +) -> Mapping[str, object]: + if request_route is None: + return {} + proxy_metadata: Final = _proxy_stamped_metadata(request_data, request_route) + return {field: proxy_metadata[field] for field in _PROXY_STAMPED_REQUEST_FIELDS if field in proxy_metadata} + + def _str_keyed_fields(value: object) -> Mapping[str, object]: return value if is_str_keyed_mapping(value) else {} @@ -455,7 +470,13 @@ class _ProxyDBLogger(CustomLogger): _metadata["used_client_oauth_token"] = _proxy_stamped_used_client_oauth_token(request_fields, request_route) existing_metadata: Final[dict] = request_data.get("metadata", None) or {} - existing_metadata.update({**_routed_metadata_fallback(request_fields), **_metadata}) + existing_metadata.update( + { + **_routed_metadata_fallback(request_fields), + **_proxy_stamped_request_fields(request_fields, request_route), + **_metadata, + } + ) litellm_metadata_bucket: Final = request_data.get("litellm_metadata") litellm_metadata: Final[ObjectMapping] = ( diff --git a/litellm/proxy/litellm_pre_call_utils.py b/litellm/proxy/litellm_pre_call_utils.py index a87fc815402..c1e541d20a9 100644 --- a/litellm/proxy/litellm_pre_call_utils.py +++ b/litellm/proxy/litellm_pre_call_utils.py @@ -379,6 +379,8 @@ LITELLM_METADATA_ROUTES: Final = ( "interactions", ) +CURSOR_CHAT_COMPLETIONS_ROUTE: Final = "/cursor/chat/completions" + LITELLM_TRACE_CONTROL_METADATA_FIELDS: Final = frozenset( { "mask_input", @@ -636,7 +638,7 @@ def _has_tag_budget(tag: "LiteLLM_TagTable | None") -> bool: return tag.litellm_budget_table.max_budget is not None -def _normalized_metadata_slot( +def normalized_metadata_slot( request_data: MutableMapping[str, object], metadata_variable_name: str ) -> dict[str, object]: """Return the request's metadata slot as a dict, normalising it in place first. @@ -653,6 +655,12 @@ def _normalized_metadata_slot( return normalized +def _holds_metadata_object(value: object) -> bool: + if value is None or isinstance(value, dict): + return True + return isinstance(value, str) and isinstance(_DECODED_JSON.validate_python(safe_json_loads(value)), dict) + + def _metadata_object_or_error(field: str, value: object) -> Mapping[str, object] | ProxyException: if value is None: return {} @@ -863,7 +871,7 @@ def _strip_router_reserved_metadata( ) -def get_metadata_variable_name(request: Request) -> str: +def get_metadata_variable_name(request: Request, data: Mapping[str, object] | None = None) -> str: """ Helper to return what the "metadata" field should be called in the request data @@ -874,7 +882,10 @@ def get_metadata_variable_name(request: Request) -> str: # Inline imports — auth_utils/route_checks participate in a proxy import cycle. from litellm.proxy.auth.auth_utils import get_request_route # noqa: PLC0415 - return metadata_variable_name_for_route(get_request_route(request)) + route: Final = get_request_route(request) + if data is None: + return metadata_variable_name_for_route(route) + return metadata_variable_name_for_request(route, data) _get_metadata_variable_name: Final = get_metadata_variable_name @@ -890,6 +901,25 @@ def metadata_variable_name_for_route(route: str) -> Literal["metadata", "litellm return "metadata" +def is_chat_completions_body(data: Mapping[str, object]) -> bool: + messages: Final = data.get("messages") + if isinstance(messages, list) and messages: + return True + return "messages" in data and "input" not in data + + +def is_cursor_responses_request(route: str, body: Mapping[str, object]) -> bool: + return CURSOR_CHAT_COMPLETIONS_ROUTE in route and not is_chat_completions_body(body) + + +def metadata_variable_name_for_request( + route: str, body: Mapping[str, object] +) -> Literal["metadata", "litellm_metadata"]: + if is_cursor_responses_request(route, body): + return "litellm_metadata" + return metadata_variable_name_for_route(route) + + def _promoted_trace_control_fields( requester_metadata: Mapping[str, object], litellm_metadata: Mapping[str, object], @@ -2309,7 +2339,9 @@ class LiteLLMProxyRequestSetup: ``_tag_max_budget_check`` resolve to ``litellm_metadata``, leaving header tags invisible to per-tag budget enforcement. """ - if any(metadata_route in route for metadata_route in LITELLM_METADATA_ROUTES): + if any(metadata_route in route for metadata_route in LITELLM_METADATA_ROUTES) or is_cursor_responses_request( + route, request_data + ): request_data.setdefault("litellm_metadata", {}) @staticmethod @@ -2380,7 +2412,7 @@ class LiteLLMProxyRequestSetup: return tags: Final = _STRING_TAGS_ADAPTER.validate_python(raw_tags) slot: Final = metadata_key if metadata_key is not None else get_metadata_variable_name_from_kwargs(request_data) - metadata: Final = _normalized_metadata_slot(request_data, slot) + metadata: Final = normalized_metadata_slot(request_data, slot) existing: Final = metadata.get("tags") if existing and not isinstance(existing, list): return @@ -2393,6 +2425,34 @@ class LiteLLMProxyRequestSetup: tags=tuple(tag for tag in tags if tag not in sent_tags) ) + @staticmethod + def merge_caller_body_tags( + request_data: MutableMapping[str, object], # mutable-ok: the request body the tag merges write into + metadata_variable_name: str | None = None, + ) -> None: + slot: Final = ( + metadata_variable_name + if metadata_variable_name is not None + else get_metadata_variable_name_from_kwargs(request_data) + ) + if slot == "metadata" or not _holds_metadata_object(request_data.get(slot)): + return + raw_caller_metadata: Final = request_data.get("metadata") + caller_metadata: Final = ( + _DECODED_JSON.validate_python(safe_json_loads(raw_caller_metadata)) + if isinstance(raw_caller_metadata, str) + else raw_caller_metadata + ) + caller_tags: Final = caller_metadata.get("tags") if isinstance(caller_metadata, dict) else None + if not isinstance(caller_tags, list) or not caller_tags: + return + metadata: Final = normalized_metadata_slot(request_data, slot) + existing_tags: Final = metadata.get("tags") + metadata["tags"] = LiteLLMProxyRequestSetup.merge_tags( + request_tags=existing_tags if isinstance(existing_tags, list) else None, + tags_to_add=caller_tags, + ) + @staticmethod def apply_key_tags_pre_auth( request_data: dict, @@ -2408,7 +2468,7 @@ class LiteLLMProxyRequestSetup: return _metadata_variable_name: Final = get_metadata_variable_name_from_kwargs(request_data) - metadata: Final = _normalized_metadata_slot(request_data, _metadata_variable_name) + metadata: Final = normalized_metadata_slot(request_data, _metadata_variable_name) existing_tags: Final = metadata.get("tags") metadata["tags"] = LiteLLMProxyRequestSetup.merge_tags( @@ -2460,7 +2520,7 @@ class LiteLLMProxyRequestSetup: # from (litellm_metadata vs metadata) so the merged tags are visible # to _tag_max_budget_check. _metadata_variable_name: Final = get_metadata_variable_name_from_kwargs(request_data) - metadata: Final = _normalized_metadata_slot(request_data, _metadata_variable_name) + metadata: Final = normalized_metadata_slot(request_data, _metadata_variable_name) existing_tags: Final = metadata.get("tags") metadata["tags"] = LiteLLMProxyRequestSetup.merge_tags( @@ -2637,7 +2697,7 @@ async def add_litellm_data_to_request( } safe_add_api_version_from_query_params(data, request) - _metadata_variable_name: Final = get_metadata_variable_name(request) + _metadata_variable_name: Final = get_metadata_variable_name(request, _REQUEST_FIELDS.validate_python(data)) if data.get(_metadata_variable_name, None) is None: data[_metadata_variable_name] = {} @@ -2974,17 +3034,7 @@ async def add_litellm_data_to_request( tags_to_add=tags, ) - _caller_body_metadata: Final = data.get("metadata") if _metadata_variable_name != "metadata" else None - _caller_body_tags: Final = ( - _caller_body_metadata.get("tags") - if isinstance(_caller_body_metadata, dict) and isinstance(_caller_body_metadata.get("tags"), list) - else None - ) - if _caller_body_tags: - data[_metadata_variable_name]["tags"] = LiteLLMProxyRequestSetup.merge_tags( # rebind-ok: matches file idiom - request_tags=data[_metadata_variable_name].get("tags"), - tags_to_add=_caller_body_tags, - ) + LiteLLMProxyRequestSetup.merge_caller_body_tags(request_data=data, metadata_variable_name=_metadata_variable_name) # caller_tags: the tags the request's headers and body carry (x-litellm-tags, body "tags", and body # "metadata.tags" / "litellm_metadata.tags" read before the mapped tags are added below), diff --git a/litellm/proxy/response_api_endpoints/endpoints.py b/litellm/proxy/response_api_endpoints/endpoints.py index 9d0d89f681e..ef59deb2e5e 100644 --- a/litellm/proxy/response_api_endpoints/endpoints.py +++ b/litellm/proxy/response_api_endpoints/endpoints.py @@ -79,6 +79,7 @@ from litellm.proxy.hooks.responses_id_security import ResponsesIDSecurity, decry from litellm.proxy.litellm_pre_call_utils import ( _CLIENT_PRICING_METADATA_FIELDS, # pyright: ignore[reportPrivateUsage] # the HTTP path's own strip list _UNTRUSTED_METADATA_CONTROL_FIELDS, # pyright: ignore[reportPrivateUsage] # the HTTP path's own strip list + is_chat_completions_body, ) from litellm.proxy.pass_through_endpoints.llm_provider_handlers.batch_attribution import ( request_tags_from_request_data, @@ -213,13 +214,6 @@ def _normalize_tool_dialect( return {**data, **{key: value for key, value in replaceable if key in data}} -def _is_chat_completions_body(data: Mapping[str, object]) -> bool: - messages: Final = data.get("messages") - if isinstance(messages, list) and messages: - return True - return "messages" in data and "input" not in data - - _CURSOR_THINKING_SEPARATOR: Final = "-thinking-" _CURSOR_FAST_SUFFIX: Final = "-fast" _CURSOR_THINKING_LEVELS: Final[frozenset[str]] = frozenset(get_args(REASONING_EFFORT)) @@ -261,7 +255,7 @@ def _resolve_cursor_model_variant( resolved: Final = {**data, "model": variant.base_model} if variant.reasoning_effort is None: return resolved - if _is_chat_completions_body(data): + if is_chat_completions_body(data): if "reasoning_effort" in data: return resolved return {**resolved, "reasoning_effort": variant.reasoning_effort} @@ -646,7 +640,7 @@ async def cursor_chat_completions( raw_body: Final = await read_request_body(request=request) - if _is_chat_completions_body(raw_body): + if is_chat_completions_body(raw_body): # Genuine chat completions body (Cursor sends these for models whose BYOK it # already fixed); delegate so behavior matches /chat/completions exactly. # Keyed on messages CONTENT, not key presence: Cursor can send a null or diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py index 9e125b95a68..b1055bb9124 100644 --- a/litellm/proxy/utils.py +++ b/litellm/proxy/utils.py @@ -250,7 +250,11 @@ from litellm.proxy.hooks.sensitive_data_routing import ( # noqa: F401, RUF100 PROXY_SensitiveDataRoutingHandler, _PROXY_SensitiveDataRoutingHandler, # pyright: ignore[reportPrivateUsage,reportUnusedImport] # backwards-compatible package export ) -from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup, add_guardrails_from_auth_metadata +from litellm.proxy.litellm_pre_call_utils import ( + LiteLLMProxyRequestSetup, + add_guardrails_from_auth_metadata, + normalized_metadata_slot, +) from litellm.proxy.management_helpers.key_settings_audit import with_settings_updated_at from litellm.proxy.policy_engine.pipeline_executor import ( PipelineExecutor, @@ -3973,9 +3977,7 @@ class ProxyLogging: ) request_data["litellm_logging_obj"] = litellm_logging_obj # rebind-ok: lifted then popped by the caller request_data.update({REJECTED_BEFORE_REQUEST_SETUP_KEY: REJECTED_BEFORE_REQUEST_SETUP}) - if "metadata" not in request_data: - request_data["metadata"] = {} - request_data["metadata"].update(user_api_key_logged_metadata) + normalized_metadata_slot(request_data, "metadata").update(user_api_key_logged_metadata) if litellm_logging_obj is not None: ## UPDATE LOGGING INPUT diff --git a/tests/integration/providers/test_cursor_responses_wire.py b/tests/integration/providers/test_cursor_responses_wire.py index d77d35f2ff8..1bd79b5b575 100644 --- a/tests/integration/providers/test_cursor_responses_wire.py +++ b/tests/integration/providers/test_cursor_responses_wire.py @@ -3,11 +3,14 @@ from __future__ import annotations import json import os import uuid +from collections.abc import Mapping +from hashlib import sha256 from importlib.resources import files from typing import Final +import httpx import pytest -from integration._support.client import Gateway, Scenario +from integration._support.client import Gateway, Scenario, eventually, list_value, object_value from integration._support.responses_vendor import same_response from integration._support.wire import Reply, Request, Wire, wire_server from integration._support.worker_readiness import wait_for_worker_readiness @@ -38,6 +41,7 @@ _CHAT_SUFFIXES: Final[tuple[tuple[str, str], ...]] = (("", ""),) + tuple( (f"-thinking-{level}", level) for level in _CHAT_THINKING_LEVELS ) _PROMPT: Final = "Inspect the requested change." +_REQUEST_TAG: Final = "cursor-integration" _INPUT: Final = [{"role": "user", "content": [{"type": "input_text", "text": _PROMPT}]}] _CUSTOM_TOOL: Final = { "type": "custom", @@ -107,16 +111,30 @@ def _stream(identity: str) -> tuple[bytes, ...]: ) + (b"data: [DONE]\n\n",) -def _chat_completion(stream: bool) -> bytes | tuple[bytes, ...]: +def _chat_completion(stream: bool, identity: str) -> bytes | tuple[bytes, ...]: if stream: - return ( - b'data: {"id":"chat_cursor_1","object":"chat.completion.chunk","created":1,"model":"gpt-5","choices":[{"index":0,"delta":{"content":"Patch inspected."},"finish_reason":null}]}\n\n', - b'data: {"id":"chat_cursor_1","object":"chat.completion.chunk","created":1,"model":"gpt-5","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}\n\n', - b"data: [DONE]\n\n", - ) + return tuple( + f"data: {json.dumps(chunk)}\n\n".encode() + for chunk in ( + { + "id": identity, + "object": "chat.completion.chunk", + "created": 1, + "model": _MODEL, + "choices": [{"index": 0, "delta": {"content": "Patch inspected."}, "finish_reason": None}], + }, + { + "id": identity, + "object": "chat.completion.chunk", + "created": 1, + "model": _MODEL, + "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}], + }, + ) + ) + (b"data: [DONE]\n\n",) return json.dumps( { - "id": "chat_cursor_1", + "id": identity, "object": "chat.completion", "created": 1, "model": _MODEL, @@ -133,11 +151,12 @@ def _chat_completion(stream: bool) -> bytes | tuple[bytes, ...]: ).encode() -def _deployment(scenario: Scenario, wire: Wire) -> str: +def _deployment(scenario: Scenario, wire: Wire, **pricing: JsonValue) -> str: model: Final = scenario.model( model=f"openai/{_MODEL}", api_key=_API_KEY, api_base=f"{wire.url}/v1", + **pricing, ) wait_for_worker_readiness(scenario.gateway, model, 1, _PROXY_WORKERS) return model @@ -167,7 +186,7 @@ def test_cursor_responses_input_uses_responses_wire_and_chat_output( assert request.method == "POST" assert request.headers["authorization"] == f"Bearer {_API_KEY}", dict(request.headers) if request.target == "/v1/chat/completions": - completion: Final = _chat_completion(stream) + completion: Final = _chat_completion(stream, raw_id) return Reply( content_type="text/event-stream" if stream else "application/json", chunks=completion if isinstance(completion, tuple) else None, @@ -175,7 +194,7 @@ def test_cursor_responses_input_uses_responses_wire_and_chat_output( ) assert request.target == "/v1/responses", request.target body: Final = _JSON_OBJECT.validate_json(request.body) - assert {key: value for key, value in body.items() if key != "metadata"} == { + assert body == { "model": _MODEL, "input": _INPUT, **reasoning, @@ -259,41 +278,69 @@ def test_cursor_responses_input_uses_responses_wire_and_chat_output( assert [request.target for request in drain_contract_requests(wire)] == ["/v1/responses"] -def test_cursor_responses_body_carries_no_proxy_metadata(gateway: Gateway) -> None: - pytest.skip( - "BUG: /cursor/chat/completions forwards proxy-internal metadata (key hash, team, user, client ip, user agent) " - "to the provider in the Responses body" - ) +@pytest.mark.parametrize( + "client_metadata", ({}, {"cursor_conversation": "integration"}), ids=("no_client_metadata", "client_metadata") +) +@pytest.mark.parametrize("stream", (False, True)) +def test_cursor_responses_body_sends_only_client_metadata_and_bills_the_key( + gateway: Gateway, + stream: bool, + client_metadata: Mapping[str, JsonValue], +) -> None: raw_id: Final = f"resp_cursor_metadata_{uuid.uuid4().hex}" + alias: Final = f"integration-cursor-{uuid.uuid4().hex}" + sent_metadata: Final = {"metadata": client_metadata} if client_metadata else {} def respond(request: Request) -> Reply: if request.method == "GET" and request.target == "/v1/models": return model_discovery_reply(_MODEL) assert request.method == "POST" and request.target == "/v1/responses", request.target + if stream: + return Reply(content_type="text/event-stream", chunks=_stream(raw_id)) return Reply(body=json.dumps(_response(raw_id)).encode()) with wire_server(respond) as wire, gateway.scenario() as scenario: model: Final = _deployment(scenario, wire) + key: Final = scenario.key(key_alias=alias) + hashed_key: Final = sha256(key.encode()).hexdigest() response: Final = gateway.request( "POST", "/cursor/chat/completions", { - "model": f"{model}-thinking-high", + "model": model, "input": _INPUT, "messages": [], "tools": [_CUSTOM_TOOL], + "stream": stream, + **sent_metadata, }, + key=key, ) assert response.status_code == 200, response.text requests: Final = drain_contract_requests(wire) assert len(requests) == 1, requests - body: Final = _JSON_OBJECT.validate_json(requests[0].body) - assert body == { + assert _JSON_OBJECT.validate_json(requests[0].body) == { "model": _MODEL, "input": _INPUT, - "reasoning": {"effort": "high"}, "tools": [_CUSTOM_TOOL], + "stream": stream, + **sent_metadata, }, requests[0].body + (row,) = eventually( + lambda: gateway.rows( + 'SELECT call_type, spend, metadata FROM "LiteLLM_SpendLogs" WHERE api_key = %s', (hashed_key,) + ), + lambda rows: len(rows) == 1, + seconds=70, + ) + assert row["call_type"] == "aresponses", row + assert object_value(row["metadata"])["user_api_key_alias"] == alias, row + assert isinstance(row["spend"], float) and row["spend"] > 0, row + eventually( + lambda: gateway.rows('SELECT spend FROM "LiteLLM_VerificationToken" WHERE token = %s', (hashed_key,)), + lambda tokens: len(tokens) == 1 and tokens[0]["spend"] == pytest.approx(row["spend"]), + seconds=70, + ) def test_cursor_stream_reports_usage_when_the_client_asks_for_it(gateway: Gateway) -> None: @@ -341,10 +388,12 @@ def test_cursor_messages_body_uses_chat_completions_wire( suffix: str, reasoning_effort: str, ) -> None: + chat_id: Final = f"chat_cursor_{uuid.uuid4().hex}" + def respond(request: Request) -> Reply: if request.method == "GET" and request.target == "/v1/models": return model_discovery_reply(_MODEL) - completion: Final = _chat_completion(stream) + completion: Final = _chat_completion(stream, chat_id) return Reply( content_type="text/event-stream" if stream else "application/json", chunks=completion if isinstance(completion, tuple) else None, @@ -367,6 +416,7 @@ def test_cursor_messages_body_uses_chat_completions_wire( "stream": stream, }, key=key, + headers={"x-litellm-tags": _REQUEST_TAG}, ) assert response.status_code == 200, response.text if stream: @@ -393,7 +443,7 @@ def test_cursor_messages_body_uses_chat_completions_wire( payload: Final = _JSON_OBJECT.validate_json(response.content) identity: Final = payload.get("id") created: Final = payload.get("created") - assert isinstance(identity, str) and same_response(identity, "chat_cursor_1"), response.text + assert isinstance(identity, str) and same_response(identity, chat_id), response.text assert isinstance(created, int) and created > 0, response.text assert payload == { "id": identity, @@ -456,3 +506,72 @@ def test_cursor_messages_body_uses_chat_completions_wire( **({"stream_options": {"include_usage": True}} if stream else {}), **({"reasoning_effort": reasoning_effort} if reasoning_effort else {}), }, requests[0].body + (row,) = eventually( + lambda: gateway.rows( + 'SELECT call_type, team_id, request_tags FROM "LiteLLM_SpendLogs" WHERE api_key = %s', + (sha256(key.encode()).hexdigest(),), + ), + lambda rows: len(rows) == 1, + seconds=70, + ) + assert (row["call_type"], row["team_id"]) == ("acompletion", team), row + assert _REQUEST_TAG in list_value(row["request_tags"]), row + + +@pytest.mark.parametrize( + ("body_shape", "metadata_as_json_string"), (("messages", False), ("input", False), ("input", True)) +) +def test_cursor_body_keeps_client_and_key_tags_and_enforces_the_client_tag_budget( + gateway: Gateway, body_shape: str, metadata_as_json_string: bool +) -> None: + chat_id: Final = f"chat_cursor_{uuid.uuid4().hex}" + response_id: Final = f"resp_cursor_tags_{uuid.uuid4().hex}" + budget_tag: Final = f"cursor-tag-budget-{uuid.uuid4().hex}" + key_tag: Final = f"cursor-key-tag-{uuid.uuid4().hex}" + client_metadata: Final = {"tags": [budget_tag]} + + def respond(request: Request) -> Reply: + if request.method == "GET" and request.target == "/v1/models": + return model_discovery_reply(_MODEL) + if request.target == "/v1/responses": + return Reply(body=json.dumps(_response(response_id)).encode()) + completion: Final = _chat_completion(False, chat_id) + return Reply(body=completion if isinstance(completion, bytes) else b"") + + def delete_tag() -> None: + gateway.post("/tag/delete", {"name": budget_tag}) + + with wire_server(respond) as wire, gateway.scenario() as scenario: + model: Final = _deployment(scenario, wire, input_cost_per_token=0.01, output_cost_per_token=0.01) + key: Final = scenario.key(models=[model], metadata={"tags": [key_tag]}) + gateway.post("/tag/new", {"name": budget_tag, "max_budget": 0.0001}) + scenario.cleanups.callback(delete_tag) + + def tagged_chat(text: str) -> httpx.Response: + return gateway.request( + "POST", + "/cursor/chat/completions", + { + "model": model, + body_shape: [{"role": "user", "content": text}], + "metadata": json.dumps(client_metadata) if metadata_as_json_string else client_metadata, + }, + key=key, + ) + + first: Final = tagged_chat("first tagged chat") + assert first.status_code == 200, first.text + (row,) = eventually( + lambda: gateway.rows( + 'SELECT request_tags FROM "LiteLLM_SpendLogs" WHERE api_key = %s', + (sha256(key.encode()).hexdigest(),), + ), + lambda rows: len(rows) == 1, + seconds=70, + ) + assert {budget_tag, key_tag} <= set(list_value(row["request_tags"])), row + blocked: Final = eventually( + lambda: tagged_chat("tag budget probe"), lambda response: response.status_code != 200, seconds=70 + ) + assert blocked.status_code == 422, blocked.text + assert "budget" in blocked.text.lower(), blocked.text diff --git a/tests/unit/integrations/test_custom_guardrail.py b/tests/unit/integrations/test_custom_guardrail.py index 614ff97553a..0483888533e 100644 --- a/tests/unit/integrations/test_custom_guardrail.py +++ b/tests/unit/integrations/test_custom_guardrail.py @@ -3699,6 +3699,20 @@ def test_get_guardrail_from_metadata() -> None: ) == [{"test-guardrail": {"extra_body": {"key": "value"}}}] + +@pytest.mark.parametrize( + "caller_metadata", + [{"guardrails": []}, {"guardrails": ["not-a-guardrail"]}, {"guardrails": None}], +) +def test_caller_metadata_guardrails_never_hide_the_proxy_guardrails_in_litellm_metadata(caller_metadata: dict) -> None: + guardrail: Final = CustomGuardrail( + guardrail_name="key-guardrail", event_hook=GuardrailEventHooks.pre_call, default_on=False + ) + request_data: Final = {"metadata": caller_metadata, "litellm_metadata": {"guardrails": ["key-guardrail"]}} + + assert guardrail.should_run_guardrail(request_data, GuardrailEventHooks.pre_call) is True + + def test_guardrail_is_in_requested_guardrails() -> None: guardrail: Final = CustomGuardrail(guardrail_name="test-guardrail") diff --git a/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py b/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py index b66e7f3d091..3aad2bda9bf 100644 --- a/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py +++ b/tests/unit/proxy/hooks/test_parallel_request_limiter_v3.py @@ -2591,7 +2591,8 @@ async def test_agent_session_rate_limit_skipped_without_session_id(): @pytest.mark.asyncio -async def test_agent_rate_limit_from_metadata_agent_id(): +@pytest.mark.parametrize("metadata_field", ("metadata", "litellm_metadata")) +async def test_agent_rate_limit_from_metadata_agent_id(metadata_field): """ Test that agent rate limits work when agent_id comes from request metadata (header) rather than from the API key. @@ -2639,7 +2640,7 @@ async def test_agent_rate_limit_from_metadata_agent_id(): cache=local_cache, data={ "model": "gpt-4", - "metadata": {"agent_id": _agent_id}, + metadata_field: {"agent_id": _agent_id}, }, call_type="", ) diff --git a/tests/unit/proxy/hooks/test_proxy_track_cost_callback.py b/tests/unit/proxy/hooks/test_proxy_track_cost_callback.py index dc46b5faf4e..8d4f8cd87ad 100644 --- a/tests/unit/proxy/hooks/test_proxy_track_cost_callback.py +++ b/tests/unit/proxy/hooks/test_proxy_track_cost_callback.py @@ -623,6 +623,56 @@ async def test_async_post_call_failure_hook_reads_used_client_oauth_token_from_t assert json.loads(payload["metadata"])["used_client_oauth_token"] is expected + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "request_route, body", + [ + ( + "/cursor/chat/completions", + { + "input": "Hello", + "metadata": {"requester_ip_address": "203.0.113.9", "user_agent": "spoofed"}, + "litellm_metadata": {"requester_ip_address": "127.0.0.1", "user_agent": "Cursor/1.0"}, + }, + ), + ( + "/v1/chat/completions", + { + "messages": [{"role": "user", "content": "Hello"}], + "metadata": {"requester_ip_address": "127.0.0.1", "user_agent": "Cursor/1.0"}, + "litellm_metadata": {"requester_ip_address": "203.0.113.9", "user_agent": "spoofed"}, + }, + ), + ], +) +async def test_async_post_call_failure_hook_logs_the_proxy_stamped_ip_and_user_agent(request_route: str, body: dict): + request_data = { + "model": "gpt-4o", + "proxy_server_request": {"request_id": "test_request_id"}, + **body, + } + + with patch( + "litellm.proxy.db.db_spend_update_writer.DBSpendUpdateWriter.update_database", + new_callable=AsyncMock, + ) as mock_update_database: + await ProxyDBLogger().async_post_call_failure_hook( + request_data=request_data, + original_exception=Exception("invalid api key"), + user_api_key_dict=UserAPIKeyAuth(api_key="test_api_key", request_route=request_route), + ) + + payload = get_logging_payload( + kwargs=mock_update_database.call_args[1]["kwargs"], + response_obj={}, + start_time=datetime.now(), + end_time=datetime.now(), + ) + assert payload["requester_ip_address"] == "127.0.0.1" + assert json.loads(payload["metadata"])["user_agent"] == "Cursor/1.0" + + @pytest.mark.asyncio async def test_async_post_call_failure_hook_bills_guardrail_cost_on_blocked_request(): """LIT-5651: a request blocked by a guardrail never reaches the LLM, but the diff --git a/tests/unit/proxy/response_api_endpoints/test_endpoints.py b/tests/unit/proxy/response_api_endpoints/test_endpoints.py index a783ac70881..1e4dded801e 100644 --- a/tests/unit/proxy/response_api_endpoints/test_endpoints.py +++ b/tests/unit/proxy/response_api_endpoints/test_endpoints.py @@ -3938,7 +3938,7 @@ class TestCursorInputArmFlattening: class TestChatCompletionsBodyDetection: def test_routing_matrix(self): - from litellm.proxy.response_api_endpoints.endpoints import _is_chat_completions_body + from litellm.proxy.litellm_pre_call_utils import is_chat_completions_body as _is_chat_completions_body assert _is_chat_completions_body({"messages": [{"role": "user", "content": "hi"}]}) is True assert _is_chat_completions_body({"messages": [{"role": "user", "content": "hi"}], "input": []}) is True diff --git a/tests/unit/proxy/test_common_request_processing.py b/tests/unit/proxy/test_common_request_processing.py index c867db2fd02..2d08609c93b 100644 --- a/tests/unit/proxy/test_common_request_processing.py +++ b/tests/unit/proxy/test_common_request_processing.py @@ -9711,6 +9711,40 @@ async def test_saved_weights_override_caller_input_and_preserve_key_precedence( assert logging.pre_call_hook.call_args.kwargs["data"].get("_router_weights") == expected +def _hook_data(**kwargs: object) -> object: + return kwargs["data"] + + +@pytest.mark.asyncio +@pytest.mark.parametrize("prompt, proxy_field, client_field", [ + ({"input": "hi"}, "litellm_metadata", "metadata"), + ({"messages": [{"role": "user", "content": "hi"}]}, "metadata", "litellm_metadata"), +]) +async def test_cursor_route_writes_proxy_fields_only_to_the_field_its_body_shape_picks( + prompt: dict[str, str | list[dict[str, str]]], + proxy_field: str, + client_field: str, +) -> None: + from litellm.proxy import proxy_server + + logging: Final = MagicMock(spec=ProxyLogging) + logging.pre_call_hook = AsyncMock(side_effect=_hook_data) + processor: Final = ProxyBaseLLMRequestProcessing(data={"model": "gpt-x", "metadata": {"client": "cursor"}, **prompt}) + data, _ = await processor.common_processing_pre_call_logic( + request=Request({"type": "http", "method": "POST", "path": "/cursor/chat/completions", "headers": []}), + general_settings={}, + user_api_key_dict=ProxyUserAPIKeyAuth(api_key="hash"), + proxy_logging_obj=logging, + proxy_config=proxy_server.ProxyConfig(), + route_type="acompletion", + logging_auth=ProxyUserAPIKeyAuth(api_key="logging-hash", key_alias="logging-alias"), + ) + assert data[proxy_field]["user_api_key_alias"] == "logging-alias" + assert "queue_time_seconds" in data[proxy_field] + assert "user_api_key_alias" not in data.get(client_field, {}) + assert "queue_time_seconds" not in data.get(client_field, {}) + assert data["metadata"]["client"] == "cursor" + class TestPerRequestModelGroupAlias: """``router_settings.model_group_alias`` on a key or team has to be resolved by the proxy: the Router resolves aliases from its own shared instance diff --git a/tests/unit/proxy/test_litellm_pre_call_utils.py b/tests/unit/proxy/test_litellm_pre_call_utils.py index 1ce61a8e9b3..2edb032b048 100644 --- a/tests/unit/proxy/test_litellm_pre_call_utils.py +++ b/tests/unit/proxy/test_litellm_pre_call_utils.py @@ -121,6 +121,23 @@ class TestGetMetadataVariableName: request = self._make_request("/chat/completions") assert get_metadata_variable_name(request) == "metadata" + @pytest.mark.parametrize( + "body, expected", + [ + ({"messages": [{"role": "user", "content": "hi"}]}, "metadata"), + ({"messages": []}, "metadata"), + ({"input": [{"role": "user", "content": "hi"}]}, "litellm_metadata"), + ({"input": "hi", "messages": []}, "litellm_metadata"), + ], + ) + def test_cursor_route_picks_the_metadata_field_by_body_shape(self, body, expected): + request = self._make_request("/cursor/chat/completions") + assert get_metadata_variable_name(request, body) == expected + + def test_cursor_route_without_a_body_keeps_the_route_answer(self): + request = self._make_request("/cursor/chat/completions") + assert get_metadata_variable_name(request) == "metadata" + def test_returns_metadata_for_completions(self): request = self._make_request("/v1/completions") assert get_metadata_variable_name(request) == "metadata" @@ -681,6 +698,45 @@ async def test_add_litellm_data_to_request_rejects_non_object_metadata(field, va assert field not in data +@pytest.mark.asyncio +@pytest.mark.parametrize("route", ["/v1/chat/completions", "/v1/responses", "/cursor/chat/completions"]) +@pytest.mark.parametrize("value, received_type", [("junk", "a string"), (5, "an integer"), ([1], "an array")]) +async def test_invalid_litellm_metadata_answers_400_when_metadata_carries_tags(route, value, received_type): + from litellm.proxy.auth.auth_checks import common_checks + from litellm.proxy.utils import ProxyLogging + + data = {"model": "gpt-4o", "input": "hi", "metadata": {"tags": ["client-tag"]}, "litellm_metadata": value} + request = _batches_request_mock() + await common_checks( + request_body=data, + team_object=None, + user_object=None, + end_user_object=None, + global_proxy_spend=None, + general_settings={}, + route=route, + llm_router=None, + proxy_logging_obj=ProxyLogging(user_api_key_cache=None), + valid_token=UserAPIKeyAuth(api_key="hashed-key"), + request=request, + ) + + with pytest.raises(ProxyException) as exc_info: + await add_litellm_data_to_request( + data=data, + request=request, + user_api_key_dict=UserAPIKeyAuth(api_key="hashed-key"), + proxy_config=MagicMock(), + general_settings={}, + version="test-version", + ) + + assert exc_info.value.code == "400" + assert exc_info.value.message == ( + f"Invalid type for 'litellm_metadata': expected an object, but got {received_type} instead." + ) + + @pytest.mark.asyncio async def test_add_litellm_data_to_request_removes_every_invalid_metadata_field_before_raising(): """When both fields are invalid, the raise for the first must not leave the @@ -11098,3 +11154,15 @@ def test_mapped_baggage_tags_keep_non_string_client_tags_as_sent(): data: Final[dict[str, object]] = {"metadata": {"tags": [1, {"team": "x"}, "client-tag", "client-tag"]}} LiteLLMProxyRequestSetup.apply_authenticated_baggage_tags(request, data) assert data["metadata"]["tags"] == [1, {"team": "x"}, "client-tag", "client-tag", "tenant:a"] + + +@pytest.mark.parametrize("caller_metadata", ({"tags": ["client-tag"]}, json.dumps({"tags": ["client-tag"]}))) +def test_caller_tags_reach_litellm_metadata_whether_metadata_is_a_dict_or_a_json_string(caller_metadata): + data: Final[dict[str, object]] = { + "input": "hi", + "metadata": caller_metadata, + "litellm_metadata": {"tags": ["key-tag"]}, + } + LiteLLMProxyRequestSetup.merge_caller_body_tags(request_data=data) + assert set(data["litellm_metadata"]["tags"]) == {"key-tag", "client-tag"} + assert data["metadata"] == caller_metadata diff --git a/tests/unit/proxy/utils/proxy_logging/test_post_call_failure_hook.py b/tests/unit/proxy/utils/proxy_logging/test_post_call_failure_hook.py index f880d3548f7..4386d5aec94 100644 --- a/tests/unit/proxy/utils/proxy_logging/test_post_call_failure_hook.py +++ b/tests/unit/proxy/utils/proxy_logging/test_post_call_failure_hook.py @@ -4,6 +4,7 @@ and ``_handle_logging_proxy_only_error``.""" from __future__ import annotations import asyncio +import json from collections.abc import Callable from datetime import datetime from types import MappingProxyType @@ -1192,3 +1193,31 @@ async def test_handle_logging_proxy_only_path_flags_requests_that_died_before_re blocked_by_guardrail.get(REJECTED_BEFORE_REQUEST_SETUP_KEY) is REJECTED_BEFORE_REQUEST_SETUP, blocked_by_guardrail["litellm_call_id"], ) == (True, "client-supplied", False, "call-1") + + +@pytest.mark.asyncio +async def test_auth_failure_on_a_responses_body_with_a_json_string_metadata_is_logged_against_the_key( + proxy_logging, monkeypatch +): + from litellm.proxy import proxy_server + + recorded: list[dict] = [] + + class _RecordingLogger(CustomLogger): + async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time): + recorded.append(kwargs) + + monkeypatch.setattr(proxy_server, "llm_router", None) + monkeypatch.setattr(litellm, "callbacks", [_RecordingLogger()]) + proxy_logging.alert_types = [] + + await proxy_logging.post_call_failure_hook( + request_data={"model": "internal-model", "input": "hi", "metadata": json.dumps({"tags": ["checkout"]})}, + original_exception=litellm.BudgetExceededError(current_cost=1.0, max_budget=0.5), + user_api_key_dict=UserAPIKeyAuth(api_key="sk-hashed", key_alias="checkout-key", request_route="/v1/responses"), + error_type=ProxyErrorTypes.auth_error, + route="/v1/responses", + ) + + assert len(recorded) == 1 + assert recorded[0]["standard_logging_object"]["metadata"]["user_api_key_alias"] == "checkout-key" From 73aaf9e3daec8bb174accdf56ecade90b34c27c8 Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 30 Oct 2026 01:16:16 -0700 Subject: [PATCH 13/23] fix(mistral): forward prompt_cache_key to the Mistral API (#47432) * fix(mistral): forward prompt_cache_key to the Mistral API * fix(mistral): send prompt_cache_key only to Mistral's API and accept reasoning_effort on devstral and code-agent models * fix(mistral): keep the upstream api_base out of the prompt_cache_key rejection message * fix(mistral): honor allowed_openai_params for prompt_cache_key and gate codestral async and env hosts * test(mistral): annotate the new adapter prompt_cache_key tests * fix(mistral): drop or reject a prompt_cache_key the Mistral API cannot take A non-ASCII prompt_cache_key gets a 500 from api.mistral.ai, and a non-string one a 422, so both now drop under drop_params and answer 400 without it, like any value the provider cannot take. The /v1/messages adapter derives no key from a non-ASCII Claude Code session for Mistral, and the 400 names the caller's provider, so a codestral deployment's error says codestral. * fix(mistral): send a reasoning_effort object as its level and keep UnsupportedParamsError on sync calls * fix(mistral): keep async codestral file uploads working by not rewriting the caller's file blocks --- .../litellm_core_utils/get_litellm_params.py | 1 + .../anthropic/messages/adapters/handler.py | 2 + .../messages/adapters/transformation.py | 30 +- litellm/llms/mistral/chat/transformation.py | 153 ++++++- litellm/llms/openai/openai.py | 5 +- litellm/main.py | 1 + litellm/types/llms/mistral.py | 19 + .../adapters/test_handler_prompt_cache_key.py | 68 ++++ .../test_mistral_chat_transformation.py | 372 +++++++++++++++--- 9 files changed, 576 insertions(+), 75 deletions(-) diff --git a/litellm/litellm_core_utils/get_litellm_params.py b/litellm/litellm_core_utils/get_litellm_params.py index 8792fd0f58a..29f96d61976 100644 --- a/litellm/litellm_core_utils/get_litellm_params.py +++ b/litellm/litellm_core_utils/get_litellm_params.py @@ -85,6 +85,7 @@ OPTIONAL_KWARGS_KEYS: Final = ( "github_copilot_user_session", "fireworks_forward_user_id", "responses_api_version", + "allowed_openai_params", PROVIDER_AFFINITY_HEADER_KWARG_KEY, } ) diff --git a/litellm/llms/anthropic/messages/adapters/handler.py b/litellm/llms/anthropic/messages/adapters/handler.py index 704bec9056f..b3b84f721c0 100644 --- a/litellm/llms/anthropic/messages/adapters/handler.py +++ b/litellm/llms/anthropic/messages/adapters/handler.py @@ -647,12 +647,14 @@ class LiteLLMMessagesToCompletionTransformationHandler: request_data["output_config"] = extra_kwargs["output_config"] custom_llm_provider: Final = extra_kwargs.get("custom_llm_provider") + api_base: Final = extra_kwargs.get("api_base") ( openai_request, tool_name_mapping, ) = ANTHROPIC_ADAPTER.translate_completion_input_params_with_tool_mapping( request_data, custom_llm_provider=custom_llm_provider if isinstance(custom_llm_provider, str) else None, + api_base=api_base if isinstance(api_base, str) else None, drop_params=drops_unsupported_params(extra_kwargs.get("drop_params")), ) diff --git a/litellm/llms/anthropic/messages/adapters/transformation.py b/litellm/llms/anthropic/messages/adapters/transformation.py index ddda759384f..8ea6293bbe6 100644 --- a/litellm/llms/anthropic/messages/adapters/transformation.py +++ b/litellm/llms/anthropic/messages/adapters/transformation.py @@ -24,6 +24,7 @@ OPENAI_MAX_TOOL_NAME_LENGTH: Final = 64 TOOL_NAME_HASH_LENGTH: Final = 8 TOOL_NAME_PREFIX_LENGTH: Final = OPENAI_MAX_TOOL_NAME_LENGTH - TOOL_NAME_HASH_LENGTH - 1 # 55 PROVIDERS_PROXYING_AN_UNKNOWN_BACKEND: Final = frozenset({"litellm_proxy"}) +PROVIDERS_SERVED_BY_THE_MISTRAL_CHAT_CONFIG: Final = frozenset({"mistral", "codestral"}) _COMPACTION_BLOCK: Final = TypeAdapter(dict[str, JsonValue]) @@ -255,7 +256,12 @@ class AnthropicAdapter: return result def translate_completion_input_params_with_tool_mapping( - self, kwargs, *, custom_llm_provider: str | None = None, drop_params: bool = True + self, + kwargs, + *, + custom_llm_provider: str | None = None, + api_base: str | None = None, + drop_params: bool = True, ) -> tuple[ChatCompletionRequest | None, dict[str, str]]: """ Translate Anthropic request params to OpenAI format, returning tool name mapping. @@ -289,6 +295,7 @@ class AnthropicAdapter: ) = LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai( anthropic_message_request=request_body, custom_llm_provider=custom_llm_provider, + api_base=api_base, drop_params=drop_params, ) @@ -1001,7 +1008,9 @@ class LiteLLMAnthropicMessagesAdapter: ) @staticmethod - def _supports_prompt_cache_key(model: str | None, custom_llm_provider: str | None) -> bool: + def _supports_prompt_cache_key( + prompt_cache_key: str, model: str | None, custom_llm_provider: str | None, api_base: str | None + ) -> bool: if not model or not custom_llm_provider: return False if custom_llm_provider in PROVIDERS_PROXYING_AN_UNKNOWN_BACKEND: @@ -1009,7 +1018,17 @@ class LiteLLMAnthropicMessagesAdapter: supported_params: Final = litellm.get_supported_openai_params( model=model, custom_llm_provider=custom_llm_provider ) - return "prompt_cache_key" in (supported_params or ()) + if "prompt_cache_key" not in (supported_params or ()): + return False + if custom_llm_provider not in PROVIDERS_SERVED_BY_THE_MISTRAL_CHAT_CONFIG: + return True + _, _, _, resolved_api_base = litellm.get_llm_provider( + model=model, custom_llm_provider=custom_llm_provider, api_base=api_base + ) + mistral_config: Final = litellm.MistralConfig() + return mistral_config.takes_prompt_cache_key_value( + prompt_cache_key + ) and mistral_config.accepts_prompt_cache_key_at(resolved_api_base) @staticmethod def _target_declares_reasoning_effort(model: str, custom_llm_provider: str | None) -> bool: @@ -1042,6 +1061,7 @@ class LiteLLMAnthropicMessagesAdapter: new_kwargs: ChatCompletionRequest, *, custom_llm_provider: str | None = None, + api_base: str | None = None, ) -> None: """Translate metadata fields from Anthropic request to OpenAI request.""" if "metadata" in anthropic_message_request: @@ -1050,7 +1070,7 @@ class LiteLLMAnthropicMessagesAdapter: new_kwargs["user"] = metadata["user_id"] prompt_cache_key: Final = prompt_cache_key_from_user_id(metadata["user_id"]) if prompt_cache_key is not None and self._supports_prompt_cache_key( - anthropic_message_request.get("model"), custom_llm_provider + prompt_cache_key, anthropic_message_request.get("model"), custom_llm_provider, api_base ): new_kwargs["prompt_cache_key"] = prompt_cache_key @@ -1244,6 +1264,7 @@ class LiteLLMAnthropicMessagesAdapter: anthropic_message_request: AnthropicMessagesRequest, *, custom_llm_provider: str | None = None, + api_base: str | None = None, preserve_midturn_system: bool = False, drop_params: bool = True, ) -> tuple[ChatCompletionRequest, dict[str, str]]: @@ -1282,6 +1303,7 @@ class LiteLLMAnthropicMessagesAdapter: anthropic_message_request=anthropic_message_request, new_kwargs=new_kwargs, custom_llm_provider=custom_llm_provider, + api_base=api_base, ) ## CONVERT TOOL CHOICE self._translate_tool_choice_to_openai( diff --git a/litellm/llms/mistral/chat/transformation.py b/litellm/llms/mistral/chat/transformation.py index 2fb9402ed03..15e348453e0 100644 --- a/litellm/llms/mistral/chat/transformation.py +++ b/litellm/llms/mistral/chat/transformation.py @@ -9,11 +9,16 @@ Docs - https://docs.mistral.ai/api/ from collections.abc import AsyncIterator, Coroutine, Iterator, Mapping, Sequence from itertools import accumulate, count from typing import TYPE_CHECKING, Final, Literal, cast, get_type_hints, overload +from urllib.parse import urlparse import httpx +from pydantic import TypeAdapter from typing_extensions import TypeIs # noqa: TID251 # TypeIs reaches typing only on 3.13 +import litellm from litellm._logging import verbose_logger +from litellm.exceptions import UnsupportedParamsError +from litellm.litellm_core_utils.core_helpers import as_str_mapping from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.prompt_templates.common_utils import ( convert_content_list_to_str, @@ -30,7 +35,15 @@ from litellm.router_utils.reasoning_effort_capability import ( nearest_declared_reasoning_effort, ) from litellm.secret_managers.main import get_secret_str -from litellm.types.llms.mistral import MistralTextBlock, MistralThinkingBlock, MistralToolCallMessage +from litellm.types.llms.mistral import ( + MistralCacheKeyParams, + MistralFilePart, + MistralReasoningEffortParams, + MistralRequestRouting, + MistralTextBlock, + MistralThinkingBlock, + MistralToolCallMessage, +) from litellm.types.llms.openai import AllMessageValues, ChatCompletionAssistantMessage, ChatCompletionTextObject from litellm.types.utils import ModelResponse, ModelResponseStream from litellm.utils import convert_to_model_response_object, supports_reasoning @@ -57,6 +70,17 @@ def _accepted_reasoning_effort(model: str, requested: str, custom_llm_provider: _NO_TOOL_CALL_TURN: Final[tuple[int, frozenset[str]]] = (-1, frozenset()) +_MISTRAL_REQUEST_ROUTING: Final = TypeAdapter(MistralRequestRouting) +_MISTRAL_CACHE_KEY_PARAMS: Final = TypeAdapter(MistralCacheKeyParams) +_MISTRAL_REASONING_EFFORT_PARAMS: Final = TypeAdapter(MistralReasoningEffortParams) +_MISTRAL_FILE_PART: Final = TypeAdapter(MistralFilePart) + + +def _reasoning_level_and_summary(reasoning_effort: object) -> tuple[object, object]: + requested: Final = as_str_mapping(reasoning_effort) + if requested is None: + return reasoning_effort, None + return requested.get("effort"), requested.get("summary") def _is_json_array(value: object) -> TypeIs[Sequence[object]]: # guard-ok: a decoded JSON array is a list @@ -88,6 +112,8 @@ class MistralConfig(OpenAIGPTConfig): - `response_format` (object or null): An object specifying the format that the model must output. Setting to { "type": "json_object" } enables JSON mode, which guarantees the message the model generates is in JSON. When using JSON mode you MUST also instruct the model to produce JSON yourself with a system or a user message. """ + _is_base_class = False + temperature: int | None = None top_p: int | None = None max_tokens: int | None = None @@ -131,6 +157,7 @@ class MistralConfig(OpenAIGPTConfig): def get_supported_openai_params(self, model: str) -> list[str]: is_magistral: Final = "magistral" in model.lower() accepts_reasoning_effort: Final = self._accepts_reasoning(model) + accepts_prompt_cache_key: Final = self.custom_llm_provider == "mistral" return [ "stream", "temperature", @@ -145,6 +172,7 @@ class MistralConfig(OpenAIGPTConfig): "parallel_tool_calls", *(("thinking",) if is_magistral else ()), *(("reasoning_effort",) if accepts_reasoning_effort else ()), + *(("prompt_cache_key",) if accepts_prompt_cache_key else ()), ] def _map_tool_choice(self, tool_choice: str) -> str: @@ -211,15 +239,84 @@ class MistralConfig(OpenAIGPTConfig): optional_params["extra_body"] = {"random_seed": value} if param == "response_format": optional_params["response_format"] = value - if param == "reasoning_effort" and "magistral" not in model.lower(): - optional_params["reasoning_effort"] = _accepted_reasoning_effort(model, value, self.custom_llm_provider) + if ( + param == "reasoning_effort" + and "magistral" not in model.lower() + and ( + reasoning_level := self._reasoning_level_to_send( + model, _MISTRAL_REASONING_EFFORT_PARAMS.validate_python(non_default_params), drop_params + ) + ) + is not None + ): + optional_params["reasoning_effort"] = reasoning_level if param in ("reasoning_effort", "thinking") and "magistral" in model.lower(): # Flag that we need to add reasoning system prompt optional_params["_add_reasoning_prompt"] = True - if param == "parallel_tool_calls": - optional_params["parallel_tool_calls"] = value + if param in ("parallel_tool_calls", "prompt_cache_key"): + optional_params[param] = value return optional_params + def _reasoning_level_to_send( + self, model: str, requested: MistralReasoningEffortParams, drop_params: bool + ) -> str | None: + level, summary = _reasoning_level_and_summary(requested.get("reasoning_effort")) + sendable: Final = ( + _accepted_reasoning_effort(model, level, self.custom_llm_provider) if isinstance(level, str) else None + ) + if sendable is not None and summary is None: + return sendable + if drop_params or litellm.drop_params is True: + return sendable + raise UnsupportedParamsError( + message=( + f"{self.custom_llm_provider} does not support parameters: ['reasoning_effort'], for model={model}, " + "since the Mistral API takes it only as a level such as 'high', with no reasoning summary. " + "To drop it, set `litellm.drop_params=True` or for proxy:\n\n`litellm_settings:\n drop_params: true`\n" + ), + llm_provider=self.custom_llm_provider, + model=model, + ) + + def accepts_prompt_cache_key_at(self, api_base: str | None) -> bool: + resolved_api_base, _ = self._get_openai_compatible_provider_info(api_base, None) + hostname: Final = urlparse(resolved_api_base).hostname or "" + return hostname == "mistral.ai" or hostname.endswith(".mistral.ai") + + @staticmethod + def takes_prompt_cache_key_value(prompt_cache_key: object) -> bool: + return isinstance(prompt_cache_key, str) and prompt_cache_key.isascii() + + def _prompt_cache_key_rejection(self, prompt_cache_key: object, api_base: str | None) -> str | None: + if not self.accepts_prompt_cache_key_at(api_base): + return "since only the Mistral API (*.mistral.ai) takes it" + if not self.takes_prompt_cache_key_value(prompt_cache_key): + return "since the Mistral API takes it only as a string of ASCII characters" + return None + + def _prompt_cache_key_withheld(self, model: str, prompt_cache_key: object, routing: MistralRequestRouting) -> bool: + requested_api_base: Final = routing.get("api_base") + api_base: Final = requested_api_base if isinstance(requested_api_base, str) else None + allowed_openai_params: Final = routing.get("allowed_openai_params") + if isinstance(allowed_openai_params, (list, tuple)) and "prompt_cache_key" in allowed_openai_params: + return False + rejection: Final = self._prompt_cache_key_rejection(prompt_cache_key, api_base) + if rejection is None: + return False + if litellm.drop_params is True or routing.get("drop_params") is True: + return True + requested_provider: Final = routing.get("custom_llm_provider") + provider: Final = requested_provider if isinstance(requested_provider, str) else self.custom_llm_provider + raise UnsupportedParamsError( + message=( + f"{provider} does not support parameters: ['prompt_cache_key'], for model={model}, {rejection}. " + "To send it anyway, pass `allowed_openai_params=['prompt_cache_key']`. To drop it, set " + "`litellm.drop_params=True` or for proxy:\n\n`litellm_settings:\n drop_params: true`\n" + ), + llm_provider=provider, + model=model, + ) + def _get_openai_compatible_provider_info(self, api_base: str | None, api_key: str | None) -> tuple[str, str | None]: # mistral is openai compatible, we just need to set this to custom_openai and have the api_base be https://api.mistral.ai api_base = ( @@ -359,24 +456,31 @@ class MistralConfig(OpenAIGPTConfig): messages = self._handle_message_with_file(messages) return messages + @staticmethod + def _with_mistral_file_id(block: Mapping[str, object]) -> Mapping[str, object]: + file_part: Final = block.get("file") if block.get("type") == "file" else None + file_id: Final = ( + _MISTRAL_FILE_PART.validate_python(file_part).get("file_id") if isinstance(file_part, Mapping) else None + ) + if not file_id: + return block + return {**{key: value for key, value in block.items() if key != "file"}, "file_id": file_id} + + @classmethod + def _with_mistral_file_ids(cls, message: AllMessageValues) -> AllMessageValues: + content: Final = message.get("content") + if not isinstance(content, list): + return message + blocks: Final[tuple[Mapping[str, object], ...]] = tuple(content) + return cast( # cast-ok: only file blocks change, to the file_id shape the Mistral API takes + AllMessageValues, {**message, "content": [cls._with_mistral_file_id(block) for block in blocks]} + ) + def _handle_message_with_file(self, messages: list[AllMessageValues]) -> list[AllMessageValues]: """ Mistral API supports only 'file_id' in message content with type 'file'. """ - for m in messages: - _content_block = m.get("content") - if _content_block and isinstance(_content_block, list): - if any(c.get("type") == "file" for c in _content_block): - # If file content is present, we get file_id from 'file' attribute of content block - # then replace 'file' with 'file_id' and assign the value of 'file_id' attribute to it. - file_contents = [c for c in _content_block if c.get("type") == "file"] - for file_content in file_contents: - file_id = file_content.get("file", {}).get("file_id") - if file_id: - # Replace 'file' with 'file_id' - file_content["file_id"] = file_id # pyright: ignore[reportGeneralTypeIssues] # legacy in-place rewrite of the block shape - file_content.pop("file", None) - return messages + return [self._with_mistral_file_ids(m) for m in messages] def _add_reasoning_system_prompt_if_needed( self, messages: list[AllMessageValues], optional_params: dict @@ -794,7 +898,16 @@ class MistralConfig(OpenAIGPTConfig): if "magistral" in model.lower() and optional_params.get("_add_reasoning_prompt", False): messages = self._add_reasoning_system_prompt_if_needed(messages, optional_params) - upstream_params: Final = {key: value for key, value in optional_params.items() if key != "client_metadata"} + cache_key_params: Final = _MISTRAL_CACHE_KEY_PARAMS.validate_python(optional_params) + withheld_params: Final = ( + ("client_metadata", "prompt_cache_key") + if "prompt_cache_key" in cache_key_params + and self._prompt_cache_key_withheld( + model, cache_key_params["prompt_cache_key"], _MISTRAL_REQUEST_ROUTING.validate_python(litellm_params) + ) + else ("client_metadata",) + ) + upstream_params: Final = {key: value for key, value in optional_params.items() if key not in withheld_params} # Call parent transform_request which handles _transform_messages return super().transform_request( diff --git a/litellm/llms/openai/openai.py b/litellm/llms/openai/openai.py index e0a23d6c99f..bef99fd0719 100644 --- a/litellm/llms/openai/openai.py +++ b/litellm/llms/openai/openai.py @@ -27,6 +27,7 @@ import litellm from litellm import LlmProviders from litellm._logging import verbose_logger from litellm.constants import DEFAULT_MAX_RETRIES +from litellm.exceptions import UnsupportedParamsError from litellm.files.types import FileContentStreamingResult from litellm.litellm_core_utils.core_helpers import set_provider_response_headers_in_hidden_params from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj @@ -971,8 +972,8 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): litellm.remove_index_from_tool_calls(messages=messages) else: raise e - except OpenAIError as e: - raise e + except (OpenAIError, UnsupportedParamsError): + raise except Exception as e: status_code: Final = getattr(e, "status_code", 500) error_headers = getattr(e, "headers", None) diff --git a/litellm/main.py b/litellm/main.py index d469aa629fe..8d6e07d6a8f 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -6199,6 +6199,7 @@ def completion( PROVIDER_AFFINITY_HEADER_KWARG_KEY, "fireworks_forward_user_id", "responses_api_version", + "allowed_openai_params", ) if key in kwargs }, diff --git a/litellm/types/llms/mistral.py b/litellm/types/llms/mistral.py index cd61f32f430..68a9d26827b 100644 --- a/litellm/types/llms/mistral.py +++ b/litellm/types/llms/mistral.py @@ -23,3 +23,22 @@ class MistralThinkingBlock(TypedDict): type: Literal["thinking"] thinking: list[MistralTextBlock] closed: NotRequired[ReadOnly[bool]] + + +class MistralRequestRouting(TypedDict, total=False): + api_base: ReadOnly[object] + drop_params: ReadOnly[object] + allowed_openai_params: ReadOnly[object] + custom_llm_provider: ReadOnly[object] + + +class MistralCacheKeyParams(TypedDict, total=False): + prompt_cache_key: ReadOnly[object] + + +class MistralReasoningEffortParams(TypedDict, total=False): + reasoning_effort: ReadOnly[object] + + +class MistralFilePart(TypedDict, total=False): + file_id: ReadOnly[object] diff --git a/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py b/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py index ed6cd6134ff..77bdacc938f 100644 --- a/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py +++ b/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py @@ -1,6 +1,7 @@ import json import os import sys +from typing import Final import pytest @@ -70,3 +71,70 @@ def test_prepare_completion_kwargs_keeps_prompt_cache_key_through_responses_rero ) assert completion_kwargs["model"] == "openai/responses/gpt-5.6-luna" assert completion_kwargs["prompt_cache_key"] == "session-abc" + + +@pytest.mark.parametrize( + ("api_base", "azure_api_base_env", "derives_key"), + [ + (None, None, True), + ("https://api.mistral.ai/v1", None, True), + ("https://my-resource.openai.azure.com/openai/v1", None, False), + (None, "https://my-resource.services.ai.azure.com/openai", False), + ], +) +def test_prepare_completion_kwargs_derives_prompt_cache_key_only_for_a_mistral_hosted_endpoint( + monkeypatch: pytest.MonkeyPatch, api_base: str | None, azure_api_base_env: str | None, derives_key: bool +) -> None: + if azure_api_base_env is None: + monkeypatch.delenv("MISTRAL_AZURE_API_BASE", raising=False) + else: + monkeypatch.setenv("MISTRAL_AZURE_API_BASE", azure_api_base_env) + + completion_kwargs: Final = _prepare( + "mistral/mistral-small-latest", {"custom_llm_provider": "mistral", "api_base": api_base} + ) + + assert completion_kwargs["user"] == CLAUDE_CODE_USER_ID + assert ("prompt_cache_key" in completion_kwargs) is derives_key + + +@pytest.mark.parametrize( + ("codestral_api_base_env", "derives_key"), + [(None, True), ("https://my-resource.services.ai.azure.com/openai/v1", False)], +) +def test_prepare_completion_kwargs_derives_prompt_cache_key_for_codestral_only_on_its_mistral_host( + monkeypatch: pytest.MonkeyPatch, codestral_api_base_env: str | None, derives_key: bool +) -> None: + monkeypatch.delenv("MISTRAL_AZURE_API_BASE", raising=False) + if codestral_api_base_env is None: + monkeypatch.delenv("CODESTRAL_API_BASE", raising=False) + else: + monkeypatch.setenv("CODESTRAL_API_BASE", codestral_api_base_env) + + completion_kwargs: Final = _prepare("codestral/codestral-latest", {"custom_llm_provider": "codestral"}) + + assert ("prompt_cache_key" in completion_kwargs) is derives_key + + +def test_prepare_completion_kwargs_derives_no_prompt_cache_key_mistral_cannot_take_from_a_non_ascii_session() -> None: + non_ascii_user_id: Final = json.dumps({"device_id": "d" * 64, "account_uuid": "", "session_id": "sitzung-ü"}) + + mistral_kwargs: Final = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs( + max_tokens=1024, + messages=MESSAGES, + model="mistral/mistral-small-latest", + metadata={"user_id": non_ascii_user_id}, + thinking=None, + extra_kwargs={"custom_llm_provider": "mistral"}, + )[0] + openai_kwargs: Final = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs( + max_tokens=1024, + messages=MESSAGES, + model="openai/gpt-5.6-luna", + metadata={"user_id": non_ascii_user_id}, + thinking=None, + extra_kwargs={"custom_llm_provider": "openai"}, + )[0] + + assert "prompt_cache_key" not in mistral_kwargs + assert openai_kwargs["prompt_cache_key"] == "sitzung-ü" diff --git a/tests/unit/llms/mistral/test_mistral_chat_transformation.py b/tests/unit/llms/mistral/test_mistral_chat_transformation.py index b76e1707805..c340f77c317 100644 --- a/tests/unit/llms/mistral/test_mistral_chat_transformation.py +++ b/tests/unit/llms/mistral/test_mistral_chat_transformation.py @@ -1,7 +1,13 @@ +import json +from collections.abc import Mapping from typing import Final, List, cast from unittest.mock import MagicMock, patch +from urllib.parse import urlparse +import httpx import pytest +import respx +from openai import AsyncOpenAI from litellm.litellm_core_utils.prompt_templates.common_utils import TOOL_RESULT_IMAGE_BOUNDARY from litellm.llms.mistral.chat.transformation import ( @@ -45,21 +51,15 @@ class TestMistralReasoningSupport: mistral_config = MistralConfig() # Test magistral model supports reasoning parameters - supported_params = mistral_config.get_supported_openai_params( - "mistral/magistral-medium-2506" - ) + supported_params = mistral_config.get_supported_openai_params("mistral/magistral-medium-2506") assert "reasoning_effort" in supported_params assert "thinking" in supported_params - supported_params_reasoning = mistral_config.get_supported_openai_params( - "mistral/mistral-medium-latest" - ) + supported_params_reasoning = mistral_config.get_supported_openai_params("mistral/mistral-medium-latest") assert "reasoning_effort" in supported_params_reasoning assert "thinking" not in supported_params_reasoning - supported_params_normal = mistral_config.get_supported_openai_params( - "mistral/mistral-large-latest" - ) + supported_params_normal = mistral_config.get_supported_openai_params("mistral/mistral-large-latest") assert "reasoning_effort" not in supported_params_normal assert "thinking" not in supported_params_normal @@ -117,6 +117,36 @@ class TestMistralReasoningSupport: ) assert optional_params["reasoning_effort"] == sent + @pytest.mark.parametrize( + ("requested", "drop_params", "sent"), + [ + ({"effort": "medium"}, False, {"reasoning_effort": "high"}), + ({"effort": "high", "summary": "auto"}, True, {"reasoning_effort": "high"}), + ({"summary": "auto"}, True, {}), + ], + ) + def test_reasoning_effort_object_is_sent_as_its_level(self, local_model_cost_map, requested, drop_params, sent): + import litellm + + optional_params = litellm.get_optional_params( + model="mistral-small-latest", + custom_llm_provider="mistral", + reasoning_effort=requested, + drop_params=drop_params, + ) + assert {key: optional_params[key] for key in ("reasoning_effort",) if key in optional_params} == sent + + @pytest.mark.parametrize("requested", [{"effort": "high", "summary": "auto"}, {"summary": "auto"}]) + def test_reasoning_effort_object_mistral_cannot_take_is_rejected(self, local_model_cost_map, requested): + import litellm + + with pytest.raises(litellm.UnsupportedParamsError, match=r"\['reasoning_effort'\]"): + litellm.get_optional_params( + model="devstral-latest", + custom_llm_provider="mistral", + reasoning_effort=requested, + ) + def test_reasoning_effort_is_forwarded_verbatim_when_the_map_declares_no_levels( self, local_model_cost_map, monkeypatch ): @@ -194,9 +224,7 @@ class TestMistralReasoningSupport: messages = [{"role": "user", "content": "What is 2+2?"}] optional_params = {"_add_reasoning_prompt": True} - result = mistral_config._add_reasoning_system_prompt_if_needed( - messages, optional_params - ) + result = mistral_config._add_reasoning_system_prompt_if_needed(messages, optional_params) # Should add a new system message at the beginning assert len(result) == 2 @@ -218,9 +246,7 @@ class TestMistralReasoningSupport: ] optional_params = {"_add_reasoning_prompt": True} - result = mistral_config._add_reasoning_system_prompt_if_needed( - messages, optional_params - ) + result = mistral_config._add_reasoning_system_prompt_if_needed(messages, optional_params) # Should modify existing system message assert len(result) == 2 @@ -251,9 +277,7 @@ class TestMistralReasoningSupport: ] optional_params = {"_add_reasoning_prompt": True} - result = mistral_config._add_reasoning_system_prompt_if_needed( - messages, optional_params - ) + result = mistral_config._add_reasoning_system_prompt_if_needed(messages, optional_params) # Should modify existing system message preserving list format assert len(result) == 2 @@ -266,10 +290,7 @@ class TestMistralReasoningSupport: # Original content should be preserved assert "You are a helpful assistant." in result[0]["content"][1]["text"] - assert ( - "You always provide detailed explanations." - in result[0]["content"][2]["text"] - ) + assert "You always provide detailed explanations." in result[0]["content"][2]["text"] assert result[1]["role"] == "user" @@ -287,9 +308,7 @@ class TestMistralReasoningSupport: ] string_params = {"_add_reasoning_prompt": True} - string_result = mistral_config._add_reasoning_system_prompt_if_needed( - string_messages, string_params - ) + string_result = mistral_config._add_reasoning_system_prompt_if_needed(string_messages, string_params) assert isinstance(string_result[0]["content"], str) assert "" in string_result[0]["content"] assert "You are helpful." in string_result[0]["content"] @@ -304,9 +323,7 @@ class TestMistralReasoningSupport: ] list_params = {"_add_reasoning_prompt": True} - list_result = mistral_config._add_reasoning_system_prompt_if_needed( - list_messages, list_params - ) + list_result = mistral_config._add_reasoning_system_prompt_if_needed(list_messages, list_params) assert isinstance(list_result[0]["content"], list) assert list_result[0]["content"][0]["type"] == "text" assert "" in list_result[0]["content"][0]["text"] @@ -319,9 +336,7 @@ class TestMistralReasoningSupport: messages = [{"role": "user", "content": "What is 2+2?"}] optional_params = {} - result = mistral_config._add_reasoning_system_prompt_if_needed( - messages, optional_params - ) + result = mistral_config._add_reasoning_system_prompt_if_needed(messages, optional_params) # Should return messages unchanged assert result == messages @@ -444,9 +459,7 @@ class TestMistralReasoningSupport: def test_mistral_streaming_chunk_preserves_thinking_blocks(): """Ensure streaming chunks keep magistral reasoning content.""" - iterator = MistralChatResponseIterator( - streaming_response=iter([]), sync_stream=True, json_mode=False - ) + iterator = MistralChatResponseIterator(streaming_response=iter([]), sync_stream=True, json_mode=False) streamed_chunk = { "id": "chunk-1", @@ -523,9 +536,7 @@ class TestMistralParallelToolCalls: def test_get_supported_openai_params_includes_parallel_tool_calls(self): """Test that parallel_tool_calls is in supported parameters.""" mistral_config = MistralConfig() - supported_params = mistral_config.get_supported_openai_params( - "mistral/mistral-large-latest" - ) + supported_params = mistral_config.get_supported_openai_params("mistral/mistral-large-latest") assert "parallel_tool_calls" in supported_params def test_transform_request_preserves_parallel_tool_calls(self): @@ -691,9 +702,7 @@ class TestMistralEmptyContentHandling: result = MistralConfig._handle_empty_content_response(response_data) - assert ( - result["choices"][0]["message"]["content"] == "Hello, how can I help you?" - ) + assert result["choices"][0]["message"]["content"] == "Hello, how can I help you?" def test_handle_empty_content_response_handles_multiple_choices(self): """Test that only the first choice is processed for empty content.""" @@ -1431,9 +1440,7 @@ def test_mistral_transform_request_keeps_an_empty_tool_message_after_its_tool_ca { "role": "assistant", "content": "", - "tool_calls": [ - {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}} - ], + "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}}], }, {"role": "tool", "tool_call_id": "call_1", "content": ""}, ], @@ -1457,9 +1464,7 @@ def test_mistral_transform_request_sends_a_null_tool_reply_as_empty_text(reply): { "role": "assistant", "content": "", - "tool_calls": [ - {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}} - ], + "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}}], }, {"role": "tool", "tool_call_id": "call_1", **reply}, ], @@ -1514,9 +1519,7 @@ def test_mistral_transform_request_hoists_tool_message_image(): { "role": "assistant", "content": "", - "tool_calls": [ - {"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}} - ], + "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}}], }, { "role": "tool", @@ -1665,3 +1668,274 @@ def test_mistral_transform_request_drops_an_empty_tool_message_whose_call_id_is_ ) assert [m.get("role") for m in request["messages"]] == ["user", "assistant"] + + +@pytest.mark.parametrize( + ("model", "custom_llm_provider"), + [ + ("mistral-large-latest", "mistral"), + ("mistral-large-4", "mistral"), + ("a-mistral-model-missing-from-the-cost-map", "mistral"), + ("codestral-latest", "codestral"), + ], +) +def test_prompt_cache_key_reaches_the_mistral_request_body( + local_model_cost_map: None, model: str, custom_llm_provider: str +) -> None: + import litellm + + optional_params: Final = litellm.get_optional_params( + model=model, custom_llm_provider=custom_llm_provider, prompt_cache_key="session-1" + ) + request: Final = MistralConfig().transform_request( + model=model, + messages=[{"role": "user", "content": "hi"}], + optional_params=optional_params, + litellm_params={}, + headers={}, + ) + + assert request["prompt_cache_key"] == "session-1" + + +AZURE_HOSTED_MISTRAL_API_BASE: Final = "https://my-resource.services.ai.azure.com/openai/v1" + + +def _transform_with_prompt_cache_key( + litellm_params: dict[str, object], prompt_cache_key: object = "session-1" +) -> Mapping[str, object]: + return cast( + Mapping[str, object], + MistralConfig().transform_request( + model="mistral-small-latest", + messages=[{"role": "user", "content": "hi"}], + optional_params={"prompt_cache_key": prompt_cache_key, "temperature": 0.2}, + litellm_params=litellm_params, + headers={}, + ), + ) + + +@pytest.mark.parametrize("api_base", ["https://api.mistral.ai/v1", "https://codestral.mistral.ai/v1"]) +def test_prompt_cache_key_reaches_a_mistral_hosted_endpoint(api_base: str) -> None: + request: Final = _transform_with_prompt_cache_key({"api_base": api_base}) + + assert request["prompt_cache_key"] == "session-1" + + +@pytest.mark.parametrize("api_base", [AZURE_HOSTED_MISTRAL_API_BASE, "https://mistral.ai.example.com/v1"]) +def test_prompt_cache_key_for_another_host_is_rejected_without_drop_params(api_base: str) -> None: + import litellm + + with pytest.raises(litellm.UnsupportedParamsError, match="prompt_cache_key") as rejection: + _transform_with_prompt_cache_key({"api_base": api_base}) + + upstream_host: Final = urlparse(api_base).hostname or api_base + assert upstream_host not in str(rejection.value) + + +UNSENDABLE_PROMPT_CACHE_KEYS: Final = ("Büro-Sitzung-42", "日本", 123, ["session-1"]) + + +@pytest.mark.parametrize("prompt_cache_key", UNSENDABLE_PROMPT_CACHE_KEYS) +def test_prompt_cache_key_the_mistral_api_cannot_take_is_rejected_without_drop_params( + prompt_cache_key: object, +) -> None: + import litellm + + with pytest.raises(litellm.UnsupportedParamsError, match="ASCII"): + _transform_with_prompt_cache_key({"api_base": "https://api.mistral.ai/v1"}, prompt_cache_key) + + +@pytest.mark.parametrize("prompt_cache_key", UNSENDABLE_PROMPT_CACHE_KEYS) +def test_prompt_cache_key_the_mistral_api_cannot_take_is_dropped_under_drop_params(prompt_cache_key: object) -> None: + request: Final = _transform_with_prompt_cache_key( + {"api_base": "https://api.mistral.ai/v1", "drop_params": True}, prompt_cache_key + ) + + assert "prompt_cache_key" not in request + assert request["temperature"] == 0.2 + + +def test_prompt_cache_key_with_ascii_punctuation_reaches_the_mistral_api() -> None: + request: Final = _transform_with_prompt_cache_key({}, "team a/session:42") + + assert request["prompt_cache_key"] == "team a/session:42" + + +def test_prompt_cache_key_allowlisted_for_another_purpose_is_still_rejected_for_another_host() -> None: + import litellm + + with pytest.raises(litellm.UnsupportedParamsError, match="allowed_openai_params"): + _transform_with_prompt_cache_key( + {"api_base": AZURE_HOSTED_MISTRAL_API_BASE, "allowed_openai_params": ["reasoning_effort"]} + ) + + +def _chat_completion_reply() -> httpx.Response: + return httpx.Response( + 200, + json={ + "id": "chatcmpl-1", + "object": "chat.completion", + "created": 0, + "model": "mistral-small-latest", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "hi"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + }, + ) + + +def test_allowlisted_prompt_cache_key_reaches_a_gateway_in_front_of_mistral(respx_mock: respx.MockRouter) -> None: + import litellm + + gateway_route: Final = respx_mock.post("https://gateway.example.com/mistral/v1/chat/completions").mock( + return_value=_chat_completion_reply() + ) + + litellm.completion( + model="mistral/mistral-small-latest", + messages=[{"role": "user", "content": "hi"}], + api_base="https://gateway.example.com/mistral/v1", + api_key="gateway-key", + prompt_cache_key="session-1", + allowed_openai_params=["prompt_cache_key"], + max_retries=0, + ) + + assert json.loads(gateway_route.calls.last.request.content)["prompt_cache_key"] == "session-1" + + +@pytest.mark.asyncio +async def test_prompt_cache_key_is_dropped_from_an_async_codestral_call_to_another_host( + respx_mock: respx.MockRouter, +) -> None: + import litellm + + azure_route: Final = respx_mock.post(f"{AZURE_HOSTED_MISTRAL_API_BASE}/chat/completions").mock( + return_value=_chat_completion_reply() + ) + + await litellm.acompletion( + model="codestral/codestral-latest", + messages=[{"role": "user", "content": "hi"}], + api_base=AZURE_HOSTED_MISTRAL_API_BASE, + api_key="azure-key", + prompt_cache_key="session-1", + drop_params=True, + max_retries=0, + client=AsyncOpenAI( + api_key="azure-key", base_url=AZURE_HOSTED_MISTRAL_API_BASE, http_client=httpx.AsyncClient() + ), + ) + + assert "prompt_cache_key" not in json.loads(azure_route.calls.last.request.content) + + +@pytest.mark.asyncio +async def test_prompt_cache_key_rejection_for_an_async_codestral_call_names_codestral( + respx_mock: respx.MockRouter, +) -> None: + import litellm + + azure_route: Final = respx_mock.post(f"{AZURE_HOSTED_MISTRAL_API_BASE}/chat/completions").mock( + return_value=_chat_completion_reply() + ) + + with pytest.raises(litellm.BadRequestError, match=r"codestral does not support parameters: \['prompt_cache_key'\]"): + await litellm.acompletion( + model="codestral/codestral-latest", + messages=[{"role": "user", "content": "hi"}], + api_base=AZURE_HOSTED_MISTRAL_API_BASE, + api_key="azure-key", + prompt_cache_key="session-1", + max_retries=0, + client=AsyncOpenAI( + api_key="azure-key", base_url=AZURE_HOSTED_MISTRAL_API_BASE, http_client=httpx.AsyncClient() + ), + ) + + assert not azure_route.called + + +@pytest.mark.asyncio +async def test_async_codestral_call_sends_an_uploaded_file_by_its_id(respx_mock: respx.MockRouter) -> None: + import litellm + + codestral_api_base: Final = "https://codestral.mistral.ai/v1" + codestral_route: Final = respx_mock.post(f"{codestral_api_base}/chat/completions").mock( + return_value=_chat_completion_reply() + ) + + await litellm.acompletion( + model="codestral/codestral-latest", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "Review this file"}, + {"type": "file", "file": {"file_id": "file-12345"}}, + ], + } + ], + api_base=codestral_api_base, + api_key="codestral-key", + max_retries=0, + client=AsyncOpenAI(api_key="codestral-key", base_url=codestral_api_base, http_client=httpx.AsyncClient()), + ) + + sent_content: Final = json.loads(codestral_route.calls.last.request.content)["messages"][0]["content"] + assert sent_content[1] == {"type": "file", "file_id": "file-12345"} + + +def test_prompt_cache_key_rejection_for_a_sync_codestral_call_is_an_unsupported_params_error( + respx_mock: respx.MockRouter, +) -> None: + import litellm + + azure_route: Final = respx_mock.post(f"{AZURE_HOSTED_MISTRAL_API_BASE}/chat/completions").mock( + return_value=_chat_completion_reply() + ) + + with pytest.raises( + litellm.UnsupportedParamsError, match=r"^litellm.UnsupportedParamsError: codestral does not support" + ): + litellm.completion( + model="codestral/codestral-latest", + messages=[{"role": "user", "content": "hi"}], + api_base=AZURE_HOSTED_MISTRAL_API_BASE, + api_key="azure-key", + prompt_cache_key="session-1", + max_retries=0, + ) + + assert not azure_route.called + + +def test_prompt_cache_key_for_another_host_is_dropped_under_drop_params() -> None: + request: Final = _transform_with_prompt_cache_key({"api_base": AZURE_HOSTED_MISTRAL_API_BASE, "drop_params": True}) + + assert "prompt_cache_key" not in request + assert request["temperature"] == 0.2 + + +def test_prompt_cache_key_for_the_azure_endpoint_set_in_the_environment_is_dropped_under_drop_params( + monkeypatch: pytest.MonkeyPatch, +) -> None: + monkeypatch.setenv("MISTRAL_AZURE_API_BASE", "https://my-resource.services.ai.azure.com/openai") + + request: Final = _transform_with_prompt_cache_key({"drop_params": True}) + + assert "prompt_cache_key" not in request + + +def test_prompt_cache_key_stays_unsupported_for_vertex_hosted_mistral(local_model_cost_map: None) -> None: + import litellm + + with pytest.raises(litellm.UnsupportedParamsError): + litellm.get_optional_params(model="mistral-medium-3", custom_llm_provider="vertex_ai", prompt_cache_key="s") + + dropped: Final = litellm.get_optional_params( + model="mistral-medium-3", custom_llm_provider="vertex_ai", prompt_cache_key="s", drop_params=True + ) + assert "prompt_cache_key" not in dropped From 4c8b90945faf69eedfbdda3d4b7dad58ce74860d Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 30 Oct 2026 01:31:39 -0700 Subject: [PATCH 14/23] fix(realtime)!: send the Gemini Live setup at the first content so session.update settings apply (#39864) * fix(realtime): send Gemini Live setup on the client's first message so session.update modalities apply Gemini and Vertex AI Live accept exactly one setup per session, and the proxy sent its default audio setup at connect, so a client's session.update (GA output_modalities or beta modalities) was dropped and every answer came back as audio. The setup now goes out on the client's first message: a session.update becomes that setup, anything else gets the model default. Bedrock Nova Sonic keeps its connect-time setup through a provider hook. Text setups label deltas response.output_text.delta, and downgrading TEXT to AUDIO on an audio-only model logs a warning * fix(realtime): keep one Gemini Live response modality and skip the "inf" token limit in the setup * fix(realtime): merge every session.update sent before the first content frame into the Gemini Live setup Gemini and Vertex Live accept one setup per connection, so the bridge used to turn only the first session.update into that setup and drop the rest without telling the client. Now every session.update that arrives before the first content frame is held, deep-merged (later values win per field), acknowledged with a session.updated built from the setup it would produce, and the merged session becomes the one setup sent with the first conversation item, audio frame, or tool result. The session events also surface the systemInstruction text as instructions, which the old string check never matched because the setup carries it as a Content object. * fix(realtime): cap the session.update merge depth so the recursive detector accepts it _merge_session_fields recurses once per nested mapping level, so CI's recursive_detector flagged it. Bound the descent at eight levels (the OpenAI session object nests four deep at most), replace anything deeper wholesale with the later session.update's value, and list the helper with the other depth-capped functions the detector allows. * fix(gemini-live): accept a string output modality and drop values Live cannot produce A bare string output_modalities is read as a one-element list and modality values Gemini Live cannot produce are dropped with a warning instead of reaching the backend, so a malformed session.update no longer costs the session. Transcription guardrails now fold their VAD disable into the first held session.update only, so the merged setup carries it exactly once. * fix(gemini realtime): answer an error for session updates Gemini Live cannot apply An output modality the model cannot produce, or a session.update sent after the first content, now gets an invalid_request_error event instead of a silent swap or drop, unless drop_params is set. The synthetic session.created reports the model's default modality, setupComplete no longer adds an extra session.updated, and merging nested session updates is linear in their size * fix(gemini realtime): send the setup before guardrail blocks and idle deadlines, mark push-to-talk turns A pre_call guardrail blocking the first typed message now goes out after the setup, so Gemini and Vertex no longer close the session with 1007. Push-to-talk sessions (turn_detection null or create_response false) wrap each audio turn in empty activityStart and activityEnd messages, which Gemini needs when automatic activity detection is off. A session with no content after 60 seconds sends its setup, since Vertex refuses a setup 300 seconds after connect. Held session.update frames past 10 MB answer session_update_too_large, and deferred mode now errors on a second session.update unless drop_params is set. * fix(realtime): ack unchanged Gemini Live session.update resends and honor null turn_detection An update that leaves the applied setup unchanged now gets session.updated instead of session_update_not_applied, so clients that resend their session with the first response keep working. Deferred setup mode merges held updates like the default mode, turn_detection null turns Gemini's activity detection off, drop_params drops a voice outside the documented Live voices, and the lazy setup deadline moves to 240 seconds * fix(realtime): keep Vertex activity detection off for null turn_detection and set up deferred sessions before text content * fix(realtime): refuse unscreenable Gemini audio, send held setup after buffered audio, reject unknown Vertex voices A realtime_input_transcription guardrail cannot screen Gemini Live audio before the answer starts, so audio input under one is refused with an audio_input_not_screenable error instead of reaching the model. In deferred setup mode a session.update that follows buffered audio is sent as setup at once. Only conversation content triggers the setup, so control events no longer lock in a default one. Vertex rejects an undocumented voice on the session.update, or drops it with drop_params, while Google AI Studio forwards any name. * fix(realtime): hold first content so a following session.update joins the Gemini setup Convert flat realtime tool schemas through Gemini's schema builder, and collapse a text and audio pair to what the model can produce on text-only models too * fix(realtime): check AI Studio Live voices like Vertex * fix(realtime): forward any voice to AI Studio Live models that accept it * fix(realtime): map OpenAI TTS voices before the Live voice check, keep outage spend rows, share the playground session config The Live voice gate now judges the Gemini voice an OpenAI TTS name maps to, so fable and onyx reach Vertex as Umbriel and Orus again. An upstream close before the held setup was sent is logged as a session, not a refusal. The playground sends one server_vad session config for typed and voice turns, and the e2e, integration, and dashboard tests follow lazy setup. * fix(realtime): take the lazy setup clock as a dependency so its tests run on a stepped clock The deadline and grace tests waited on real 50 ms timers, and the merge cost check compared wall-clock runs. The session now reads an injected clock, time.monotonic by default, the two tests step it, and the merge cost check moves to the CodSpeed benchmarks. The Nova Sonic style handler test passes the backend through wrap_backend, which main added after the test was written --- litellm/constants.py | 2 + .../litellm_core_utils/realtime_streaming.py | 330 ++++++++- .../llms/base_llm/realtime/transformation.py | 32 +- litellm/llms/bedrock/realtime/handler.py | 20 +- .../llms/bedrock/realtime/transformation.py | 19 +- litellm/llms/custom_httpx/llm_http_handler.py | 11 +- .../llms/gemini/realtime/transformation.py | 317 +++++++-- .../llms/vertex_ai/realtime/transformation.py | 17 +- litellm/types/llms/gemini.py | 16 +- litellm/utils.py | 2 +- .../test_realtime_session_merge_benchmarks.py | 21 + .../code_coverage_tests/recursive_detector.py | 1 + .../realtime/realtime_client.py | 3 + .../realtime/test_realtime_e2e.py | 4 +- ..._transcription_guardrail_session_update.py | 24 +- .../test_realtime_streaming.py | 670 +++++++++++++++++- .../custom_httpx/test_llm_http_handler.py | 189 +++++ .../test_gemini_realtime_transformation.py | 310 +++++++- .../test_vertex_ai_realtime_transformation.py | 110 +++ .../chat_ui/RealtimePlayground.test.tsx | 18 + .../components/chat_ui/RealtimePlayground.tsx | 65 +- 21 files changed, 1993 insertions(+), 188 deletions(-) create mode 100644 tests/benchmarks/test_realtime_session_merge_benchmarks.py diff --git a/litellm/constants.py b/litellm/constants.py index e7c9c5dcbfb..8e80c07b4ed 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -420,6 +420,8 @@ REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS: Final = get_env_float( "REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS", 20.0 ) BEDROCK_REALTIME_CLOSE_TIMEOUT_SECONDS: Final = get_env_float("BEDROCK_REALTIME_CLOSE_TIMEOUT_SECONDS", 5.0) +REALTIME_LAZY_SETUP_DEADLINE_SECONDS: Final = 240.0 +REALTIME_LAZY_SETUP_CONTENT_GRACE_SECONDS: Final = 1.0 # RFC 6455 caps the close frame payload at 125 bytes, 2 of which carry the status code WEBSOCKET_CLOSE_REASON_MAX_BYTES: Final = 123 diff --git a/litellm/litellm_core_utils/realtime_streaming.py b/litellm/litellm_core_utils/realtime_streaming.py index 484bacdc72c..4db22e9d5bf 100644 --- a/litellm/litellm_core_utils/realtime_streaming.py +++ b/litellm/litellm_core_utils/realtime_streaming.py @@ -1,5 +1,6 @@ import asyncio import json +import time import traceback from collections.abc import Callable, Coroutine, Mapping, MutableMapping, Sequence from dataclasses import dataclass @@ -15,13 +16,19 @@ from litellm._logging import redact_internal_details_from_client_message, verbos from litellm._uuid import uuid from litellm.constants import ( BEDROCK_REALTIME_PENDING_OPENING_MESSAGE_SCOPE_KEY, + REALTIME_LAZY_SETUP_CONTENT_GRACE_SECONDS, + REALTIME_LAZY_SETUP_DEADLINE_SECONDS, REALTIME_METERED_USAGE_KEY, REALTIME_SESSION_FAILURE_LOGGED_KEY, REALTIME_SESSION_SUCCESS_LOGGED_KEY, ) from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from litellm.llms.base_llm.realtime.transcription_protocol import RealtimeTranscriptionProtocolError -from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig, RealtimeBackend +from litellm.llms.base_llm.realtime.transformation import ( + BaseRealtimeConfig, + RealtimeBackend, + RealtimeSessionUpdateRejection, +) from litellm.types.llms.openai import ( OpenAIRealtimeEvents, OpenAIRealtimeOutputItemDone, @@ -125,6 +132,65 @@ class _LoggingWorker(Protocol): _DECODED_JSON: Final = TypeAdapter(object) +_SESSION_FIELDS: Final = TypeAdapter(dict[str, object]) +_NO_SESSION_FIELDS: Final[Mapping[str, object]] = MappingProxyType({}) +_SESSION_MERGE_MAX_DEPTH: Final = 8 + + +class _SessionUpdateFrame(TypedDict): + type: ReadOnly[str] + session: ReadOnly[Mapping[str, object]] + + +def _merge_session_fields( + base: Mapping[str, object], update: Mapping[str, object], depth: int = 0 +) -> Mapping[str, object]: + def merged_value(key: str) -> object: + if key not in update: + return base[key] + value: Final = update[key] + base_value: Final = base.get(key) + if depth < _SESSION_MERGE_MAX_DEPTH and isinstance(value, Mapping) and isinstance(base_value, Mapping): + return _merge_session_fields( + _SESSION_FIELDS.validate_python(base_value), _SESSION_FIELDS.validate_python(value), depth + 1 + ) + return update[key] + + return MappingProxyType({key: merged_value(key) for key in dict.fromkeys((*base, *update))}) + + +def _session_update_frame(session: Mapping[str, object]) -> str: + frame: Final[_SessionUpdateFrame] = {"type": "session.update", "session": session} + return json.dumps(frame, default=_SESSION_FIELDS.validate_python) + + +_HELD_SESSION_UPDATE_TOO_LARGE: Final = RealtimeSessionUpdateRejection( + code="session_update_too_large", + message=( + "The session.update frames sent before the first conversation content add up to more than 10 MB, " + "so this session.update was not applied" + ), + param="session", +) + + +def _unscreened_audio_error_event(client_event_id: object) -> Mapping[str, object]: + return { + "type": "error", + "event_id": f"event_{uuid.uuid4()}", + "error": { + "type": "invalid_request_error", + "code": "audio_input_not_screenable", + "message": ( + "A realtime_input_transcription guardrail is on, and Gemini Live only transcribes a turn after it " + "starts answering it, so audio input was not sent. Send text with conversation.item.create instead" + ), + "param": None, + "event_id": client_event_id if isinstance(client_event_id, str) else None, + }, + } + + def _decode_json_object(payload: str) -> Mapping[str, object]: """Decode a realtime frame into its top-level field mapping.""" return json.loads(payload) @@ -243,8 +309,17 @@ class RealTimeStreaming: event_normalizer: RealtimeEventNormalizer | None = None, usage_meter: RealtimeUsageMeter | None = None, logging_worker: _LoggingWorker = GLOBAL_LOGGING_WORKER, + lazy_setup_deadline_seconds: float = REALTIME_LAZY_SETUP_DEADLINE_SECONDS, + lazy_setup_content_grace_seconds: float = REALTIME_LAZY_SETUP_CONTENT_GRACE_SECONDS, + clock: Callable[[], float] = time.monotonic, ): self.websocket: _ClientWebSocket = websocket + self._clock: Final = clock + self._lazy_setup_deadline: Final = clock() + lazy_setup_deadline_seconds + self._lazy_setup_content_grace_seconds: Final = lazy_setup_content_grace_seconds + self._held_client_content: tuple[str, ...] = () + self._held_client_content_bytes: int = 0 + self._held_client_content_deadline: float | None = None self._replay_scope: Final = client_websocket_scope(websocket) self.backend_ws = backend_ws self.logging_obj = logging_obj @@ -300,6 +375,10 @@ class RealTimeStreaming: # Gemini Live rejects a follow-up BidiGenerateContentSetup once any # content (realtimeInput / clientContent / toolResponse) has been sent. self._content_sent_after_setup: bool = False + self._held_session_update: Mapping[str, object] | None = None + self._applied_session_fields: Mapping[str, object] | None = None + self._held_session_update_bytes: int = 0 + self._unscreened_audio_refused: bool = False # Whether this is a transcription-only session (session.type == "transcription", # e.g. gpt-realtime-whisper). Such sessions must not be sent response.create and # their input_audio_transcription.completed usage drives duration-based cost. @@ -337,6 +416,17 @@ class RealTimeStreaming: ] ) _CLIENT_AUDIO_BUFFER_COMMIT_TYPES = frozenset(["input_audio_buffer.commit", "input_audio_buffer.end"]) + _CLIENT_CONTENT_TYPES = frozenset( + [ + "conversation.item.create", + "response.create", + "input_audio_buffer.append", + "input_audio_buffer.commit", + "input_audio_buffer.end", + ] + ) + _HELD_CONTENT_TYPES: Final = frozenset(["conversation.item.create", "input_audio_buffer.append"]) + _DEFERRED_SETUP_HELD_CONTENT_TYPES: Final = frozenset(["conversation.item.create"]) _AUDIO_FORMAT_MAP: dict[str, dict[str, str | int]] = { "pcm16": {"type": "audio/pcm", "rate": 24000}, "g711_ulaw": {"type": "audio/G711-ulaw", "rate": 8000}, @@ -377,7 +467,7 @@ class RealTimeStreaming: message_obj: dict[str, Any] = cast(dict[str, Any], message) else: message_obj = cast(dict[str, Any], json.loads(message)) - self._collect_tool_calls_from_response_done(cast(dict, message_obj)) + self._collect_tool_calls_from_response_done(message_obj) if not self._should_store_message(message_obj): return try: @@ -502,7 +592,7 @@ class RealTimeStreaming: self.store_message(flush_event) self._capture_transcription_usage(flush_event) - def _collect_tool_calls_from_response_done(self, event_obj: dict | OpenAIRealtimeEvents) -> None: + def _collect_tool_calls_from_response_done(self, event_obj: Mapping[str, object] | OpenAIRealtimeEvents) -> None: """Extract function_call items from response.done events for spend logging.""" try: if event_obj.get("type") != "response.done": @@ -607,6 +697,167 @@ class RealTimeStreaming: self._usage_meter.observe_client_event(sent_event) return True + def _builds_setup_lazily(self) -> bool: + if self.provider_config is None: + return False + return self.provider_config.builds_setup_from_session_update() + + def _holds_session_updates_until_first_content(self) -> bool: + return self.session_configuration_request is None and self._builds_setup_lazily() + + async def _rejected_setup_session_update(self, message: str) -> bool: + provider_config: Final = self.provider_config + if provider_config is None or not provider_config.builds_setup_from_session_update(): + return False + client_event: Final = _decode_json_object(message) + session: Final = client_event.get("session") + update: Final = _SESSION_FIELDS.validate_python(session) if isinstance(session, Mapping) else _NO_SESSION_FIELDS + if self._leaves_applied_setup_unchanged(update): + await self._acknowledge_session_setup(self._applied_session_fields or _NO_SESSION_FIELDS) + return True + rejection: Final = provider_config.setup_session_update_rejection( + self.model, update, setup_sent=self.session_configuration_request is not None + ) + if rejection is None: + return False + await self.websocket.send_text(json.dumps(rejection.error_event(client_event.get("event_id")))) + return True + + def _held_session_update_message(self) -> str: + return _session_update_frame(self._held_session_update or _NO_SESSION_FIELDS) + + async def _hold_session_update(self, message: str) -> None: + client_event: Final = _decode_json_object(message) + if self._held_session_update_bytes + len(message.encode("utf-8")) > RealTimeStreaming._MAX_BUFFERED_BYTES: + await self.websocket.send_text( + json.dumps(_HELD_SESSION_UPDATE_TOO_LARGE.error_event(client_event.get("event_id"))) + ) + return + session: Final = client_event.get("session") + update: Final = _SESSION_FIELDS.validate_python(session) if isinstance(session, Mapping) else _NO_SESSION_FIELDS + held: Final = _merge_session_fields(self._held_session_update or _NO_SESSION_FIELDS, update) + self._held_session_update = held + self._held_session_update_bytes = len(_session_update_frame(held).encode("utf-8")) + await self._acknowledge_session_setup(held) + + def _setup_preview(self, session: Mapping[str, object]) -> str | None: + if self.provider_config is None: + return None + preview: Final = self.provider_config.transform_realtime_request( + _session_update_frame(session), self.model, None + ) + return next( + (frame for frame in preview if isinstance(frame, str) and "setup" in _decode_json_object(frame)), None + ) + + def _leaves_applied_setup_unchanged(self, update: Mapping[str, object]) -> bool: + applied: Final = self._applied_session_fields + if self.session_configuration_request is None or applied is None: + return False + applied_setup: Final = self._setup_preview(applied) + merged_setup: Final = self._setup_preview(_merge_session_fields(applied, update)) + return ( + applied_setup is not None + and merged_setup is not None + and _decode_json_object(applied_setup) == _decode_json_object(merged_setup) + ) + + async def _acknowledge_session_setup(self, session: Mapping[str, object]) -> None: + if self.provider_config is None: + return + setup: Final = self._setup_preview(session) + if setup is None: + return + acknowledged: Final = self.provider_config.transform_session_created_event( + self.model, str(self.logging_obj.litellm_trace_id), setup + ) + if acknowledged is None: + return + event: Final[Mapping[str, object]] = {**acknowledged, "type": "session.updated"} + event_str: Final = json.dumps(event) + self.store_message(event_str) + await self._send_event_to_client(event, event_str) + + async def _send_held_session_update_as_setup(self) -> None: + held: Final = self._held_session_update + if held is None: + return + message: Final = self._held_session_update_message() + self._held_session_update = None + self._held_session_update_bytes = 0 + if await self._send_to_backend(message): + self._applied_session_fields = held + await self._send_held_client_content() + + def _holds_client_content(self, msg_type: str | None, message: str) -> bool: + held_types: Final = ( + RealTimeStreaming._DEFERRED_SETUP_HELD_CONTENT_TYPES + if self._uses_deferred_backend_setup() + else RealTimeStreaming._HELD_CONTENT_TYPES + ) + return ( + msg_type in held_types + and self._held_session_update is not None + and self._holds_session_updates_until_first_content() + and not self._pending_messages_until_setup + and len(self._held_client_content) < RealTimeStreaming._MAX_BUFFERED_MESSAGES + and self._held_client_content_bytes + len(message.encode("utf-8")) <= RealTimeStreaming._MAX_BUFFERED_BYTES + ) + + def _hold_client_content(self, message: str) -> None: + if self._held_client_content_deadline is None: + self._held_client_content_deadline = self._clock() + self._lazy_setup_content_grace_seconds + self._held_client_content = (*self._held_client_content, message) + self._held_client_content_bytes += len(message.encode("utf-8")) + + async def _send_held_client_content(self) -> None: + held: Final = self._held_client_content + self._held_client_content = () + self._held_client_content_bytes = 0 + self._held_client_content_deadline = None + for message in held: + await self._send_to_backend(message) + + def _setup_deadline(self) -> float: + content_deadline: Final = self._held_client_content_deadline + return ( + self._lazy_setup_deadline if content_deadline is None else min(self._lazy_setup_deadline, content_deadline) + ) + + async def _send_setup_before_content(self, msg_type: str | None) -> None: + if self._holds_session_updates_until_first_content(): + await self._send_held_session_update_as_setup() + await self._send_default_setup_before_first_client_message(msg_type) + + async def _receive_client_message_before_setup_deadline(self) -> str | None: + if not self._holds_session_updates_until_first_content(): + return await self._receive_client_message() + receive: Final = asyncio.ensure_future(self._receive_client_message()) + try: + done, _ = await asyncio.wait((receive,), timeout=self._setup_deadline() - self._clock()) + if not done: + await self._send_setup_before_content(None) + return await receive + finally: + receive.cancel() + + async def _send_default_setup_before_first_client_message(self, msg_type: str | None) -> None: + if self.session_configuration_request is not None or self.provider_config is None: + return + if msg_type == "session.update" or ( + self._uses_deferred_backend_setup() + and (msg_type in RealTimeStreaming._CLIENT_AUDIO_BUFFER_TYPES or self._pending_messages_until_setup) + ): + return + default_setup: Final = self.provider_config.session_configuration_request(self.model) + if not default_setup: + return + setup: Final = self.maybe_inject_guardrail_auto_response_disable(default_setup) + await self.backend_ws.send(setup) + self._cache_session_configuration_request(setup) + self._applied_session_fields = _NO_SESSION_FIELDS + await self._send_held_client_content() + def _enforce_transcription_session_model(self, message: str) -> str: """Force client transcription session updates to the authorized model. @@ -851,6 +1102,20 @@ class RealTimeStreaming: self._guardrail_cancel_event_ids = self._guardrail_cancel_event_ids | {event_id} await self._send_to_backend(json.dumps({"type": "response.cancel", "event_id": event_id})) + async def _send_blocked_turn_to_backend(self, pre_block_backend_message: str | None, item_id: str | None) -> None: + if pre_block_backend_message is None and self._is_transcription_session: + return + await self._send_setup_before_content("conversation.item.create") + # Deliver any caller-supplied backend message FIRST so that + # protocol contracts requiring a specific ordering (e.g. + # Gemini Live's mandatory ``toolResponse`` after a + # ``toolCall``) are honored before the guardrail's own + # clientContent / cancel messages are sent. + if pre_block_backend_message is not None: + await self._send_to_backend(pre_block_backend_message) + if not self._is_transcription_session: + await self._withdraw_blocked_turn(item_id) + async def _withdraw_blocked_turn(self, item_id: str | None) -> None: await self._send_guardrail_cancel() if item_id is not None and self._deletes_conversation_items: @@ -958,6 +1223,21 @@ class RealTimeStreaming: """Transcription-only sessions have no assistant turn to gate and reject realtime session updates.""" return not self._is_transcription_session and self._has_audio_transcription_guardrails() + def _refuses_unscreened_audio(self, msg_type: str | None) -> bool: + return ( + msg_type in RealTimeStreaming._CLIENT_AUDIO_BUFFER_TYPES + and self._builds_setup_lazily() + and self._should_disable_vad_auto_response() + ) + + async def _refuse_unscreened_audio(self, msg_type: str | None, message: str) -> None: + if self._unscreened_audio_refused and msg_type not in RealTimeStreaming._CLIENT_AUDIO_BUFFER_COMMIT_TYPES: + return + self._unscreened_audio_refused = True + await self.websocket.send_text( + json.dumps(_unscreened_audio_error_event(_decode_json_object(message).get("event_id"))) + ) + async def run_realtime_guardrails( self, transcript: str, @@ -1030,15 +1310,7 @@ class RealTimeStreaming: # Use realtime_violation_message if configured; fall back to guardrail error text. error_msg = getattr(callback, "realtime_violation_message", None) or safe_msg - # Deliver any caller-supplied backend message FIRST so that - # protocol contracts requiring a specific ordering (e.g. - # Gemini Live's mandatory ``toolResponse`` after a - # ``toolCall``) are honored before the guardrail's own - # clientContent / cancel messages are sent. - if pre_block_backend_message is not None: - await self._send_to_backend(pre_block_backend_message) - if not self._is_transcription_session: - await self._withdraw_blocked_turn(item_id) + await self._send_blocked_turn_to_backend(pre_block_backend_message, item_id) # Send the policy violation hint (shows as small gray status text in UI). await self.websocket.send_text( json.dumps( @@ -1136,6 +1408,9 @@ class RealTimeStreaming: break finally: self._flushing_pending_messages_until_setup = False + if self.session_created_sent_to_client and self._builds_setup_lazily(): + await self._maybe_send_guardrail_turn_detection_update() + continue if self.session_created_sent_to_client: # A synthetic session.created (with placeholder defaults) was # already forwarded to the client when we connected. The @@ -1295,7 +1570,11 @@ class RealTimeStreaming: return BackendClose(code=1011, reason="proxy failed while relaying the upstream websocket") def _backend_refused_session(self, close: BackendClose) -> bool: - return close.code != 1000 and not self._backend_sent_frames + return ( + close.code != 1000 + and not self._backend_sent_frames + and not self._holds_session_updates_until_first_content() + ) async def _log_session_end(self, transport_refusal: Exception | None) -> None: provider_refusal: Final = None if self.provider_config is None else self.provider_config.session_refusal() @@ -1488,7 +1767,7 @@ class RealTimeStreaming: client_event: _ClientEventFrame try: while True: - message = await self._receive_client_message() + message = await self._receive_client_message_before_setup_deadline() if message is None: return ClientLoopExit.CLIENT_DISCONNECTED @@ -1600,6 +1879,7 @@ class RealTimeStreaming: if ( msg_type == "session.update" and self.session_configuration_request is None + and self._held_session_update is None and not self._guardrail_turn_detection_update_sent and self._should_disable_vad_auto_response() and ( @@ -1663,12 +1943,25 @@ class RealTimeStreaming: except (json.JSONDecodeError, AttributeError): pass + if self._refuses_unscreened_audio(msg_type): + await self._refuse_unscreened_audio(msg_type, message) + continue + ## LOGGING # Log after any in-place modifications (GA remap, guardrail # turn_detection injection) so audit logs reflect what we # actually forward to the backend. self.store_input(message=message) + if self._holds_client_content(msg_type, message): + self._hold_client_content(message) + continue + if self._held_client_content and msg_type != "session.update": + await self._send_setup_before_content(msg_type) + + if msg_type in RealTimeStreaming._CLIENT_CONTENT_TYPES and self._uses_deferred_backend_setup(): + await self._send_setup_before_content(msg_type) + if self._should_buffer_client_message_until_setup(message): self._buffer_pending_message_until_setup(message) continue @@ -1697,6 +1990,15 @@ class RealTimeStreaming: # ``_maybe_send_guardrail_turn_detection_update`` would retry. if msg_type == "session.update" and self.provider_config is None: await self._send_session_gate_once() + if msg_type == "session.update" and await self._rejected_setup_session_update(message): + continue + if msg_type == "session.update" and self._holds_session_updates_until_first_content(): + await self._hold_session_update(message) + if self._pending_messages_until_setup: + await self._send_held_session_update_as_setup() + continue + if msg_type in RealTimeStreaming._CLIENT_CONTENT_TYPES: + await self._send_setup_before_content(msg_type) try: sent = await self._send_to_backend(message) except RealtimeTranscriptionProtocolError as e: diff --git a/litellm/llms/base_llm/realtime/transformation.py b/litellm/llms/base_llm/realtime/transformation.py index f1539d16b30..ff714bb908b 100644 --- a/litellm/llms/base_llm/realtime/transformation.py +++ b/litellm/llms/base_llm/realtime/transformation.py @@ -1,10 +1,12 @@ +import dataclasses from abc import ABC, abstractmethod from collections.abc import Mapping, Sequence from types import TracebackType -from typing import TYPE_CHECKING, Any, Protocol +from typing import TYPE_CHECKING, Any, Literal, Protocol import httpx +from litellm._uuid import uuid from litellm.types.llms.openai import OpenAIRealtimeStreamSessionEvents from litellm.types.realtime import ( RealtimeInputAudioTranscriptionUsage, @@ -22,6 +24,26 @@ else: LiteLLMLoggingObj = Any +@dataclasses.dataclass(frozen=True, slots=True) +class RealtimeSessionUpdateRejection: + code: Literal["invalid_type", "invalid_value", "session_update_not_applied", "session_update_too_large"] + message: str + param: str + + def error_event(self, client_event_id: object) -> dict[str, object]: + return { + "type": "error", + "event_id": f"event_{uuid.uuid4()}", + "error": { + "type": "invalid_request_error", + "code": self.code, + "message": self.message, + "param": self.param, + "event_id": client_event_id if isinstance(client_event_id, str) else None, + }, + } + + class RealtimeBackend(Protocol): async def __aenter__(self) -> object: ... @@ -93,6 +115,14 @@ class BaseRealtimeConfig(ABC): def session_configuration_request(self, model: str) -> str | None: # message sent to setup the realtime session return None + def builds_setup_from_session_update(self) -> bool: + return False + + def setup_session_update_rejection( + self, model: str, session: Mapping[str, object], *, setup_sent: bool + ) -> RealtimeSessionUpdateRejection | None: + return None + def unbilled_usage_on_session_close(self, model: str) -> RealtimeInputAudioTranscriptionUsage | None: return None diff --git a/litellm/llms/bedrock/realtime/handler.py b/litellm/llms/bedrock/realtime/handler.py index e9b2f26e977..00622a1e287 100644 --- a/litellm/llms/bedrock/realtime/handler.py +++ b/litellm/llms/bedrock/realtime/handler.py @@ -17,7 +17,6 @@ from pydantic import JsonValue, TypeAdapter import litellm from litellm._logging import redact_string, verbose_proxy_logger -from litellm._uuid import uuid from litellm.constants import ( BEDROCK_REALTIME_CLOSE_TIMEOUT_SECONDS, BEDROCK_REALTIME_COMMITTED_FAILURE_SCOPE_KEY, @@ -31,13 +30,14 @@ from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from litellm.litellm_core_utils.realtime_streaming import DefaultLoggedRealTimeEventTypes +from litellm.llms.base_llm.realtime.transformation import RealtimeSessionUpdateRejection from litellm.types.llms.bedrock import AwsAuthParams from litellm.types.llms.openai import OpenAIRealtimeEvents from litellm.types.realtime import RealtimeResponseTransformInput from ..base_aws_llm import BaseAWSLLM, run_aws_signing from ..common_utils import BedrockError -from .transformation import BedrockRealtimeConfig, BedrockSessionUpdateRejection, requested_modalities +from .transformation import BedrockRealtimeConfig, requested_modalities _CLIENT_MESSAGE_ADAPTER: Final[TypeAdapter[JsonValue]] = TypeAdapter(JsonValue) @@ -256,23 +256,11 @@ async def _ack_session_update( async def _reject_session_update( client_ws: RealtimeClientWebSocket, parsed_client_message: Mapping[str, JsonValue], - rejection: BedrockSessionUpdateRejection, + rejection: RealtimeSessionUpdateRejection, ) -> bool: """Send an error event for a session.update Bedrock cannot apply; False means the client is gone.""" - client_event_id: Final = parsed_client_message.get("event_id") - error_event: Final = { - "type": "error", - "event_id": f"event_{uuid.uuid4()}", - "error": { - "type": "invalid_request_error", - "code": rejection.code, - "message": rejection.message, - "param": rejection.param, - "event_id": client_event_id if isinstance(client_event_id, str) else None, - }, - } try: - await client_ws.send_text(json.dumps(error_event)) + await client_ws.send_text(json.dumps(rejection.error_event(parsed_client_message.get("event_id")))) except Exception as e: # noqa: BLE001 # any send failure means the client is gone verbose_proxy_logger.debug("Client to Bedrock forwarding ended: %s", e, exc_info=True) return False diff --git a/litellm/llms/bedrock/realtime/transformation.py b/litellm/llms/bedrock/realtime/transformation.py index 952d8858f78..25498fef46f 100644 --- a/litellm/llms/bedrock/realtime/transformation.py +++ b/litellm/llms/bedrock/realtime/transformation.py @@ -11,7 +11,7 @@ import math import uuid as uuid_lib from collections.abc import Mapping, Sequence from types import MappingProxyType -from typing import Final, Literal +from typing import Final import httpx from pydantic import ConfigDict, JsonValue, TypeAdapter @@ -19,7 +19,7 @@ from pydantic import ConfigDict, JsonValue, TypeAdapter from litellm._logging import verbose_logger from litellm._uuid import uuid from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj -from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig +from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig, RealtimeSessionUpdateRejection from litellm.llms.bedrock.common_utils import BedrockError from litellm.llms.bedrock.realtime.trigger_audio import ready_trigger_pcm from litellm.types.llms.base import LiteLLMBaseModel @@ -196,13 +196,6 @@ def _malformed_settings(session_config: Mapping[str, JsonValue]) -> tuple[tuple[ return tuple((name, expected) for name, expected, is_malformed in malformed if is_malformed) -@dataclasses.dataclass(frozen=True, slots=True) -class BedrockSessionUpdateRejection: - code: Literal["invalid_type", "invalid_value", "session_update_not_applied"] - message: str - param: str - - @dataclasses.dataclass(frozen=True, slots=True) class BedrockLockedSessionSettings: """The session settings Bedrock Nova Sonic locks in once the prompt starts.""" @@ -533,7 +526,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig): max_tokens=max_tokens, ) - def session_update_rejection(self, json_message: Mapping[str, JsonValue]) -> BedrockSessionUpdateRejection | None: + def session_update_rejection(self, json_message: Mapping[str, JsonValue]) -> RealtimeSessionUpdateRejection | None: """The error for a session.update this session cannot apply, or None when it applies.""" session_config: Final = json_message.get("session") if not isinstance(session_config, Mapping): @@ -541,7 +534,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig): malformed: Final = () if self.drop_params else _malformed_settings(session_config) if malformed: name, expected = malformed[0] - return BedrockSessionUpdateRejection( + return RealtimeSessionUpdateRejection( code="invalid_type", message=( f"Invalid type for 'session.{name}': Bedrock Nova Sonic needs {expected} there. " @@ -551,7 +544,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig): ) unplayable_voice: Final = None if self.drop_params else _unplayable_voice(session_config) if unplayable_voice is not None: - return BedrockSessionUpdateRejection( + return RealtimeSessionUpdateRejection( code="invalid_value", message=( f"Invalid value for 'session.voice': '{unplayable_voice}' is not a Bedrock Nova Sonic voice. " @@ -562,7 +555,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig): if not self.prompt_started: return None if self._apply_locked_settings(session_config, self._started_settings) != self._started_settings: - return BedrockSessionUpdateRejection( + return RealtimeSessionUpdateRejection( code="session_update_not_applied", message="Bedrock Nova Sonic cannot change tools, instructions, voice, output audio format, temperature or max output tokens after the session has started, so this session.update was not applied", param="session", diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index c92315e53b3..61fc23589ac 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -6973,14 +6973,19 @@ class BaseLLMHTTPHandler: ), ) - # Auto-send session setup if the provider requires it (e.g. - # Gemini/Vertex AI Live needs a `setup` before any realtime_input). + # Auto-send session setup at connect unless the provider builds it from the + # client's first session.update (Gemini/Vertex AI Live accept exactly one + # setup, so RealTimeStreaming sends theirs on the first client message). # Build the streaming handler first so a transcription guardrail's # auto-response disable can be folded into this one setup: Gemini # rejects a second setup, so a follow-up disable would be dropped # and the guardrail bypassed. _session_config: str | None = None - if provider_config.requires_session_configuration(): + sends_setup_at_connect: Final = ( + provider_config.requires_session_configuration() + and not provider_config.builds_setup_from_session_update() + ) + if sends_setup_at_connect: _session_config = provider_config.session_configuration_request(model) if _session_config: _session_config = realtime_streaming.maybe_inject_guardrail_auto_response_disable( diff --git a/litellm/llms/gemini/realtime/transformation.py b/litellm/llms/gemini/realtime/transformation.py index 6f778ab2761..3134dfa1e4b 100644 --- a/litellm/llms/gemini/realtime/transformation.py +++ b/litellm/llms/gemini/realtime/transformation.py @@ -7,6 +7,7 @@ from collections import OrderedDict from collections.abc import Mapping, Sequence from typing import Final, cast +from pydantic import TypeAdapter from typing_extensions import ReadOnly, Required, TypedDict import litellm @@ -15,8 +16,9 @@ from litellm._uuid import uuid from litellm.litellm_core_utils.core_helpers import as_str_mapping from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.url_utils import join_api_base -from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig +from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig, RealtimeSessionUpdateRejection from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import ( + OPENAI_TO_GEMINI_TTS_VOICES, VertexGeminiConfig, ) from litellm.responses.litellm_completion_transformation.transformation import ( @@ -76,27 +78,82 @@ MAP_GEMINI_FIELD_TO_OPENAI_EVENT: Final[dict[str, OpenAIRealtimeEventTypes | Res # Keys the main transform loop handles; siblings like ``usageMetadata`` are skipped. _KNOWN_GEMINI_TOP_LEVEL_KEYS: Final[set] = {map_key.split(".", 1)[0] for map_key in MAP_GEMINI_FIELD_TO_OPENAI_EVENT} +_GEMINI_LIVE_RESPONSE_MODALITIES: Final = frozenset({"TEXT", "AUDIO"}) OPENAI_STOCK_REALTIME_VOICES: Final[frozenset[str]] = frozenset( {"alloy", "ash", "ballad", "cedar", "coral", "echo", "marin", "sage", "shimmer", "verse"} ) +# Per https://docs.cloud.google.com/vertex-ai/generative-ai/docs/live-api/configure-language-voice (2026-10-28) +GEMINI_LIVE_DOCUMENTED_VOICES: Final[frozenset[str]] = frozenset( + { + "achernar", + "achird", + "algenib", + "algieba", + "alnilam", + "aoede", + "autonoe", + "callirrhoe", + "charon", + "despina", + "enceladus", + "erinome", + "fenrir", + "gacrux", + "iapetus", + "kore", + "laomedeia", + "leda", + "orus", + "puck", + "pulcherrima", + "rasalgethi", + "sadachbia", + "sadaltager", + "schedar", + "sulafat", + "umbriel", + "vindemiatrix", + "zephyr", + "zubenelgenubi", + } +) -def _gemini_live_speech_config(voice: object) -> Mapping[str, object] | None: + +def _is_undocumented_live_voice(voice: object) -> bool: + return ( + isinstance(voice, str) + and voice.lower() not in OPENAI_STOCK_REALTIME_VOICES + and OPENAI_TO_GEMINI_TTS_VOICES.get(voice.lower(), voice).lower() not in GEMINI_LIVE_DOCUMENTED_VOICES + ) + + +def _gemini_live_speech_config(voice: object, drop_undocumented: bool) -> Mapping[str, object] | None: """Build the Gemini Live speechConfig for a client-requested voice. - OpenAI stock voice names have no Gemini equivalent and Gemini Live closes - the session on an unknown voice, so they are dropped with a warning and - the model keeps its default voice. Every other name is forwarded verbatim. + OpenAI stock voice names have no Gemini equivalent, so they are always + dropped, and so is any name outside Google's documented Live voices when + ``drop_undocumented`` is set, keeping the model's default voice. Every + other name is forwarded verbatim. """ - if isinstance(voice, str) and voice.lower() in OPENAI_STOCK_REALTIME_VOICES: + if not isinstance(voice, str): + return VertexGeminiConfig().map_audio_params({"voice": voice}) + if voice.lower() in OPENAI_STOCK_REALTIME_VOICES: verbose_logger.warning( "Gemini Realtime: voice %s is an OpenAI voice with no Gemini equivalent; " "dropping it so the session keeps the model's default voice.", voice, ) return None + if drop_undocumented and _is_undocumented_live_voice(voice): + verbose_logger.warning( + "Gemini Realtime: voice %s is not a documented Gemini Live voice; " + "drop_params is set, so it is dropped and the session keeps the model's default voice.", + voice, + ) + return None return VertexGeminiConfig().map_audio_params({"voice": voice}) @@ -142,16 +199,34 @@ def _base64_decoded_byte_count(data: str) -> int: return max(len(data) * 3 // 4 - padding, 0) +_JSON_ARRAY: Final = TypeAdapter(tuple[object, ...]) + + +def _chat_completion_tool(tool: object) -> dict[str, object]: + fields: Final = as_str_mapping(tool) or {} + if fields.get("type") != "function" or "function" in fields: + return {**fields} + return {"type": "function", "function": {key: value for key, value in fields.items() if key != "type"}} + + +def _system_instruction_text(system_instruction: HttpxContentType | None) -> str | None: + parts: Final = system_instruction.get("parts", []) if system_instruction else [] + texts: Final = tuple(text for part in parts if isinstance(text := part.get("text"), str)) + return "\n".join(texts) if texts else None + + class GeminiRealtimeConfig(BaseRealtimeConfig): _TOOL_CALL_ID_TO_NAME_MAX = 256 # LRU cap for call_id→name mapping - def __init__(self): + def __init__(self, drop_params: bool = False): super().__init__() + self.drop_params = drop_params self._tool_call_id_to_name: OrderedDict[str, str] = OrderedDict() # Gemini Live sometimes emits usageMetadata in a standalone frame between # turns; buffer it here so the next response.done carries the token counts. self._pending_usage_metadata: dict | None = None self._unbilled_input_audio_bytes: int = 0 + self._manual_activity_open: bool = False def is_setup_message(self, msg_obj: dict) -> bool: return "setup" in msg_obj @@ -167,6 +242,10 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): """Google AI Studio closes the socket (1007) on padding or silence tuning next to ``disabled: true``.""" return True + def _rejects_undocumented_voices(self, model: str) -> bool: + """Google AI Studio closes the socket (1007) on an unknown voice only for its native-audio Live models.""" + return bool(GeminiRealtimeConfig._model_cost_entry(model).get("gemini_native_audio")) + @staticmethod def _usage_detail_alias(details: object, defaults: dict[str, int]) -> dict[str, object]: detail_mapping: Final = as_str_mapping(details) @@ -272,17 +351,26 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): except (json.JSONDecodeError, TypeError, AttributeError): return False + def _start_manual_activity(self, session_configuration_request: str | None) -> tuple[str, ...]: + if self._manual_activity_open or not self._manual_turn_detection_enabled(session_configuration_request): + return () + self._manual_activity_open = True + activity_start: Final[BidiGenerateContentRealtimeInput] = {"activityStart": {}} + return (json.dumps({"realtimeInput": activity_start}),) + + def _end_manual_activity(self) -> tuple[str, ...]: + if not self._manual_activity_open: + return () + self._manual_activity_open = False + activity_end: Final[BidiGenerateContentRealtimeInput] = {"activityEnd": {}} + return (json.dumps({"realtimeInput": activity_end}),) + def _handle_input_audio_buffer_commit_or_end(self, session_configuration_request: str | None) -> list[str]: """Map OpenAI buffer commit/end to Gemini Live turn-boundary signals.""" if self._manual_turn_detection_enabled(session_configuration_request): - realtime_input_dict: BidiGenerateContentRealtimeInput = { - "activityEnd": True, - } - verbose_logger.debug("Gemini Realtime: Sending activityEnd realtimeInput to backend") - else: - realtime_input_dict = {"audioStreamEnd": True} - verbose_logger.debug("Gemini Realtime: Sending audioStreamEnd realtimeInput to backend") - return [json.dumps({"realtimeInput": realtime_input_dict})] + return list(self._end_manual_activity()) + audio_stream_end: Final[BidiGenerateContentRealtimeInput] = {"audioStreamEnd": True} + return [json.dumps({"realtimeInput": audio_stream_end})] def map_automatic_turn_detection(self, value: OpenAIRealtimeTurnDetection) -> AutomaticActivityDetection: """Map OpenAI ``server_vad`` to Gemini ``automaticActivityDetection``. @@ -325,7 +413,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): "voice", ] - def map_openai_params(self, optional_params: dict, non_default_params: dict) -> dict: + def map_openai_params(self, optional_params: dict, non_default_params: dict, *, model: str) -> dict: if "generationConfig" not in optional_params: optional_params["generationConfig"] = {} for key, value in non_default_params.items(): @@ -333,12 +421,12 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): optional_params["systemInstruction"] = HttpxContentType(role="user", parts=[{"text": value}]) elif key == "temperature": optional_params["generationConfig"]["temperature"] = value - elif key == "max_response_output_tokens": + elif key == "max_response_output_tokens" and isinstance(value, int): optional_params["generationConfig"]["maxOutputTokens"] = value elif key == "modalities": - optional_params["generationConfig"]["responseModalities"] = [ - modality.upper() for modality in cast(list[str], value) - ] + optional_params["generationConfig"]["responseModalities"] = list( + GeminiRealtimeConfig._requested_response_modalities(cast(object, value)) + ) elif key == "tools": from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import ( VertexGeminiConfig, @@ -347,34 +435,43 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): vertex_gemini_config = VertexGeminiConfig() # Tools should be at the top level of setup, not inside generationConfig optional_params["tools"] = vertex_gemini_config.map_function( - value=value, optional_params=optional_params + value=[_chat_completion_tool(tool) for tool in _JSON_ARRAY.validate_python(value)], + optional_params=optional_params, ) elif key == "input_audio_transcription" and value is not None: optional_params["inputAudioTranscription"] = {} - elif key == "turn_detection" and value is not None: - value_typed = cast(OpenAIRealtimeTurnDetection, value) - if ( - isinstance(value_typed, dict) - and value_typed.get("type") == "semantic_vad" - and "create_response" not in value_typed - ): - # Pipecat/OpenAI GA semantic VAD — skip; Gemini uses its own VAD. - # Only skip when there is no create_response override so that - # a guardrail-injected create_response:false is not dropped. - continue - transformed_audio_activity_config = self.map_automatic_turn_detection(value_typed) - if transformed_audio_activity_config: - optional_params["realtimeInputConfig"] = BidiGenerateContentRealtimeInputConfig( - automaticActivityDetection=transformed_audio_activity_config - ) + elif key == "turn_detection": + realtime_input_config = self._realtime_input_config(cast(object, value)) + if realtime_input_config is not None: + optional_params["realtimeInputConfig"] = realtime_input_config elif key == "voice": - speech_config = _gemini_live_speech_config(value) + speech_config = _gemini_live_speech_config( + value, self.drop_params and self._rejects_undocumented_voices(model) + ) if speech_config: optional_params["generationConfig"]["speechConfig"] = speech_config if len(optional_params["generationConfig"]) == 0: optional_params.pop("generationConfig") return optional_params + def _realtime_input_config(self, turn_detection: object) -> BidiGenerateContentRealtimeInputConfig | None: + if turn_detection is None: + return BidiGenerateContentRealtimeInputConfig(automaticActivityDetection={"disabled": True}) + value_typed: Final = cast(OpenAIRealtimeTurnDetection, turn_detection) + if ( + isinstance(value_typed, dict) + and value_typed.get("type") == "semantic_vad" + and "create_response" not in value_typed + ): + # Pipecat/OpenAI GA semantic VAD — skip; Gemini uses its own VAD. + # Only skip when there is no create_response override so that + # a guardrail-injected create_response:false is not dropped. + return None + transformed_audio_activity_config: Final = self.map_automatic_turn_detection(value_typed) + if not transformed_audio_activity_config: + return None + return BidiGenerateContentRealtimeInputConfig(automaticActivityDetection=transformed_audio_activity_config) + @staticmethod def _extract_turn_detection(session: dict) -> dict | None: """Extract turn_detection from a session.update payload. @@ -423,6 +520,8 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): if isinstance(input_cfg, dict): if "input_audio_transcription" not in normalized and "transcription" in input_cfg: normalized["input_audio_transcription"] = input_cfg["transcription"] + if "turn_detection" not in normalized and "turn_detection" in input_cfg: + normalized["turn_detection"] = input_cfg["turn_detection"] output_cfg: Final = audio.get("output") if isinstance(output_cfg, dict) and output_cfg.get("voice"): normalized["voice"] = output_cfg["voice"] @@ -434,7 +533,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): return normalized @staticmethod - def _model_cost_entry(model: str) -> dict: + def _model_cost_entry(model: str) -> Mapping[str, object]: entry = litellm.model_cost.get(model) if entry is None: stripped: Final = model.split("/", 1)[-1] @@ -455,18 +554,122 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): return "TEXT" if GeminiRealtimeConfig._is_text_only_live_model(model) else "AUDIO" @staticmethod - def _coerce_response_modalities(model: str, modalities: Sequence[object]) -> tuple[str, ...]: - """Swap responseModalities a Live model cannot produce: TEXT to AUDIO for - audio-only models, AUDIO to TEXT for text-only ones (e.g. transcribe-live).""" - normalized: Final = tuple( - modality.upper() if isinstance(modality, str) else str(modality).upper() for modality in modalities + def _requested_response_modalities(value: object) -> tuple[str, ...]: + if isinstance(value, str): + return (value.upper(),) + if not isinstance(value, (list, tuple)): + return () + return tuple(str(modality).upper() for modality in _JSON_ARRAY.validate_python(value)) + + @staticmethod + def _unsupported_response_modalities(model: str, requested: tuple[str, ...]) -> tuple[str, ...]: + unknown: Final = tuple(modality for modality in requested if modality not in _GEMINI_LIVE_RESPONSE_MODALITIES) + if unknown: + return unknown + if GeminiRealtimeConfig._is_audio_only_live_model(model) and frozenset(requested) == frozenset(("TEXT",)): + return ("TEXT",) + if GeminiRealtimeConfig._is_text_only_live_model(model) and frozenset(requested) == frozenset(("AUDIO",)): + return ("AUDIO",) + return () + + def setup_session_update_rejection( + self, model: str, session: Mapping[str, object], *, setup_sent: bool + ) -> RealtimeSessionUpdateRejection | None: + if self.drop_params: + return None + if setup_sent: + return RealtimeSessionUpdateRejection( + code="session_update_not_applied", + message=( + "Gemini Live accepts one session configuration per connection, and this session already sent " + "its configuration, so this session.update was not applied. Put every setting in session.update " + "before the first conversation content, or set drop_params to drop it instead" + ), + param="session", + ) + return self._modality_rejection(model, session) or self._voice_rejection(model, session) + + @staticmethod + def _modality_rejection(model: str, session: Mapping[str, object]) -> RealtimeSessionUpdateRejection | None: + param: Final = "output_modalities" if "output_modalities" in session else "modalities" + unsupported: Final = GeminiRealtimeConfig._unsupported_response_modalities( + model, GeminiRealtimeConfig._requested_response_modalities(session.get(param)) ) + if not unsupported: + return None + return RealtimeSessionUpdateRejection( + code="invalid_value", + message=( + f"Invalid value for 'session.{param}': {model} cannot produce " + f"{', '.join(modality.lower() for modality in unsupported)} output. Set drop_params to drop it instead" + ), + param=f"session.{param}", + ) + + def _voice_rejection(self, model: str, session: Mapping[str, object]) -> RealtimeSessionUpdateRejection | None: + audio_output: Final = as_str_mapping((as_str_mapping(session.get("audio")) or {}).get("output")) or {} + output_voice: Final = audio_output.get("voice") + voice: Final = output_voice or session.get("voice") + if not (self._rejects_undocumented_voices(model) and _is_undocumented_live_voice(voice)): + return None + param: Final = "session.audio.output.voice" if output_voice else "session.voice" + return RealtimeSessionUpdateRejection( + code="invalid_value", + message=( + f"Invalid value for '{param}': {voice} is not a Gemini Live voice {model} offers. " + "Set drop_params to drop it instead" + ), + param=param, + ) + + @staticmethod + def _coerce_response_modalities(model: str, modalities: Sequence[object]) -> tuple[str, ...]: + """Drop modalities Gemini Live cannot produce at all, swap the ones this + model cannot (TEXT to AUDIO for audio-only models, AUDIO to TEXT for + text-only ones such as transcribe-live), then keep a single modality + because Gemini Live rejects a setup naming two.""" + return GeminiRealtimeConfig._single_response_modality( + GeminiRealtimeConfig._swap_unsupported_response_modalities( + model, GeminiRealtimeConfig._known_response_modalities(model, modalities) + ) + ) + + @staticmethod + def _known_response_modalities(model: str, modalities: Sequence[object]) -> tuple[str, ...]: + normalized: Final = tuple(str(modality).upper() for modality in modalities) + known: Final = tuple(modality for modality in normalized if modality in _GEMINI_LIVE_RESPONSE_MODALITIES) + if len(known) < len(normalized): + verbose_logger.warning( + "Gemini Live: %s only produces TEXT or AUDIO; ignoring the requested %s", + model, + ", ".join(modality for modality in normalized if modality not in _GEMINI_LIVE_RESPONSE_MODALITIES), + ) + return known or (GeminiRealtimeConfig._default_response_modality(model),) + + @staticmethod + def _swap_unsupported_response_modalities(model: str, normalized: tuple[str, ...]) -> tuple[str, ...]: if GeminiRealtimeConfig._is_audio_only_live_model(model) and "TEXT" in normalized: + verbose_logger.warning( + "Gemini Live: %s only produces audio; downgrading the requested TEXT response modality to AUDIO", + model, + ) return tuple(modality for modality in normalized if modality != "TEXT") or ("AUDIO",) if GeminiRealtimeConfig._is_text_only_live_model(model) and "AUDIO" in normalized: return tuple(modality for modality in normalized if modality != "AUDIO") or ("TEXT",) return normalized + @staticmethod + def _single_response_modality(modalities: tuple[str, ...]) -> tuple[str, ...]: + if len(modalities) <= 1: + return modalities + chosen: Final = "AUDIO" if "AUDIO" in modalities else modalities[0] + verbose_logger.warning( + "Gemini Live: at most one response modality is allowed per session; using %s for the requested %s", + chosen, + ", ".join(modalities), + ) + return (chosen,) + @staticmethod def _finalize_gemini_live_setup(model: str, setup: dict[str, object]) -> dict[str, object]: generation_config: Final = setup.get("generationConfig") @@ -504,7 +707,9 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): # would be silently dropped because ``map_openai_params`` only # recognises the flat OpenAI-beta key names. session_payload = self._normalize_session_payload_for_mapping(session_payload) - new_overrides: Final = self.map_openai_params(optional_params={}, non_default_params=session_payload) + new_overrides: Final = self.map_openai_params( + optional_params={}, non_default_params=session_payload, model=model + ) if session_configuration_request is None: generation_config: Final = new_overrides.setdefault("generationConfig", {}) @@ -618,14 +823,13 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): message_str = str(message) raise ValueError(f"Invalid JSON message: {message_str}") - messages: Final[list[str]] = [] msg_type: Final = json_message.get("type") if msg_type == "session.update": return self._handle_session_update(json_message, model, session_configuration_request) if msg_type == "response.create": - return [] # Gemini responds automatically; nothing to forward + return list(self._end_manual_activity()) if msg_type == "conversation.item.create": return self._handle_conversation_item(json_message) @@ -643,8 +847,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): gemini_msg: Final = json.dumps({"realtimeInput": realtime_input_dict}) verbose_logger.debug("Gemini Realtime: Sending audio realtimeInput to backend") - messages.append(gemini_msg) - return messages + return [*self._start_manual_activity(session_configuration_request), gemini_msg] if msg_type in ("input_audio_buffer.commit", "input_audio_buffer.end"): return self._handle_input_audio_buffer_commit_or_end(session_configuration_request) @@ -660,22 +863,21 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): logging_session_id: str, session_configuration_request: str | None = None, ) -> OpenAIRealtimeStreamSessionEvents: - if session_configuration_request: - session_configuration_request_dict: BidiGenerateContentSetup = _parse_setup(session_configuration_request) - else: - session_configuration_request_dict = {} + session_configuration_request_dict: Final = _parse_setup( + session_configuration_request or self.session_configuration_request(model) + ) _model: Final = session_configuration_request_dict.get("model") or model generation_config: Final = session_configuration_request_dict.get("generationConfig", {}) or {} gemini_modalities: Final = generation_config.get("responseModalities", ["AUDIO"]) _modalities: Final = [modality.lower() for modality in cast(list[str], gemini_modalities)] - _system_instruction: Final = session_configuration_request_dict.get("systemInstruction") + instructions: Final = _system_instruction_text(session_configuration_request_dict.get("systemInstruction")) session: Final = OpenAIRealtimeStreamSession( id=logging_session_id, modalities=_modalities, ) - if _system_instruction is not None and isinstance(_system_instruction, str): - session["instructions"] = _system_instruction + if instructions is not None: + session["instructions"] = instructions if _model is not None and isinstance(_model, str): # Strip Vertex/AI Studio path prefixes to expose the bare model name. if "/models/" in _model: @@ -1662,6 +1864,9 @@ class GeminiRealtimeConfig(BaseRealtimeConfig): # Deferred setup opt-in: litellm.gemini_live_defer_setup = True return not litellm.gemini_live_defer_setup + def builds_setup_from_session_update(self) -> bool: + return True + def session_configuration_request(self, model: str) -> str: """ diff --git a/litellm/llms/vertex_ai/realtime/transformation.py b/litellm/llms/vertex_ai/realtime/transformation.py index 13f547b3b92..7635cd32a06 100644 --- a/litellm/llms/vertex_ai/realtime/transformation.py +++ b/litellm/llms/vertex_ai/realtime/transformation.py @@ -39,8 +39,8 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig): (they require async I/O) and injected at construction time. """ - def __init__(self, access_token: str, project: str, location: str) -> None: - super().__init__() + def __init__(self, access_token: str, project: str, location: str, drop_params: bool = False) -> None: + super().__init__(drop_params=drop_params) self._access_token = access_token self._project = project self._location = location @@ -51,6 +51,9 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig): def _rejects_vad_tuning_with_detection_off(self) -> bool: return False + def _rejects_undocumented_voices(self, model: str) -> bool: + return True + # ------------------------------------------------------------------ # URL # ------------------------------------------------------------------ @@ -163,7 +166,7 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig): # settings would be silently dropped because ``map_openai_params`` only # recognises the flat OpenAI-beta key names. session_params = self._normalize_session_payload_for_mapping(session_params) - setup_config: Final = self.map_openai_params(optional_params={}, non_default_params=session_params) + setup_config: Final = self.map_openai_params(optional_params={}, non_default_params=session_params, model=model) # Use full Vertex AI model path setup_config["model"] = self._vertex_model_path(model) @@ -177,15 +180,16 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig): # ``silence_duration_ms``). ``map_automatic_turn_detection`` sets # ``disabled=True`` whenever ``create_response`` is absent or # ``False``. Force ``disabled=False`` only when the client did - # not explicitly request ``create_response: False`` — that path + # not explicitly request ``create_response: False`` or + # ``turn_detection: null`` — the first of those paths # is how transcription guardrails suppress automatic responses, # and overriding it here would silently bypass the guardrail. # Vertex Live has no "VAD on, no auto-response" mode, so callers # that need that behaviour must accept that VAD is off. client_turn_detection: Final = session_params.get("turn_detection") client_disabled_auto_response: Final = ( - isinstance(client_turn_detection, dict) and client_turn_detection.get("create_response") is False - ) + "turn_detection" in session_params and client_turn_detection is None + ) or (isinstance(client_turn_detection, dict) and client_turn_detection.get("create_response") is False) realtime_input_config: Final = setup_config.setdefault("realtimeInputConfig", {}) automatic_detection: Final = realtime_input_config.setdefault("automaticActivityDetection", {}) if not client_disabled_auto_response: @@ -261,4 +265,5 @@ def vertex_realtime_config( access_token=access_token, project=project, location=VertexBase.get_vertex_region(vertex_region=location, model=model), + drop_params=drop_params, ) diff --git a/litellm/types/llms/gemini.py b/litellm/types/llms/gemini.py index 2ad81e64e05..67d90eefab0 100644 --- a/litellm/types/llms/gemini.py +++ b/litellm/types/llms/gemini.py @@ -81,6 +81,14 @@ class BidiGenerateContentServerMessage(TypedDict, total=False): """Output only. The setup complete message.""" +class BidiGenerateContentActivityStart(TypedDict): + pass + + +class BidiGenerateContentActivityEnd(TypedDict): + pass + + class BidiGenerateContentRealtimeInput(TypedDict, total=False): text: str """The text to be sent to the model.""" @@ -94,11 +102,11 @@ class BidiGenerateContentRealtimeInput(TypedDict, total=False): audioStreamEnd: bool """Output only. If true, indicates that the audio stream has ended.""" - activityStart: bool - """Output only. If true, indicates that the activity has started.""" + activityStart: BidiGenerateContentActivityStart + """Marks the start of user activity when automatic activity detection is disabled.""" - activityEnd: bool - """Output only. If true, indicates that the activity has ended.""" + activityEnd: BidiGenerateContentActivityEnd + """Marks the end of user activity when automatic activity detection is disabled.""" StartOfSpeechSensitivityEnum = Literal[ diff --git a/litellm/utils.py b/litellm/utils.py index e85e089289a..6f56cc65f45 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -10911,7 +10911,7 @@ class ProviderConfigManager: if LlmProviders.GEMINI == provider: from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig - return GeminiRealtimeConfig() + return GeminiRealtimeConfig(drop_params=drop_params) if LlmProviders.META == provider: from litellm.llms.meta.realtime.transformation import MetaRealtimeConfig diff --git a/tests/benchmarks/test_realtime_session_merge_benchmarks.py b/tests/benchmarks/test_realtime_session_merge_benchmarks.py new file mode 100644 index 00000000000..29ee1b638ad --- /dev/null +++ b/tests/benchmarks/test_realtime_session_merge_benchmarks.py @@ -0,0 +1,21 @@ +from typing import Final + +import pytest + +from litellm.litellm_core_utils.realtime_streaming import _merge_session_fields + +NESTED_LEVELS: Final = 8 +LEAVES: Final = 2000 + + +def _nested_session(depth: int) -> dict[str, object]: + return {"child": _nested_session(depth - 1)} if depth else {f"k{index}": index for index in range(LEAVES)} + + +@pytest.mark.benchmark +def test_merging_a_session_update_that_repeats_every_nested_level() -> None: + session: Final = _nested_session(NESTED_LEVELS) + + merged: Final = _merge_session_fields(session, session) + + assert merged == session diff --git a/tests/code_coverage_tests/recursive_detector.py b/tests/code_coverage_tests/recursive_detector.py index c943ff2cbd0..9b61c1d39ae 100644 --- a/tests/code_coverage_tests/recursive_detector.py +++ b/tests/code_coverage_tests/recursive_detector.py @@ -85,6 +85,7 @@ IGNORE_FUNCTIONS = [ "_sort_processed_sets", # bounded by the nesting depth of the log-record extra it walks (a finite JSON tree, no cycles possible). "scrub_json_strings", # max depth set (MAX_SCRUB_DEPTH); fails closed by returning "[Filtered]" for anything nested past the cap. "_without_requester_ip_address", # walks only the fresh tree _sanitize_request_body_for_spend_logs_payload returns (mapping cycles already cut), so it never goes deeper than that sanitizer did. + "_merge_session_fields", # max depth set (_SESSION_MERGE_MAX_DEPTH); deeper mappings are replaced wholesale by the later session.update. ] diff --git a/tests/e2e/llm_translation/realtime/realtime_client.py b/tests/e2e/llm_translation/realtime/realtime_client.py index 4d8336775ef..6bd5d61d721 100644 --- a/tests/e2e/llm_translation/realtime/realtime_client.py +++ b/tests/e2e/llm_translation/realtime/realtime_client.py @@ -47,6 +47,7 @@ class RealtimeProvider: id: str alias: str litellm_params: LiteLLMParamsBody + modalities: tuple[str, ...] = ("text",) PROVIDERS = ( @@ -75,6 +76,7 @@ PROVIDERS = ( model="gemini/gemini-3.1-flash-live-preview", api_key="os.environ/GEMINI_API_KEY", ), + modalities=("audio",), ), RealtimeProvider( "vertex_ai", @@ -84,6 +86,7 @@ PROVIDERS = ( vertex_location="us-central1", vertex_credentials="os.environ/VERTEXAI_CREDENTIALS", ), + modalities=("audio",), ), RealtimeProvider( "bedrock", diff --git a/tests/e2e/llm_translation/realtime/test_realtime_e2e.py b/tests/e2e/llm_translation/realtime/test_realtime_e2e.py index f5592abb34d..3b7ba3a2430 100644 --- a/tests/e2e/llm_translation/realtime/test_realtime_e2e.py +++ b/tests/e2e/llm_translation/realtime/test_realtime_e2e.py @@ -124,7 +124,8 @@ def test_text_conversation( session.send( SessionUpdate( session=SessionConfig( - instructions="You are a terse assistant. Reply in one short sentence." + modalities=list(provider.modalities), + instructions="You are a terse assistant. Reply in one short sentence.", ) ) ) @@ -159,6 +160,7 @@ def test_tool_call_round_trip( session.send( SessionUpdate( session=SessionConfig( + modalities=list(provider.modalities), tools=[WEATHER_TOOL], tool_choice="auto", instructions=( diff --git a/tests/integration/providers/test_realtime_transcription_guardrail_session_update.py b/tests/integration/providers/test_realtime_transcription_guardrail_session_update.py index ad58a400cec..2712d6e30cd 100644 --- a/tests/integration/providers/test_realtime_transcription_guardrail_session_update.py +++ b/tests/integration/providers/test_realtime_transcription_guardrail_session_update.py @@ -355,7 +355,13 @@ def _voice_scenario_created_after_the_first_frame(transcript: str) -> RealtimeRe def _live_scenario() -> RealtimeResponse: - return RealtimeResponse(content_type="application/x-realtime", events=()) + return RealtimeResponse( + content_type="application/x-realtime", + events=( + {"serverContent": {"modelTurn": {"parts": [{"text": "scripted $REQUEST_ID"}]}}}, + {"serverContent": {"turnComplete": True}}, + ), + ) def _realtime_update(turn_detection: JsonValue) -> dict[str, JsonValue]: @@ -1348,10 +1354,14 @@ def test_gemini_ai_studio_voice_session_with_vad_tuning_connects_under_the_guard _ws_base(_owned_url(guardrail_proxy)), f"model={model}", key, - (_step(_update_frame(update), until="session.updated"),), + ( + _step(_update_frame(update), until="session.updated"), + _step(_user_text_item(CLEAN_TRANSCRIPT), RESPONSE_CREATE, until=RESPONSE_DONE), + ), ) observed: Final = _observed(upstream_url, handle.scenario_id) - assert session.types == ("session.created", "session.updated"), session + assert session.types[:2] == ("session.created", "session.updated"), session + assert session.types[-1] == RESPONSE_DONE, session assert _setup_detection(observed) == {"disabled": True}, _sent(observed) @@ -1365,10 +1375,14 @@ def test_vertex_live_voice_session_keeps_vad_tuning_under_the_guardrail(guardrai _ws_base(_owned_url(guardrail_proxy)), f"model={model}", key, - (_step(_update_frame(_realtime_update(TUNED_SERVER_VAD)), until="session.updated"),), + ( + _step(_update_frame(_realtime_update(TUNED_SERVER_VAD)), until="session.updated"), + _step(_user_text_item(CLEAN_TRANSCRIPT), RESPONSE_CREATE, until=RESPONSE_DONE), + ), ) observed: Final = _observed(upstream_url, handle.scenario_id) - assert session.types == ("session.created", "session.updated"), session + assert session.types[:2] == ("session.created", "session.updated"), session + assert session.types[-1] == RESPONSE_DONE, session assert _setup_detection(observed) == GEMINI_TUNED_DETECTION, _sent(observed) diff --git a/tests/unit/litellm_core_utils/test_realtime_streaming.py b/tests/unit/litellm_core_utils/test_realtime_streaming.py index a8543ebcfc2..ba76138a268 100644 --- a/tests/unit/litellm_core_utils/test_realtime_streaming.py +++ b/tests/unit/litellm_core_utils/test_realtime_streaming.py @@ -1,7 +1,7 @@ import asyncio import base64 import json -from collections.abc import Coroutine, Mapping, Sequence +from collections.abc import Callable, Coroutine, Mapping, Sequence from dataclasses import dataclass from typing import Final, Literal from unittest.mock import AsyncMock, MagicMock, patch @@ -14,6 +14,7 @@ from websockets.frames import Close import litellm from litellm.constants import ( BEDROCK_REALTIME_PENDING_OPENING_MESSAGE_SCOPE_KEY, + REALTIME_LAZY_SETUP_DEADLINE_SECONDS, REALTIME_METERED_USAGE_KEY, REALTIME_SESSION_FAILURE_LOGGED_KEY, REALTIME_SESSION_SUCCESS_LOGGED_KEY, @@ -21,6 +22,7 @@ from litellm.constants import ( from litellm.integrations.custom_guardrail import CustomGuardrail from litellm.litellm_core_utils.realtime_streaming import ( RealTimeStreaming, + _merge_session_fields, client_sent_openai_beta_realtime_header, ) from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj @@ -1255,11 +1257,15 @@ async def test_client_ack_caches_setup_to_prevent_duplicate_session_update_setup side_effect=[ json.dumps({"type": "session.update", "session": {"tools": []}}), json.dumps({"type": "session.update", "session": {"tools": []}}), + _TEXT_ITEM_CREATE, + _RESPONSE_CREATE, Exception("client done"), ] ) + websocket.send_text = AsyncMock() provider_config = MagicMock() + provider_config.setup_session_update_rejection = MagicMock(return_value=None) def _transform(message: str, model: str, session_configuration_request=None): if session_configuration_request is None: @@ -2285,6 +2291,7 @@ async def test_guardrail_turn_detection_injected_into_first_session_update_defer logging_obj.success_handler = MagicMock() provider_config = MagicMock() + provider_config.builds_setup_from_session_update = MagicMock(return_value=False) transformed_messages = [] def mock_transform(msg, model, session_config): @@ -2351,6 +2358,7 @@ async def test_guardrail_turn_detection_injection_tolerates_non_dict_value( logging_obj.success_handler = MagicMock() provider_config = MagicMock() + provider_config.builds_setup_from_session_update = MagicMock(return_value=False) transformed_messages = [] def mock_transform(msg, model, session_config): @@ -2413,6 +2421,7 @@ async def test_subsequent_session_update_cannot_reenable_vad_when_guardrails_act logging_obj.success_handler = MagicMock() provider_config = MagicMock() + provider_config.setup_session_update_rejection = MagicMock(return_value=None) transformed_messages = [] def mock_transform(msg, model, session_config): @@ -2465,6 +2474,7 @@ async def test_follow_up_setup_updates_cached_session_configuration_request(): logging_obj.success_handler = MagicMock() provider_config = MagicMock() + provider_config.setup_session_update_rejection = MagicMock(return_value=None) follow_up_setup = json.dumps( { "setup": { @@ -2498,6 +2508,636 @@ async def test_follow_up_setup_updates_cached_session_configuration_request(): assert streaming.session_configuration_request == follow_up_setup +def _gemini_default_mode_streaming( + monkeypatch: pytest.MonkeyPatch, + client_messages: Sequence[str], + model: str = "gemini-live-2.5-flash-native-audio", + drop_params: bool = False, + defer_setup: bool = False, +) -> tuple[RealTimeStreaming, MagicMock]: + monkeypatch.setattr(litellm, "gemini_live_defer_setup", defer_setup, raising=False) + from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig + + client_ws = MagicMock() + client_ws.receive_text = AsyncMock(side_effect=[*client_messages, ConnectionClosed(None, None)]) + client_ws.send_text = AsyncMock() + backend_ws = MagicMock() + backend_ws.send = AsyncMock() + streaming = RealTimeStreaming( + client_ws, + backend_ws, + MagicMock(), + provider_config=GeminiRealtimeConfig(drop_params=drop_params), + model=model, + ) + return streaming, backend_ws + + +def _events_sent_to_client(streaming: RealTimeStreaming) -> list[dict]: + return [json.loads(call.args[0]) for call in streaming.websocket.send_text.await_args_list] + + +_TEXT_ITEM_CREATE = json.dumps( + { + "type": "conversation.item.create", + "item": {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "hi"}]}, + } +) + +_RESPONSE_CREATE = json.dumps({"type": "response.create"}) + + +@pytest.mark.asyncio +async def test_default_mode_sends_default_setup_before_first_non_session_client_message(monkeypatch): + """Gemini Live setup is no longer pre-sent at connect, so a client that never + sends session.update still gets the provider default setup ahead of its + first content frame.""" + streaming, backend_ws = _gemini_default_mode_streaming(monkeypatch, [_TEXT_ITEM_CREATE]) + + await streaming.client_ack_messages() + + frames = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["generationConfig"]["responseModalities"] == ["AUDIO"] + assert streaming.session_configuration_request == backend_ws.send.await_args_list[0].args[0] + + +@pytest.mark.asyncio +async def test_default_mode_session_update_first_builds_the_only_setup(monkeypatch): + session_update = json.dumps({"type": "session.update", "session": {"modalities": ["text"]}}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [session_update, _TEXT_ITEM_CREATE, _RESPONSE_CREATE], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["generationConfig"]["responseModalities"] == ["TEXT"] + + +@pytest.mark.asyncio +async def test_default_mode_merges_every_session_update_sent_before_the_first_content_frame(monkeypatch): + """Clients configure a session over several session.update messages. Gemini + Live takes one setup, so every update sent before the first content frame is + merged into that setup instead of only the first one applying.""" + instructions_update = json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + modalities_update = json.dumps({"type": "session.update", "session": {"modalities": ["text"]}}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [instructions_update, modalities_update, _TEXT_ITEM_CREATE, _RESPONSE_CREATE], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["generationConfig"]["responseModalities"] == ["TEXT"] + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + + +@pytest.mark.asyncio +async def test_default_mode_merges_nested_audio_settings_across_session_updates(monkeypatch): + vad_update = json.dumps( + {"type": "session.update", "session": {"audio": {"input": {"turn_detection": {"silence_duration_ms": 1500}}}}} + ) + voice_update = json.dumps({"type": "session.update", "session": {"audio": {"output": {"voice": "Kore"}}}}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [vad_update, voice_update, _TEXT_ITEM_CREATE, _RESPONSE_CREATE], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + setup = json.loads(backend_ws.send.await_args_list[0].args[0])["setup"] + assert setup["realtimeInputConfig"]["automaticActivityDetection"]["silenceDurationMs"] == 1500 + assert setup["generationConfig"]["speechConfig"]["voiceConfig"]["prebuiltVoiceConfig"]["voiceName"] == "Kore" + + +def test_merge_session_fields_replaces_nested_mappings_wholesale_past_the_depth_cap(): + def nested(levels: int, key: str) -> dict: + return {key: levels, **({"child": nested(levels - 1, key)} if levels else {})} + + def level(node, depth: int): + return node if depth == 0 else level(node["child"], depth - 1) + + merged = _merge_session_fields(nested(10, "keep"), nested(10, "new")) + + assert all({"keep", "new"} <= set(level(merged, depth)) for depth in range(9)) + assert "keep" not in level(merged, 9) + assert level(merged, 9)["new"] == 1 + + +@pytest.mark.asyncio +async def test_default_mode_acknowledges_each_held_session_update_before_any_backend_frame(monkeypatch): + """A client that waits for session.updated before sending content must not + hang while its updates are held, so each one is acknowledged right away + with the modalities the merged setup will carry.""" + instructions_update = json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + modalities_update = json.dumps({"type": "session.update", "session": {"modalities": ["text"]}}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [instructions_update, modalities_update], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + acks = [json.loads(call.args[0]) for call in streaming.websocket.send_text.await_args_list] + assert [ack["type"] for ack in acks] == ["session.updated", "session.updated"] + assert [ack["session"].get("output_modalities") or ack["session"].get("modalities") for ack in acks] == [ + ["audio"], + ["text"], + ] + assert acks[1]["session"]["instructions"] == "Be brief." + backend_ws.send.assert_not_awaited() + + +_LATE_INSTRUCTIONS_UPDATE = json.dumps( + {"type": "session.update", "event_id": "evt_late", "session": {"instructions": "Be brief."}} +) + + +@pytest.mark.asyncio +async def test_default_mode_answers_a_session_update_after_the_first_content_with_an_error(monkeypatch): + streaming, backend_ws = _gemini_default_mode_streaming(monkeypatch, [_TEXT_ITEM_CREATE, _LATE_INSTRUCTIONS_UPDATE]) + + await streaming.client_ack_messages() + + errors: Final = [event["error"] for event in _events_sent_to_client(streaming) if event["type"] == "error"] + assert [(error["code"], error["event_id"]) for error in errors] == [("session_update_not_applied", "evt_late")] + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + + +@pytest.mark.asyncio +async def test_default_mode_drops_a_late_session_update_quietly_under_drop_params(monkeypatch): + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [_TEXT_ITEM_CREATE, _LATE_INSTRUCTIONS_UPDATE], drop_params=True + ) + + await streaming.client_ack_messages() + + assert _events_sent_to_client(streaming) == [] + assert backend_ws.send.await_count == 2 + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("model", "requested_output"), + [ + ("gemini-live-2.5-flash-native-audio", {"modalities": ["text"]}), + ("gemini-live-2.5-flash", {"output_modalities": ["image"]}), + ], +) +async def test_default_mode_rejects_an_output_modality_the_model_cannot_produce(monkeypatch, model, requested_output): + update: Final = json.dumps( + {"type": "session.update", "session": {**requested_output, "instructions": "Be brief."}} + ) + streaming, backend_ws = _gemini_default_mode_streaming(monkeypatch, [update, _TEXT_ITEM_CREATE], model=model) + + await streaming.client_ack_messages() + + events: Final = _events_sent_to_client(streaming) + assert [(event["type"], event["error"]["code"]) for event in events] == [("error", "invalid_value")] + setup: Final = json.loads(backend_ws.send.await_args_list[0].args[0])["setup"] + assert "systemInstruction" not in setup + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("modalities", "drop_params"), + [(["text", "audio"], False), (["text"], True)], + ids=["openai-default-pair", "text-dropped-under-drop-params"], +) +async def test_default_mode_applies_the_update_with_audio_output_on_an_audio_only_model( + monkeypatch, modalities, drop_params +): + update: Final = json.dumps( + {"type": "session.update", "session": {"modalities": modalities, "instructions": "Be brief."}} + ) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [update, _TEXT_ITEM_CREATE, _RESPONSE_CREATE], drop_params=drop_params + ) + + await streaming.client_ack_messages() + + assert [event["type"] for event in _events_sent_to_client(streaming)] == ["session.updated"] + setup: Final = json.loads(backend_ws.send.await_args_list[0].args[0])["setup"] + assert setup["generationConfig"]["responseModalities"] == ["AUDIO"] + assert setup["systemInstruction"]["parts"][0]["text"] == "Be brief." + + +@pytest.mark.asyncio +async def test_default_mode_sends_no_extra_session_updated_when_gemini_confirms_the_setup(monkeypatch): + """Each held session.update was already acknowledged, so Gemini's setupComplete + must not reach the client as one more session.updated nobody asked for.""" + monkeypatch.setattr(litellm, "gemini_live_defer_setup", False, raising=False) + from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig + + model: Final = "gemini-live-2.5-flash-native-audio" + config: Final = GeminiRealtimeConfig() + client_ws: Final = MagicMock() + client_ws.send_text = AsyncMock() + backend_ws: Final = MagicMock() + backend_ws.send = AsyncMock() + backend_ws.recv = AsyncMock(side_effect=[json.dumps({"setupComplete": {}}).encode(), ConnectionClosed(None, None)]) + logging_obj: Final = _mock_logging() + logging_obj.litellm_trace_id = "trace_lazy" + logging_obj.async_success_handler = AsyncMock() + logging_obj.success_handler = MagicMock() + streaming: Final = RealTimeStreaming(client_ws, backend_ws, logging_obj, provider_config=config, model=model) + streaming.session_created_sent_to_client = True + streaming.session_configuration_request = config.session_configuration_request(model) + + await streaming.backend_to_client_send_messages() + + assert [json.loads(call.args[0])["type"] for call in client_ws.send_text.await_args_list] == [] + + +@pytest.mark.asyncio +async def test_default_mode_folds_transcription_guardrail_vad_disable_into_the_merged_setup(monkeypatch): + """A transcription guardrail gates every turn, so its auto-response disable + has to land in the one setup built from the held session.updates, injected + once rather than into every update that gets merged.""" + monkeypatch.setattr(litellm, "callbacks", [_transcription_guardrail()]) + debug = MagicMock() + monkeypatch.setattr("litellm.litellm_core_utils.realtime_streaming.verbose_logger.debug", debug) + instructions_update = json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + vad_update = json.dumps( + {"type": "session.update", "session": {"turn_detection": {"type": "server_vad", "silence_duration_ms": 900}}} + ) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [instructions_update, vad_update, _TEXT_ITEM_CREATE, _RESPONSE_CREATE], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["realtimeInputConfig"]["automaticActivityDetection"] == {"disabled": True} + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + assert streaming._guardrail_turn_detection_update_sent is True + injected = [call for call in debug.call_args_list if str(call.args[0]).startswith("Injected turn_detection")] + assert len(injected) == 1 + + +def _blocking_text_guardrail() -> CustomGuardrail: + class _BlockingTextGuardrail(CustomGuardrail): + async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None): + if any("zebrafruit" in text for text in inputs.get("texts", [])): + raise ValueError("Content blocked") + return inputs + + return _BlockingTextGuardrail( + guardrail_name="test_text_guard", event_hook=GuardrailEventHooks.pre_call, default_on=True + ) + + +@pytest.mark.asyncio +async def test_default_mode_sends_the_setup_before_a_guardrail_blocks_the_first_message(monkeypatch): + """Gemini Live closes the socket when content arrives before the setup, so a + guardrail that blocks the first typed message must not reach the backend first.""" + monkeypatch.setattr(litellm, "callbacks", [_blocking_text_guardrail()]) + instructions_update = json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + blocked_item = json.dumps( + { + "type": "conversation.item.create", + "item": {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "zebrafruit"}]}, + } + ) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [instructions_update, blocked_item], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + assert "zebrafruit" not in json.dumps(frames[1]) + errors: Final = [event["error"]["type"] for event in _events_sent_to_client(streaming) if event["type"] == "error"] + assert errors == ["guardrail_violation"] + + +class _SteppedClock: + def __init__(self, epoch: float) -> None: + self.now = epoch + + def __call__(self) -> float: + return self.now + + +def _client_ws_receiving(*steps: Callable[[], Coroutine[object, object, str]]) -> MagicMock: + remaining: Final = iter(steps) + + async def receive_text() -> str: + return await next(remaining)() + + client_ws: Final = MagicMock() + client_ws.receive_text = receive_text + client_ws.send_text = AsyncMock() + return client_ws + + +@pytest.mark.asyncio +@pytest.mark.parametrize("epoch", [0.0, 1e9]) +async def test_default_mode_sends_the_held_setup_when_no_content_arrives_before_the_deadline(monkeypatch, epoch): + """Vertex AI refuses a setup sent minutes after connecting, so an idle session + sends what it holds at the deadline and answers any later session.update as late.""" + monkeypatch.setattr(litellm, "gemini_live_defer_setup", False, raising=False) + from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig + + clock: Final = _SteppedClock(epoch) + setup_sent: Final = asyncio.Event() + + async def instructions_update_then_the_deadline_passes() -> str: + clock.now += REALTIME_LAZY_SETUP_DEADLINE_SECONDS + return json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + + async def late_update_once_the_setup_is_out() -> str: + await setup_sent.wait() + return json.dumps( + {"type": "session.update", "event_id": "evt_late", "session": {"instructions": "Answer in French."}} + ) + + async def disconnect() -> str: + raise ConnectionClosed(None, None) + + backend_ws: Final = MagicMock() + backend_ws.send = AsyncMock(side_effect=lambda frame: setup_sent.set()) + streaming: Final = RealTimeStreaming( + _client_ws_receiving(instructions_update_then_the_deadline_passes, late_update_once_the_setup_is_out, disconnect), + backend_ws, + MagicMock(), + provider_config=GeminiRealtimeConfig(), + model="gemini-live-2.5-flash", + clock=clock, + ) + + await asyncio.wait_for(streaming.client_ack_messages(), timeout=5) + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup"] + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + events: Final = _events_sent_to_client(streaming) + assert [(event["type"], event.get("error", {}).get("code")) for event in events] == [ + ("session.updated", None), + ("error", "session_update_not_applied"), + ] + + +@pytest.mark.asyncio +async def test_default_mode_refuses_held_session_updates_past_the_size_cap(monkeypatch): + def instructions_update(text: str) -> str: + return json.dumps({"type": "session.update", "event_id": f"evt_{text[0]}", "session": {"instructions": text}}) + + half_the_cap: Final = RealTimeStreaming._MAX_BUFFERED_BYTES // 2 + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, + [instructions_update("a" * half_the_cap), instructions_update("b" * half_the_cap), _TEXT_ITEM_CREATE, _RESPONSE_CREATE], + model="gemini-live-2.5-flash", + ) + + await streaming.client_ack_messages() + + events: Final = _events_sent_to_client(streaming) + assert [(event["type"], event.get("error", {}).get("event_id")) for event in events] == [ + ("session.updated", None), + ("error", "evt_b"), + ] + assert events[1]["error"]["code"] == "session_update_too_large" + setup: Final = json.loads(backend_ws.send.await_args_list[0].args[0])["setup"] + assert setup["systemInstruction"]["parts"][0]["text"].startswith("a") + + +@pytest.mark.asyncio +@pytest.mark.parametrize("defer_setup", [False, True]) +async def test_a_session_update_between_the_first_items_and_response_create_joins_the_setup(monkeypatch, defer_setup): + modalities_update: Final = json.dumps({"type": "session.update", "session": {"modalities": ["text"]}}) + instructions_update: Final = json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, + [modalities_update, _TEXT_ITEM_CREATE, instructions_update, _RESPONSE_CREATE], + model="gemini-live-2.5-flash", + defer_setup=defer_setup, + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["generationConfig"]["responseModalities"] == ["TEXT"] + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + assert [event["type"] for event in _events_sent_to_client(streaming)] == ["session.updated", "session.updated"] + + +@pytest.mark.asyncio +async def test_default_mode_applies_a_session_update_that_arrives_while_audio_streams_in(monkeypatch): + modalities_update: Final = json.dumps({"type": "session.update", "session": {"modalities": ["text"]}}) + append: Final = json.dumps({"type": "input_audio_buffer.append", "audio": "AAAA"}) + instructions_update: Final = json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + commit: Final = json.dumps({"type": "input_audio_buffer.commit"}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [modalities_update, append, instructions_update, commit], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames][:2] == ["setup", "realtimeInput"] + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + assert frames[1]["realtimeInput"]["audio"]["data"] == "AAAA" + assert "error" not in [event["type"] for event in _events_sent_to_client(streaming)] + + +@pytest.mark.asyncio +@pytest.mark.parametrize("epoch", [0.0, 1e9]) +async def test_held_items_go_out_with_the_setup_when_no_response_create_follows(monkeypatch, epoch): + monkeypatch.setattr(litellm, "gemini_live_defer_setup", False, raising=False) + from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig + + item_sent: Final = asyncio.Event() + + async def instructions_update() -> str: + return json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}) + + async def text_item() -> str: + return _TEXT_ITEM_CREATE + + async def disconnect_once_the_item_is_out() -> str: + await item_sent.wait() + raise ConnectionClosed(None, None) + + backend_ws: Final = MagicMock() + backend_ws.send = AsyncMock( + side_effect=lambda frame: item_sent.set() if "clientContent" in json.loads(frame) else None + ) + streaming: Final = RealTimeStreaming( + _client_ws_receiving(instructions_update, text_item, disconnect_once_the_item_is_out), + backend_ws, + MagicMock(), + provider_config=GeminiRealtimeConfig(), + model="gemini-live-2.5-flash", + lazy_setup_content_grace_seconds=0, + clock=_SteppedClock(epoch), + ) + + await asyncio.wait_for(streaming.client_ack_messages(), timeout=5) + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + + +@pytest.mark.asyncio +async def test_default_mode_acknowledges_an_unchanged_session_update_after_content(monkeypatch): + """Pipecat resends its unchanged session.update right after its first + conversation items and treats any error event as fatal.""" + session_update: Final = json.dumps( + {"type": "session.update", "session": {"modalities": ["text"], "instructions": "Be brief."}} + ) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [session_update, _TEXT_ITEM_CREATE, _RESPONSE_CREATE, session_update], model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + events: Final = _events_sent_to_client(streaming) + assert [event["type"] for event in events] == ["session.updated", "session.updated"] + assert events[0]["session"] == events[1]["session"] + + +@pytest.mark.asyncio +async def test_deferred_setup_merges_session_updates_until_the_first_content(monkeypatch): + """Deferred setup merges every session.update sent before the first content, + the same as the default mode, and answers a changed one after it as late.""" + monkeypatch.setattr(litellm, "gemini_live_defer_setup", True, raising=False) + from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig + + client_ws: Final = MagicMock() + client_ws.receive_text = AsyncMock( + side_effect=[ + json.dumps({"type": "session.update", "session": {"modalities": ["text"]}}), + json.dumps({"type": "session.update", "session": {"instructions": "Be brief."}}), + _TEXT_ITEM_CREATE, + _RESPONSE_CREATE, + json.dumps( + {"type": "session.update", "event_id": "evt_late", "session": {"instructions": "Answer in French."}} + ), + ConnectionClosed(None, None), + ] + ) + client_ws.send_text = AsyncMock() + backend_ws: Final = MagicMock() + backend_ws.send = AsyncMock() + streaming: Final = RealTimeStreaming( + client_ws, backend_ws, MagicMock(), provider_config=GeminiRealtimeConfig(), model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["generationConfig"]["responseModalities"] == ["TEXT"] + assert frames[0]["setup"]["systemInstruction"]["parts"][0]["text"] == "Be brief." + errors: Final = [event["error"] for event in _events_sent_to_client(streaming) if event["type"] == "error"] + assert [(error["code"], error["event_id"]) for error in errors] == [("session_update_not_applied", "evt_late")] + + +@pytest.mark.asyncio +async def test_deferred_setup_sends_the_default_setup_before_text_content_that_comes_first(monkeypatch): + """Gemini Live closes the socket on content sent before any setup, so text that arrives + before a session.update in deferred mode gets the model's default setup first.""" + monkeypatch.setattr(litellm, "gemini_live_defer_setup", True, raising=False) + from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig + + client_ws: Final = MagicMock() + client_ws.receive_text = AsyncMock( + side_effect=[ + _TEXT_ITEM_CREATE, + json.dumps({"type": "session.update", "event_id": "evt_late", "session": {"modalities": ["text"]}}), + ConnectionClosed(None, None), + ] + ) + client_ws.send_text = AsyncMock() + backend_ws: Final = MagicMock() + backend_ws.send = AsyncMock() + streaming: Final = RealTimeStreaming( + client_ws, backend_ws, MagicMock(), provider_config=GeminiRealtimeConfig(), model="gemini-live-2.5-flash" + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + errors: Final = [event["error"] for event in _events_sent_to_client(streaming) if event["type"] == "error"] + assert [(error["code"], error["event_id"]) for error in errors] == [("session_update_not_applied", "evt_late")] + + +@pytest.mark.asyncio +async def test_deferred_setup_sends_a_session_update_that_follows_buffered_audio_at_once(monkeypatch): + append: Final = json.dumps({"type": "input_audio_buffer.append", "audio": "AAAA"}) + commit: Final = json.dumps({"type": "input_audio_buffer.commit"}) + session_update: Final = json.dumps({"type": "session.update", "session": {"instructions": "Be a pirate."}}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, [append, commit, session_update], model="gemini-live-2.5-flash", defer_setup=True + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup"] + assert "Be a pirate." in json.dumps(frames[0]["setup"]["systemInstruction"]) + + +@pytest.mark.parametrize("defer_setup", [False, True]) +@pytest.mark.asyncio +async def test_a_transcription_guardrail_refuses_gemini_audio_it_cannot_screen(monkeypatch, defer_setup): + monkeypatch.setattr(litellm, "callbacks", [_transcription_guardrail()]) + session_update: Final = json.dumps( + {"type": "session.update", "session": {"turn_detection": {"type": "server_vad", "create_response": False}}} + ) + append: Final = json.dumps({"type": "input_audio_buffer.append", "audio": "AAAA"}) + commit: Final = json.dumps({"type": "input_audio_buffer.commit", "event_id": "commit_1"}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, + [session_update, append, append, commit, json.dumps({"type": "response.create"}), _TEXT_ITEM_CREATE], + model="gemini-live-2.5-flash", + defer_setup=defer_setup, + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + refusals: Final = [ + event["error"]["event_id"] + for event in _events_sent_to_client(streaming) + if event["type"] == "error" and event["error"]["code"] == "audio_input_not_screenable" + ] + assert refusals == [None, "commit_1"] + + +@pytest.mark.parametrize("defer_setup", [False, True]) +@pytest.mark.parametrize("control_type", ["response.cancel", "conversation.item.delete", "input_audio_buffer.clear"]) +@pytest.mark.asyncio +async def test_a_control_event_before_session_update_leaves_the_setup_open(monkeypatch, defer_setup, control_type): + control_event: Final = json.dumps({"type": control_type, "item_id": "item_x"}) + session_update: Final = json.dumps({"type": "session.update", "session": {"modalities": ["text"]}}) + streaming, backend_ws = _gemini_default_mode_streaming( + monkeypatch, + [control_event, session_update, _TEXT_ITEM_CREATE, _RESPONSE_CREATE], + model="gemini-live-2.5-flash", + defer_setup=defer_setup, + ) + + await streaming.client_ack_messages() + + frames: Final = [json.loads(call.args[0]) for call in backend_ws.send.await_args_list] + assert [next(iter(frame)) for frame in frames] == ["setup", "clientContent"] + assert frames[0]["setup"]["generationConfig"]["responseModalities"] == ["TEXT"] + assert [event["type"] for event in _events_sent_to_client(streaming) if event["type"] == "error"] == [] + + @pytest.mark.asyncio async def test_deferred_setup_buffers_audio_until_backend_setup_complete(monkeypatch): """Pipecat may send audio before session.update when setup is deferred.""" @@ -2533,14 +3173,18 @@ async def test_deferred_setup_buffers_audio_until_backend_setup_complete(monkeyp @pytest.mark.asyncio -async def test_deferred_setup_sends_session_update_before_buffered_audio(monkeypatch): +async def test_deferred_setup_sends_the_held_session_update_when_more_audio_arrives(monkeypatch): monkeypatch.setattr(litellm, "gemini_live_defer_setup", True, raising=False) from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig client_ws = MagicMock() + client_ws.send_text = AsyncMock() audio_msg = json.dumps({"type": "input_audio_buffer.append", "audio": "AA=="}) session_update = json.dumps({"type": "session.update", "session": {"modalities": ["audio"]}}) - client_ws.receive_text = AsyncMock(side_effect=[audio_msg, session_update, ConnectionClosed(None, None)]) + later_audio_msg = json.dumps({"type": "input_audio_buffer.append", "audio": "AQ=="}) + client_ws.receive_text = AsyncMock( + side_effect=[audio_msg, session_update, later_audio_msg, ConnectionClosed(None, None)] + ) backend_ws = MagicMock() backend_ws.send = AsyncMock() logging_obj: Final = _mock_logging() @@ -2560,7 +3204,7 @@ async def test_deferred_setup_sends_session_update_before_buffered_audio(monkeyp sent_payload = json.loads(backend_ws.send.await_args_list[0].args[0]) assert "setup" in sent_payload assert "realtimeInput" not in sent_payload - assert streaming._pending_messages_until_setup == [audio_msg] + assert streaming._pending_messages_until_setup == [audio_msg, later_audio_msg] @pytest.mark.asyncio @@ -3523,6 +4167,24 @@ async def test_upstream_refusal_after_a_synthetic_session_created_still_logs_a_f assert session.logging.logged_sessions == () +class _LazySetupRealtimeConfig(_PassthroughRealtimeConfig): + def builds_setup_from_session_update(self) -> bool: + return True + + +@pytest.mark.asyncio +async def test_upstream_outage_before_a_held_setup_was_sent_logs_the_session_not_a_refusal(): + upstream_close: Final = ConnectionClosed(Close(1011, "upstream outage"), None) + session: Final = _relay_session( + _client_ws_that_never_sends(), _backend_ws_closing_with(upstream_close), _LazySetupRealtimeConfig() + ) + + await session.run() + + assert session.logging.logged_failures == () + assert len(session.logging.logged_sessions) == 1 + + @pytest.mark.asyncio async def test_upstream_close_after_relayed_events_still_logs_the_session_as_success(): client_ws: Final = _client_ws_that_never_sends() diff --git a/tests/unit/llms/custom_httpx/test_llm_http_handler.py b/tests/unit/llms/custom_httpx/test_llm_http_handler.py index 794ac2d26d5..23d7feb2557 100644 --- a/tests/unit/llms/custom_httpx/test_llm_http_handler.py +++ b/tests/unit/llms/custom_httpx/test_llm_http_handler.py @@ -14,6 +14,7 @@ import httpcore import httpx import pytest import respx +import websockets.exceptions from botocore.credentials import RefreshableCredentials from h2.events import ConnectionTerminated @@ -2787,6 +2788,194 @@ async def test_async_realtime_error_event_send_failure_still_closes(): assert client_ws.events == [("close", (1011, "Internal server error: vertex token refresh exploded"))] +_GEMINI_LIVE_SETUP_COMPLETE = json.dumps({"setupComplete": {}}) +_GEMINI_LIVE_TEXT_TURN = json.dumps({"serverContent": {"modelTurn": {"parts": [{"text": "Blue."}]}}}) +_GEMINI_LIVE_TURN_COMPLETE = json.dumps({"serverContent": {"turnComplete": True}}) + + +class _ScriptedRealtimeBackend: + """Backend socket that answers the first setup frame with the scripted replies + and closes once the client is done, so both bridge loops run to completion.""" + + def __init__(self, replies_after_setup): + self.sent = [] + self.closed = asyncio.Event() + self._replies_after_setup = list(replies_after_setup) + self._inbox = asyncio.Queue() + + async def __aenter__(self): + return self + + async def __aexit__(self, exc_type, exc, tb): + return None + + async def send(self, payload): + self.sent.append(payload) + if "setup" in json.loads(payload): + for reply in self._replies_after_setup: + self._inbox.put_nowait(reply) + + def close_after_replies(self): + self._inbox.put_nowait(None) + + async def recv(self, decode=True): + reply = await self._inbox.get() + if reply is None: + self.closed.set() + raise websockets.exceptions.ConnectionClosedOK(None, None) + return reply + + +class _ScriptedRealtimeClient: + def __init__(self, messages, backend): + self.scope = {"headers": []} + self.exceptions = websockets.exceptions + self.sent = [] + self._messages = list(messages) + self._backend = backend + + async def send_text(self, payload): + self.sent.append(payload) + + async def receive_text(self): + if self._messages: + return self._messages.pop(0) + self._backend.close_after_replies() + await self._backend.closed.wait() + raise websockets.exceptions.ConnectionClosedOK(None, None) + + async def close(self, code=None, reason=None): + return None + + +def _realtime_logging_obj(): + logging_obj = Mock() + logging_obj.litellm_trace_id = "trace-lit-6974" + logging_obj.async_success_handler = AsyncMock() + logging_obj.dispatch_failure_handlers = AsyncMock() + return logging_obj + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "model, session, backend_modality, client_modality, backend_turn, drop_params", + [ + ( + "gemini-live-2.5-flash", + {"type": "realtime", "output_modalities": ["text"]}, + "TEXT", + "text", + [_GEMINI_LIVE_TEXT_TURN], + False, + ), + ("gemini-live-2.5-flash", {"modalities": ["text"]}, "TEXT", "text", [_GEMINI_LIVE_TEXT_TURN], False), + ( + "gemini-live-2.5-flash-native-audio", + {"type": "realtime", "output_modalities": ["text"]}, + "AUDIO", + "audio", + [], + True, + ), + ], +) +async def test_async_realtime_vertex_first_session_update_builds_the_only_setup( + model, session, backend_modality, client_modality, backend_turn, drop_params: bool, monkeypatch +): + """Regression for text-only Vertex Live sessions answered in audio: the + connect-time default setup (AUDIO) used to go out before the client's + session.update, and Gemini accepts only one setup per connection, so the + requested modality was dropped and billed at the audio rate. The setup now + goes out on the client's first message, built from session.update when that + comes first. Under drop_params an audio-only model still gets AUDIO, and the + client learns the downgrade from session.updated.""" + from litellm.llms.vertex_ai.realtime.transformation import VertexAIRealtimeConfig + + monkeypatch.setattr(litellm, "gemini_live_defer_setup", False, raising=False) + backend = _ScriptedRealtimeBackend([_GEMINI_LIVE_SETUP_COMPLETE, *backend_turn, _GEMINI_LIVE_TURN_COMPLETE]) + client_ws = _ScriptedRealtimeClient( + [ + json.dumps({"type": "session.update", "session": session}), + json.dumps( + { + "type": "conversation.item.create", + "item": { + "type": "message", + "role": "user", + "content": [{"type": "input_text", "text": "What color is the sky? One word."}], + }, + } + ), + json.dumps({"type": "response.create"}), + ], + backend, + ) + handler = BaseLLMHTTPHandler() + + with patch.object(handler, "_open_realtime_backend_ws", AsyncMock(return_value=backend)): + await asyncio.wait_for( + handler.async_realtime( + model=model, + websocket=client_ws, + logging_obj=_realtime_logging_obj(), + provider_config=VertexAIRealtimeConfig( + access_token="fake-token", project="fake-project", location="us-central1", drop_params=drop_params + ), + headers={}, + ), + timeout=10, + ) + + backend_frames = [json.loads(frame) for frame in backend.sent] + setups = [frame["setup"] for frame in backend_frames if "setup" in frame] + assert len(setups) == 1 + assert "setup" in backend_frames[0] + assert setups[0]["generationConfig"]["responseModalities"] == [backend_modality] + assert any("clientContent" in frame for frame in backend_frames[1:]) + + client_events = [json.loads(frame) for frame in client_ws.sent] + event_types = [event["type"] for event in client_events] + assert event_types[:2] == ["session.created", "session.updated"] + assert client_events[1]["session"]["modalities"] == [client_modality] + if backend_turn: + assert "response.output_text.delta" in event_types + assert "response.output_audio_transcript.delta" not in event_types + + +@pytest.mark.asyncio +async def test_async_realtime_sends_setup_at_connect_when_provider_cannot_build_it_from_session_update(): + """Bedrock Nova Sonic style providers keep the connect-time setup and get no + synthetic session.created.""" + setup = json.dumps({"event": {"sessionStart": {}}}) + backend = _ScriptedRealtimeBackend([]) + client_ws = _ScriptedRealtimeClient([], backend) + provider_config = Mock() + provider_config.get_complete_url.return_value = "wss://backend.example/live" + provider_config.validate_environment.return_value = {} + provider_config.requires_session_configuration.return_value = True + provider_config.builds_setup_from_session_update.return_value = False + provider_config.session_configuration_request.return_value = setup + provider_config.unbilled_usage_on_session_close.return_value = None + provider_config.open_backend = AsyncMock(return_value=None) + provider_config.wrap_backend.side_effect = lambda opened: opened + handler = BaseLLMHTTPHandler() + + with patch.object(handler, "_open_realtime_backend_ws", AsyncMock(return_value=backend)): + await asyncio.wait_for( + handler.async_realtime( + model="amazon.nova-sonic-v1:0", + websocket=client_ws, + logging_obj=_realtime_logging_obj(), + provider_config=provider_config, + headers={}, + ), + timeout=10, + ) + + assert backend.sent == [setup] + assert client_ws.sent == [] + + class _JSONBodyAudioTranscriptionConfig(BaseAudioTranscriptionConfig): def get_supported_openai_params(self, model): return [] diff --git a/tests/unit/llms/gemini/realtime/test_gemini_realtime_transformation.py b/tests/unit/llms/gemini/realtime/test_gemini_realtime_transformation.py index 017f279d6c4..460a50ed61a 100644 --- a/tests/unit/llms/gemini/realtime/test_gemini_realtime_transformation.py +++ b/tests/unit/llms/gemini/realtime/test_gemini_realtime_transformation.py @@ -1,3 +1,4 @@ +import base64 import json from collections.abc import Mapping from typing import Final, cast @@ -7,7 +8,7 @@ import pytest import litellm -from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig +from litellm.llms.gemini.realtime.transformation import GEMINI_LIVE_DOCUMENTED_VOICES, GeminiRealtimeConfig from litellm.types.llms.gemini import BidiGenerateContentServerMessage @@ -439,6 +440,51 @@ def test_gemini_realtime_session_update_with_tools(): assert "parameters" in function_decl +def _schema_nodes(schema: object): + if isinstance(schema, dict): + yield schema + for value in schema.values(): + yield from _schema_nodes(value) + elif isinstance(schema, list): + for value in schema: + yield from _schema_nodes(value) + + +def test_gemini_session_update_converts_flat_realtime_tool_schemas_to_gemini_schemas(): + session_update: Final = { + "type": "session.update", + "session": { + "tools": [ + { + "type": "function", + "name": "lookup_weather", + "description": "Look up the weather in a city.", + "parameters": { + "type": "object", + "$defs": { + "City": {"type": "object", "properties": {"name": {"type": "string"}}, "required": ["name"]} + }, + "properties": {"city": {"$ref": "#/$defs/City"}, "note": {"type": ["string", "null"]}}, + "required": ["city"], + }, + } + ] + }, + } + + messages: Final = GeminiRealtimeConfig().transform_realtime_request( + json.dumps(session_update), "gemini-2.5-flash", session_configuration_request=None + ) + + declaration: Final = json.loads(messages[0])["setup"]["tools"][0]["function_declarations"][0] + assert declaration["name"] == "lookup_weather" + parameters: Final = declaration["parameters"] + assert parameters["properties"]["city"]["properties"]["name"]["type"] == "string" + nodes: Final = tuple(_schema_nodes(parameters)) + assert not any("$defs" in node or "$ref" in node for node in nodes) + assert all(isinstance(node["type"], str) for node in nodes if "type" in node) + + def test_gemini_session_update_defaults_to_audio_modality(): config = GeminiRealtimeConfig() @@ -613,6 +659,35 @@ def test_gemini_audio_only_live_models_drop_text_from_text_audio_combo(patch_gem assert setup["generationConfig"]["responseModalities"] == ["AUDIO"] +@pytest.mark.parametrize("output_modalities", ["text", ["text"]]) +def test_gemini_session_update_accepts_a_single_string_output_modality(output_modalities): + """Regression: a bare string was iterated per character, so output_modalities + "text" reached Vertex as responseModalities ["T"], which it rejects with 1007.""" + config = GeminiRealtimeConfig() + session_update = {"type": "session.update", "session": {"type": "realtime", "output_modalities": output_modalities}} + + messages = config.transform_realtime_request( + json.dumps(session_update), "gemini-live-2.5-flash", session_configuration_request=None + ) + + assert json.loads(messages[0])["setup"]["generationConfig"]["responseModalities"] == ["TEXT"] + + +@pytest.mark.parametrize( + ("output_modalities", "expected"), + [(["video", "text"], ["TEXT"]), (["video"], ["AUDIO"]), (None, ["AUDIO"])], +) +def test_gemini_session_update_drops_modalities_live_cannot_produce(output_modalities, expected): + config = GeminiRealtimeConfig() + session_update = {"type": "session.update", "session": {"type": "realtime", "output_modalities": output_modalities}} + + messages = config.transform_realtime_request( + json.dumps(session_update), "gemini-live-2.5-flash", session_configuration_request=None + ) + + assert json.loads(messages[0])["setup"]["generationConfig"]["responseModalities"] == expected + + def test_gemini_non_live_model_preserves_text_modality(): config = GeminiRealtimeConfig() session_update = { @@ -1396,22 +1471,71 @@ def test_gemini_input_audio_buffer_clear_is_local_noop(): assert messages == [] -def test_gemini_input_audio_buffer_commit_maps_to_activity_end_when_manual_vad(): +_MANUAL_VAD_SETUP = json.dumps({"setup": {"realtimeInputConfig": {"automaticActivityDetection": {"disabled": True}}}}) +_APPEND = json.dumps({"type": "input_audio_buffer.append", "audio": base64.b64encode(b"\x00\x01" * 80).decode()}) + + +def _realtime_input_keys(config: GeminiRealtimeConfig, client_event: str, setup: str | None) -> list[str]: + frames = config.transform_realtime_request(client_event, "gemini-live-2.5-flash-native-audio", setup) + return [key for frame in frames for key in json.loads(frame)["realtimeInput"]] + + +@pytest.mark.parametrize("turn_end", ["input_audio_buffer.commit", "response.create"]) +def test_gemini_manual_vad_wraps_each_audio_turn_in_activity_start_and_end(turn_end): + """With activity detection off, Gemini Live answers only after activityEnd closes an + activityStart, and it closes the socket on any other shape of those markers.""" config = GeminiRealtimeConfig() - setup = { - "setup": { - "realtimeInputConfig": { - "automaticActivityDetection": {"disabled": True}, - } - } - } - messages = config.transform_realtime_request( - json.dumps({"type": "input_audio_buffer.commit"}), - "gemini-live-2.5-flash-native-audio", - session_configuration_request=json.dumps(setup), + end = json.dumps({"type": turn_end}) + + turns = [ + [_realtime_input_keys(config, event, _MANUAL_VAD_SETUP) for event in (_APPEND, _APPEND, end, end)] + for _ in range(2) + ] + + assert turns == [[["activityStart", "audio"], ["audio"], ["activityEnd"], []]] * 2 + + +def test_gemini_manual_vad_activity_markers_are_empty_messages(): + config = GeminiRealtimeConfig() + commit = json.dumps({"type": "input_audio_buffer.commit"}) + + first_append = config.transform_realtime_request(_APPEND, "gemini-live-2.5-flash", _MANUAL_VAD_SETUP) + turn_end = config.transform_realtime_request(commit, "gemini-live-2.5-flash", _MANUAL_VAD_SETUP) + + assert json.loads(first_append[0]) == {"realtimeInput": {"activityStart": {}}} + assert [json.loads(frame) for frame in turn_end] == [{"realtimeInput": {"activityEnd": {}}}] + + +@pytest.mark.parametrize( + ("session", "expected_keys"), + [ + ({"turn_detection": None}, [["activityStart", "audio"], ["activityEnd"]]), + ({"type": "realtime", "audio": {"input": {"turn_detection": None}}}, [["activityStart", "audio"], ["activityEnd"]]), + ({"type": "realtime", "audio": {"input": {}}}, [["audio"], ["audioStreamEnd"]]), + ], +) +def test_gemini_null_turn_detection_makes_the_client_mark_each_turn(session, expected_keys): + """OpenAI's turn_detection null turns server VAD off, so the client commits each turn itself.""" + config = GeminiRealtimeConfig() + setup = config.transform_realtime_request( + json.dumps({"type": "session.update", "session": session}), "gemini-live-2.5-flash-native-audio", None ) - assert len(messages) == 1 - assert json.loads(messages[0]) == {"realtimeInput": {"activityEnd": True}} + commit = json.dumps({"type": "input_audio_buffer.commit"}) + + keys = [_realtime_input_keys(config, event, setup[0]) for event in (_APPEND, commit)] + + assert keys == expected_keys + + +def test_gemini_automatic_vad_sends_no_activity_markers(): + config = GeminiRealtimeConfig() + commit = json.dumps({"type": "input_audio_buffer.commit"}) + response_create = json.dumps({"type": "response.create"}) + + keys = [_realtime_input_keys(config, event, None) for event in (_APPEND, commit)] + + assert keys == [["audio"], ["audioStreamEnd"]] + assert config.transform_realtime_request(response_create, "gemini-live-2.5-flash", None) == [] def test_gemini_tool_call_id_to_name_evicts_oldest_when_capped(): @@ -1883,13 +2007,75 @@ def test_map_openai_params_drops_stock_voice_case_insensitively(): """ cfg = GeminiRealtimeConfig() - dropped = cfg.map_openai_params(optional_params={}, non_default_params={"voice": "Alloy"}) + dropped = cfg.map_openai_params(optional_params={}, non_default_params={"voice": "Alloy"}, model="gemini-3.8-live") assert "speechConfig" not in dropped.get("generationConfig", {}) - passthrough = cfg.map_openai_params(optional_params={}, non_default_params={"voice": "Kore"}) + passthrough = cfg.map_openai_params( + optional_params={}, non_default_params={"voice": "Kore"}, model="gemini-3.8-live" + ) assert passthrough["generationConfig"]["speechConfig"]["voiceConfig"]["prebuiltVoiceConfig"]["voiceName"] == "Kore" +_DOCUMENTED_VOICE: Final = sorted(GEMINI_LIVE_DOCUMENTED_VOICES)[0].capitalize() + + +@pytest.mark.parametrize("drop_params", [False, True]) +def test_google_ai_studio_forwards_a_documented_voice(drop_params): + config: Final = GeminiRealtimeConfig(drop_params=drop_params) + mapped: Final = config.map_openai_params( + optional_params={}, non_default_params={"voice": _DOCUMENTED_VOICE}, model="gemini-3.8-live" + ) + + speech_config: Final = mapped["generationConfig"]["speechConfig"] + assert speech_config["voiceConfig"]["prebuiltVoiceConfig"]["voiceName"] == _DOCUMENTED_VOICE + assert config.setup_session_update_rejection("gemini-3.8-live", {"voice": _DOCUMENTED_VOICE}, setup_sent=False) is None + + +@pytest.mark.parametrize( + ("session", "param"), + [ + ({"voice": "Nova"}, "session.voice"), + ({"type": "realtime", "audio": {"output": {"voice": "Nova"}}}, "session.audio.output.voice"), + ], +) +def test_google_ai_studio_rejects_an_undocumented_voice_on_the_session_update(session, param): + rejection: Final = GeminiRealtimeConfig().setup_session_update_rejection( + "gemini-2.5-flash-native-audio-preview-12-2025", session, setup_sent=False + ) + + assert rejection is not None + assert (rejection.code, rejection.param) == ("invalid_value", param) + + +def _setup_speech_config(config: GeminiRealtimeConfig, model: str, session: dict[str, object]) -> object: + messages: Final = config.transform_realtime_request( + json.dumps({"type": "session.update", "session": session}), model, session_configuration_request=None + ) + return json.loads(messages[0])["setup"].get("generationConfig", {}).get("speechConfig") + + +@pytest.mark.parametrize( + "session", + [{"voice": "Nova"}, {"type": "realtime", "audio": {"output": {"voice": "Nova"}}}], +) +def test_google_ai_studio_drops_an_undocumented_voice_with_drop_params(session): + config: Final = GeminiRealtimeConfig(drop_params=True) + model: Final = "gemini-2.5-flash-native-audio-preview-12-2025" + + assert config.setup_session_update_rejection(model, session, setup_sent=False) is None + assert _setup_speech_config(config, model, session) is None + + +@pytest.mark.parametrize("drop_params", [False, True]) +@pytest.mark.parametrize("model", ["gemini-3.8-live", "gemini-3.1-flash-live-preview"]) +def test_google_ai_studio_forwards_any_voice_to_models_that_accept_it(model, drop_params): + config: Final = GeminiRealtimeConfig(drop_params=drop_params) + + assert config.setup_session_update_rejection(model, {"voice": "Nova"}, setup_sent=False) is None + speech_config: Final = _setup_speech_config(config, model, {"voice": "Nova"}) + assert speech_config == {"voiceConfig": {"prebuiltVoiceConfig": {"voiceName": "Nova"}}} + + @pytest.fixture(autouse=False) def patch_gemini_transcribe_live_cost_map_entry(monkeypatch): """Inject the gemini-3.5-transcribe-live registry entry locally. @@ -1954,6 +2140,55 @@ def test_gemini_transcribe_live_coerces_audio_modality_to_text(modalities, patch assert setup["generationConfig"]["responseModalities"] == ["TEXT"] +@pytest.mark.parametrize("model", ["gemini-3.5-transcribe-live", "gemini/gemini-3.5-transcribe-live"]) +def test_gemini_transcribe_live_synthetic_session_created_reports_text_output( + model, patch_gemini_transcribe_live_cost_map_entry +): + session = GeminiRealtimeConfig().transform_session_created_event(model, "sess_transcribe")["session"] + + assert session["modalities"] == ["text"] + + +@pytest.mark.parametrize( + ("model", "modalities", "producible"), + [ + ("gemini-3.5-transcribe-live", ["audio"], False), + ("gemini-3.5-transcribe-live", ["audio", "text"], True), + ("gemini-live-2.5-flash-native-audio", ["text"], False), + ("gemini-live-2.5-flash-native-audio", ["text", "audio"], True), + ], +) +@pytest.mark.parametrize("drop_params", [False, True]) +def test_gemini_live_rejects_only_output_modalities_the_model_cannot_produce_at_all( + model, + modalities, + producible, + drop_params, + patch_gemini_transcribe_live_cost_map_entry, + patch_gemini_audio_cost_map_entries, +): + rejection = GeminiRealtimeConfig(drop_params=drop_params).setup_session_update_rejection( + model, {"output_modalities": modalities}, setup_sent=False + ) + + assert (rejection is None) is (drop_params or producible) + assert rejection is None or rejection.param == "session.output_modalities" + + +@pytest.mark.parametrize("drop_params", [False, True]) +def test_gemini_provider_realtime_config_carries_drop_params(drop_params): + from litellm.types.utils import LlmProviders + from litellm.utils import ProviderConfigManager + + model = "gemini-live-2.5-flash-native-audio" + config = ProviderConfigManager.get_provider_realtime_config( + model=model, provider=LlmProviders.GEMINI, drop_params=drop_params + ) + + rejection = config.setup_session_update_rejection(model, {"modalities": ["text"]}, setup_sent=False) + assert (rejection is None) is drop_params + + def test_gemini_chat_model_with_text_output_modalities_keeps_audio_eager_setup( patch_gemini_transcribe_live_cost_map_entry, ): @@ -2169,6 +2404,47 @@ def test_unbilled_usage_on_session_close_flushes_trailing_audio(patch_gemini_tra assert config.unbilled_usage_on_session_close("gemini-3.5-transcribe-live") is None +def test_gemini_realtime_two_output_modalities_collapse_to_audio(): + """Gemini Live closes the socket with 1007 when the setup names two response + modalities, so a GA session.update asking for ["text", "audio"] must reach the + backend as AUDIO only, the same choice the beta shim makes.""" + config = GeminiRealtimeConfig() + session_update = { + "type": "session.update", + "session": {"output_modalities": ["text", "audio"], "instructions": "Be brief."}, + } + + messages = config.transform_realtime_request( + json.dumps(session_update), + "gemini-live-2.5-flash", + session_configuration_request=None, + ) + + assert len(messages) == 1 + assert json.loads(messages[0])["setup"]["generationConfig"]["responseModalities"] == ["AUDIO"] + + +def test_gemini_realtime_inf_max_response_output_tokens_is_not_forwarded(): + """OpenAI's default max_response_output_tokens is the string "inf", which Gemini + rejects with 1007 (TYPE_INT32), so only integer limits map to maxOutputTokens.""" + config = GeminiRealtimeConfig() + + def generation_config_for(limit): + session_update = { + "type": "session.update", + "session": {"modalities": ["text"], "max_response_output_tokens": limit}, + } + messages = config.transform_realtime_request( + json.dumps(session_update), + "gemini-live-2.5-flash", + session_configuration_request=None, + ) + return json.loads(messages[0])["setup"]["generationConfig"] + + assert "maxOutputTokens" not in generation_config_for("inf") + assert generation_config_for(200)["maxOutputTokens"] == 200 + + def _grounded_live_frame(grounding_metadata: Mapping[str, object] | None) -> Mapping[str, object]: """One Live server frame. Grounding metadata and usageMetadata arrive together, as Vertex sends them.""" from typing import Final diff --git a/tests/unit/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py b/tests/unit/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py index 0bfff870d15..7f0106616ce 100644 --- a/tests/unit/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py +++ b/tests/unit/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py @@ -18,6 +18,7 @@ import websockets.exceptions # registers websockets.exceptions on the websocket import litellm from litellm.llms.vertex_ai.common_utils import get_vertex_base_url +from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import OPENAI_TO_GEMINI_TTS_VOICES from litellm.llms.vertex_ai.realtime.transformation import VertexAIRealtimeConfig # --------------------------------------------------------------------------- @@ -179,6 +180,25 @@ def test_vertex_audio_only_live_model_coerces_text_modality_to_audio( assert setup["generationConfig"]["responseModalities"] == ["AUDIO"] +@pytest.mark.parametrize("drop_params", [False, True]) +def test_vertex_realtime_config_carries_drop_params(drop_params, patch_native_audio_cost_map_entry): + from litellm.llms.vertex_ai.realtime.transformation import vertex_realtime_config + + cfg: Final = vertex_realtime_config( + _NATIVE_AUDIO_MODEL, + access_token="tok", + resolve_access_token=AsyncMock(return_value="tok"), + project="my-proj", + location="us-central1", + drop_params=drop_params, + ) + + rejection: Final = cfg.setup_session_update_rejection( + _NATIVE_AUDIO_MODEL, {"modalities": ["text"]}, setup_sent=False + ) + assert (rejection is None) is drop_params + + def test_vertex_session_update_normalizes_ga_remapped_fields( patch_native_audio_cost_map_entry, ): @@ -253,6 +273,26 @@ def test_vertex_keeps_turn_detection_tuning_when_auto_response_is_off(): } + +@pytest.mark.parametrize( + ("session", "disabled"), + [ + ({"turn_detection": None}, True), + ({"type": "realtime", "audio": {"input": {"turn_detection": None}}}, True), + ({"type": "realtime", "audio": {"input": {}}}, False), + ], +) +def test_vertex_turns_activity_detection_off_for_null_turn_detection(session, disabled): + """OpenAI's turn_detection null is push-to-talk, so Vertex must wait for the client's activity markers.""" + cfg: Final = VertexAIRealtimeConfig(access_token="tok", project="my-proj", location="us-central1") + messages: Final = cfg.transform_realtime_request( + json.dumps({"type": "session.update", "session": session}), + _NATIVE_AUDIO_MODEL, + session_configuration_request=None, + ) + setup: Final = json.loads(messages[0])["setup"] + assert setup["realtimeInputConfig"]["automaticActivityDetection"]["disabled"] is disabled + # --------------------------------------------------------------------------- # Round-trip test: text-in / text-out via RealTimeStreaming # --------------------------------------------------------------------------- @@ -607,3 +647,73 @@ def test_vertex_native_audio_unmapped_voice_passes_through(patch_native_audio_co generation_config = json.loads(messages[0])["setup"]["generationConfig"] assert generation_config["speechConfig"]["voiceConfig"]["prebuiltVoiceConfig"]["voiceName"] == "Kore" + + +def test_vertex_session_update_two_output_modalities_collapse_to_audio(): + """Vertex Live rejects a setup naming two response modalities, so a GA + ["text", "audio"] request must be sent as AUDIO only.""" + cfg = VertexAIRealtimeConfig(access_token="tok", project="my-proj", location="global") + session_update = { + "type": "session.update", + "session": {"output_modalities": ["text", "audio"], "instructions": "Be brief."}, + } + + messages = cfg.transform_realtime_request( + json.dumps(session_update), + "gemini-live-2.5-flash", + session_configuration_request=None, + ) + + assert len(messages) == 1 + assert json.loads(messages[0])["setup"]["generationConfig"]["responseModalities"] == ["AUDIO"] + + +@pytest.mark.parametrize( + ("session", "param"), + [ + ({"voice": "Nova"}, "session.voice"), + ({"type": "realtime", "audio": {"output": {"voice": "Nova"}}}, "session.audio.output.voice"), + ], +) +def test_vertex_rejects_an_undocumented_voice_on_the_session_update(session, param): + cfg: Final = VertexAIRealtimeConfig(access_token="tok", project="my-proj", location="us-central1") + + rejection: Final = cfg.setup_session_update_rejection("gemini-live-2.5-flash", session, setup_sent=False) + + assert rejection is not None + assert (rejection.code, rejection.param) == ("invalid_value", param) + + +@pytest.mark.parametrize("voice", ["Kore", "alloy"]) +def test_vertex_accepts_documented_and_openai_stock_voices_on_the_session_update(voice): + cfg: Final = VertexAIRealtimeConfig(access_token="tok", project="my-proj", location="us-central1") + + assert cfg.setup_session_update_rejection("gemini-live-2.5-flash", {"voice": voice}, setup_sent=False) is None + + +def test_vertex_drops_an_undocumented_voice_with_drop_params(): + cfg: Final = VertexAIRealtimeConfig(access_token="tok", project="my-proj", location="us-central1", drop_params=True) + session_update: Final = {"type": "session.update", "session": {"voice": "Nova", "instructions": "Be brief."}} + + assert cfg.setup_session_update_rejection("gemini-live-2.5-flash", session_update["session"], setup_sent=False) is None + messages: Final = cfg.transform_realtime_request( + json.dumps(session_update), "gemini-live-2.5-flash", session_configuration_request=None + ) + assert "speechConfig" not in json.loads(messages[0])["setup"].get("generationConfig", {}) + + +@pytest.mark.parametrize("drop_params", [False, True]) +@pytest.mark.parametrize("voice", ["fable", "onyx"]) +def test_vertex_sends_the_documented_voice_an_openai_tts_voice_maps_to(voice, drop_params): + cfg: Final = VertexAIRealtimeConfig( + access_token="tok", project="my-proj", location="us-central1", drop_params=drop_params + ) + session_update: Final = {"type": "session.update", "session": {"voice": voice}} + + assert cfg.setup_session_update_rejection("gemini-live-2.5-flash", session_update["session"], setup_sent=False) is None + messages: Final = cfg.transform_realtime_request( + json.dumps(session_update), "gemini-live-2.5-flash", session_configuration_request=None + ) + assert json.loads(messages[0])["setup"]["generationConfig"]["speechConfig"] == { + "voiceConfig": {"prebuiltVoiceConfig": {"voiceName": OPENAI_TO_GEMINI_TTS_VOICES[voice]}} + } diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.test.tsx index 117796d658e..56135c779e9 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.test.tsx @@ -173,6 +173,24 @@ describe("RealtimePlayground", () => { expect(update.session.type).toBe("realtime"); }); + it("starts recording on the configuration the session was created with", async () => { + const user = userEvent.setup(); + render(); + + await connect(user); + await act(async () => { + latestSocket().emit({ type: "session.created" }); + }); + await user.click(screen.getByTitle("Start recording")); + + const updates = latestSocket() + .sent.map((raw) => JSON.parse(raw)) + .filter((payload) => payload.type === "session.update"); + expect(updates).toHaveLength(2); + expect(updates[1]).toEqual(updates[0]); + expect(updates[0].session.turn_detection).toEqual({ type: "server_vad" }); + }); + it("sends what was typed and then asks for a response", async () => { const user = userEvent.setup(); render(); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.tsx b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.tsx index 01151203e65..408cbd58593 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/playground/components/chat_ui/RealtimePlayground.tsx @@ -22,6 +22,20 @@ interface RealtimePlaygroundProps { selectedGuardrails?: string[]; } +const sessionUpdate = (voice: string) => + JSON.stringify({ + type: "session.update", + session: { + type: "realtime", + modalities: ["text", "audio"], + voice, + input_audio_format: "pcm16", + output_audio_format: "pcm16", + input_audio_transcription: { model: "gpt-4o-mini-transcribe" }, + turn_detection: { type: "server_vad" }, + }, + }); + const RealtimePlayground: React.FC = ({ accessToken, selectedModel, @@ -125,21 +139,7 @@ const RealtimePlayground: React.FC = ({ const type = data.type; if (type === "session.created") { - // GA: session.type is required ("realtime" | "transcription") - ws.send( - JSON.stringify({ - type: "session.update", - session: { - type: "realtime", - modalities: ["text", "audio"], - voice: selectedVoice, - input_audio_format: "pcm16", - output_audio_format: "pcm16", - input_audio_transcription: { model: "gpt-4o-mini-transcribe" }, - turn_detection: null, - }, - }), - ); + ws.send(sessionUpdate(selectedVoice)); } else if (type === "session.updated") { // session configured } else if ( @@ -231,22 +231,7 @@ const RealtimePlayground: React.FC = ({ const startRecording = useCallback(async () => { if (!wsRef.current || wsRef.current.readyState !== WebSocket.OPEN) return; - // Switch to server VAD mode for voice input - // GA: session.type is required - wsRef.current.send( - JSON.stringify({ - type: "session.update", - session: { - type: "realtime", - modalities: ["text", "audio"], - voice: selectedVoice, - input_audio_format: "pcm16", - output_audio_format: "pcm16", - input_audio_transcription: { model: "gpt-4o-mini-transcribe" }, - turn_detection: { type: "server_vad" }, - }, - }), - ); + wsRef.current.send(sessionUpdate(selectedVoice)); try { const stream = await navigator.mediaDevices.getUserMedia({ audio: true }); @@ -301,7 +286,7 @@ const RealtimePlayground: React.FC = ({ } catch (err: any) { addMessage("status", `Microphone error: ${err.message}`); } - }, [addMessage]); + }, [addMessage, selectedVoice]); const stopRecording = useCallback(() => { processorRef.current?.disconnect(); @@ -317,21 +302,7 @@ const RealtimePlayground: React.FC = ({ if (!wsRef.current || wsRef.current.readyState !== WebSocket.OPEN) return; if (configureSessionRef.current) return; configureSessionRef.current = true; - // GA: session.type is required - wsRef.current.send( - JSON.stringify({ - type: "session.update", - session: { - type: "realtime", - modalities: ["text", "audio"], - voice: selectedVoice, - input_audio_format: "pcm16", - output_audio_format: "pcm16", - input_audio_transcription: { model: "gpt-4o-mini-transcribe" }, - turn_detection: null, - }, - }), - ); + wsRef.current.send(sessionUpdate(selectedVoice)); }, [selectedVoice]); const sendTextMessage = useCallback(() => { From 373f1752d30c70f8d2c87022e677f4c0ec58e14e Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 30 Oct 2026 02:21:39 -0700 Subject: [PATCH 15/23] fix(anthropic): send PDF image_url content to Claude as document blocks (#46984) * fix(anthropic): convert PDF image_url content to documents Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(prompt_templates): strip content-type parameters from fetched media Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> * style(prompt_templates): format MIME normalization Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../prompt_templates/common_utils.py | 6 +- .../prompt_templates/factory.py | 29 +++----- ...llm_core_utils_prompt_templates_factory.py | 74 +++++++++++++++++++ .../litellm_core_utils/test_image_handling.py | 31 ++++++++ 4 files changed, 119 insertions(+), 21 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index b3f89b33d1d..862e61cb631 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -2023,12 +2023,14 @@ def infer_content_type_from_url_and_content( """ from litellm.litellm_core_utils.token_counter import get_image_type + media_type: Final = current_content_type.split(";", 1)[0].strip().lower() if current_content_type else None + # If we have a valid content type that's not generic, use it - if current_content_type and current_content_type not in [ + if media_type and media_type not in [ "binary/octet-stream", "application/octet-stream", ]: - return current_content_type + return media_type # Extension to MIME type mapping # Supports images, documents, and other common file types diff --git a/litellm/litellm_core_utils/prompt_templates/factory.py b/litellm/litellm_core_utils/prompt_templates/factory.py index 834d26b4fe8..de896cde4dc 100644 --- a/litellm/litellm_core_utils/prompt_templates/factory.py +++ b/litellm/litellm_core_utils/prompt_templates/factory.py @@ -912,9 +912,9 @@ def create_anthropic_image_param( image_url_input: str | dict, format: str | None = None, is_bedrock_invoke: bool = False, -) -> AnthropicMessagesImageParam: +) -> AnthropicMessagesImageParam | AnthropicMessagesDocumentParam: """ - Create an AnthropicMessagesImageParam from an image URL input. + Create an Anthropic content block from an image URL input. Supports both URL references (for HTTP/HTTPS URLs) and base64 encoding. """ @@ -933,14 +933,7 @@ def create_anthropic_image_param( if is_bedrock_invoke or image_url.startswith("http://"): base64_url: Final = convert_url_to_base64(url=image_url) image_chunk = convert_to_anthropic_image_obj(openai_image_url=base64_url, format=format) - return AnthropicMessagesImageParam( - type="image", - source=AnthropicContentParamSource( - type="base64", - media_type=image_chunk["media_type"], - data=image_chunk["data"], - ), - ) + return _anthropic_content_element_factory(image_chunk) else: # HTTPS URL - pass directly for regular Anthropic return AnthropicMessagesImageParam( @@ -953,14 +946,7 @@ def create_anthropic_image_param( else: # Convert to base64 for data URIs or other formats image_chunk = convert_to_anthropic_image_obj(openai_image_url=image_url, format=format) - return AnthropicMessagesImageParam( - type="image", - source=AnthropicContentParamSource( - type="base64", - media_type=image_chunk["media_type"], - data=image_chunk["data"], - ), - ) + return _anthropic_content_element_factory(image_chunk) # The following XML functions will be deprecated once JSON schema support is available on Bedrock and Vertex @@ -1910,7 +1896,12 @@ def convert_to_anthropic_tool_result( anthropic_content_element=_anthropic_image_param, original_content_element=content, ) - anthropic_content_list.append(cast(AnthropicMessagesImageParam, _anthropic_image_param)) + anthropic_content_list.append( + cast( + AnthropicMessagesImageParam | AnthropicMessagesDocumentParam, + _anthropic_image_param, + ) + ) elif content["type"] == "tool_reference": anthropic_content_list.append(ToolReference(type="tool_reference", tool_name=content["tool_name"])) elif content["type"] == "file": diff --git a/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py b/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py index 53be3608979..778c9d68be5 100644 --- a/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py +++ b/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py @@ -1246,6 +1246,80 @@ def test_anthropic_messages_pt_keeps_signed_thinking_block(): assert thinking_blocks[0]["thinking"] == "genuine anthropic reasoning" +@pytest.mark.parametrize("llm_provider", ("anthropic", "vertex_ai")) +def test_anthropic_messages_pt_converts_pdf_image_url_data_uri_to_document(llm_provider: str) -> None: + pdf_data = "JVBERi0xLjQK" + result: Final = anthropic_messages_pt( + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "summarize"}, + {"type": "image_url", "image_url": {"url": f"data:application/pdf;base64,{pdf_data}"}}, + ], + } + ], + model="claude-opus-5-5", + llm_provider=llm_provider, + ) + + assert result[0]["content"][1] == { + "type": "document", + "source": {"type": "base64", "media_type": "application/pdf", "data": pdf_data}, + } + + +@pytest.mark.parametrize("llm_provider", ("anthropic", "vertex_ai")) +def test_anthropic_messages_pt_uses_pdf_format_for_generic_mime_data_uri(llm_provider: str) -> None: + pdf_data = "JVBERi0xLjQK" + result: Final = anthropic_messages_pt( + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "summarize"}, + { + "type": "image_url", + "image_url": { + "url": f"data:application/octet-stream;base64,{pdf_data}", + "format": "application/pdf", + }, + }, + ], + } + ], + model="claude-opus-5-5", + llm_provider=llm_provider, + ) + + assert result[0]["content"][1] == { + "type": "document", + "source": {"type": "base64", "media_type": "application/pdf", "data": pdf_data}, + } + + +def test_anthropic_messages_pt_keeps_png_image_url_as_image() -> None: + image_data = "iVBORw0KGgo=" + result: Final = anthropic_messages_pt( + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "describe"}, + {"type": "image_url", "image_url": f"data:image/png;base64,{image_data}"}, + ], + } + ], + model="claude-opus-5-5", + llm_provider="anthropic", + ) + + assert result[0]["content"][1] == { + "type": "image", + "source": {"type": "base64", "media_type": "image/png", "data": image_data}, + } + + EMPTY_USER_MESSAGES: Final = pytest.mark.parametrize( "empty_user_message", [{"role": "user", "content": None}, {"role": "user", "content": []}, {"role": "user"}], diff --git a/tests/unit/litellm_core_utils/test_image_handling.py b/tests/unit/litellm_core_utils/test_image_handling.py index 9d9aef0c22a..d1bebde9917 100644 --- a/tests/unit/litellm_core_utils/test_image_handling.py +++ b/tests/unit/litellm_core_utils/test_image_handling.py @@ -288,6 +288,37 @@ def test_convert_url_to_base64_keeps_a_real_content_type(monkeypatch): assert result.startswith("data:image/jpeg;base64,") +@pytest.mark.parametrize( + ("url", "content_type", "expected"), + [ + ( + "https://docs.example/document.pdf", + "application/pdf; qs=0.001", + "data:application/pdf;base64,cGRm", + ), + ( + "https://img.example/image.png", + "image/png; charset=binary", + "data:image/png;base64,cGRm", + ), + ( + "https://docs.example/document.pdf", + "application/octet-stream; charset=binary", + "data:application/pdf;base64,cGRm", + ), + ], +) +def test_convert_url_to_base64_uses_bare_content_type( + monkeypatch: pytest.MonkeyPatch, + url: str, + content_type: str, + expected: str, +) -> None: + monkeypatch.setattr(litellm, "module_level_client", _stub_image_client(b"pdf", content_type)) + + assert convert_url_to_base64(url) == expected + + def test_convert_url_to_base64_raises_when_no_content_type_is_determinable(monkeypatch): monkeypatch.setattr( litellm, From d1758cda91a663b4e00b35af0921389bc6e2c95e Mon Sep 17 00:00:00 2001 From: simonp9 <349053457+simonp9@users.noreply.github.com> Date: Fri, 30 Oct 2026 02:40:01 -0700 Subject: [PATCH 16/23] fix(router)!: keep a deployment's own model_id over the request body's (#49045) * fix(router): keep a deployment's own model_id over the request body's A request body model_id overrode the deployment's pinned model_id, so a client could move SageMaker chat and embedding calls to another inference component. The router now drops the request's model_id when the deployment sets one. * fix(router): let deployments opt back in to a client model_id and stop client component headers on SageMaker chat Drop the request's model_id after the router defaults are applied, so a default_litellm_params model_id cannot win over the deployment either, and apply the same precedence on the sync text_completion path. A deployment that lists model_id in configurable_clientside_auth_params keeps the client's value. SageMaker chat drops any case variant of a client's X-Amzn-SageMaker-Inference-Component header when the deployment sets a component * fix(router): keep a SageMaker deployment's component when it pins it through extra_headers * fix(router): read the SageMaker component header as a pin only on SageMaker deployments * fix(sagemaker): send the inference component only as a header, never as a body model_id SageMaker picks the component from X-Amzn-SageMaker-Inference-Component, so the chat container body and legacy parameters no longer carry model_id. Nova endpoints accept a fixed body field list, so a pinned Nova deployment could reject the extra field. * fix(health): probe header-pinned SageMaker deployments at their component Health checks bypass the router, so a legacy sagemaker/ deployment pinned only through its inference component header sent no component on /health while routed requests now send it. Also update the two SageMaker integration tests that still asserted the old precedence and the dropped body field. * fix(health): apply the router's model_id pin rule to health probes Health probes now share the router's pin function, so a deployment that lets clients set model_id gets no header pin and its own model_id wins without reading the header. Header keys that are not strings are skipped when reading the inference component header instead of failing the call. * fix(sagemaker): keep header keys that aren't strings and report a string opt-in in health checks The chat transformation now keeps extra_headers keys that aren't strings, so such a deployment fails the way it did before instead of answering 500, and the model_id pin checks configurable_clientside_auth_params with isinstance, so a string there reaches the health probe's own error capture again. --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/llms/sagemaker/chat/transformation.py | 14 +- litellm/llms/sagemaker/common_utils.py | 33 ++- .../sagemaker/completion/transformation.py | 6 +- litellm/proxy/health_check.py | 8 +- litellm/router.py | 7 + litellm/router_utils/common_utils.py | 11 + .../providers/test_sagemaker_chat_wire.py | 2 +- .../test_sagemaker_embedding_wire.py | 4 +- .../test_sagemaker_chat_transformation.py | 20 ++ .../test_sagemaker_completion_handler.py | 21 ++ .../proxy/test_health_check_max_tokens.py | 82 +++++++ tests/unit/test_router/test_router.py | 219 ++++++++++++++++++ 12 files changed, 414 insertions(+), 13 deletions(-) diff --git a/litellm/llms/sagemaker/chat/transformation.py b/litellm/llms/sagemaker/chat/transformation.py index 4a59d5f1802..76eafcd6bb6 100644 --- a/litellm/llms/sagemaker/chat/transformation.py +++ b/litellm/llms/sagemaker/chat/transformation.py @@ -28,7 +28,7 @@ from litellm.types.llms.openai import AllMessageValues from litellm.types.utils import LlmProviders from ...openai.chat.gpt_transformation import OpenAIGPTConfig -from ..common_utils import AWSEventStreamDecoder, SagemakerError +from ..common_utils import INFERENCE_COMPONENT_HEADER, OBJECT_KEYED_MAPPING, AWSEventStreamDecoder, SagemakerError if TYPE_CHECKING: from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj @@ -59,7 +59,12 @@ class SagemakerChatConfig(OpenAIGPTConfig, BaseAWSLLM): inference_component_name: Final = optional_params.get("model_id") if not isinstance(inference_component_name, str): return headers - return {**headers, "X-Amzn-SageMaker-Inference-Component": inference_component_name} + other_headers: Final = { + name: value + for name, value in OBJECT_KEYED_MAPPING.validate_python(headers).items() + if not isinstance(name, str) or name.lower() != INFERENCE_COMPONENT_HEADER.lower() + } + return {**other_headers, INFERENCE_COMPONENT_HEADER: inference_component_name} def transform_request( self, @@ -76,9 +81,8 @@ class SagemakerChatConfig(OpenAIGPTConfig, BaseAWSLLM): litellm_params=litellm_params, headers=headers, ) - container_request: Final = { - key: value for key, value in request.items() if key not in self.aws_authentication_params - } + header_only_params: Final = frozenset((*self.aws_authentication_params, "model_id")) + container_request: Final = {key: value for key, value in request.items() if key not in header_only_params} served_model_name: Final = litellm_params.get("hf_model_name") if not isinstance(served_model_name, str): return container_request diff --git a/litellm/llms/sagemaker/common_utils.py b/litellm/llms/sagemaker/common_utils.py index eaa805e7769..bb7719a19a1 100644 --- a/litellm/llms/sagemaker/common_utils.py +++ b/litellm/llms/sagemaker/common_utils.py @@ -14,9 +14,10 @@ from typing_extensions import ReadOnly, Required, TypedDict from litellm import verbose_logger from litellm.llms.base_llm.chat.transformation import BaseLLMException from litellm.types.utils import GenericStreamingChunk as GChunk -from litellm.types.utils import StreamingChatCompletionChunk +from litellm.types.utils import LlmProviders, StreamingChatCompletionChunk STR_KEYED_MAPPING: Final = TypeAdapter(Mapping[str, object], config=ConfigDict(hide_input_in_errors=True)) +OBJECT_KEYED_MAPPING: Final = TypeAdapter(Mapping[object, object], config=ConfigDict(hide_input_in_errors=True)) @with_config(ConfigDict(extra="allow", strict=True, hide_input_in_errors=True)) @@ -167,6 +168,36 @@ def get_sagemaker_response_stream_shape(): return _load_sagemaker_response_stream_shape() +INFERENCE_COMPONENT_HEADER: Final = "X-Amzn-SageMaker-Inference-Component" + + +_SAGEMAKER_PROVIDERS: Final = frozenset( + provider.value for provider in (LlmProviders.SAGEMAKER, LlmProviders.SAGEMAKER_CHAT, LlmProviders.SAGEMAKER_NOVA) +) + + +def _deployment_provider(litellm_params: Mapping[str, object]) -> str | None: + custom_llm_provider: Final = litellm_params.get("custom_llm_provider") + if isinstance(custom_llm_provider, str) and custom_llm_provider: + return custom_llm_provider + model: Final = litellm_params.get("model") + return model.split("/", 1)[0] if isinstance(model, str) else None + + +def deployment_inference_component(litellm_params: Mapping[str, object]) -> str | None: + headers: Final = litellm_params.get("extra_headers") + if _deployment_provider(litellm_params) not in _SAGEMAKER_PROVIDERS or not isinstance(headers, Mapping): + return None + return next( + ( + value + for name, value in OBJECT_KEYED_MAPPING.validate_python(headers).items() + if isinstance(name, str) and name.lower() == INFERENCE_COMPONENT_HEADER.lower() and isinstance(value, str) + ), + None, + ) + + def with_inference_component_hint(error_message: str) -> str: if "Inference Component Name header is required" in error_message: return f"{error_message}\n pass in via `litellm.embedding(..., model_id={{InferenceComponentName}})`" diff --git a/litellm/llms/sagemaker/completion/transformation.py b/litellm/llms/sagemaker/completion/transformation.py index a69846c816f..5f6bef4246e 100644 --- a/litellm/llms/sagemaker/completion/transformation.py +++ b/litellm/llms/sagemaker/completion/transformation.py @@ -23,7 +23,7 @@ from litellm.types.llms.openai import AllMessageValues from litellm.types.utils import ModelResponse, Usage from litellm.utils import token_counter -from ..common_utils import SagemakerError +from ..common_utils import STR_KEYED_MAPPING, SagemakerError if TYPE_CHECKING: from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj @@ -153,7 +153,9 @@ class SagemakerConfig(BaseConfig): litellm_params: dict, headers: dict, ) -> dict: - inference_params: Final = optional_params.copy() + inference_params: Final = { + key: value for key, value in STR_KEYED_MAPPING.validate_python(optional_params).items() if key != "model_id" + } stream: Final = inference_params.pop("stream", False) data: Final[dict] = {"parameters": inference_params} if stream is True: diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py index 73c2fc35492..20d920f1104 100644 --- a/litellm/proxy/health_check.py +++ b/litellm/proxy/health_check.py @@ -32,7 +32,7 @@ from litellm.router_strategy.auto_router.auto_router_model_naming import ( classify_strategy_router_model, strategy_router_dependencies, ) -from litellm.router_utils.common_utils import is_model_capability_field +from litellm.router_utils.common_utils import deployment_model_id_pin, is_model_capability_field from litellm.types.utils import secret_bearing_wif_litellm_params, server_owned_wif_litellm_params # Provider routing and server-owned federation or OAuth token-exchange fields. Allowed for proxy @@ -855,6 +855,8 @@ def update_litellm_params_for_health_check( `deployment_model` (default: the incoming `litellm_params` model), so a `health_check_model` or a `model` in `health_check_params` is judged on its own. Keys already set through `health_check_params` are kept, and `modalities` set there without "audio" turns the audio fields off + - sends a SageMaker deployment's `X-Amzn-SageMaker-Inference-Component` extra header as `model_id` when the + deployment sets none, so legacy `sagemaker/` probes reach that component the way routed requests do """ incoming_model: Final = _PROBE_PARAMS_ADAPTER.validate_python(litellm_params).get("model") base_model_owner: Final = deployment_model or (incoming_model if isinstance(incoming_model, str) else None) @@ -935,9 +937,11 @@ def update_litellm_params_for_health_check( audio_output_params: Final = _health_check_audio_output_params( _MODEL_INFO_ADAPTER.validate_python(model_info), probe_params, mode, base_model_owner ) + model_id_pin: Final = deployment_model_id_pin(probe_params) + model_id_params: Final = {} if model_id_pin is None else {"model_id": model_id_pin} return { key: value - for key, value in {**audio_output_params, **probe_params}.items() + for key, value in {**audio_output_params, **model_id_params, **probe_params}.items() if not is_model_capability_field(key) } diff --git a/litellm/router.py b/litellm/router.py index 20af2e5649e..af2181cedb5 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -255,6 +255,7 @@ from litellm.router_utils.clientside_credential_handler import ( from litellm.router_utils.common_utils import ( MID_STREAM_CONTINUATION_KWARG, as_retry_skipped_deployment_ids, + deployment_model_id_pin, deployments_the_caller_may_use, filter_team_based_models, filter_web_search_deployments, @@ -5316,6 +5317,9 @@ class Router(DeprecatedRouterMembers): ): kwargs[STREAM_USAGE_INJECTED_FIELD] = True self._update_kwargs_with_default_litellm_params(kwargs=kwargs, metadata_variable_name=metadata_variable_name) + pinned_model_id: Final = deployment_model_id_pin(deployment_context) + if pinned_model_id is not None: + kwargs["model_id"] = pinned_model_id # rebind-ok: this method's contract is filling the caller's kwargs hop_kwargs: Final = _FALLBACK_HOP_ADAPTER.validate_python(kwargs) if _is_fallback_hop(hop_kwargs): EncryptedContentAffinityCheck.strip_reasoning_the_targets_cannot_decrypt( @@ -6242,6 +6246,9 @@ class Router(DeprecatedRouterMembers): kwargs[k] = v elif k == "metadata": kwargs[k].update(v) + pinned_model_id: Final = deployment_model_id_pin(_LITELLM_PARAMS_ADAPTER.validate_python(data)) + if pinned_model_id is not None: + kwargs["model_id"] = pinned_model_id # rebind-ok: this call's own kwargs, filled like the defaults # call via litellm.completion() return litellm.text_completion( diff --git a/litellm/router_utils/common_utils.py b/litellm/router_utils/common_utils.py index 2451422b5e6..a06860e7786 100644 --- a/litellm/router_utils/common_utils.py +++ b/litellm/router_utils/common_utils.py @@ -16,6 +16,7 @@ from litellm.exceptions import BadRequestError from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap from litellm.litellm_core_utils.sensitive_data_masker import mask_sensitive_structure +from litellm.llms.sagemaker.common_utils import deployment_inference_component from litellm.types.router import ( CredentialLiteLLMParams, Deployment, @@ -522,6 +523,16 @@ PROVIDER_SPECIFIC_CAPABILITY_FIELDS: Final = frozenset(ProviderSpecificModelInfo _LITELLM_PARAMS_ADAPTER: Final = TypeAdapter(Mapping[str, object]) +def deployment_model_id_pin(deployment_litellm_params: Mapping[str, object]) -> object | None: + clientside_params: Final = deployment_litellm_params.get("configurable_clientside_auth_params") + if isinstance(clientside_params, list | tuple) and "model_id" in clientside_params: + return None + own_model_id: Final = deployment_litellm_params.get("model_id") + if own_model_id is not None: + return own_model_id + return deployment_inference_component(deployment_litellm_params) + + def is_model_capability_field(name: str) -> bool: return name in PROVIDER_SPECIFIC_CAPABILITY_FIELDS or ( name.startswith("supports_") and name in GetModelCostMap.loaded_model_cost_map_fields() diff --git a/tests/integration/providers/test_sagemaker_chat_wire.py b/tests/integration/providers/test_sagemaker_chat_wire.py index fdf0d94f7c7..784e43f85e2 100644 --- a/tests/integration/providers/test_sagemaker_chat_wire.py +++ b/tests/integration/providers/test_sagemaker_chat_wire.py @@ -243,7 +243,7 @@ def test_sagemaker_chat_signs_the_inference_component_header_and_sends_hf_model_ ) assert "x-amzn-sagemaker-inference-component" in signed_headers.removeprefix("SignedHeaders=").split(";") assert JSON_OBJECT.validate_json(request.body) == { - **container_body(wire.url, PROMPT, stream=False, model_id=_INFERENCE_COMPONENT), + **container_body(wire.url, PROMPT, stream=False), "model": SERVED_MODEL, }, request.body diff --git a/tests/integration/providers/test_sagemaker_embedding_wire.py b/tests/integration/providers/test_sagemaker_embedding_wire.py index 74f3c8c8a00..332f473aeaf 100644 --- a/tests/integration/providers/test_sagemaker_embedding_wire.py +++ b/tests/integration/providers/test_sagemaker_embedding_wire.py @@ -668,12 +668,12 @@ def test_an_extra_body_that_is_not_a_populated_object_adds_nothing_to_the_body( assert _body(request) == {"input": [text]}, request.body -def test_a_caller_sent_model_id_overrides_the_deployments_inference_component(proxy: Gateway, wire: Wire) -> None: +def test_a_caller_sent_model_id_cannot_override_the_deployments_inference_component(proxy: Gateway, wire: Wire) -> None: text: Final = _text() response: Final = _embed(proxy, "sm-openai", [text], model_id="caller-chosen-component") assert response.status_code == 200, response.text request: Final = _only_request(wire) - assert request.headers[COMPONENT_HEADER] == "caller-chosen-component", dict(request.headers) + assert request.headers[COMPONENT_HEADER] == COMPONENT, dict(request.headers) assert _body(request) == {"input": [text], "model": SERVED_MODEL}, request.body diff --git a/tests/unit/llms/sagemaker/test_sagemaker_chat_transformation.py b/tests/unit/llms/sagemaker/test_sagemaker_chat_transformation.py index f8caa8d5fed..356528d8909 100644 --- a/tests/unit/llms/sagemaker/test_sagemaker_chat_transformation.py +++ b/tests/unit/llms/sagemaker/test_sagemaker_chat_transformation.py @@ -343,6 +343,26 @@ def test_body_carries_no_aws_credentials( assert client.request_headers["X-Amz-Security-Token"] == "test-session-token" +@pytest.mark.parametrize("model", ["sagemaker_chat/my-endpoint", "sagemaker_nova/my-endpoint"]) +def test_inference_component_travels_only_in_the_header(monkeypatch: pytest.MonkeyPatch, model: str): + client = _invoke_sagemaker_chat(monkeypatch, model=model, model_id="my-inference-component") + + assert client.request_headers[_INFERENCE_COMPONENT_HEADER] == "my-inference-component" + assert "model_id" not in client.request_body + + +def test_inference_component_header_keeps_header_keys_that_are_not_strings(): + headers = SagemakerChatConfig().validate_environment( + headers={7: "other", "x-amzn-sagemaker-inference-component": "client-component"}, + model="my-endpoint", + messages=[], + optional_params={"model_id": "my-inference-component"}, + litellm_params={}, + ) + + assert headers == {7: "other", _INFERENCE_COMPONENT_HEADER: "my-inference-component"} + + def test_body_model_stays_the_endpoint_name_when_hf_model_name_is_unset(monkeypatch): """Without `hf_model_name` the body must keep the model it has today.""" client = _invoke_sagemaker_chat(monkeypatch) diff --git a/tests/unit/llms/sagemaker/test_sagemaker_completion_handler.py b/tests/unit/llms/sagemaker/test_sagemaker_completion_handler.py index 0980c9f5c66..5e40c2ebb60 100644 --- a/tests/unit/llms/sagemaker/test_sagemaker_completion_handler.py +++ b/tests/unit/llms/sagemaker/test_sagemaker_completion_handler.py @@ -372,6 +372,27 @@ def test_completion_sagemaker_non_stream( assert response.choices[0].message.content == "SageMaker reply" +@pytest.mark.usefixtures("httpx_transport") +@pytest.mark.usefixtures("fake_provider_credentials") +def test_completion_sagemaker_sends_the_inference_component_only_as_a_header(respx_mock: respx.MockRouter) -> None: + url: Final = "https://runtime.sagemaker.us-west-2.amazonaws.com/endpoints/my-endpoint/invocations" + route: Final = respx_mock.post(url).mock( + return_value=httpx.Response(200, json={"generated_text": "SageMaker reply"}) + ) + litellm.completion( + model="sagemaker/my-endpoint", + messages=[{"role": "user", "content": "hi"}], + temperature=0.2, + model_id="my-inference-component", + aws_region_name="us-west-2", + num_retries=0, + ) + + request: Final = route.calls[0].request + assert request.headers["X-Amzn-SageMaker-Inference-Component"] == "my-inference-component" + assert json.loads(request.content) == {"inputs": "hi", "parameters": {"temperature": 0.2}} + + @pytest.mark.usefixtures("httpx_transport") @pytest.mark.usefixtures("fake_provider_credentials") def test_completion_sagemaker_non_stream_with_aws_params(respx_mock: respx.MockRouter) -> None: diff --git a/tests/unit/proxy/test_health_check_max_tokens.py b/tests/unit/proxy/test_health_check_max_tokens.py index df0429316ed..26d9fe2fb3d 100644 --- a/tests/unit/proxy/test_health_check_max_tokens.py +++ b/tests/unit/proxy/test_health_check_max_tokens.py @@ -1,5 +1,6 @@ import asyncio import copy +import io import json import logging from typing import Final @@ -1564,3 +1565,84 @@ async def test_a_failing_health_probe_leaves_the_deployment_serving_requests(mon assert "error" in probe assert reply.choices[0].message.content == "ok" + + +@pytest.mark.parametrize( + ("component_params", "probed_component"), + [ + ({"extra_headers": {"X-Amzn-SageMaker-Inference-Component": "header-component"}}, "header-component"), + ( + {"model_id": "own-component", "extra_headers": {"X-Amzn-SageMaker-Inference-Component": "header-component"}}, + "own-component", + ), + ( + {"model_id": "own-component", "extra_headers": {1: "other", "X-Amzn-SageMaker-Inference-Component": "x"}}, + "own-component", + ), + ( + {"extra_headers": {1: "other", "X-Amzn-SageMaker-Inference-Component": "header-component"}}, + "header-component", + ), + ( + { + "configurable_clientside_auth_params": ["model_id"], + "extra_headers": {"X-Amzn-SageMaker-Inference-Component": "header-component"}, + }, + None, + ), + ], + ids=[ + "header_pinned", + "own_model_id_beats_the_header", + "own_model_id_beside_a_non_string_header_key", + "header_pinned_beside_a_non_string_header_key", + "client_configurable_model_id_takes_no_header_pin", + ], +) +@pytest.mark.asyncio +async def test_run_model_health_check_probes_a_legacy_sagemaker_embedding_at_the_deployments_component( + component_params: dict[str, object], probed_component: str | None +) -> None: + embedding_reply: Final = {"object": "list", "data": [{"object": "embedding", "index": 0, "embedding": [0.1]}]} + invoke_endpoint: Final = MagicMock(return_value={"Body": io.BytesIO(json.dumps(embedding_reply).encode())}) + client: Final = MagicMock(invoke_endpoint=invoke_endpoint) + deployment: Final = { + "litellm_params": { + "model": "sagemaker/openai/embed-endpoint", + "aws_region_name": "us-east-1", + "aws_access_key_id": "test-key", + "aws_secret_access_key": "test-secret", + **component_params, + }, + "model_info": {"mode": "embedding"}, + } + + with patch("boto3.Session", return_value=MagicMock(client=MagicMock(return_value=client))): + result: Final = await hc_module._run_model_health_check(deployment) + + assert "error" not in result, result + assert invoke_endpoint.call_count == 1 + assert invoke_endpoint.call_args.kwargs.get("InferenceComponentName") == probed_component + + +@pytest.mark.asyncio +async def test_run_model_health_check_reports_why_a_string_configurable_clientside_auth_params_fails() -> None: + invoke_endpoint: Final = MagicMock() + deployment: Final = { + "litellm_params": { + "model": "sagemaker/openai/embed-endpoint", + "model_id": "own-component", + "aws_region_name": "us-east-1", + "aws_access_key_id": "test-key", + "aws_secret_access_key": "test-secret", + }, + "model_info": {"mode": "embedding", "health_check_params": {"configurable_clientside_auth_params": "model_id"}}, + } + + client: Final = MagicMock(invoke_endpoint=invoke_endpoint) + + with patch("boto3.Session", return_value=MagicMock(client=MagicMock(return_value=client))): + result: Final = await hc_module._run_model_health_check(deployment) + + assert "configurable_clientside_auth_params" in result["error"] + invoke_endpoint.assert_not_called() diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index a0a7dd78193..74959155380 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -4,6 +4,7 @@ import copy import functools import gc import itertools +import io import json import linecache import logging @@ -31523,6 +31524,224 @@ def test_router_keeps_supported_fallback_shapes( assert router.fallbacks == expected_fallbacks +_SAGEMAKER_ROUTER_CREDENTIALS: Final = MappingProxyType( + {"aws_region_name": "us-east-1", "aws_access_key_id": "test-key", "aws_secret_access_key": "test-secret"} +) + +_PINNED_COMPONENT: Final = MappingProxyType({"model_id": "deployment-component"}) +_HEADER_PINNED_COMPONENT: Final = MappingProxyType( + {"extra_headers": {"x-amzn-sagemaker-inference-component": "deployment-component"}} +) + +_INFERENCE_COMPONENT_PRECEDENCE: Final = pytest.mark.parametrize( + ("deployment_params", "sent_component"), + [ + (_PINNED_COMPONENT, "deployment-component"), + (_HEADER_PINNED_COMPONENT, "deployment-component"), + (MappingProxyType({}), "client-component"), + ( + MappingProxyType({**_PINNED_COMPONENT, "configurable_clientside_auth_params": ["model_id"]}), + "client-component", + ), + ], + ids=[ + "deployment_pins_the_component", + "deployment_pins_the_component_with_a_header", + "deployment_leaves_the_component_to_the_client", + "deployment_lets_the_client_override_its_component", + ], +) + + +def _sagemaker_router( + model: str, + deployment_params: Mapping[str, object], + mode: str, + router_defaults: Mapping[str, str] = MappingProxyType({}), + model_name: str = "sagemaker-model", +) -> Router: + return Router( + model_list=[ + { + "model_name": model_name, + "litellm_params": {"model": model, **_SAGEMAKER_ROUTER_CREDENTIALS, **deployment_params}, + "model_info": {"mode": mode}, + } + ], + default_litellm_params=dict(router_defaults), + ) + + +_CHAT_COMPLETION_REPLY: Final = MappingProxyType( + { + "id": "chatcmpl-1", + "object": "chat.completion", + "created": 1, + "model": "served-model", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "hi"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + } +) + + +def _sagemaker_chat_route(respx_mock: respx.MockRouter) -> respx.Route: + return respx_mock.post( + "https://runtime.sagemaker.us-east-1.amazonaws.com/endpoints/chat-endpoint/invocations" + ).respond(json=dict(_CHAT_COMPLETION_REPLY)) + + +@_INFERENCE_COMPONENT_PRECEDENCE +@pytest.mark.asyncio +async def test_a_client_model_id_cannot_move_a_sagemaker_chat_call_off_the_deployments_component( + respx_mock: respx.MockRouter, + _router_httpx_transport: None, + deployment_params: Mapping[str, object], + sent_component: str, +) -> None: + route: Final = _sagemaker_chat_route(respx_mock) + router: Final = _sagemaker_router("sagemaker_chat/chat-endpoint", deployment_params, "chat") + + await router.acompletion( + model="sagemaker-model", messages=[{"role": "user", "content": "hi"}], model_id="client-component" + ) + + assert route.calls.last.request.headers["X-Amzn-SageMaker-Inference-Component"] == sent_component + + +@_INFERENCE_COMPONENT_PRECEDENCE +def test_a_client_model_id_cannot_move_a_sync_sagemaker_text_completion_off_the_deployments_component( + respx_mock: respx.MockRouter, deployment_params: Mapping[str, object], sent_component: str +) -> None: + route: Final = _sagemaker_chat_route(respx_mock) + router: Final = _sagemaker_router( + "sagemaker_chat/chat-endpoint", deployment_params, "completion", model_name="sagemaker_chat/chat-endpoint" + ) + + router.text_completion(model="sagemaker_chat/chat-endpoint", prompt="hi", model_id="client-component") + + assert route.calls.last.request.headers["X-Amzn-SageMaker-Inference-Component"] == sent_component + + +@pytest.mark.parametrize( + "client_header", + [ + MappingProxyType({"extra_headers": {"x-amzn-sagemaker-inference-component": "client-component"}}), + MappingProxyType({"headers": {"x-amzn-sagemaker-inference-component": "client-component"}}), + MappingProxyType( + { + "provider_specific_header": { + "custom_llm_provider": "sagemaker_chat", + "extra_headers": {"x-amzn-sagemaker-inference-component": "client-component"}, + } + } + ), + ], + ids=["extra_headers", "headers", "provider_specific_header"], +) +@pytest.mark.parametrize( + "deployment_pin", [_PINNED_COMPONENT, _HEADER_PINNED_COMPONENT], ids=["model_id_pin", "header_pin"] +) +@pytest.mark.asyncio +async def test_a_client_inference_component_header_cannot_move_a_pinned_sagemaker_chat_call( + respx_mock: respx.MockRouter, + _router_httpx_transport: None, + client_header: Mapping[str, object], + deployment_pin: Mapping[str, object], +) -> None: + route: Final = _sagemaker_chat_route(respx_mock) + router: Final = _sagemaker_router("sagemaker_chat/chat-endpoint", deployment_pin, "chat") + + await router.acompletion(model="sagemaker-model", messages=[{"role": "user", "content": "hi"}], **client_header) + + assert route.calls.last.request.headers["X-Amzn-SageMaker-Inference-Component"] == "deployment-component" + + +@pytest.mark.parametrize( + "request_component", + [MappingProxyType({}), MappingProxyType({"model_id": "deployment-component"})], + ids=["request_without_model_id", "request_with_the_pinned_model_id"], +) +@pytest.mark.asyncio +async def test_a_router_default_model_id_cannot_move_a_sagemaker_call_off_the_deployments_component( + respx_mock: respx.MockRouter, _router_httpx_transport: None, request_component: Mapping[str, str] +) -> None: + route: Final = _sagemaker_chat_route(respx_mock) + router: Final = _sagemaker_router( + "sagemaker_chat/chat-endpoint", + _PINNED_COMPONENT, + "chat", + router_defaults=MappingProxyType({"model_id": "default-component"}), + ) + + await router.acompletion(model="sagemaker-model", messages=[{"role": "user", "content": "hi"}], **request_component) + + assert route.calls.last.request.headers["X-Amzn-SageMaker-Inference-Component"] == "deployment-component" + + +@_INFERENCE_COMPONENT_PRECEDENCE +@pytest.mark.asyncio +async def test_a_client_model_id_cannot_move_a_sagemaker_embedding_call_off_the_deployments_component( + deployment_params: Mapping[str, object], sent_component: str +) -> None: + component_embeddings: Final = MappingProxyType({"deployment-component": [0.1], "client-component": [0.2]}) + + def invoke_endpoint(**request: str | bytes) -> Mapping[str, io.BytesIO]: + served: Final = component_embeddings[str(request["InferenceComponentName"])] + return { + "Body": io.BytesIO( + json.dumps({"object": "list", "data": [{"object": "embedding", "index": 0, "embedding": served}]}).encode() + ) + } + + client: Final = MagicMock(invoke_endpoint=invoke_endpoint) + router: Final = _sagemaker_router("sagemaker/openai/embed-endpoint", deployment_params, "embedding") + + with patch("boto3.Session", return_value=MagicMock(client=MagicMock(return_value=client))): + response: Final = await router.aembedding(model="sagemaker-model", input=["hello"], model_id="client-component") + + assert response.data[0]["embedding"] == component_embeddings[sent_component] + + +@pytest.mark.asyncio +async def test_a_header_pin_holds_when_the_deployment_names_sagemaker_through_custom_llm_provider( + respx_mock: respx.MockRouter, _router_httpx_transport: None +) -> None: + route: Final = _sagemaker_chat_route(respx_mock) + router: Final = _sagemaker_router( + "chat-endpoint", MappingProxyType({**_HEADER_PINNED_COMPONENT, "custom_llm_provider": "sagemaker_chat"}), "chat" + ) + + await router.acompletion( + model="sagemaker-model", messages=[{"role": "user", "content": "hi"}], model_id="client-component" + ) + + assert route.calls.last.request.headers["X-Amzn-SageMaker-Inference-Component"] == "deployment-component" + + +@pytest.mark.asyncio +async def test_a_component_header_on_a_non_sagemaker_deployment_leaves_the_clients_model_id_alone( + respx_mock: respx.MockRouter, _router_httpx_transport: None +) -> None: + route: Final = respx_mock.post("https://gateway.local/v1/chat/completions").respond(json=dict(_CHAT_COMPLETION_REPLY)) + router: Final = Router( + model_list=[ + { + "model_name": "gateway-model", + "litellm_params": { + "model": "openai/served-model", + "api_base": "https://gateway.local/v1", + "api_key": "test-key", + "extra_headers": {"X-Amzn-SageMaker-Inference-Component": "gateway-component"}, + }, + } + ] + ) + + await router.acompletion(model="gateway-model", messages=[{"role": "user", "content": "hi"}], model_id="client-value") + + assert json.loads(route.calls.last.request.content)["model_id"] == "client-value" + + def test_router_warns_when_a_deployment_enforces_endpoints_without_a_list(caplog: pytest.LogCaptureFixture) -> None: with caplog.at_level(logging.WARNING, logger="LiteLLM Router"): router: Final = Router( From 080e643edc2db94cf34824fe851721c320099bd7 Mon Sep 17 00:00:00 2001 From: tessellark <349031000+tessellark@users.noreply.github.com> Date: Fri, 30 Oct 2026 03:19:25 -0700 Subject: [PATCH 17/23] fix(proxy): apply new model rows on every worker as soon as they are written (#49034) * fix(proxy): apply published model rows on every worker before the debounced resync Config sync messages now carry the ids of model rows a create, update, or upsert wrote. Every worker that receives one loads those rows by id right away, including while its debounced full resync is still waiting, so a group added one deployment at a time is complete on every worker as soon as the write returns. Deletes, bulk writes, and messages from older pods carry no ids and keep the full resync path * fix(proxy): load published model rows from the primary under the reconcile lock Read the by-id rows through the writer so a lagging read replica cannot put an old row back, and read them inside MODEL_RECONCILE_LOCK so a concurrent delete or reload stays the last word. Look the ids up in bounded chunks, reload DB credentials only when a row names one this worker lacks, and leave auto-router rows to the debounced full resync. Add an integration test that boots two proxies and needs the peer to fall back to a just-added backup within 4 s. * test(proxy): cover a published model row that names no credential * fix(proxy): keep the unknown model name read-through on the replica and outside the lock Only the by-id load that follows a published write reads the primary inside the reconcile lock. A request for a model name a worker does not know reads through the replica routing before taking the lock again, as on main, so unknown names put no load on the primary and do not wait behind a reload. * fix(proxy): keep access group edits from dropping staying deployments and batch published config ids * fix(proxy): read access group rows from the primary so quick edits are not lost to a lagging replica * fix(proxy): install ordinary rows queued with an auto router and check access group edits before writing A batch of published model ids that held one auto router skipped every row in it, so ordinary rows written at the same moment waited for the full reload. The by-id load now leaves only the auto router rows to that reload. An access group update now reads every deployment from the primary, answers 404 for an unknown group and 400 for an unknown model id or name before it writes anything, then removes the group from the deployments that leave before it adds it to the ones that join, so a failed write leaves the group narrower than asked, never wider. * fix(proxy): write an access group update in one transaction after a fresh read The update now checks model names first, reads every deployment from the primary, and commits every join and leave in one batch right after that read, so a failed write changes nothing and a field another admin set during validation is kept. Other workers hear about the rows only after the commit. * fix(proxy): lock the model rows while an access group update reads and rewrites them * test(proxy): run the model resync tests on the real installer and a real router The by-id resync tests now install rows through ProxyConfig's real installer and credential loader on a real Router, faking only the database, and the config sync test row model is frozen. * fix(proxy): lock only the rows an access group update touches and answer 404 before 400 An access group update locked every model row, so a lock on any unrelated row stalled it for 30 s and then failed it with a 500, and unrelated model edits on other proxies waited behind it. It now takes a per-group advisory lock, locks only the rows that carry the group or that it selects, and reads them again once locked. The group existence check runs before the model name check again, so a missing group answers 404 as on main. * fix(proxy): never wait on a locked row in an access group update The update now takes its row locks with NOWAIT. When another edit holds one of the group's rows, it rolls back, holding nothing, and tries again for about 10 s before answering 409. Holding rows while waiting deadlocked against team deletes and let a concurrent model edit undo the update's change. The per-group lock also moves to the two-key advisory space, apart from team ids. * fix(proxy): retry access group row locks for 30 s with SKIP LOCKED and lock only the rows model_ids select * fix(proxy): take the access group lock with pg_try_advisory_xact_lock and retry while another edit holds it --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/proxy/_lazy_openapi_snapshot.json | 2 +- .../proxy/common_utils/config_sync_pubsub.py | 136 +++- .../common_utils/registry_read_through.py | 126 ++- ...model_access_group_management_endpoints.py | 315 ++++++-- .../model_management_endpoints.py | 6 +- litellm/proxy/proxy_server.py | 2 + ...backup_deployment_serves_peer_fallbacks.py | 109 +++ .../common_utils/test_config_sync_pubsub.py | 482 +++++++++--- .../test_registry_read_through.py | 347 +++++++++ .../test_access_group_management.py | 735 +++++++++++++++++- .../test_model_management_endpoints.py | 6 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 5 +- 12 files changed, 2049 insertions(+), 222 deletions(-) create mode 100644 tests/integration/management/test_added_backup_deployment_serves_peer_fallbacks.py diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index bc64a4c7ef0..c76b9be7c33 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -1648,7 +1648,7 @@ }, "/access_group/{access_group}/update": { "put": { - "description": "Update an access group's model names.\n\nThis will:\n1. Remove the access group from all current deployments\n2. Add the access group to all deployments for the new model_names list\n\nExample:\n```bash\ncurl -X PUT 'http://localhost:4000/access_group/production-models/update' \\\n -H \"Authorization: Bearer $LITELLM_MASTER_KEY\" \\\n -H 'Content-Type: application/json' \\\n -d '{\n \"model_names\": [\"gpt-4\", \"claude-3-sonnet\"]\n }'\n```\n\nParameters:\n- access_group: str - The access group name (URL path parameter)\n- model_names: List[str] - New list of model groups to include\n\nReturns:\n- NewModelGroupResponse with the updated access group details\n\nRaises:\n- HTTPException 400: If any model names don't exist\n- HTTPException 404: If access group not found", + "description": "Update an access group's model names.\n\nThis will:\n1. Add the access group to all deployments for the new model_names list\n2. Remove the access group from every other deployment\n\nExample:\n```bash\ncurl -X PUT 'http://localhost:4000/access_group/production-models/update' \\\n -H \"Authorization: Bearer $LITELLM_MASTER_KEY\" \\\n -H 'Content-Type: application/json' \\\n -d '{\n \"model_names\": [\"gpt-4\", \"claude-3-sonnet\"]\n }'\n```\n\nParameters:\n- access_group: str - The access group name (URL path parameter)\n- model_names: List[str] - New list of model groups to include\n\nReturns:\n- NewModelGroupResponse with the updated access group details\n\nRaises:\n- HTTPException 400: If any model names don't exist\n- HTTPException 404: If access group not found\n- HTTPException 409: If another edit holds the group or one of its models for 30 s", "operationId": "update_access_group_access_group__access_group__update_put", "parameters": [ { diff --git a/litellm/proxy/common_utils/config_sync_pubsub.py b/litellm/proxy/common_utils/config_sync_pubsub.py index da74851b2d8..e9aaa0e81e4 100644 --- a/litellm/proxy/common_utils/config_sync_pubsub.py +++ b/litellm/proxy/common_utils/config_sync_pubsub.py @@ -2,9 +2,12 @@ import asyncio import json import random import time -from collections.abc import Awaitable, Callable +from collections.abc import AsyncIterator, Awaitable, Callable from dataclasses import asdict, dataclass -from typing import TYPE_CHECKING, Final, Protocol, cast # noqa: TID251 # untyped prisma/redis boundary needs cast +from itertools import chain +from typing import TYPE_CHECKING, Final, Protocol, TypeAlias, cast # noqa: TID251 # untyped prisma/redis boundary + +from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError from litellm._logging import verbose_proxy_logger from litellm.repositories.prisma_protocols import RowT_co, TableActions @@ -106,24 +109,69 @@ _pubsub_capable_client: Final = pubsub_capable_client @dataclass(frozen=True, slots=True) class _ConfigChangeMessage: object_type: str + model_ids: tuple[str, ...] = () -def _config_change_message_json(object_type: str) -> str: - return json.dumps(asdict(_ConfigChangeMessage(object_type=object_type))) +@dataclass(frozen=True, slots=True) +class _PubSubEnvelope: + data: str | bytes -async def publish_config_change(redis_cache: "RedisCache | None", object_type: str) -> None: +class _WrittenModelRow(BaseModel): + model_config = ConfigDict(from_attributes=True, frozen=True) + + model_id: str + + +_CONFIG_CHANGE_MESSAGE: Final = TypeAdapter(_ConfigChangeMessage) +_PUBSUB_ENVELOPE: Final = TypeAdapter(_PubSubEnvelope) +_MODEL_TABLE_NAME: Final = "litellm_proxymodeltable" +_TARGETED_WRITE_ACTION_NAMES: Final[frozenset[str]] = frozenset({"create", "update", "upsert"}) + +PublishConfigChange: TypeAlias = Callable[[str, tuple[str, ...]], Awaitable[None]] + + +def _config_change_message_json(object_type: str, model_ids: tuple[str, ...]) -> str: + return json.dumps(asdict(_ConfigChangeMessage(object_type=object_type, model_ids=model_ids))) + + +def parse_config_change(message: object) -> _ConfigChangeMessage | None: + try: + payload: Final = ( + message if isinstance(message, (str, bytes)) else _PUBSUB_ENVELOPE.validate_python(message).data + ) + return _CONFIG_CHANGE_MESSAGE.validate_json(payload) + except ValidationError: + return None + + +def written_model_ids(object_type: str, action_name: str, result: object) -> tuple[str, ...]: + if object_type != _MODEL_TABLE_NAME or action_name not in _TARGETED_WRITE_ACTION_NAMES: + return () + try: + return (_WrittenModelRow.model_validate(result).model_id,) + except ValidationError: + return () + + +async def publish_config_change( + redis_cache: "RedisCache | None", object_type: str, model_ids: tuple[str, ...] = () +) -> None: if redis_cache is None: return try: client: Final = pubsub_capable_client(redis_cache) - await client.publish(config_sync_channel(redis_cache), _config_change_message_json(object_type)) + await client.publish(config_sync_channel(redis_cache), _config_change_message_json(object_type, model_ids)) except Exception as e: # noqa: BLE001 # best-effort publish; writes must never fail on redis errors verbose_proxy_logger.warning("config sync publish for %s failed: %s", object_type, e) -async def publish_config_change_for_object_type(object_type: str) -> None: - await publish_config_change(redis_cache=coordination_redis_cache(), object_type=object_type) +async def publish_config_change_for_object_type(object_type: str, model_ids: tuple[str, ...] = ()) -> None: + await publish_config_change(redis_cache=coordination_redis_cache(), object_type=object_type, model_ids=model_ids) + + +async def publish_model_rows_written(model_ids: tuple[str, ...]) -> None: + await publish_config_change_for_object_type(_MODEL_TABLE_NAME, model_ids) async def publish_config_param_change(param_name: str) -> None: @@ -139,7 +187,7 @@ async def publish_config_param_change(param_name: str) -> None: class _PublishOnWriteActions: __slots__ = ("_actions", "_object_type", "_publish") - def __init__(self, actions: object, object_type: str, publish: Callable[[str], Awaitable[None]]) -> None: + def __init__(self, actions: object, object_type: str, publish: PublishConfigChange) -> None: self._actions = actions self._object_type = object_type self._publish = publish @@ -157,7 +205,7 @@ class _PublishOnWriteActions: **kwargs: object, # kwargs-ok: transparent passthrough to untyped prisma action ) -> object: result: Final = await write_action(*args, **kwargs) - await publish(object_type) + await publish(object_type, written_model_ids(object_type, name, result)) return result return _write_then_publish @@ -166,7 +214,7 @@ class _PublishOnWriteActions: def wrap_table_actions_for_config_sync( actions: "TableActions[RowT_co]", table_name: str, - publish: Callable[[str], Awaitable[None]] = publish_config_change_for_object_type, + publish: PublishConfigChange = publish_config_change_for_object_type, ) -> "TableActions[RowT_co]": if table_name not in _CONFIG_SYNCED_TABLE_NAMES: return actions @@ -174,8 +222,14 @@ def wrap_table_actions_for_config_sync( return cast("TableActions[RowT_co]", wrapped) # cast-ok: dynamic write-through proxy keeps the wrapped row type +def _published_model_ids(messages: tuple[object, ...]) -> tuple[str, ...]: + changes: Final = (parse_config_change(message) for message in messages) + return tuple(dict.fromkeys(chain.from_iterable(change.model_ids for change in changes if change is not None))) + + class ConfigSyncSubscriber: __slots__ = ( + "_apply_model_ids", "_backoff_initial_seconds", "_backoff_max_seconds", "_debounce_seconds", @@ -195,6 +249,7 @@ class ConfigSyncSubscriber: self, redis_cache: "RedisCache", resync_callbacks: tuple[Callable[[], Awaitable[None]], ...], + apply_model_ids: Callable[[tuple[str, ...]], Awaitable[None]] | None = None, debounce_seconds: float = CONFIG_SYNC_DEBOUNCE_SECONDS, jitter_max_seconds: float = CONFIG_SYNC_JITTER_MAX_SECONDS, min_resync_interval_seconds: float = CONFIG_SYNC_MIN_RESYNC_INTERVAL_SECONDS, @@ -206,6 +261,7 @@ class ConfigSyncSubscriber: ) -> None: self._redis_cache = redis_cache self._resync_callbacks = resync_callbacks + self._apply_model_ids = apply_model_ids self._debounce_seconds = debounce_seconds self._jitter_max_seconds = jitter_max_seconds self._min_resync_interval_seconds = min_resync_interval_seconds @@ -266,32 +322,64 @@ class ConfigSyncSubscriber: message = await pubsub.get_message(ignore_subscribe_messages=True, timeout=_POLL_TIMEOUT_SECONDS) if message is None: continue + await self._apply(pubsub, message) await self._debounce_and_resync(pubsub) - async def _debounce_and_resync(self, pubsub: _ConfigSyncPubSub) -> None: - await self._sleep(self._debounce_seconds + self._rng.uniform(0.0, self._jitter_max_seconds)) - await self._wait_for_min_resync_interval() + async def _debounce_and_resync(self, pubsub: ConfigSyncPubSub) -> None: + await self._apply_until( + pubsub, self._monotonic() + self._debounce_seconds + self._rng.uniform(0.0, self._jitter_max_seconds) + ) + await self._apply_until(pubsub, self._next_resync_at()) await self._drain_pending(pubsub) await self._run_resync_callbacks() self._last_resync_at = self._monotonic() - async def _wait_for_min_resync_interval(self) -> None: + def _next_resync_at(self) -> float: if self._last_resync_at is None: - return - seconds_until_next_resync = self._min_resync_interval_seconds - (self._monotonic() - self._last_resync_at) - if seconds_until_next_resync <= 0: - return - verbose_proxy_logger.debug( - "config sync resync throttled for %.1fs to cap fleet-wide reload rate", - seconds_until_next_resync, - ) - await self._sleep(seconds_until_next_resync) + return self._monotonic() + next_resync_at: Final = self._last_resync_at + self._min_resync_interval_seconds + throttled_for: Final = next_resync_at - self._monotonic() + if throttled_for > 0: + verbose_proxy_logger.debug( + "config sync resync throttled for %.1fs to cap fleet-wide reload rate", + throttled_for, + ) + return next_resync_at + + async def _apply_until(self, pubsub: ConfigSyncPubSub, deadline: float) -> None: + while True: + remaining = deadline - self._monotonic() + if remaining <= 0: + return + message = await pubsub.get_message( + ignore_subscribe_messages=True, timeout=min(remaining, _POLL_TIMEOUT_SECONDS) + ) + if message is not None: + await self._apply(pubsub, message) @staticmethod async def _drain_pending(pubsub: ConfigSyncPubSub) -> None: while await pubsub.get_message(ignore_subscribe_messages=True, timeout=0) is not None: pass + @staticmethod + async def _ready_messages(pubsub: ConfigSyncPubSub) -> AsyncIterator[object]: + while (message := await pubsub.get_message(ignore_subscribe_messages=True, timeout=0)) is not None: + yield message + + async def _apply(self, pubsub: ConfigSyncPubSub, message: object) -> None: + apply: Final = self._apply_model_ids + if apply is None: + return + ready: Final = (message, *[ready_message async for ready_message in self._ready_messages(pubsub)]) + model_ids: Final = _published_model_ids(ready) + if not model_ids: + return + try: + await apply(model_ids) + except Exception as e: # noqa: BLE001 # the debounced full resync retries what a targeted apply could not + verbose_proxy_logger.warning("config sync targeted apply of %s failed: %s", model_ids, e) + async def _run_resync_callbacks(self) -> None: for callback in self._resync_callbacks: try: diff --git a/litellm/proxy/common_utils/registry_read_through.py b/litellm/proxy/common_utils/registry_read_through.py index 22df11429c0..487c6feb68d 100644 --- a/litellm/proxy/common_utils/registry_read_through.py +++ b/litellm/proxy/common_utils/registry_read_through.py @@ -15,12 +15,13 @@ or the first routed request on this replica would pick a tier it has not loaded. import asyncio import time from collections.abc import Awaitable, Callable, Mapping, Sequence -from typing import TYPE_CHECKING, Final +from typing import TYPE_CHECKING, Final, Literal, TypeAlias from pydantic import TypeAdapter, ValidationError from litellm._logging import verbose_proxy_logger from litellm.caching.in_memory_cache import InMemoryCache +from litellm.litellm_core_utils.credential_accessor import CredentialAccessor from litellm.proxy.common_utils.encrypt_decrypt_utils import decrypt_value_helper if TYPE_CHECKING: @@ -33,8 +34,14 @@ if TYPE_CHECKING: ) from litellm.integrations.custom_guardrail import CustomGuardrail + from litellm.repositories.prisma_protocols import TableActions from litellm.types.agents import AgentResponse + _ModelTable: TypeAlias = TableActions[prisma_models.LiteLLM_ProxyModelTable] + _ModelRows: TypeAlias = Sequence[prisma_models.LiteLLM_ProxyModelTable] + +ModelLoadOutcome: TypeAlias = Literal["loaded", "missing", "no_router", "auto_router"] + READ_THROUGH_MISS_TTL_SECONDS: Final = 2.0 READ_THROUGH_RESYNC_WINDOW_SECONDS: Final = 5.0 READ_THROUGH_MAX_RESYNCS_PER_WINDOW: Final = 20 @@ -123,47 +130,114 @@ async def _resync_model_deployments(model_name: str) -> bool: return False prisma_client: Final = proxy_server.prisma_client assert prisma_client is not None - table: Final = ModelRepository(prisma_client).table - name_filter: Final[LiteLLM_ProxyModelTableWhereInput] = {"model_name": model_name} - id_filter: Final[LiteLLM_ProxyModelTableWhereInput] = {"model_id": model_name} - rows: Final = await table.find_many(where=name_filter) or await table.find_many(where=id_filter) + rows: Final = await _find_rows_by_name_or_id(model_name, ModelRepository(prisma_client).table) if not rows: return False - router: Final = proxy_server.llm_router - if router is None: - await proxy_server.proxy_config.add_deployment( - prisma_client=prisma_client, proxy_logging_obj=proxy_server.proxy_logging_obj - ) - return proxy_server.llm_router is not None - if _holds_an_auto_router(rows): - await proxy_server.proxy_config.add_deployment( - prisma_client=prisma_client, proxy_logging_obj=proxy_server.proxy_logging_obj - ) - return _model_is_loaded(model_name) async with proxy_server.MODEL_RECONCILE_LOCK: + outcome: Final = await _install_model_rows(rows, _always) + if outcome in ("no_router", "auto_router"): + await _reload_every_model() + return outcome == "loaded" or _model_is_loaded(model_name) + + +async def resync_model_deployments_by_ids(model_ids: tuple[str, ...]) -> None: + from litellm.proxy import proxy_server + from litellm.proxy.db.routing_prisma_wrapper import WriterPinnedClient + from litellm.repositories.model_repository import ModelRepository + + if not _db_backed_registries_enabled("models"): + return + prisma_client: Final = proxy_server.prisma_client + assert prisma_client is not None + writer_table: Final = ModelRepository(WriterPinnedClient(prisma_client.db)).table + async with proxy_server.MODEL_RECONCILE_LOCK: + rows: Final = await _find_rows_by_ids(model_ids, writer_table) + _ = await _install_model_rows(_without_auto_routers(rows), _names_an_unloaded_credential) + + +async def _find_rows_by_name_or_id(model_name: str, table: "_ModelTable") -> "_ModelRows": + name_filter: Final[LiteLLM_ProxyModelTableWhereInput] = {"model_name": model_name} + id_filter: Final[LiteLLM_ProxyModelTableWhereInput] = {"model_id": model_name} + return await table.find_many(where=name_filter) or await table.find_many(where=id_filter) + + +async def _find_rows_by_ids(model_ids: tuple[str, ...], table: "_ModelTable") -> "_ModelRows": + from litellm.repositories.chunked_in import find_many_in + + return list(await find_many_in(table, "model_id", model_ids)) + + +async def _install_model_rows( + rows: "_ModelRows", + needs_credentials: Callable[["_ModelRows"], bool], +) -> ModelLoadOutcome: + from litellm.proxy import proxy_server + + router: Final = proxy_server.llm_router + if not rows: + return "missing" + if router is None: + return "no_router" + if _holds_an_auto_router(rows): + return "auto_router" + if needs_credentials(rows): + prisma_client: Final = proxy_server.prisma_client + assert prisma_client is not None await proxy_server.proxy_config.get_credentials(prisma_client=prisma_client) - proxy_server.proxy_config._add_deployment(db_models=rows) - proxy_server.llm_model_list = router.registry().deployments() + proxy_server.proxy_config._add_deployment(db_models=rows) + proxy_server.llm_model_list = router.registry().deployments() + return "loaded" + + +async def _reload_every_model() -> None: + from litellm.proxy import proxy_server + + prisma_client: Final = proxy_server.prisma_client + assert prisma_client is not None + await proxy_server.proxy_config.add_deployment( + prisma_client=prisma_client, proxy_logging_obj=proxy_server.proxy_logging_obj + ) + + +def _always(rows: "_ModelRows") -> bool: return True -def _holds_an_auto_router(rows: Sequence["prisma_models.LiteLLM_ProxyModelTable"]) -> bool: +def _holds_an_auto_router(rows: "_ModelRows") -> bool: return any(_names_an_auto_router(row.litellm_params) for row in rows) +def _without_auto_routers(rows: "_ModelRows") -> "_ModelRows": + return list(row for row in rows if not _names_an_auto_router(row.litellm_params)) + + +def _names_an_unloaded_credential(rows: "_ModelRows") -> bool: + return any(_names_an_unloaded_credential_param(row.litellm_params) for row in rows) + + _STORED_LITELLM_PARAMS: Final = TypeAdapter(Mapping[str, object]) -def _names_an_auto_router(stored_litellm_params: object) -> bool: +def _decrypted_param(stored_litellm_params: object, key: str) -> str | None: try: params: Final = _STORED_LITELLM_PARAMS.validate_python(stored_litellm_params) except ValidationError: - return False - stored_model: Final = params.get("model") - if not isinstance(stored_model, str): - return False - model: Final = decrypt_value_helper(value=stored_model, key="model", return_original_value=True) - return isinstance(model, str) and model.startswith("auto_router/") + return None + stored_value: Final = params.get(key) + if not isinstance(stored_value, str): + return None + value: Final = decrypt_value_helper(value=stored_value, key=key, return_original_value=True) + return value if isinstance(value, str) else None + + +def _names_an_auto_router(stored_litellm_params: object) -> bool: + model: Final = _decrypted_param(stored_litellm_params, "model") + return model is not None and model.startswith("auto_router/") + + +def _names_an_unloaded_credential_param(stored_litellm_params: object) -> bool: + credential_name: Final = _decrypted_param(stored_litellm_params, "litellm_credential_name") + return credential_name is not None and CredentialAccessor.find_credential(credential_name) is None async def _resync_guardrails(guardrail_name: str) -> bool: diff --git a/litellm/proxy/management_endpoints/model_access_group_management_endpoints.py b/litellm/proxy/management_endpoints/model_access_group_management_endpoints.py index 2d88882a2ac..3a2de12d108 100644 --- a/litellm/proxy/management_endpoints/model_access_group_management_endpoints.py +++ b/litellm/proxy/management_endpoints/model_access_group_management_endpoints.py @@ -12,23 +12,30 @@ Endpoints here: - DELETE /access_group/{access_group}/budget - Clear an access group's shared budget """ +import asyncio +import contextlib +import itertools import json -from collections.abc import Mapping, Sequence -from datetime import datetime +import time +from collections.abc import Awaitable, Callable, Iterator, Mapping, Sequence +from datetime import datetime, timedelta from types import MappingProxyType from typing import TYPE_CHECKING, Annotated, Any, Final, Protocol from fastapi import APIRouter, Depends, HTTPException +from pydantic import BaseModel, ConfigDict, TypeAdapter from typing_extensions import ReadOnly, TypedDict from litellm._logging import verbose_proxy_logger from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.auth.user_api_key_auth import user_api_key_auth +from litellm.proxy.common_utils.config_sync_pubsub import publish_model_rows_written from litellm.proxy.common_utils.user_api_key_cache import ( UserApiKeyCache, model_access_group_cache_key, model_access_group_registry_cache_key, ) +from litellm.proxy.db.routing_prisma_wrapper import WriterPinnedClient, writer_wrapper from litellm.proxy.management_endpoints.common_utils import validate_budget_duration # Clear cache and reload models to pick up the access group changes @@ -116,6 +123,28 @@ class _ModelTableClient(Protocol): async def update(self, *, where: Mapping[str, object], data: Mapping[str, object]) -> object: ... +class _ModelTableBatchActions(Protocol): + def update(self, *, where: Mapping[str, object], data: Mapping[str, object]) -> None: ... + + +class _ModelTableBatch(Protocol): + @property + def litellm_proxymodeltable(self) -> _ModelTableBatchActions: ... + + +class _ModelTableTxActions(Protocol): + async def find_many(self) -> Sequence[_DeploymentRow]: ... + + +class _ModelTableTx(Protocol): + @property + def litellm_proxymodeltable(self) -> _ModelTableTxActions: ... + + async def query_raw(self, query: str, *args: object) -> object: ... + + def batch_(self) -> contextlib.AbstractAsyncContextManager[_ModelTableBatch]: ... + + class _BudgetRow(Protocol): @property def budget_id(self) -> str: ... @@ -168,7 +197,42 @@ class _ModelAccessGroupBudgetTableClient(Protocol): def _model_table(prisma_client: PrismaClient) -> _ModelTableClient: - return ModelRepository(prisma_client).table + return ModelRepository(WriterPinnedClient(prisma_client.db)).table + + +_TRY_LOCK_ACCESS_GROUP: Final = ( + "SELECT pg_try_advisory_xact_lock(hashtext('model_access_group'), hashtext($1)) AS locked" +) +_LOCK_MODEL_ROWS: Final = ( + 'SELECT model_id FROM "LiteLLM_ProxyModelTable" WHERE model_id = ANY($1::text[]) ' + "ORDER BY model_id FOR UPDATE SKIP LOCKED" +) +_MEMBERSHIP_TX_TIMEOUT: Final = timedelta(seconds=30) +_LOCK_RETRY_BUDGET_SECONDS: Final = _MEMBERSHIP_TX_TIMEOUT.total_seconds() +_LOCK_RETRY_BACKOFF_SECONDS: Final = (0.05, 0.1, 0.2, 0.4, 0.8) +_LOCK_RETRY_LONGEST_DELAY_SECONDS: Final = 1.0 + + +class _LockedModelRow(BaseModel): + model_config = ConfigDict(frozen=True) + + model_id: str + + +_LOCKED_MODEL_ROWS: Final = TypeAdapter(tuple[_LockedModelRow, ...]) + + +class _AccessGroupLock(BaseModel): + model_config = ConfigDict(frozen=True) + + locked: bool + + +_ACCESS_GROUP_LOCK: Final = TypeAdapter(tuple[_AccessGroupLock]) + + +def _model_table_tx(prisma_client: PrismaClient) -> contextlib.AbstractAsyncContextManager[_ModelTableTx]: + return writer_wrapper(prisma_client.db).tx(timeout=_MEMBERSHIP_TX_TIMEOUT) def _model_access_group_budget_table(prisma_client: PrismaClient) -> _ModelAccessGroupBudgetTableClient: @@ -379,11 +443,10 @@ async def _tag_deployment_with_access_group( """Write `access_group` into one deployment's model_info; returns the (model_id, updated model_info) pair when a write happened, None when the deployment already carried the group.""" - updated_model_info, was_modified = add_access_group_to_deployment( - model_info=dict(_readable_model_info_or_raise(model_id=model_id, model_info=model_info)), - access_group=access_group, + updated_model_info: Final = _model_info_with_access_group( + model_id=model_id, model_info=model_info, access_group=access_group ) - if not was_modified: + if updated_model_info is None: return None await _model_table(prisma_client).update( where={"model_id": model_id}, @@ -393,6 +456,141 @@ async def _tag_deployment_with_access_group( return (model_id, updated_model_info) +def _model_info_with_access_group(model_id: str, model_info: object, access_group: str) -> Mapping[str, object] | None: + updated_model_info, was_modified = add_access_group_to_deployment( + model_info=dict(_readable_model_info_or_raise(model_id=model_id, model_info=model_info)), + access_group=access_group, + ) + return updated_model_info if was_modified else None + + +def _model_info_without_access_group( + model_id: str, model_info: object, access_group: str +) -> Mapping[str, object] | None: + updated_model_info, was_modified = remove_access_group_from_deployment( + model_info=dict(_readable_model_info_or_raise(model_id=model_id, model_info=model_info)), + access_group=access_group, + ) + return updated_model_info if was_modified else None + + +def _membership_rewrite(deployment: _DeploymentRow, selected: bool, access_group: str) -> Mapping[str, object] | None: + if selected: + return _model_info_with_access_group( + model_id=deployment.model_id, model_info=deployment.model_info, access_group=access_group + ) + return _model_info_without_access_group( + model_id=deployment.model_id, model_info=deployment.model_info, access_group=access_group + ) + + +def _membership_rewrites( + deployments: Sequence[_DeploymentRow], selected_model_ids: frozenset[str], access_group: str +) -> tuple[tuple[str, Mapping[str, object]], ...]: + rewrites: Final = ( + ( + deployment.model_id, + _membership_rewrite( + deployment=deployment, + selected=deployment.model_id in selected_model_ids, + access_group=access_group, + ), + ) + for deployment in deployments + ) + return tuple((model_id, model_info) for model_id, model_info in rewrites if model_info is not None) + + +def _carries_access_group(model_info: object, access_group: str) -> bool: + access_groups: Final = (model_info_as_mapping(model_info) or {}).get("access_groups") + return isinstance(access_groups, list) and access_group in access_groups + + +def _model_ids_to_lock( + deployments: Sequence[_DeploymentRow], access_group: str, data: UpdateModelGroupRequest +) -> frozenset[str]: + selected_model_ids: Final = frozenset(data.model_ids or ()) + selected_model_names: Final = frozenset(() if data.model_ids else data.model_names or ()) + return frozenset( + deployment.model_id + for deployment in deployments + if deployment.model_id in selected_model_ids + or deployment.model_name in selected_model_names + or _carries_access_group(deployment.model_info, access_group) + ) + + +async def _raise_404_unless_the_primary_carries(access_group: str, prisma_client: PrismaClient) -> None: + if access_group not in _access_groups_of_deployments(await _model_table(prisma_client).find_many()): + raise _http_error(404, f"Access group '{access_group}' not found") + + +async def _rewrite_membership_unless_locked( + prisma_client: PrismaClient, access_group: str, data: UpdateModelGroupRequest +) -> tuple[tuple[_DeploymentRow, ...], tuple[tuple[str, Mapping[str, object]], ...]] | None: + async with _model_table_tx(prisma_client) as tx: + (group_lock,) = _ACCESS_GROUP_LOCK.validate_python(await tx.query_raw(_TRY_LOCK_ACCESS_GROUP, access_group)) + if not group_lock.locked: + return None + model_ids_to_lock: Final = _model_ids_to_lock( + deployments=await tx.litellm_proxymodeltable.find_many(), access_group=access_group, data=data + ) + locked_rows: Final = _LOCKED_MODEL_ROWS.validate_python( + await tx.query_raw(_LOCK_MODEL_ROWS, sorted(model_ids_to_lock)) + ) + if frozenset(row.model_id for row in locked_rows) != model_ids_to_lock: + return None + deployments: Final = tuple( + deployment + for deployment in await tx.litellm_proxymodeltable.find_many() + if deployment.model_id in model_ids_to_lock + ) + if access_group not in _access_groups_of_deployments(deployments): + raise HTTPException( + status_code=404, + detail={"error": f"Access group '{access_group}' not found"}, + ) + selected_deployments: Final = _selected_deployments_or_400(deployments=deployments, data=data) + rewrites: Final = _membership_rewrites( + deployments=deployments, + selected_model_ids=frozenset(deployment.model_id for deployment in selected_deployments), + access_group=access_group, + ) + if rewrites: + async with tx.batch_() as batch: + for model_id, model_info in rewrites: + batch.litellm_proxymodeltable.update( + where={"model_id": model_id}, data={"model_info": json.dumps(model_info)} + ) + return selected_deployments, rewrites + + +def _lock_retry_delays() -> Iterator[float]: + return itertools.chain(_LOCK_RETRY_BACKOFF_SECONDS, itertools.repeat(_LOCK_RETRY_LONGEST_DELAY_SECONDS)) + + +async def _rewrite_membership_once_unlocked( + prisma_client: PrismaClient, + access_group: str, + data: UpdateModelGroupRequest, + retry_budget_seconds: float = _LOCK_RETRY_BUDGET_SECONDS, + sleep: Callable[[float], Awaitable[None]] = asyncio.sleep, + clock: Callable[[], float] = time.monotonic, +) -> tuple[tuple[_DeploymentRow, ...], tuple[tuple[str, Mapping[str, object]], ...]]: + deadline: Final = clock() + retry_budget_seconds + for delay in _lock_retry_delays(): + if ( + rewritten := await _rewrite_membership_unless_locked( + prisma_client=prisma_client, access_group=access_group, data=data + ) + ) is not None: + return rewritten + if clock() + delay > deadline: + break + await sleep(delay) + raise _http_error(409, f"Another edit holds access group '{access_group}' or one of its models, try again") + + def _readable_model_info_or_raise(model_id: str, model_info: object) -> Mapping[str, object]: """These helpers rewrite the model_info column wholesale, so a present-but-unreadable value must refuse loudly rather than be silently replaced with a fresh object; an @@ -412,11 +610,10 @@ async def _strip_access_group_from_deployment( """Remove `access_group` from one deployment's model_info; returns the (model_id, updated model_info) pair when a write happened, None when the deployment did not carry the group.""" - updated_model_info, was_modified = remove_access_group_from_deployment( - model_info=dict(_readable_model_info_or_raise(model_id=model_id, model_info=model_info)), - access_group=access_group, + updated_model_info: Final = _model_info_without_access_group( + model_id=model_id, model_info=model_info, access_group=access_group ) - if not was_modified: + if updated_model_info is None: return None await _model_table(prisma_client).update( where={"model_id": model_id}, @@ -493,6 +690,35 @@ async def update_specific_deployments_with_access_group( return tuple(pair for pair in tagged if pair is not None) +def _selected_deployments_or_400( + deployments: Sequence[_DeploymentRow], data: UpdateModelGroupRequest +) -> tuple[_DeploymentRow, ...]: + if data.model_ids: + known_model_ids: Final = frozenset(deployment.model_id for deployment in deployments) + unknown_model_id: Final = next( + (model_id for model_id in data.model_ids if model_id not in known_model_ids), None + ) + if unknown_model_id is not None: + raise HTTPException( + status_code=400, + detail={"error": f"Deployment with model_id '{unknown_model_id}' not found in Database."}, + ) + selected_model_ids: Final = frozenset(data.model_ids) + return tuple(deployment for deployment in deployments if deployment.model_id in selected_model_ids) + model_names: Final = data.model_names or () + known_model_names: Final = frozenset(deployment.model_name for deployment in deployments) + unknown_model_name: Final = next((name for name in model_names if name not in known_model_names), None) + if unknown_model_name is not None: + raise HTTPException( + status_code=400, + detail={ + "error": f"Can't find model '{unknown_model_name}' in Database. Access group management is only supported for database models." + }, + ) + selected_model_names: Final = frozenset(model_names) + return tuple(deployment for deployment in deployments if deployment.model_name in selected_model_names) + + async def _find_deployment_or_400(model_id: str, prisma_client: PrismaClient) -> object: deployment: Final = await _model_table(prisma_client).find_unique(where={"model_id": model_id}) if deployment is None: @@ -536,10 +762,10 @@ async def get_all_access_groups_from_db( Returns: Dict[str, AccessGroupInfo]: Dictionary mapping access_group name to info """ - # Get all deployments - deployments: Final = await ModelRepository(prisma_client).table.find_many() + return dict(_access_groups_of_deployments(await ModelRepository(prisma_client).table.find_many())) - # Build access group map + +def _access_groups_of_deployments(deployments: Sequence[_DeploymentRow]) -> Mapping[str, AccessGroupInfo]: model_names_by_group: Final[dict[str, list[str]]] = {} for deployment in deployments: @@ -835,8 +1061,8 @@ async def update_access_group( Update an access group's model names. This will: - 1. Remove the access group from all current deployments - 2. Add the access group to all deployments for the new model_names list + 1. Add the access group to all deployments for the new model_names list + 2. Remove the access group from every other deployment Example: ```bash @@ -858,6 +1084,7 @@ async def update_access_group( Raises: - HTTPException 400: If any model names don't exist - HTTPException 404: If access group not found + - HTTPException 409: If another edit holds the group or one of its models for 30 s """ from litellm.proxy.proxy_server import llm_router, prisma_client @@ -879,26 +1106,14 @@ async def update_access_group( detail={"error": "Either model_names or model_ids must be provided and non-empty"}, ) - use_model_ids: Final = has_model_ids - - # Validation: Check if access group exists try: - access_groups_map: Final = await get_all_access_groups_from_db(prisma_client=prisma_client) - if access_group not in access_groups_map: - raise HTTPException( - status_code=404, - detail={"error": f"Access group '{access_group}' not found"}, - ) + await _raise_404_unless_the_primary_carries(access_group=access_group, prisma_client=prisma_client) except HTTPException: raise except Exception as e: - raise HTTPException( - status_code=500, - detail={"error": f"Failed to check access group existence: {e}"}, - ) + raise _http_error(500, f"Failed to check access group existence: {e}") - # Validation: Check if all new models exist (only if using model_names path) - if not use_model_ids and has_model_names: + if not has_model_ids and has_model_names: assert data.model_names is not None missing_models: Final = await _missing_models_after_read_through( model_names=data.model_names, @@ -912,43 +1127,19 @@ async def update_access_group( ) try: - # Step 1: Remove access group from ALL DB deployments (skip config models) - all_deployments: Final = await _model_table(prisma_client).find_many() - - stripped: Final = [ - await _strip_access_group_from_deployment( - model_id=deployment.model_id, - model_info=deployment.model_info, - access_group=access_group, - prisma_client=prisma_client, - ) - for deployment in all_deployments - ] - stripped_pairs: Final = tuple(pair for pair in stripped if pair is not None) - - # Step 2: Add access group using the appropriate method - if use_model_ids: - assert data.model_ids is not None - updated_pairs = await update_specific_deployments_with_access_group( - model_ids=data.model_ids, - access_group=access_group, - prisma_client=prisma_client, - ) - else: - assert data.model_names is not None - updated_pairs = await update_deployments_with_access_group( - model_names=data.model_names, - access_group=access_group, - prisma_client=prisma_client, - ) - models_updated: Final = len(updated_pairs) + selected_deployments, rewrites = await _rewrite_membership_once_unlocked( + prisma_client=prisma_client, access_group=access_group, data=data + ) + if rewrites: + await publish_model_rows_written(tuple(model_id for model_id, _ in rewrites)) + models_updated: Final = len(selected_deployments) # Clear cache and reload models to pick up the access group changes live_before_reload: Final = live_model_ids_snapshot() reload_outcome: Final = await clear_cache() _raise_http_if_reload_degraded_serving( before=live_before_reload, - written_models=list({**dict(stripped_pairs), **dict(updated_pairs)}.items()), + written_models=rewrites, access_group=access_group, still_desired=reload_outcome.still_desired, live_after=reload_outcome.live_after, diff --git a/litellm/proxy/management_endpoints/model_management_endpoints.py b/litellm/proxy/management_endpoints/model_management_endpoints.py index cdfe3c90182..74fa8e3b3c1 100644 --- a/litellm/proxy/management_endpoints/model_management_endpoints.py +++ b/litellm/proxy/management_endpoints/model_management_endpoints.py @@ -779,7 +779,11 @@ async def _auto_router_capability_slot( limit=limit, ) yield tables.litellm_proxymodeltable - await publish_config_change(redis_cache=coordination_redis_cache(), object_type="litellm_proxymodeltable") + await publish_config_change( + redis_cache=coordination_redis_cache(), + object_type="litellm_proxymodeltable", + model_ids=() if model_id is None else (model_id,), + ) ENFORCE_RPM_TPM_ON_MODEL_ADD_SETTING: Final = "enforce_rpm_tpm_on_model_add" diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index b0ccf74c63b..3e004f1ced7 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -549,6 +549,7 @@ from litellm.proxy.common_utils.periodic_reload_schedule import ( write_reload_interval, ) from litellm.proxy.common_utils.proxy_state import ProxyState +from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids from litellm.proxy.common_utils.reset_budget_job import ResetBudgetJob from litellm.proxy.common_utils.responses_stream_errors import ResponsesStreamErrorState from litellm.proxy.common_utils.scheduled_job_stagger import ( @@ -8750,6 +8751,7 @@ class ProxyConfig: subscriber: Final = ConfigSyncSubscriber( redis_cache=redis_cache, resync_callbacks=(_resync_config_from_db,), + apply_model_ids=resync_model_deployments_by_ids, ) self.config_sync_subscriber = subscriber subscriber.start() diff --git a/tests/integration/management/test_added_backup_deployment_serves_peer_fallbacks.py b/tests/integration/management/test_added_backup_deployment_serves_peer_fallbacks.py new file mode 100644 index 00000000000..942bed6c165 --- /dev/null +++ b/tests/integration/management/test_added_backup_deployment_serves_peer_fallbacks.py @@ -0,0 +1,109 @@ +import uuid +from collections.abc import Iterator +from contextlib import contextmanager +from pathlib import Path +from typing import Final + +import httpx +import pytest +from pydantic import TypeAdapter +from redis import Redis + +from tests.integration._support.client import Gateway, eventually, list_value, object_value, string_value +from tests.integration._support.database import scratch_database +from tests.integration._support.process import graceful_stop_seconds, owned_proxy +from tests.integration._support.redis_process import owned_redis + +pytestmark: Final = [pytest.mark.timeout(int(4 * graceful_stop_seconds() + 120))] + +COORDINATION_CONFIG: Final = Path(__file__).resolve().parents[1] / "coordination_redis_proxy_config.yaml" +CONFIG_SYNC_CHANNEL: Final = "litellm_proxy.config_change" +SUBSCRIBER_COUNTS: Final = TypeAdapter(tuple[tuple[str, int], ...]) +PERIODIC_RELOAD_SECONDS: Final = 3600 +PEER_PICKUP_SECONDS: Final = 4 +PROVIDER_KEY: Final = "integration-provider-key" + + +@contextmanager +def _proxy_pair(gateway: Gateway, directory: Path) -> Iterator[tuple[Gateway, Gateway]]: + with scratch_database("integration_model_add_sync") as database, owned_redis(directory) as coordination: + shared_stores: Final = { + "DATABASE_URL": database, + "REDIS_HOST": coordination.host, + "REDIS_PORT": str(coordination.port), + "PROXY_CONFIG_RELOAD_INTERVAL_SECONDS": str(PERIODIC_RELOAD_SECONDS), + } + with ( + owned_proxy(gateway, directory / "writer", shared_stores, config=COORDINATION_CONFIG) as writer, + owned_proxy(gateway, directory / "peer", shared_stores, config=COORDINATION_CONFIG) as peer, + Redis(host=coordination.host, port=coordination.port, socket_timeout=2) as redis, + ): + eventually( + lambda: SUBSCRIBER_COUNTS.validate_python(redis.pubsub_numsub(CONFIG_SYNC_CHANNEL))[0][1], + lambda count: count >= 2, + seconds=30, + ) + yield writer, peer + + +@contextmanager +def _failing_upstream_model(gateway: Gateway, upstream_model: str) -> Iterator[None]: + script: Final = f"{gateway.upstream_url}/__scripts/{upstream_model}" + httpx.post(script, json={"statuses": [500]}, timeout=5, trust_env=False).raise_for_status() + try: + yield + finally: + httpx.delete(script, timeout=5, trust_env=False).raise_for_status() + + +def _add_deployment(writer: Gateway, group: str, upstream_model: str, order: int) -> None: + writer.post( + "/model/new", + { + "model_name": group, + "litellm_params": { + "model": f"openai/{upstream_model}", + "api_key": PROVIDER_KEY, + "api_base": f"{writer.upstream_url}/v1", + "order": order, + "num_retries": 0, + }, + }, + ) + + +def _listed_models(proxy: Gateway) -> frozenset[str]: + return frozenset(string_value(object_value(entry)["id"]) for entry in list_value(proxy.get("/v1/models")["data"])) + + +def _chat(proxy: Gateway, group: str) -> httpx.Response: + return proxy.request( + "POST", + "/v1/chat/completions", + {"model": group, "messages": [{"role": "user", "content": f"order fallback {uuid.uuid4().hex}"}]}, + ) + + +def test_a_backup_added_on_one_proxy_catches_the_peers_failing_primary_right_away( + gateway: Gateway, tmp_path: Path +) -> None: + group: Final = f"integration-{uuid.uuid4().hex}" + primary: Final = f"integration-primary-{uuid.uuid4().hex}" + backup: Final = f"integration-backup-{uuid.uuid4().hex}" + with _proxy_pair(gateway, tmp_path) as (writer, peer), _failing_upstream_model(gateway, primary): + _add_deployment(writer, group, primary, order=1) + eventually(lambda: _listed_models(peer), lambda listed: group in listed, seconds=30) + + _add_deployment(writer, group, backup, order=2) + on_writer: Final = _chat(writer, group) + on_peer: Final = eventually( + lambda: _chat(peer, group), + lambda response: response.status_code == 200, + seconds=PEER_PICKUP_SECONDS, + return_last_on_timeout=True, + ) + + assert on_writer.status_code == 200, on_writer.text + assert object_value(on_writer.json())["model"] == backup, on_writer.text + assert on_peer.status_code == 200, on_peer.text + assert object_value(on_peer.json())["model"] == backup, on_peer.text diff --git a/tests/unit/proxy/common_utils/test_config_sync_pubsub.py b/tests/unit/proxy/common_utils/test_config_sync_pubsub.py index 92cfb6a71e5..cbb37c964c6 100644 --- a/tests/unit/proxy/common_utils/test_config_sync_pubsub.py +++ b/tests/unit/proxy/common_utils/test_config_sync_pubsub.py @@ -5,6 +5,7 @@ from typing import Callable, Coroutine, Final, Iterable, List, Optional, Tuple from unittest.mock import AsyncMock, MagicMock, patch import pytest +from pydantic import BaseModel, ConfigDict from redis.asyncio import Redis import litellm @@ -102,8 +103,8 @@ class _ScriptedPubSubClient: class _QueuePubSub: - def __init__(self, initial_messages: Iterable[str] = ()) -> None: - self.queue: "asyncio.Queue[str]" = asyncio.Queue() + def __init__(self, initial_messages: Iterable[object] = ()) -> None: + self.queue: "asyncio.Queue[object]" = asyncio.Queue() for message in initial_messages: self.queue.put_nowait(message) self.subscribed_channels: List[str] = [] @@ -112,7 +113,7 @@ class _QueuePubSub: async def subscribe(self, *channels: str) -> None: self.subscribed_channels.extend(channels) - async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> Optional[str]: + async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> Optional[object]: if timeout == 0: try: return self.queue.get_nowait() @@ -128,7 +129,7 @@ class _QueuePubSub: class _BrokenPubSub(_QueuePubSub): - async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> Optional[str]: + async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> Optional[object]: raise ConnectionError("connection lost") @@ -138,11 +139,11 @@ class _CloseFailingBrokenPubSub(_BrokenPubSub): class _EmptyPollsThenMessagePubSub(_QueuePubSub): - def __init__(self, empty_polls: int, initial_messages: Iterable[str] = ()) -> None: + def __init__(self, empty_polls: int, initial_messages: Iterable[object] = ()) -> None: super().__init__(initial_messages=initial_messages) self.remaining_empty_polls = empty_polls - async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> Optional[str]: + async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> Optional[object]: if timeout != 0 and self.remaining_empty_polls > 0: self.remaining_empty_polls -= 1 return None @@ -157,6 +158,38 @@ class _FakeClock: return self.now +class _VirtualTimePubSub(_QueuePubSub): + """Timed polls move the injected clock forward instead of waiting on the wall clock. + + A delivery scheduled at a clock time is returned by the first poll whose window + reaches it, with the clock set to that time, so a test asserts when each resync + and targeted apply ran in subscriber time. + """ + + def __init__(self, clock: _FakeClock, deliveries: Iterable[Tuple[float, object]] = ()) -> None: + super().__init__() + self._clock = clock + self.pending: List[Tuple[float, object]] = sorted(deliveries, key=lambda delivery: delivery[0]) + + async def get_message(self, *, ignore_subscribe_messages: bool, timeout: float) -> Optional[object]: + await asyncio.sleep(0) + due_by = self._clock.now + timeout + if self.pending and self.pending[0][0] <= due_by: + delivered_at, message = self.pending.pop(0) + self._clock.now = max(self._clock.now, delivered_at) + return message + self._clock.now = due_by + return None + + +def _change(*model_ids: str) -> str: + return json.dumps({"object_type": "litellm_proxymodeltable", "model_ids": list(model_ids)}) + + +def _redis_delivery(message: str) -> dict: + return {"type": "message", "pattern": None, "channel": b"litellm_proxy.config_change", "data": message.encode()} + + class _ScriptedPubSubRedisClient(Redis): def __init__(self, pubsubs: Iterable[_QueuePubSub]) -> None: self._scripted_pubsubs = iter(pubsubs) @@ -204,7 +237,16 @@ async def test_publish_sends_object_type_json_on_channel() -> None: assert len(client.published) == 1 channel, message = client.published[0] assert channel == "litellm_proxy.config_change" - assert json.loads(message) == {"object_type": "litellm_proxymodeltable"} + assert json.loads(message) == {"object_type": "litellm_proxymodeltable", "model_ids": []} + + +async def test_publish_carries_the_written_model_ids() -> None: + client = _RecordingRedisClient() + cache = _FakeRedisCache(client) + + await publish_config_change(redis_cache=cache, object_type="litellm_proxymodeltable", model_ids=("m-1", "m-2")) + + assert json.loads(client.published[0][1]) == {"object_type": "litellm_proxymodeltable", "model_ids": ["m-1", "m-2"]} async def test_publish_uses_namespaced_channel() -> None: @@ -235,7 +277,7 @@ async def test_publish_reaches_cluster_derived_pubsub_clients() -> None: await publish_config_change(redis_cache=cache, object_type="litellm_proxymodeltable") assert client.published == [ - (CONFIG_SYNC_CHANNEL, json.dumps({"object_type": "litellm_proxymodeltable"})) + (CONFIG_SYNC_CHANNEL, json.dumps({"object_type": "litellm_proxymodeltable", "model_ids": []})) ] @@ -312,29 +354,210 @@ class _MaxJitterRandom(random.Random): return b -async def test_debounce_sleep_adds_jitter_from_injected_rng() -> None: - pubsub = _QueuePubSub(initial_messages=[json.dumps({"object_type": "litellm_proxymodeltable"})]) - cache = _FakeRedisCache(_ScriptedPubSubRedisClient([pubsub])) - sleeps: List[float] = [] - fired = asyncio.Event() +def _virtual_time_subscriber( + pubsub: _VirtualTimePubSub, + clock: _FakeClock, + events: List[str], + resynced: asyncio.Event, + *, + debounce_seconds: float = 0.0, + jitter_max_seconds: float = 0.0, + min_resync_interval_seconds: float = 10.0, + rng: Optional[random.Random] = None, + expected_resyncs: int = 1, + apply_error: Optional[Exception] = None, +) -> ConfigSyncSubscriber: + async def resync() -> None: + events.append(f"resync@{clock.now}") + if sum(event.startswith("resync@") for event in events) >= expected_resyncs: + resynced.set() - async def recording_sleep(seconds: float) -> None: - sleeps.append(seconds) + async def apply(model_ids: Tuple[str, ...]) -> None: + events.append(f"apply:{','.join(model_ids)}@{clock.now}") + if apply_error is not None: + raise apply_error - subscriber = ConfigSyncSubscriber( - redis_cache=cache, - resync_callbacks=(_recording_callback([], "resync", fired),), - debounce_seconds=1.0, - jitter_max_seconds=4.0, - rng=_MaxJitterRandom(), - sleep=recording_sleep, + return ConfigSyncSubscriber( + redis_cache=_FakeRedisCache(_ScriptedPubSubRedisClient([pubsub])), + resync_callbacks=(resync,), + apply_model_ids=apply, + debounce_seconds=debounce_seconds, + jitter_max_seconds=jitter_max_seconds, + min_resync_interval_seconds=min_resync_interval_seconds, + rng=rng, + monotonic=clock, ) - subscriber.start() - await asyncio.wait_for(fired.wait(), timeout=5) - await subscriber.stop() - assert sleeps == [5.0] +async def _run_until_resynced(subscriber: ConfigSyncSubscriber, resynced: asyncio.Event) -> None: + subscriber.start() + try: + await asyncio.wait_for(resynced.wait(), timeout=5) + finally: + await subscriber.stop() + + +async def test_debounce_waits_out_the_jitter_from_injected_rng() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, _change())]) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber( + pubsub, clock, events, resynced, debounce_seconds=1.0, jitter_max_seconds=4.0, rng=_MaxJitterRandom() + ) + + await _run_until_resynced(subscriber, resynced) + + assert events == ["resync@1005.0"] + + +async def test_published_model_ids_apply_on_receipt_while_the_full_resync_still_debounces() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, _change("m-1"))]) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber( + pubsub, clock, events, resynced, debounce_seconds=1.0, jitter_max_seconds=4.0, rng=_MaxJitterRandom() + ) + + await _run_until_resynced(subscriber, resynced) + + assert events == ["apply:m-1@1000.0", "resync@1005.0"] + + +async def test_model_ids_published_inside_the_throttle_wait_apply_without_waiting() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub( + clock, deliveries=[(1000.0, _change()), (1004.0, _change("m-2")), (1006.0, _change("m-3", "m-4"))] + ) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced, expected_resyncs=2) + + await _run_until_resynced(subscriber, resynced) + + assert events == ["resync@1000.0", "apply:m-2@1004.0", "apply:m-3,m-4@1006.0", "resync@1010.0"] + + +async def test_model_ids_already_waiting_apply_together_in_one_call() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub( + clock, deliveries=[(1000.0, _change("m-1")), (1000.0, _change("m-2", "m-1")), (1000.0, _change("m-3"))] + ) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced) + + await _run_until_resynced(subscriber, resynced) + + assert events == ["apply:m-1,m-2,m-3@1000.0", "resync@1000.0"] + + +async def test_redis_shaped_message_applies_its_model_ids() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, _redis_delivery(_change("m-1")))]) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced) + + await _run_until_resynced(subscriber, resynced) + + assert events == ["apply:m-1@1000.0", "resync@1000.0"] + + +async def test_messages_without_usable_model_ids_only_schedule_the_full_resync() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub( + clock, + deliveries=[ + (1000.0, "change"), + (1004.0, json.dumps({"object_type": "litellm_proxymodeltable"})), + (1005.0, json.dumps({"object_type": "litellm_proxymodeltable", "model_ids": "m-1"})), + (1006.0, {"type": "message", "pattern": None, "channel": b"c", "data": 7}), + (1007.0, None), + ], + ) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced, expected_resyncs=2) + + await _run_until_resynced(subscriber, resynced) + + assert events == ["resync@1000.0", "resync@1010.0"] + + +async def test_failing_targeted_apply_still_runs_the_full_resync() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, _change("m-1"))]) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced, apply_error=RuntimeError("db down")) + + await _run_until_resynced(subscriber, resynced) + + assert events == ["apply:m-1@1000.0", "resync@1000.0"] + + +async def test_subscriber_without_an_apply_callback_only_resyncs() -> None: + clock = _FakeClock() + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, _change("m-1"))]) + resyncs: List[str] = [] + fired = asyncio.Event() + subscriber = ConfigSyncSubscriber( + redis_cache=_FakeRedisCache(_ScriptedPubSubRedisClient([pubsub])), + resync_callbacks=(_recording_callback(resyncs, "resync", fired),), + debounce_seconds=0.0, + jitter_max_seconds=0.0, + monotonic=clock, + ) + + await _run_until_resynced(subscriber, fired) + + assert resyncs == ["resync"] + + +async def test_peer_holds_every_published_row_before_its_next_full_resync() -> None: + publisher = _RecordingRedisClient() + cache = _FakeRedisCache(publisher) + await publish_config_change(redis_cache=cache, object_type="litellm_proxymodeltable", model_ids=("m-order-1",)) + await publish_config_change(redis_cache=cache, object_type="litellm_proxymodeltable", model_ids=("m-order-2",)) + on_the_wire = [_redis_delivery(message) for _, message in publisher.published] + clock = _FakeClock() + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, on_the_wire[0]), (1007.0, on_the_wire[1])]) + db_rows_written_at = {"m-order-1": 1000.0, "m-order-2": 1007.0} + registry: set = set() + snapshots: List[Tuple[str, float, frozenset]] = [] + resynced = asyncio.Event() + + async def apply(model_ids: Tuple[str, ...]) -> None: + registry.update(model_ids) + snapshots.append(("apply", clock.now, frozenset(registry))) + + async def resync() -> None: + registry.update(row for row, written_at in db_rows_written_at.items() if written_at <= clock.now) + snapshots.append(("resync", clock.now, frozenset(registry))) + if sum(kind == "resync" for kind, _, _ in snapshots) == 2: + resynced.set() + + subscriber = ConfigSyncSubscriber( + redis_cache=_FakeRedisCache(_ScriptedPubSubRedisClient([pubsub])), + resync_callbacks=(resync,), + apply_model_ids=apply, + debounce_seconds=1.0, + jitter_max_seconds=5.0, + min_resync_interval_seconds=10.0, + rng=_MaxJitterRandom(), + monotonic=clock, + ) + + await _run_until_resynced(subscriber, resynced) + + assert snapshots == [ + ("apply", 1000.0, frozenset({"m-order-1"})), + ("resync", 1006.0, frozenset({"m-order-1"})), + ("apply", 1007.0, frozenset({"m-order-1", "m-order-2"})), + ("resync", 1016.0, frozenset({"m-order-1", "m-order-2"})), + ] def test_default_jitter_window_is_nonzero() -> None: @@ -345,92 +568,43 @@ def test_default_min_resync_interval_caps_reload_rate() -> None: assert CONFIG_SYNC_MIN_RESYNC_INTERVAL_SECONDS > CONFIG_SYNC_JITTER_MAX_SECONDS -def _throttled_subscriber( - cache: object, - events: List[str], - fired: asyncio.Event, - clock: _FakeClock, - min_resync_interval_seconds: float = 10.0, -) -> ConfigSyncSubscriber: - async def recording_sleep(seconds: float) -> None: - events.append(f"sleep:{seconds}") - await asyncio.sleep(0) - - async def resync() -> None: - events.append("resync") - fired.set() - - return ConfigSyncSubscriber( - redis_cache=cache, - resync_callbacks=(resync,), - debounce_seconds=0.0, - jitter_max_seconds=0.0, - min_resync_interval_seconds=min_resync_interval_seconds, - sleep=recording_sleep, - monotonic=clock, - ) - - async def test_resync_arriving_inside_min_interval_waits_out_the_remainder() -> None: - pubsub = _QueuePubSub() - cache = _FakeRedisCache(_ScriptedPubSubRedisClient([pubsub])) - events: List[str] = [] - fired = asyncio.Event() clock = _FakeClock() - subscriber = _throttled_subscriber(cache=cache, events=events, fired=fired, clock=clock) + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, "change"), (1004.0, "change")]) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced, expected_resyncs=2) - subscriber.start() - pubsub.queue.put_nowait("change") - await asyncio.wait_for(fired.wait(), timeout=5) - fired.clear() - clock.now += 4.0 - pubsub.queue.put_nowait("change") - await asyncio.wait_for(fired.wait(), timeout=5) - await subscriber.stop() + await _run_until_resynced(subscriber, resynced) - assert events == ["sleep:0.0", "resync", "sleep:0.0", "sleep:6.0", "resync"] + assert events == ["resync@1000.0", "resync@1010.0"] async def test_resync_after_min_interval_elapsed_is_not_throttled() -> None: - pubsub = _QueuePubSub() - cache = _FakeRedisCache(_ScriptedPubSubRedisClient([pubsub])) - events: List[str] = [] - fired = asyncio.Event() clock = _FakeClock() - subscriber = _throttled_subscriber(cache=cache, events=events, fired=fired, clock=clock) + pubsub = _VirtualTimePubSub(clock, deliveries=[(1000.0, "change"), (1030.0, "change")]) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced, expected_resyncs=2) - subscriber.start() - pubsub.queue.put_nowait("change") - await asyncio.wait_for(fired.wait(), timeout=5) - fired.clear() - clock.now += 30.0 - pubsub.queue.put_nowait("change") - await asyncio.wait_for(fired.wait(), timeout=5) - await subscriber.stop() + await _run_until_resynced(subscriber, resynced) - assert events == ["sleep:0.0", "resync", "sleep:0.0", "resync"] + assert events == ["resync@1000.0", "resync@1030.0"] async def test_writes_during_the_throttle_wait_collapse_into_the_next_resync() -> None: - pubsub = _QueuePubSub() - cache = _FakeRedisCache(_ScriptedPubSubRedisClient([pubsub])) - events: List[str] = [] - fired = asyncio.Event() clock = _FakeClock() - subscriber = _throttled_subscriber(cache=cache, events=events, fired=fired, clock=clock) + pubsub = _VirtualTimePubSub( + clock, deliveries=[(1000.0, "change")] + [(1004.0 + offset, "change") for offset in range(5)] + ) + events: List[str] = [] + resynced = asyncio.Event() + subscriber = _virtual_time_subscriber(pubsub, clock, events, resynced, expected_resyncs=2) - subscriber.start() - pubsub.queue.put_nowait("change") - await asyncio.wait_for(fired.wait(), timeout=5) - fired.clear() - for _ in range(5): - pubsub.queue.put_nowait("change") - await asyncio.wait_for(fired.wait(), timeout=5) - await asyncio.sleep(0.1) - await subscriber.stop() + await _run_until_resynced(subscriber, resynced) - assert events.count("resync") == 2 - assert pubsub.queue.empty() + assert events == ["resync@1000.0", "resync@1010.0"] + assert pubsub.pending == [] async def test_polls_without_messages_do_not_trigger_resyncs() -> None: @@ -646,13 +820,39 @@ class _AllWritesTableActions: return action -def _recording_publish(calls: List[Tuple[str, str]]) -> Callable[[str], Coroutine[None, None, None]]: - async def publish(object_type: str) -> None: +def _recording_publish(calls: List[Tuple[str, str]]) -> Callable[[str, Tuple[str, ...]], Coroutine[None, None, None]]: + async def publish(object_type: str, model_ids: Tuple[str, ...]) -> None: calls.append(("publish", object_type)) return publish +def _recording_publish_with_ids( + calls: List[Tuple[str, Tuple[str, ...]]], +) -> Callable[[str, Tuple[str, ...]], Coroutine[None, None, None]]: + async def publish(object_type: str, model_ids: Tuple[str, ...]) -> None: + calls.append((object_type, model_ids)) + + return publish + + +class _ModelRow(BaseModel): + model_config = ConfigDict(frozen=True) + + model_id: str + model_name: str + + +class _ModelRowTableActions: + def __getattr__(self, name: str) -> Callable[..., Coroutine[None, None, object]]: + async def action(*args: object, **kwargs: object) -> object: + if name.endswith("_many"): + return 2 + return _ModelRow(model_id="m-1", model_name="gpt-5.2") + + return action + + def test_wrapper_passes_through_unsynced_tables() -> None: actions = object() @@ -722,6 +922,46 @@ async def test_wrapper_publishes_for_every_write_action(action_name: str) -> Non assert publish_calls == [("publish", "litellm_guardrailstable")] +@pytest.mark.parametrize( + "action_name, expected_ids", + [ + ("create", ("m-1",)), + ("update", ("m-1",)), + ("upsert", ("m-1",)), + ("delete", ()), + ("create_many", ()), + ("update_many", ()), + ("delete_many", ()), + ], +) +async def test_wrapper_publishes_the_written_model_id_for_row_writes( + action_name: str, expected_ids: Tuple[str, ...] +) -> None: + publish_calls: List[Tuple[str, Tuple[str, ...]]] = [] + wrapped = wrap_table_actions_for_config_sync( + actions=_ModelRowTableActions(), + table_name="litellm_proxymodeltable", + publish=_recording_publish_with_ids(publish_calls), + ) + + await getattr(wrapped, action_name)(data={}) + + assert publish_calls == [("litellm_proxymodeltable", expected_ids)] + + +async def test_wrapper_publishes_no_model_ids_for_other_tables() -> None: + publish_calls: List[Tuple[str, Tuple[str, ...]]] = [] + wrapped = wrap_table_actions_for_config_sync( + actions=_ModelRowTableActions(), + table_name="litellm_guardrailstable", + publish=_recording_publish_with_ids(publish_calls), + ) + + await wrapped.create(data={}) + + assert publish_calls == [("litellm_guardrailstable", ())] + + async def test_model_repository_write_publishes_via_live_coordination_cache() -> None: from litellm.proxy import proxy_server from litellm.proxy.proxy_server import _set_redis_usage_cache @@ -747,7 +987,7 @@ async def test_model_repository_write_publishes_via_live_coordination_cache() -> assert len(client.published) == 1 channel, message = client.published[0] assert channel == CONFIG_SYNC_CHANNEL - assert json.loads(message) == {"object_type": "litellm_proxymodeltable"} + assert json.loads(message) == {"object_type": "litellm_proxymodeltable", "model_ids": ["m-1"]} async def test_ui_settings_write_publishes_via_live_coordination_cache() -> None: @@ -774,7 +1014,7 @@ async def test_ui_settings_write_publishes_via_live_coordination_cache() -> None assert len(client.published) == 1 channel, message = client.published[0] assert channel == CONFIG_SYNC_CHANNEL - assert json.loads(message) == {"object_type": "litellm_uisettings"} + assert json.loads(message) == {"object_type": "litellm_uisettings", "model_ids": []} async def _publish_calls_for_invalidated_param(param_name: str) -> List[Tuple[str, str]]: @@ -799,7 +1039,7 @@ async def test_invalidate_config_param_publishes_params_a_resync_applies(param_n assert len(published) == 1 channel, message = published[0] assert channel == CONFIG_SYNC_CHANNEL - assert json.loads(message) == {"object_type": param_name} + assert json.loads(message) == {"object_type": param_name, "model_ids": []} @pytest.mark.parametrize("param_name", _STARTUP_ONLY_CONFIG_PARAM_NAMES) @@ -932,6 +1172,46 @@ async def test_proxy_config_subscriber_resyncs_deployments_only() -> None: assert subscriber._task is None +async def test_proxy_config_subscriber_applies_published_model_ids_from_the_db( + monkeypatch: pytest.MonkeyPatch, +) -> None: + from litellm.proxy import proxy_server + from litellm.proxy.proxy_server import ProxyConfig + + rows = [MagicMock()] + prisma_client = MagicMock() + prisma_client.db.litellm_proxymodeltable.find_many = AsyncMock(return_value=rows) + router = MagicMock() + router.registry.return_value.deployments.return_value = ["deployment"] + installed: List[object] = [] + + def install(db_models: object) -> None: + installed.append(db_models) + + monkeypatch.setattr(proxy_server, "prisma_client", prisma_client) + monkeypatch.setattr(proxy_server, "store_model_in_db", True) + monkeypatch.setattr(proxy_server, "llm_router", router) + monkeypatch.setattr(proxy_server, "llm_model_list", None) + monkeypatch.setattr(proxy_server.proxy_config, "get_credentials", AsyncMock()) + monkeypatch.setattr(proxy_server.proxy_config, "_add_deployment", install) + cache = _FakeRedisCache(_ScriptedPubSubRedisClient([_QueuePubSub()])) + config = ProxyConfig() + + config.start_config_sync_subscriber(prisma_client=prisma_client, proxy_logging_obj=MagicMock(), redis_cache=cache) + subscriber = config.config_sync_subscriber + assert subscriber is not None + apply = subscriber._apply_model_ids + assert apply is not None + await apply(("m-1", "m-2")) + await config.stop_config_sync_subscriber() + + prisma_client.db.litellm_proxymodeltable.find_many.assert_awaited_once_with( + where={"model_id": {"in": ["m-1", "m-2"]}} + ) + assert installed == [rows] + assert proxy_server.llm_model_list == ["deployment"] + + async def test_proxy_config_does_not_start_subscriber_without_coordination_redis() -> None: from litellm.proxy.proxy_server import ProxyConfig diff --git a/tests/unit/proxy/common_utils/test_registry_read_through.py b/tests/unit/proxy/common_utils/test_registry_read_through.py index 04729167707..da7b1f497c9 100644 --- a/tests/unit/proxy/common_utils/test_registry_read_through.py +++ b/tests/unit/proxy/common_utils/test_registry_read_through.py @@ -1,4 +1,5 @@ import asyncio +from collections.abc import Mapping, Sequence from typing import TYPE_CHECKING, Final import pytest @@ -6,7 +7,10 @@ import pytest from litellm.proxy.common_utils.registry_read_through import RegistryReadThrough if TYPE_CHECKING: + from litellm import Router + from litellm.proxy._types import LiteLLM_ProxyModelTable from litellm.proxy.agent_endpoints.agent_registry import AgentRegistry + from litellm.types.utils import CredentialItem def nothing_loaded(_key: str) -> bool: @@ -487,6 +491,178 @@ async def test_resync_model_deployments_loads_db_credentials_before_reconciling_ installed.assert_called_once_with(db_models=rows, credential={"api_key": "sk-from-db"}) +@pytest.fixture +def stored_params_salt(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setenv("LITELLM_SALT_KEY", "sk-salt-stored-model-rows") + + +def _stored_model_row(model_id: str, stored_params: Mapping[str, str]) -> "LiteLLM_ProxyModelTable": + from litellm.proxy._types import LiteLLM_ProxyModelTable + from litellm.proxy.common_utils.encrypt_decrypt_utils import encrypt_value_helper + + return LiteLLM_ProxyModelTable( + model_id=model_id, + model_name="gpt-failover-group", + litellm_params={key: encrypt_value_helper(value) for key, value in stored_params.items()}, + model_info={"id": model_id}, + ) + + +def _backup_row(model_id: str, model: str = "openai/gpt-backup") -> "LiteLLM_ProxyModelTable": + return _stored_model_row(model_id, {"model": model, "api_key": "sk-backup"}) + + +def _router_serving(*model_ids: str) -> "Router": + from litellm import Router + + return Router( + model_list=[ + { + "model_name": "gpt-failover-group", + "litellm_params": {"model": "openai/gpt-primary", "api_key": "sk-primary"}, + "model_info": {"id": model_id}, + } + for model_id in model_ids + ] + ) + + +def _served_ids(deployments: "Sequence[Mapping[str, Mapping[str, object]]] | None") -> frozenset[object]: + return frozenset(deployment["model_info"]["id"] for deployment in deployments or ()) + + +class _ProxyModelTable: + def __init__(self, rows: "tuple[LiteLLM_ProxyModelTable, ...]") -> None: + self.rows = rows + self.reads = 0 + + async def find_many( + self, where: Mapping[str, Mapping[str, Sequence[str]]] + ) -> "tuple[LiteLLM_ProxyModelTable, ...]": + self.reads += 1 + return tuple(row for row in self.rows if row.model_id in where["model_id"]["in"]) + + async def create(self, data: Mapping[str, object]) -> None: + raise AssertionError(f"a model load wrote a row: {data}") + + +class _CredentialsTable: + def __init__(self, rows: "tuple[CredentialItem, ...]") -> None: + self.rows = rows + + async def find_many(self) -> "tuple[CredentialItem, ...]": + return self.rows + + +class _Db: + def __init__( + self, litellm_proxymodeltable: _ProxyModelTable, litellm_credentialstable: _CredentialsTable | None = None + ) -> None: + self.litellm_proxymodeltable = litellm_proxymodeltable + self.litellm_credentialstable = litellm_credentialstable or _CredentialsTable(()) + self.other_tables_read: list[str] = [] + + def __getattr__(self, table_name: str) -> object: + self.other_tables_read.append(table_name) + raise AttributeError(table_name) + + +class _PrismaClientWithoutConfigRows: + def __init__(self, db: object) -> None: + self.db = db + + async def get_generic_data(self, key: str, value: str, table_name: str) -> None: + return None + + +def _serve_db_models(monkeypatch: pytest.MonkeyPatch, db: object, router: "Router | None") -> None: + import litellm.proxy.proxy_server as proxy_server + + monkeypatch.setattr(proxy_server, "prisma_client", _PrismaClientWithoutConfigRows(db)) + monkeypatch.setattr(proxy_server, "store_model_in_db", True) + monkeypatch.setattr(proxy_server, "llm_router", router) + monkeypatch.setattr(proxy_server, "llm_model_list", None) + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("stored_params_salt") +async def test_resync_model_deployments_by_ids_installs_only_the_published_rows( + monkeypatch: pytest.MonkeyPatch, +) -> None: + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + + router: Final = _router_serving("primary-id") + _serve_db_models( + monkeypatch, _Db(_ProxyModelTable((_backup_row("m-1"), _backup_row("m-2"), _backup_row("m-3")))), router + ) + + await resync_model_deployments_by_ids(("m-1", "m-2")) + + assert _served_ids(router.registry().deployments()) == {"primary-id", "m-1", "m-2"} + assert _served_ids(proxy_server.llm_model_list) == {"primary-id", "m-1", "m-2"} + assert not proxy_server.MODEL_RECONCILE_LOCK.locked() + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("stored_params_salt") +async def test_resync_model_deployments_by_ids_leaves_the_router_alone_for_deleted_rows( + monkeypatch: pytest.MonkeyPatch, +) -> None: + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + + router: Final = _router_serving("primary-id") + _serve_db_models(monkeypatch, _Db(_ProxyModelTable(())), router) + + await resync_model_deployments_by_ids(("deleted-id",)) + + assert _served_ids(router.registry().deployments()) == {"primary-id"} + assert proxy_server.llm_model_list is None + assert not proxy_server.MODEL_RECONCILE_LOCK.locked() + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("stored_params_salt") +async def test_resync_model_deployments_by_ids_without_a_router_leaves_the_reload_to_the_debounced_resync( + monkeypatch: pytest.MonkeyPatch, +) -> None: + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + + table: Final = _ProxyModelTable((_backup_row("m-1"),)) + db: Final = _Db(table) + _serve_db_models(monkeypatch, db, None) + + await resync_model_deployments_by_ids(("m-1",)) + + assert proxy_server.llm_router is None + assert table.reads == 1 + assert db.other_tables_read == [] + + +@pytest.mark.asyncio +async def test_resync_model_deployments_by_ids_respects_supported_db_objects( + monkeypatch: pytest.MonkeyPatch, +) -> None: + from unittest.mock import AsyncMock, MagicMock + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + + prisma_client: Final = MagicMock() + prisma_client.db.litellm_proxymodeltable.find_many = AsyncMock( + side_effect=AssertionError("db hit for an object type this replica does not load") + ) + monkeypatch.setattr(proxy_server, "prisma_client", prisma_client) + monkeypatch.setattr(proxy_server, "store_model_in_db", True) + monkeypatch.setattr(proxy_server, "general_settings", {"supported_db_objects": ["guardrails"]}) + + await resync_model_deployments_by_ids(("gated-out-id",)) + + prisma_client.db.litellm_proxymodeltable.find_many.assert_not_awaited() + + @pytest.mark.asyncio async def test_resync_model_deployments_respects_supported_db_objects(monkeypatch): from unittest.mock import AsyncMock, MagicMock @@ -786,3 +962,174 @@ async def test_resync_model_deployments_reconciles_every_db_row_when_the_missed_ assert await _resync_model_deployments(router_name) is router_holds_the_name_after_reconcile reconciled.assert_called_once_with(prisma_client=prisma_client, proxy_logging_obj=proxy_server.proxy_logging_obj) + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("stored_params_salt") +async def test_resync_model_deployments_by_ids_reads_the_writer_not_a_lagging_replica( + monkeypatch: pytest.MonkeyPatch, +) -> None: + from unittest.mock import MagicMock + + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + from litellm.proxy.db.prisma_client import PrismaWrapper + from litellm.proxy.db.routing_prisma_wrapper import RoutingPrismaWrapper + + writer: Final = MagicMock(name="writer_prisma") + writer.litellm_proxymodeltable = _ProxyModelTable((_backup_row("backup-id"),)) + replica_table: Final = _ProxyModelTable(()) + replica: Final = MagicMock(name="lagging_replica_prisma") + replica.litellm_proxymodeltable = replica_table + router: Final = _router_serving("primary-id") + _serve_db_models( + monkeypatch, + RoutingPrismaWrapper( + writer=PrismaWrapper(original_prisma=writer, iam_token_db_auth=False), + reader=PrismaWrapper(original_prisma=replica, iam_token_db_auth=False), + ), + router, + ) + + await resync_model_deployments_by_ids(("backup-id",)) + + assert router.registry().has_deployment("backup-id") + assert replica_table.reads == 0 + + +@pytest.mark.asyncio +async def test_unknown_model_name_read_through_reads_the_replica_without_waiting_for_a_reload( + monkeypatch: pytest.MonkeyPatch, +) -> None: + from types import SimpleNamespace + from unittest.mock import AsyncMock, MagicMock + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.common_utils.registry_read_through import _resync_model_deployments + from litellm.proxy.db.prisma_client import PrismaWrapper + from litellm.proxy.db.routing_prisma_wrapper import RoutingPrismaWrapper + + writer: Final = MagicMock(name="writer_prisma") + writer.litellm_proxymodeltable = SimpleNamespace(find_many=AsyncMock(return_value=[]), create=AsyncMock()) + replica: Final = MagicMock(name="replica_prisma") + replica.litellm_proxymodeltable = SimpleNamespace(find_many=AsyncMock(return_value=[]), create=AsyncMock()) + prisma_client: Final = MagicMock() + prisma_client.db = RoutingPrismaWrapper( + writer=PrismaWrapper(original_prisma=writer, iam_token_db_auth=False), + reader=PrismaWrapper(original_prisma=replica, iam_token_db_auth=False), + ) + reload_in_progress: Final = asyncio.Lock() + monkeypatch.setattr(proxy_server, "prisma_client", prisma_client) + monkeypatch.setattr(proxy_server, "store_model_in_db", True) + monkeypatch.setattr(proxy_server, "llm_router", MagicMock()) + monkeypatch.setattr(proxy_server, "MODEL_RECONCILE_LOCK", reload_in_progress) + + async with reload_in_progress: + found: Final = await asyncio.wait_for(_resync_model_deployments("no-such-model"), timeout=2) + + assert found is False + writer.litellm_proxymodeltable.find_many.assert_not_awaited() + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("stored_params_salt") +async def test_resync_model_deployments_by_ids_installs_the_row_a_concurrent_reload_left_current( + monkeypatch: pytest.MonkeyPatch, +) -> None: + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + + table: Final = _ProxyModelTable((_backup_row("m-1", model="openai/gpt-before-update"),)) + router: Final = _router_serving("primary-id") + _serve_db_models(monkeypatch, _Db(table), router) + reload_in_progress: Final = asyncio.Lock() + monkeypatch.setattr(proxy_server, "MODEL_RECONCILE_LOCK", reload_in_progress) + + async with reload_in_progress: + targeted_load: Final = asyncio.create_task(resync_model_deployments_by_ids(("m-1",))) + for _ in range(5): + await asyncio.sleep(0) + assert not router.registry().has_deployment("m-1") + table.rows = (_backup_row("m-1", model="openai/gpt-after-update"),) + await targeted_load + + installed: Final = router.registry().deployment("m-1") + assert installed is not None + assert installed.litellm_params.model == "openai/gpt-after-update" + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("stored_params_salt") +@pytest.mark.parametrize( + ("stored_credential_name", "credentials_on_this_worker", "api_key_this_worker_uses"), + [ + ("shared-openai", (), "sk-from-db"), + ("shared-openai", ("another-openai",), "sk-from-db"), + ("shared-openai", ("shared-openai",), "sk-on-this-worker"), + (None, (), None), + ], +) +async def test_resync_model_deployments_by_ids_reloads_credentials_only_for_one_this_worker_lacks( + monkeypatch: pytest.MonkeyPatch, + stored_credential_name: str | None, + credentials_on_this_worker: tuple[str, ...], + api_key_this_worker_uses: str | None, +) -> None: + import litellm + from litellm.litellm_core_utils.credential_accessor import CredentialAccessor + from litellm.proxy.common_utils.encrypt_decrypt_utils import encrypt_value_helper + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + from litellm.types.utils import CredentialItem + + model_params: Final = {"model": "openai/gpt-backup"} + row: Final = _stored_model_row( + "backup-id", + model_params if stored_credential_name is None else model_params | {"litellm_credential_name": stored_credential_name}, + ) + stored_credential: Final = CredentialItem( + credential_name="shared-openai", + credential_values={"api_key": encrypt_value_helper("sk-from-db")}, + credential_info={}, + ) + router: Final = _router_serving("primary-id") + monkeypatch.setattr( + litellm, + "credential_list", + [ + CredentialItem(credential_name=name, credential_values={"api_key": "sk-on-this-worker"}, credential_info={}) + for name in credentials_on_this_worker + ], + ) + _serve_db_models(monkeypatch, _Db(_ProxyModelTable((row,)), _CredentialsTable((stored_credential,))), router) + + await resync_model_deployments_by_ids(("backup-id",)) + + credential: Final = CredentialAccessor.find_credential("shared-openai") + assert (credential.credential_values.get("api_key") if credential else None) == api_key_this_worker_uses + assert router.registry().has_deployment("backup-id") + + +@pytest.mark.asyncio +@pytest.mark.usefixtures("stored_params_salt") +async def test_resync_model_deployments_by_ids_installs_the_ordinary_rows_queued_with_an_auto_router( + monkeypatch: pytest.MonkeyPatch, +) -> None: + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.common_utils.registry_read_through import resync_model_deployments_by_ids + + table: Final = _ProxyModelTable( + ( + _stored_model_row("auto-router-id", {"model": "auto_router/complexity-router"}), + _backup_row("backup-id"), + ) + ) + db: Final = _Db(table) + router: Final = _router_serving("primary-id") + _serve_db_models(monkeypatch, db, router) + + await resync_model_deployments_by_ids(("auto-router-id", "backup-id")) + + assert _served_ids(router.registry().deployments()) == {"primary-id", "backup-id"} + assert _served_ids(proxy_server.llm_model_list) == {"primary-id", "backup-id"} + assert table.reads == 1 + assert db.other_tables_read == [] + assert not proxy_server.MODEL_RECONCILE_LOCK.locked() diff --git a/tests/unit/proxy/management_endpoints/test_access_group_management.py b/tests/unit/proxy/management_endpoints/test_access_group_management.py index c36576beaa9..d79af0af519 100644 --- a/tests/unit/proxy/management_endpoints/test_access_group_management.py +++ b/tests/unit/proxy/management_endpoints/test_access_group_management.py @@ -687,33 +687,152 @@ class _FakeAccessGroupBudgetTable: return self.rows.pop(access_group_name, None) +class _FakeClock: + def __init__(self) -> None: + self.now = 0.0 + self.slept: list[float] = [] + + def __call__(self) -> float: + return self.now + + async def sleep(self, seconds: float) -> None: + self.slept.append(seconds) + self.now += seconds + + class _FakeModelTable: - def __init__(self, journal: list[str], deployments) -> None: + def __init__(self, journal: list[str], deployments, failing_model_ids=frozenset()) -> None: self.journal = journal self.deployments = list(deployments) + self.failing_model_ids = failing_model_ids self.updates: list[tuple[dict, dict]] = [] + self.rows_locked_by_another_session: frozenset[str] = frozenset() + self.group_locked_by_another_session = False + self.edits_the_other_session_commits: dict[str, dict] = {} + self.other_session_commits_after_refusals: int | None = None + self.lock_refusals = 0 async def find_many(self, where=None, **kwargs): - return list(self.deployments) + model_names = (where or {}).get("model_name", {}).get("in") + return [d for d in self.deployments if model_names is None or d.model_name in model_names] + + def lock_group(self, wait): + if not self.group_locked_by_another_session: + return True + if wait: + raise TimeoutError("lock_timeout expired waiting for the access group lock") + self._refuse_lock() + return False + + def lock_rows(self, model_ids, skip_locked): + held = self.rows_locked_by_another_session.intersection(model_ids) + if not held: + return list(model_ids) + if not skip_locked: + raise TimeoutError(f"timed out waiting for the row lock on {sorted(held)}") + self._refuse_lock() + return [model_id for model_id in model_ids if model_id not in held] + + def _refuse_lock(self): + self.lock_refusals += 1 + if self.lock_refusals == self.other_session_commits_after_refusals: + self._other_session_commits() + + def _other_session_commits(self): + self.deployments = [ + _FakeDeployment( + model_id=d.model_id, + model_name=d.model_name, + model_info=self.edits_the_other_session_commits.get(d.model_id, d.model_info), + ) + for d in self.deployments + ] + self.rows_locked_by_another_session = frozenset() + self.group_locked_by_another_session = False async def find_unique(self, where, include=None): return next((d for d in self.deployments if d.model_id == where["model_id"]), None) async def update(self, where, data, include=None): + if where["model_id"] in self.failing_model_ids: + raise RuntimeError(f"database write to {where['model_id']} failed") self.journal.append(f"model_table.update:{where['model_id']}") self.updates.append((dict(where), dict(data))) return None + async def create(self, data, **kwargs): + raise AssertionError("access group endpoints never create deployments") + + +class _FakeModelTableBatch: + """Queues model-table updates and applies them all or none when the batch closes, like a Prisma batch.""" + + def __init__(self, model_table: _FakeModelTable) -> None: + self.model_table = model_table + self.queued: list[tuple[dict, dict]] = [] + self.litellm_proxymodeltable = SimpleNamespace(update=self._queue_update) + + def _queue_update(self, where, data): + self.queued.append((dict(where), dict(data))) + + async def __aenter__(self): + return self + + async def __aexit__(self, exc_type, exc, tb): + if exc_type is not None: + return False + failing = [where["model_id"] for where, _ in self.queued if where["model_id"] in self.model_table.failing_model_ids] + if failing: + raise RuntimeError(f"batch write to {failing} failed, nothing applied") + for where, data in self.queued: + await self.model_table.update(where=where, data=data) + return False + + +class _FakeModelTableTx: + """An interactive transaction on the primary that journals the row lock and the read it guards.""" + + def __init__(self, journal: list[str], model_table: _FakeModelTable) -> None: + self.journal = journal + self.model_table = model_table + self.litellm_proxymodeltable = SimpleNamespace(find_many=self._find_many) + + async def query_raw(self, query, *args): + if "advisory_xact_lock" in query: + self.journal.append(f"access_group.lock:{args[0]}") + return [{"locked": self.model_table.lock_group(wait="pg_try_" not in query)}] + if "FOR UPDATE" in query: + model_ids = sorted(args[0] if args else (d.model_id for d in self.model_table.deployments)) + self.journal.append(f"model_table.lock:{','.join(model_ids)}") + locked = self.model_table.lock_rows(model_ids, skip_locked="SKIP LOCKED" in query) + return [{"model_id": model_id} for model_id in locked] + self.journal.append(f"model_table.query:{query}") + return [] + + async def _find_many(self): + self.journal.append("model_table.read") + return await self.model_table.find_many() + + def batch_(self): + return _FakeModelTableBatch(self.model_table) + + async def __aenter__(self): + return self + + async def __aexit__(self, exc_type, exc, tb): + return False + class _FakePrismaClient: - def __init__(self, journal: list[str], deployments=()) -> None: + def __init__(self, journal: list[str], deployments=(), failing_model_ids=frozenset()) -> None: self.budget_table = _FakeBudgetTable(journal) self.access_group_budget_table = _FakeAccessGroupBudgetTable(journal, self.budget_table) - self.model_table = _FakeModelTable(journal, deployments) + self.model_table = _FakeModelTable(journal, deployments, failing_model_ids) self.db = SimpleNamespace( litellm_budgettable=self.budget_table, litellm_modelaccessgroupbudgettable=self.access_group_budget_table, litellm_proxymodeltable=self.model_table, + tx=lambda timeout=None: _FakeModelTableTx(journal, self.model_table), ) def jsonify_object(self, data): @@ -784,6 +903,12 @@ def _proxy_with_stubbed_reload(prisma): yield +def _serve_model_groups(router, model_groups): + router.registry.return_value.names.side_effect = lambda model_name=None, *, kind="groups", **_: ( + list(model_groups) if kind == "groups" else [] + ) + + def _eviction_journal(access_group): """Both auth cache keys, in the order a write path has to evict them.""" from litellm.proxy.common_utils.user_api_key_cache import ( @@ -1290,3 +1415,605 @@ async def test_deleting_an_access_group_that_never_had_a_budget_still_evicts(): assert response.models_updated == 1 assert prisma.access_group_budget_table.rows == {} _assert_evicted_after_write(journal, "prod-models", "access_group_budget.delete:prod-models") + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "selection", + [{"model_ids": ["stays", "joins"]}, {"model_names": ["gpt-stays", "gpt-joins"]}], +) +async def test_updating_an_access_group_writes_only_the_deployments_that_join_or_leave_it(selection): + import json + + import litellm.proxy.proxy_server as proxy_server + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ], + ) + + with _proxy_with_stubbed_reload(prisma): + _serve_model_groups(proxy_server.llm_router, ["gpt-stays", "gpt-leaves", "gpt-joins"]) + response = await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(**selection), + user_api_key_dict=_admin(), + ) + + written_access_groups = { + where["model_id"]: json.loads(data["model_info"])["access_groups"] + for where, data in prisma.model_table.updates + } + assert written_access_groups == {"joins": ["prod-models"], "leaves": []} + assert response.models_updated == 2 + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "selection", + [{"model_ids": ["stays", "joins"]}, {"model_names": ["gpt-stays", "gpt-joins"]}], +) +async def test_updating_an_access_group_reads_the_primary_not_a_lagging_replica(selection): + import json + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.db.prisma_client import PrismaWrapper + from litellm.proxy.db.routing_prisma_wrapper import RoutingPrismaWrapper + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + primary = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ], + ) + replica_before_the_last_edit = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves", access_groups=()), + _deployment(model_id="joins", model_name="gpt-joins"), + ], + ) + primary.db = RoutingPrismaWrapper( + writer=PrismaWrapper(original_prisma=primary.db, iam_token_db_auth=False), + reader=PrismaWrapper(original_prisma=replica_before_the_last_edit.db, iam_token_db_auth=False), + ) + + with _proxy_with_stubbed_reload(primary): + _serve_model_groups(proxy_server.llm_router, ["gpt-stays", "gpt-leaves", "gpt-joins"]) + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(**selection), + user_api_key_dict=_admin(), + ) + + written_access_groups = { + where["model_id"]: json.loads(data["model_info"])["access_groups"] + for where, data in primary.model_table.updates + } + assert written_access_groups == {"joins": ["prod-models"], "leaves": []} + + +@pytest.mark.asyncio +@pytest.mark.parametrize("failing_model_id", ["leaves", "joins"]) +async def test_a_failed_access_group_update_leaves_every_deployment_unchanged(failing_model_id): + from fastapi import HTTPException + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + deployments = [ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ] + prisma = _FakePrismaClient([], deployments=deployments, failing_model_ids=frozenset({failing_model_id})) + + with _proxy_with_stubbed_reload(prisma): + _serve_model_groups(proxy_server.llm_router, ["gpt-stays", "gpt-leaves", "gpt-joins"]) + with pytest.raises(HTTPException) as raised: + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays", "joins"]), + user_api_key_dict=_admin(), + ) + + assert raised.value.status_code == 500 + assert prisma.model_table.updates == [] + + +@pytest.mark.asyncio +async def test_an_access_group_update_keeps_a_model_info_field_another_admin_set_while_names_were_validated(): + import json + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ], + ) + + def another_admin_edits_joins_while_names_are_checked(model_name=None, *, kind="groups", **_): + if kind != "groups": + return [] + prisma.model_table.deployments = [ + deployment + if deployment.model_id != "joins" + else _FakeDeployment( + model_id="joins", + model_name="gpt-joins", + model_info={"access_groups": [], "description": "set by another admin"}, + ) + for deployment in prisma.model_table.deployments + ] + return ["gpt-stays", "gpt-joins"] + + with _proxy_with_stubbed_reload(prisma): + proxy_server.llm_router.registry.return_value.names.side_effect = ( + another_admin_edits_joins_while_names_are_checked + ) + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_names=["gpt-stays", "gpt-joins"]), + user_api_key_dict=_admin(), + ) + + written = {where["model_id"]: json.loads(data["model_info"]) for where, data in prisma.model_table.updates} + assert written["joins"]["description"] == "set by another admin" + assert written["joins"]["access_groups"] == ["prod-models"] + + +@pytest.mark.asyncio +async def test_an_access_group_update_locks_the_rows_it_reads_and_tells_other_workers_after_the_commit(): + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + journal: list[str] = [] + prisma = _FakePrismaClient( + journal, + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ], + ) + + async def record_publish(model_ids): + journal.append(f"publish:{','.join(sorted(model_ids))}") + + with ( + _proxy_with_stubbed_reload(prisma), + patch( # test-quality-ok: the endpoint publishes through its module-level import; Redis is the real seam + "litellm.proxy.management_endpoints.model_access_group_management_endpoints.publish_model_rows_written", + new=record_publish, + ), + ): + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays", "joins"]), + user_api_key_dict=_admin(), + ) + + assert journal == [ + "access_group.lock:prod-models", + "model_table.read", + "model_table.lock:joins,leaves,stays", + "model_table.read", + "model_table.update:leaves", + "model_table.update:joins", + "publish:joins,leaves", + ] + + +@pytest.mark.asyncio +async def test_an_access_group_update_that_changes_no_membership_writes_and_announces_nothing(): + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + journal: list[str] = [] + prisma = _FakePrismaClient( + journal, + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="also-stays", model_name="gpt-also-stays"), + _deployment(model_id="outside", model_name="gpt-outside", access_groups=()), + ], + ) + + async def record_publish(model_ids): + journal.append(f"publish:{','.join(sorted(model_ids))}") + + with ( + _proxy_with_stubbed_reload(prisma), + patch( # test-quality-ok: the endpoint publishes through its module-level import; Redis is the real seam + "litellm.proxy.management_endpoints.model_access_group_management_endpoints.publish_model_rows_written", + new=record_publish, + ), + ): + response = await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays", "also-stays"]), + user_api_key_dict=_admin(), + ) + + assert response.models_updated == 2 + assert journal == [ + "access_group.lock:prod-models", + "model_table.read", + "model_table.lock:also-stays,stays", + "model_table.read", + ] + + +@pytest.mark.asyncio +@pytest.mark.parametrize("ignored_model_names", [None, ["gpt-unrelated"]]) +async def test_an_access_group_update_goes_through_while_another_session_holds_a_model_row_outside_the_group( + ignored_model_names, +): + import json + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + _deployment(model_id="unrelated", model_name="gpt-unrelated", access_groups=("other-group",)), + ], + ) + prisma.model_table.rows_locked_by_another_session = frozenset({"unrelated"}) + + with _proxy_with_stubbed_reload(prisma): + response = await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays", "joins"], model_names=ignored_model_names), + user_api_key_dict=_admin(), + ) + + written_access_groups = { + where["model_id"]: json.loads(data["model_info"])["access_groups"] + for where, data in prisma.model_table.updates + } + assert written_access_groups == {"joins": ["prod-models"], "leaves": []} + assert response.models_updated == 2 + + +@pytest.mark.asyncio +async def test_an_access_group_update_retries_a_locked_row_and_keeps_the_edit_committed_there(): + import json + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ], + ) + prisma.model_table.rows_locked_by_another_session = frozenset({"joins"}) + prisma.model_table.edits_the_other_session_commits = { + "joins": {"access_groups": [], "description": "saved while the update waited"} + } + prisma.model_table.other_session_commits_after_refusals = 1 + + with _proxy_with_stubbed_reload(prisma): + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays", "joins"]), + user_api_key_dict=_admin(), + ) + + written = {where["model_id"]: json.loads(data["model_info"]) for where, data in prisma.model_table.updates} + assert written == {"joins": {"access_groups": ["prod-models"], "description": "saved while the update waited"}} + + +@pytest.mark.asyncio +async def test_an_access_group_emptied_while_the_update_retried_a_locked_row_answers_404_and_writes_nothing(): + from fastapi import HTTPException + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="leaves", model_name="gpt-leaves"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ], + ) + prisma.model_table.rows_locked_by_another_session = frozenset({"leaves"}) + prisma.model_table.edits_the_other_session_commits = {"leaves": {"access_groups": []}} + prisma.model_table.other_session_commits_after_refusals = 1 + + with _proxy_with_stubbed_reload(prisma), pytest.raises(HTTPException) as raised: + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["joins"]), + user_api_key_dict=_admin(), + ) + + assert raised.value.status_code == 404 + assert prisma.model_table.updates == [] + + +def _another_session_holds_a_member_row(model_table): + model_table.rows_locked_by_another_session = frozenset({"leaves"}) + + +def _another_edit_holds_the_group(model_table): + model_table.group_locked_by_another_session = True + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "hold", [_another_session_holds_a_member_row, _another_edit_holds_the_group], ids=["member_row", "group"] +) +async def test_an_access_group_update_answers_409_and_writes_nothing_while_another_session_keeps_a_lock_it_needs( + hold, +): + from fastapi import HTTPException + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + _rewrite_membership_once_unlocked, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves"), + ], + ) + hold(prisma.model_table) + clock = _FakeClock() + + with pytest.raises(HTTPException) as raised: + await _rewrite_membership_once_unlocked( + prisma_client=prisma, + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays"]), + retry_budget_seconds=0.5, + sleep=clock.sleep, + clock=clock, + ) + + assert raised.value.status_code == 409 + assert 0 < clock.now <= 0.5 + assert prisma.model_table.lock_refusals == len(clock.slept) + 1 + assert prisma.model_table.updates == [] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "hold", [_another_session_holds_a_member_row, _another_edit_holds_the_group], ids=["member_row", "group"] +) +async def test_an_access_group_update_waits_out_a_lock_held_for_over_ten_seconds_and_writes_the_edit(hold): + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + _rewrite_membership_once_unlocked, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="leaves", model_name="gpt-leaves"), + ], + ) + hold(prisma.model_table) + prisma.model_table.other_session_commits_after_refusals = 20 + clock = _FakeClock() + + _, rewrites = await _rewrite_membership_once_unlocked( + prisma_client=prisma, + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays"]), + sleep=clock.sleep, + clock=clock, + ) + + assert clock.now > 10 + assert [model_id for model_id, _ in rewrites] == ["leaves"] + assert [where["model_id"] for where, _ in prisma.model_table.updates] == ["leaves"] + + +@pytest.mark.asyncio +async def test_updating_a_missing_access_group_answers_404_before_checking_the_model_names(): + from fastapi import HTTPException + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], deployments=[_deployment(model_id="other", model_name="gpt-other", access_groups=("other-group",))] + ) + + with _proxy_with_stubbed_reload(prisma): + _serve_model_groups(proxy_server.llm_router, ["gpt-other"]) + with pytest.raises(HTTPException) as raised: + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_names=["no-such-model"]), + user_api_key_dict=_admin(), + ) + + assert raised.value.status_code == 404 + assert prisma.model_table.updates == [] + + +@pytest.mark.asyncio +async def test_an_access_group_update_answers_500_when_the_deployments_cannot_be_read(): + from fastapi import HTTPException + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient([], deployments=[_deployment(model_id="stays")]) + prisma.model_table.find_many = AsyncMock(side_effect=RuntimeError("database unreachable")) + + with _proxy_with_stubbed_reload(prisma), pytest.raises(HTTPException) as raised: + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["stays"]), + user_api_key_dict=_admin(), + ) + + assert raised.value.status_code == 500 + assert prisma.model_table.updates == [] + + +@pytest.mark.asyncio +@pytest.mark.parametrize("model_ids", [["no-such-deployment"], ["joins", "no-such-deployment"]]) +async def test_updating_an_access_group_with_an_unknown_model_id_writes_no_deployment(model_ids): + from fastapi import HTTPException + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[_deployment(model_id="stays"), _deployment(model_id="joins", access_groups=())], + ) + + with _proxy_with_stubbed_reload(prisma), pytest.raises(HTTPException) as raised: + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=model_ids), + user_api_key_dict=_admin(), + ) + + assert raised.value.status_code == 400 + assert prisma.model_table.updates == [] + + +@pytest.mark.asyncio +async def test_updating_an_access_group_no_deployment_carries_answers_404_and_writes_nothing(): + from fastapi import HTTPException + + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient([], deployments=[_deployment(model_id="other", access_groups=("other-group",))]) + + with _proxy_with_stubbed_reload(prisma), pytest.raises(HTTPException) as raised: + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_ids=["other"]), + user_api_key_dict=_admin(), + ) + + assert raised.value.status_code == 404 + assert prisma.model_table.updates == [] + + +@pytest.mark.asyncio +async def test_updating_an_access_group_with_a_model_name_only_the_config_serves_writes_no_deployment(): + from fastapi import HTTPException + + import litellm.proxy.proxy_server as proxy_server + from litellm.proxy.management_endpoints.model_access_group_management_endpoints import ( + update_access_group, + ) + from litellm.types.proxy.management_endpoints.model_management_endpoints import ( + UpdateModelGroupRequest, + ) + + prisma = _FakePrismaClient( + [], + deployments=[ + _deployment(model_id="stays", model_name="gpt-stays"), + _deployment(model_id="joins", model_name="gpt-joins", access_groups=()), + ], + ) + + with _proxy_with_stubbed_reload(prisma): + _serve_model_groups(proxy_server.llm_router, ["gpt-stays", "gpt-joins", "config-only-model"]) + with pytest.raises(HTTPException) as raised: + await update_access_group( + access_group="prod-models", + data=UpdateModelGroupRequest(model_names=["gpt-joins", "config-only-model"]), + user_api_key_dict=_admin(), + ) + + assert raised.value.status_code == 400 + assert prisma.model_table.updates == [] diff --git a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py index 14d93f368c1..872d1607351 100644 --- a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py @@ -6508,7 +6508,11 @@ class TestStrategyRouterWriteValidation: 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") + published.assert_awaited_once_with( + redis_cache=None, + object_type="litellm_proxymodeltable", + model_ids=(model_id,) if model_id else (), + ) (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 == (AUTO_ROUTER_CAPABILITY_SLOT_LOCK_KEY,) diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 13a064517a4..7d9e9839a5f 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -699,8 +699,8 @@ export interface paths { * @description Update an access group's model names. * * This will: - * 1. Remove the access group from all current deployments - * 2. Add the access group to all deployments for the new model_names list + * 1. Add the access group to all deployments for the new model_names list + * 2. Remove the access group from every other deployment * * Example: * ```bash @@ -722,6 +722,7 @@ export interface paths { * Raises: * - HTTPException 400: If any model names don't exist * - HTTPException 404: If access group not found + * - HTTPException 409: If another edit holds the group or one of its models for 30 s */ put: operations["update_access_group_access_group__access_group__update_put"]; post?: never; From c0ded7b2824e7af9d8e5e8008fdf35d1afec0c3c Mon Sep 17 00:00:00 2001 From: mika-lt <349045450+mika-lt@users.noreply.github.com> Date: Fri, 30 Oct 2026 03:47:59 -0700 Subject: [PATCH 18/23] fix(responses)!: send a dict reasoning_effort the same way on every Responses bridge (#49082) * fix(responses): send a dict reasoning_effort the same way on every Responses bridge Azure and ChatGPT bridges dropped reasoning mode and context unless the dict also carried summary. Every bridge now runs the reasoning step the OpenAI bridge uses, and each Responses config declares the reasoning values its API accepts, so drop_params drops what a backend refuses. Perplexity ignores mode and context, so it answers 400 for them without drop_params. * fix(responses): refuse ignored reasoning fields on xAI and Fireworks, resolve openai/ names for OpenRouter * fix(responses): refuse a reasoning mode Copilot ignores, keep catalog modes, return refusals as values Apply additional_drop_params before the refusal check, keep a null reasoning field instead of refusing it, let a cost-map row's reasoning_modes and reasoning_contexts override a backend's declared values, and declare GitHub Copilot's accepted values from a live probe * fix(responses): read Azure base_model and deployment rows, accept standard mode, refuse only mode and context * fix(responses): send only effort and summary to an Azure deployment litellm cannot identify under drop_params Azure and Azure AI deployment names are user-chosen, so when neither the name, model_info.base_model, nor a cost map row resolves the model, drop_params now drops mode and context instead of forwarding a pro the model may refuse. Names resolve by their last path segment, so a regional base_model like azure/eu/gpt-5.4 follows gpt-5.4's rows * fix(responses): resolve deployment names by model family and keep reasoning_modes per deployment A custom Azure deployment name now resolves to the longest OpenAI catalog prefix of its base_model or name, an unresolved name forwards mode and context, and a deployment's model_info reasoning_modes and reasoning_contexts apply only to that deployment instead of every sibling sharing its model name. litellm_proxy forwards every reasoning field. * fix(responses): find the model family anywhere in an Azure deployment name and keep only effort and summary for unknown gpt-5 or gpt-6 names under drop_params * fix(responses): look up OpenAI model names inside deployment names only on Azure Databricks endpoints like databricks-gpt-5-5-pro spell versions with hyphens and other backends name local models freely, so only Azure and Azure AI search inside the name, and a run that cuts a hyphen-spelled version never matches * fix(responses): judge Azure AI reasoning fields by the request api_base --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../llms/azure/responses/transformation.py | 7 + .../llms/base_llm/responses/transformation.py | 28 +- .../llms/chatgpt/responses/transformation.py | 4 + .../fireworks_ai/responses/transformation.py | 10 + .../responses/transformation.py | 7 + .../litellm_proxy/responses/transformation.py | 4 + .../perplexity/responses/transformation.py | 11 + litellm/llms/xai/responses/transformation.py | 10 + litellm/main.py | 26 +- litellm/responses/utils.py | 186 +++++-- litellm/types/utils.py | 2 + tests/unit/test_main.py | 457 ++++++++++++++++++ 12 files changed, 716 insertions(+), 36 deletions(-) diff --git a/litellm/llms/azure/responses/transformation.py b/litellm/llms/azure/responses/transformation.py index 09777da8669..2b757bb10d9 100644 --- a/litellm/llms/azure/responses/transformation.py +++ b/litellm/llms/azure/responses/transformation.py @@ -14,6 +14,7 @@ from litellm.constants import AZURE_DEFAULT_RESPONSES_API_VERSION from litellm.litellm_core_utils.url_utils import encode_url_path_segment from litellm.llms.azure.chat.gpt_reasoning_transformation import AzureOpenAIGPT5Config from litellm.llms.azure.common_utils import BaseAzureLLM +from litellm.llms.base_llm.responses.transformation import SUMMARY_ONLY_REASONING_FIELD_VALUES, ReasoningFieldValues from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig from litellm.types.llms.openai import * from litellm.types.responses.main import * @@ -92,6 +93,12 @@ class AzureOpenAIResponsesAPIConfig(OpenAIResponsesAPIConfig): def _effort_resolves_to_none(model: str, effort: str | None) -> bool: return AzureOpenAIGPT5Config.effort_resolves_to_none(model, effort) + def deployment_names_wrap_openai_models(self) -> bool: + return True + + def unresolved_model_reasoning_field_values(self, model: str) -> ReasoningFieldValues | None: + return SUMMARY_ONLY_REASONING_FIELD_VALUES if self._is_gpt_5_model(model) else None + def get_supported_openai_params(self, model: str) -> list: """ Azure Responses API does not support context_management (compaction). diff --git a/litellm/llms/base_llm/responses/transformation.py b/litellm/llms/base_llm/responses/transformation.py index 792b14090b8..32d0e408aba 100644 --- a/litellm/llms/base_llm/responses/transformation.py +++ b/litellm/llms/base_llm/responses/transformation.py @@ -1,7 +1,7 @@ import types from abc import ABC, abstractmethod from collections.abc import Mapping -from typing import TYPE_CHECKING, Any, Final, Literal, TypeAlias, cast +from typing import TYPE_CHECKING, Any, Final, Literal, TypeAlias, cast, get_args import httpx @@ -27,6 +27,20 @@ else: BaseLLMException = Any ReasoningSummaryLevel: TypeAlias = Literal["auto", "concise", "detailed"] +ReasoningFieldValues: TypeAlias = Mapping[str, frozenset[str | None] | None] +REASONING_SUMMARY_VALUES: Final = frozenset((*get_args(ReasoningSummaryLevel), None)) +SUMMARY_ONLY_REASONING_FIELD_VALUES: Final[ReasoningFieldValues] = types.MappingProxyType( + {"summary": REASONING_SUMMARY_VALUES} +) +STANDARD_MODE_WITHOUT_CONTEXT_REASONING_FIELD_VALUES: Final[ReasoningFieldValues] = types.MappingProxyType( + {**SUMMARY_ONLY_REASONING_FIELD_VALUES, "mode": frozenset({"standard"})} +) +STANDARD_MODE_REASONING_FIELD_VALUES: Final[ReasoningFieldValues] = types.MappingProxyType( + {**STANDARD_MODE_WITHOUT_CONTEXT_REASONING_FIELD_VALUES, "context": None} +) +ANY_REASONING_FIELD_VALUES: Final[ReasoningFieldValues] = types.MappingProxyType( + {**SUMMARY_ONLY_REASONING_FIELD_VALUES, "mode": None, "context": None} +) class BaseResponsesAPIConfig(ABC): @@ -100,6 +114,18 @@ class BaseResponsesAPIConfig(ABC): """The ``reasoning.summary`` LiteLLM's opt-in auto summary asks this provider for.""" return "detailed" + def reasoning_field_values(self) -> ReasoningFieldValues | None: + return None + + def deployment_names_wrap_openai_models(self) -> bool: + return False + + def unresolved_model_reasoning_field_values(self, model: str) -> ReasoningFieldValues | None: + return None + + def ignores_unsupported_reasoning_fields(self) -> bool: + return False + @abstractmethod def map_openai_params( self, diff --git a/litellm/llms/chatgpt/responses/transformation.py b/litellm/llms/chatgpt/responses/transformation.py index 72b5cde21d1..26a5a3f150b 100644 --- a/litellm/llms/chatgpt/responses/transformation.py +++ b/litellm/llms/chatgpt/responses/transformation.py @@ -11,6 +11,7 @@ from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response impo safe_convert_created_field, ) from litellm.litellm_core_utils.url_utils import join_api_base +from litellm.llms.base_llm.responses.transformation import STANDARD_MODE_REASONING_FIELD_VALUES, ReasoningFieldValues from litellm.llms.custom_httpx.sse_lines import split_sse_lines from litellm.llms.openai.common_utils import OpenAIError from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig @@ -295,3 +296,6 @@ class ChatGPTResponsesAPIConfig(OpenAIResponsesAPIConfig): def supports_native_websocket(self) -> bool: """ChatGPT does not support native WebSocket for Responses API""" return False + + def reasoning_field_values(self) -> ReasoningFieldValues: + return STANDARD_MODE_REASONING_FIELD_VALUES diff --git a/litellm/llms/fireworks_ai/responses/transformation.py b/litellm/llms/fireworks_ai/responses/transformation.py index 86d9df9618a..463b9cbfc9a 100644 --- a/litellm/llms/fireworks_ai/responses/transformation.py +++ b/litellm/llms/fireworks_ai/responses/transformation.py @@ -6,6 +6,10 @@ from urllib.parse import unquote import httpx from litellm.litellm_core_utils.url_utils import join_api_base +from litellm.llms.base_llm.responses.transformation import ( + STANDARD_MODE_WITHOUT_CONTEXT_REASONING_FIELD_VALUES, + ReasoningFieldValues, +) from litellm.llms.fireworks_ai.common_utils import ( FIREWORKS_FORWARD_USER_ID_PARAM, get_fireworks_forwarded_user_id, @@ -129,3 +133,9 @@ class FireworksAIResponsesAPIConfig(OpenAIResponsesAPIConfig): def supports_native_file_search(self) -> bool: return False + + def reasoning_field_values(self) -> ReasoningFieldValues: + return STANDARD_MODE_WITHOUT_CONTEXT_REASONING_FIELD_VALUES + + def ignores_unsupported_reasoning_fields(self) -> bool: + return True diff --git a/litellm/llms/github_copilot/responses/transformation.py b/litellm/llms/github_copilot/responses/transformation.py index b055d388dd5..b87a3d543e7 100644 --- a/litellm/llms/github_copilot/responses/transformation.py +++ b/litellm/llms/github_copilot/responses/transformation.py @@ -24,6 +24,7 @@ from litellm._logging import verbose_logger from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH from litellm.exceptions import AuthenticationError from litellm.litellm_core_utils.url_utils import join_api_base +from litellm.llms.base_llm.responses.transformation import STANDARD_MODE_REASONING_FIELD_VALUES, ReasoningFieldValues from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig from litellm.types.llms.openai import ( ResponseInputParam, @@ -437,6 +438,12 @@ class GithubCopilotResponsesAPIConfig(OpenAIResponsesAPIConfig): return False + def reasoning_field_values(self) -> ReasoningFieldValues: + return STANDARD_MODE_REASONING_FIELD_VALUES + + def ignores_unsupported_reasoning_fields(self) -> bool: + return True + def supports_native_websocket(self) -> bool: """GitHub Copilot does not support native WebSocket for Responses API""" return False diff --git a/litellm/llms/litellm_proxy/responses/transformation.py b/litellm/llms/litellm_proxy/responses/transformation.py index fc94eb30504..ed62e132c99 100644 --- a/litellm/llms/litellm_proxy/responses/transformation.py +++ b/litellm/llms/litellm_proxy/responses/transformation.py @@ -12,6 +12,7 @@ from pydantic import TypeAdapter, ValidationError from litellm.litellm_core_utils.url_utils import join_api_base from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.llms.base_llm.responses.transformation import ANY_REASONING_FIELD_VALUES, ReasoningFieldValues from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig from litellm.secret_managers.main import get_secret_str from litellm.types.utils import LlmProviders @@ -42,6 +43,9 @@ class LiteLLMProxyResponsesAPIConfig(OpenAIResponsesAPIConfig): def custom_llm_provider(self) -> LlmProviders: return LlmProviders.LITELLM_PROXY + def reasoning_field_values(self) -> ReasoningFieldValues: + return ANY_REASONING_FIELD_VALUES + def get_complete_url( self, api_base: str | None, diff --git a/litellm/llms/perplexity/responses/transformation.py b/litellm/llms/perplexity/responses/transformation.py index 58f71028451..afd27b0f696 100644 --- a/litellm/llms/perplexity/responses/transformation.py +++ b/litellm/llms/perplexity/responses/transformation.py @@ -9,6 +9,7 @@ The only provider quirks: Ref: https://docs.perplexity.ai/api-reference/responses-post """ +from types import MappingProxyType from typing import Any, Final import httpx @@ -17,6 +18,7 @@ from pydantic import TypeAdapter from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.url_utils import join_api_base from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.llms.base_llm.responses.transformation import STANDARD_MODE_REASONING_FIELD_VALUES, ReasoningFieldValues from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig, drops_params from litellm.secret_managers.main import get_secret_str from litellm.types.llms.openai import ResponseInputParam, ResponsesAPIResponse @@ -24,6 +26,9 @@ from litellm.types.router import GenericLiteLLMParams from litellm.types.utils import LlmProviders _DECODED_JSON: Final = TypeAdapter(object) +PERPLEXITY_REASONING_FIELD_VALUES: Final[ReasoningFieldValues] = MappingProxyType( + {**STANDARD_MODE_REASONING_FIELD_VALUES, "context": frozenset({"auto", "current_turn", "all_turns"})} +) class PerplexityResponsesConfig(OpenAIResponsesAPIConfig): @@ -127,3 +132,9 @@ class PerplexityResponsesConfig(OpenAIResponsesAPIConfig): def supports_native_websocket(self) -> bool: """Perplexity does not support native WebSocket for Responses API""" return False + + def reasoning_field_values(self) -> ReasoningFieldValues: + return PERPLEXITY_REASONING_FIELD_VALUES + + def ignores_unsupported_reasoning_fields(self) -> bool: + return True diff --git a/litellm/llms/xai/responses/transformation.py b/litellm/llms/xai/responses/transformation.py index b6ea516fb79..c907fde9926 100644 --- a/litellm/llms/xai/responses/transformation.py +++ b/litellm/llms/xai/responses/transformation.py @@ -10,6 +10,10 @@ from litellm._logging import verbose_logger from litellm.constants import XAI_API_BASE from litellm.exceptions import AuthenticationError from litellm.litellm_core_utils.url_utils import join_api_base +from litellm.llms.base_llm.responses.transformation import ( + STANDARD_MODE_WITHOUT_CONTEXT_REASONING_FIELD_VALUES, + ReasoningFieldValues, +) from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig from litellm.llms.xai.common_utils import XAIModelInfo, xai_reported_cost_in_usd from litellm.secret_managers.main import get_secret_str @@ -290,3 +294,9 @@ class XAIResponsesAPIConfig(OpenAIResponsesAPIConfig): def supports_native_websocket(self) -> bool: """XAI does not support native WebSocket for Responses API""" return False + + def reasoning_field_values(self) -> ReasoningFieldValues: + return STANDARD_MODE_WITHOUT_CONTEXT_REASONING_FIELD_VALUES + + def ignores_unsupported_reasoning_fields(self) -> bool: + return True diff --git a/litellm/main.py b/litellm/main.py index 8d6e07d6a8f..54afe654cec 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -6267,6 +6267,9 @@ def completion( # reasoningSummary/reasoning_summary without tools (AI SDK) that the first # (early) check doesn't cover. _reasoning_summary_for_bridge: Final = peek_reasoning_summary_aliases(optional_params) + typed_deployment_model_info: Final = str_keyed_mapping( + cast(object, model_info) # cast-ok: model_info is the untyped dict kwarg + ) if responses_api_model_info.get("mode") != "responses": responses_api_model_info, model = responses_api_bridge_check( model=model, @@ -6276,9 +6279,7 @@ def completion( reasoning_effort=bridge_reasoning_effort(resolved_reasoning_effort, optional_params.get("extra_body")), reasoning_summary=_reasoning_summary_for_bridge, api_base=api_base, - deployment_model_info=str_keyed_mapping( - cast(object, model_info) # cast-ok: model_info is the untyped dict kwarg - ), + deployment_model_info=typed_deployment_model_info, ) # Use base_model (the true underlying model) for Azure model-type @@ -6290,19 +6291,28 @@ def completion( optional_params, rs_val = strip_reasoning_summary_aliases_from_optional_params(optional_params) - if custom_llm_provider == "openai" and isinstance(resolved_reasoning_effort, dict): - from litellm.responses.utils import openai_bridge_reasoning + if isinstance(resolved_reasoning_effort, dict): + from litellm.responses.utils import ( + RefusedReasoningFields, + bridge_reasoning, + raise_refused_reasoning_fields, + ) - optional_params["reasoning_effort"] = openai_bridge_reasoning( + bridged_reasoning: Final = bridge_reasoning( requested=resolved_reasoning_effort, effort_kept="reasoning_effort" in optional_params, summary_alias=rs_val, model=model, + base_model=base_model if isinstance(base_model, str) else None, + custom_llm_provider=custom_llm_provider, + api_base=api_base, + deployment_model_info=typed_deployment_model_info, drop_params=drop_unsupported_reasoning, additional_drop_params=routing_kwargs.get("additional_drop_params"), ) - elif isinstance(resolved_reasoning_effort, dict) and "summary" in resolved_reasoning_effort: - optional_params["reasoning_effort"] = resolved_reasoning_effort + if isinstance(bridged_reasoning, RefusedReasoningFields): + raise_refused_reasoning_fields(bridged_reasoning, model, custom_llm_provider) + optional_params["reasoning_effort"] = bridged_reasoning elif rs_val is not None: eff: Final = optional_params.get("reasoning_effort", resolved_reasoning_effort) if isinstance(eff, dict): diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py index b0db0516658..1a97c1cb48c 100644 --- a/litellm/responses/utils.py +++ b/litellm/responses/utils.py @@ -1,10 +1,11 @@ import base64 import os import re -from collections.abc import Callable, Collection, Iterable, Mapping, Sequence +from collections.abc import Callable, Collection, Iterable, Iterator, Mapping, Sequence +from dataclasses import dataclass from functools import reduce from types import MappingProxyType -from typing import Any, Final, Optional, TypeVar, Union, cast, get_args, get_type_hints, overload +from typing import Any, Final, NoReturn, Optional, TypeVar, Union, cast, get_type_hints, overload from pydantic import BaseModel, TypeAdapter from typing_extensions import TypedDict as ExtraItemsTypedDict @@ -15,7 +16,13 @@ from litellm._logging import verbose_logger from litellm.litellm_core_utils.core_helpers import as_str_mapping, attribute_or_none, iter_items from litellm.litellm_core_utils.dot_notation_indexing import delete_nested_value, is_nested_path from litellm.litellm_core_utils.get_provider_specific_headers import ProviderSpecificHeaderUtils -from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig, ReasoningSummaryLevel +from litellm.llms.base_llm.responses.transformation import ( + ANY_REASONING_FIELD_VALUES, + REASONING_SUMMARY_VALUES, + BaseResponsesAPIConfig, + ReasoningFieldValues, + ReasoningSummaryLevel, +) from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig from litellm.types.llms.openai import ( AllMessageValues, @@ -1411,9 +1418,6 @@ def reasoning_with_summary_default( return {**reasoning} if summary is None else {**reasoning, "summary": summary} -_SUMMARY_VALUES: Final = frozenset((*get_args(ReasoningSummaryLevel), None)) - - def _openai_model_info(model: str) -> ModelInfo | None: from litellm.utils import ModelInfoLookupMiss, lookup_model_info_for_registration @@ -1425,24 +1429,115 @@ def _declared_values(values: Sequence[str] | None) -> frozenset[str | None] | No return None if values is None else frozenset(values) -def _openai_reasoning_field_values(model: str) -> Mapping[str, frozenset[str | None] | None]: +_REFUSABLE_REASONING_FIELDS: Final = frozenset({"mode", "context"}) + + +def _openai_reasoning_field_values(model: str) -> ReasoningFieldValues: model_info: Final = _openai_model_info(model) if model_info is None: - return MappingProxyType({"summary": _SUMMARY_VALUES, "mode": None, "context": None}) + return ANY_REASONING_FIELD_VALUES if not model_info.get("supports_reasoning"): return MappingProxyType({}) return MappingProxyType( { - "summary": _SUMMARY_VALUES, + "summary": REASONING_SUMMARY_VALUES, "mode": _declared_values(model_info.get("reasoning_modes")), "context": _declared_values(model_info.get("reasoning_contexts")), } ) -def _openai_accepts_reasoning_field( - key: str, value: object, values_by_field: Mapping[str, frozenset[str | None] | None] -) -> bool: +_CATALOG_REASONING_FIELDS: Final = (("mode", "reasoning_modes"), ("context", "reasoning_contexts")) + + +def _catalog_values(values: object) -> frozenset[str | None]: + return frozenset(value for value in iter_items(values) if isinstance(value, str)) + + +def _declared_reasoning_values(row: Mapping[str, object] | None) -> Mapping[str, frozenset[str | None]]: + if row is None: + return {} + return { + field: _catalog_values(row[key]) for field, key in _CATALOG_REASONING_FIELDS if isinstance(row.get(key), list) + } + + +def _with_declared_reasoning_values( + defaults: ReasoningFieldValues | None, + model: str, + custom_llm_provider: str, + deployment_model_info: Mapping[str, object] | None, +) -> ReasoningFieldValues | None: + catalog: Final = cast(Mapping[str, object], litellm.model_cost) # cast-ok: read-only view of the legacy registry + catalog_row: Final = as_str_mapping(catalog.get(f"{custom_llm_provider}/{model.removeprefix('responses/')}")) + overrides: Final = { + **_declared_reasoning_values(catalog_row), + **_declared_reasoning_values(deployment_model_info), + } + if not overrides: + return defaults + return MappingProxyType({**(ANY_REASONING_FIELD_VALUES if defaults is None else defaults), **overrides}) + + +def _is_version_number(part: str) -> bool: + return part.isdigit() and len(part) <= 2 + + +def _splits_a_version(parts: Sequence[str], end: int) -> bool: + return end < len(parts) and _is_version_number(parts[end - 1]) and _is_version_number(parts[end]) + + +def _hyphen_runs_longest_first(name: str) -> Iterator[str]: + parts: Final = name.rsplit("/", 1)[-1].split("-") + for length in range(len(parts), 0, -1): + yield from ( + "-".join(parts[start : start + length]) + for start in range(len(parts) - length + 1) + if not _splits_a_version(parts, start + length) + ) + + +def _openai_catalog_family(name: str) -> str | None: + runs: Final = _hyphen_runs_longest_first(name) + return next((run for run in runs if OpenAIGPTConfig.is_openai_catalog_model(run)), None) + + +def _openai_catalog_name(name: str) -> str | None: + last_segment: Final = name.rsplit("/", 1)[-1] + return last_segment if OpenAIGPTConfig.is_openai_catalog_model(last_segment) else None + + +def _openai_catalog_model(model: str, base_model: str | None, names_wrap_models: bool) -> str | None: + find: Final = _openai_catalog_family if names_wrap_models else _openai_catalog_name + families: Final = (find(name) for name in (base_model, model) if name) + return next((family for family in families if family is not None), None) + + +def _default_reasoning_field_values( + model: str, base_model: str | None, responses_config: BaseResponsesAPIConfig | None +) -> ReasoningFieldValues | None: + declared: Final = None if responses_config is None else responses_config.reasoning_field_values() + if declared is not None: + return declared + names_wrap_models: Final = responses_config is not None and responses_config.deployment_names_wrap_openai_models() + catalog_model: Final = _openai_catalog_model(model, base_model, names_wrap_models) + if catalog_model is not None: + return _openai_reasoning_field_values(catalog_model) + return None if responses_config is None else responses_config.unresolved_model_reasoning_field_values(model) + + +def _reasoning_field_values( + model: str, + base_model: str | None, + custom_llm_provider: str, + responses_config: BaseResponsesAPIConfig | None, + deployment_model_info: Mapping[str, object] | None, +) -> ReasoningFieldValues | None: + defaults: Final = _default_reasoning_field_values(model, base_model, responses_config) + return _with_declared_reasoning_values(defaults, model, custom_llm_provider, deployment_model_info) + + +def _accepts_reasoning_field(key: str, value: object, values_by_field: ReasoningFieldValues) -> bool: if key == "effort": return True if key not in values_by_field: @@ -1451,11 +1546,8 @@ def _openai_accepts_reasoning_field( return allowed is None or (isinstance(value, str | None) and value in allowed) -def _openai_accepted_reasoning(reasoning: Mapping[str, object], model: str) -> Mapping[str, object]: - values_by_field: Final = _openai_reasoning_field_values(model) - return { - key: value for key, value in reasoning.items() if _openai_accepts_reasoning_field(key, value, values_by_field) - } +def _accepted_reasoning(reasoning: Mapping[str, object], values_by_field: ReasoningFieldValues) -> Mapping[str, object]: + return {key: value for key, value in reasoning.items() if _accepts_reasoning_field(key, value, values_by_field)} def _additional_drop_paths(additional_drop_params: object) -> tuple[str, ...]: @@ -1464,26 +1556,66 @@ def _additional_drop_paths(additional_drop_params: object) -> tuple[str, ...]: return tuple(path for path in iter_items(additional_drop_params) if isinstance(path, str)) -def openai_bridge_reasoning( +@dataclass(frozen=True, slots=True) +class RefusedReasoningFields: + refused: tuple[str, ...] + + +def raise_refused_reasoning_fields(refusal: RefusedReasoningFields, model: str, custom_llm_provider: str) -> NoReturn: + raise litellm.UnsupportedParamsError( + message=( + f"{custom_llm_provider} does not support parameters: " + f"{[f'reasoning_effort.{key}' for key in refusal.refused]}, for model={model}. " + "To drop these, set `litellm.drop_params=True` or for proxy:\n\n`litellm_settings:\n drop_params: true`\n" + ), + llm_provider=custom_llm_provider, + model=model, + ) + + +def bridge_reasoning( *, requested: Mapping[str, object], effort_kept: bool, summary_alias: object, model: str, + base_model: str | None, + custom_llm_provider: str, + api_base: str | None, + deployment_model_info: Mapping[str, object] | None, drop_params: bool, additional_drop_params: object, -) -> Mapping[str, object] | None: +) -> Mapping[str, object] | RefusedReasoningFields | None: + from litellm.utils import ProviderConfigManager + drop_paths: Final = _additional_drop_paths(additional_drop_params) if "reasoning_effort" in drop_paths: return None kept: Final = requested if effort_kept else {key: value for key, value in requested.items() if key != "effort"} - with_summary: Final = reasoning_with_summary_default(kept, summary_alias, model, "openai") - accepted: Final = ( - _openai_accepted_reasoning(with_summary, model) - if drop_params and OpenAIGPTConfig.is_openai_catalog_model(model) - else {**with_summary} + with_summary: Final = reasoning_with_summary_default(kept, summary_alias, model, custom_llm_provider) + sent: Final = ( + as_str_mapping( + _apply_nested_drop_params({"reasoning_effort": {**with_summary}}, list(drop_paths)).get("reasoning_effort") + ) + or {} ) - remaining: Final = as_str_mapping( - _apply_nested_drop_params({"reasoning_effort": accepted}, list(drop_paths)).get("reasoning_effort") + responses_config: Final = ProviderConfigManager.get_provider_responses_api_config( + custom_llm_provider, model, api_base ) - return remaining or None + values_by_field: Final = _reasoning_field_values( + model, base_model, custom_llm_provider, responses_config, deployment_model_info + ) + accepted: Final = sent if values_by_field is None else _accepted_reasoning(sent, values_by_field) + refused: Final = tuple( + key + for key, value in sent.items() + if key in _REFUSABLE_REASONING_FIELDS and key not in accepted and value is not None + ) + if ( + refused + and not drop_params + and responses_config is not None + and responses_config.ignores_unsupported_reasoning_fields() + ): + return RefusedReasoningFields(refused) + return (accepted if drop_params else sent) or None diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 5905aaa3380..6a758961c7b 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -4155,6 +4155,8 @@ DEPLOYMENT_SCOPED_CAPABILITY_FIELDS: Final[frozenset[str]] = frozenset( "supports_assistant_prefill", "supports_bedrock_native_structured_output", "supports_pdf_input", + "reasoning_modes", + "reasoning_contexts", } ) diff --git a/tests/unit/test_main.py b/tests/unit/test_main.py index b2ae6b43d26..b0bf1cb879e 100644 --- a/tests/unit/test_main.py +++ b/tests/unit/test_main.py @@ -2463,6 +2463,463 @@ _BRIDGED_RESPONSE_ANSWER: Final = json.dumps( ).encode() +@pytest.fixture +def non_openai_responses_bridge( + restore_model_registry, respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, tmp_path: Path +) -> Callable[[str], tuple[str, Mapping[str, str], respx.Route]]: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.delenv("LITELLM_REASONING_AUTO_SUMMARY", raising=False) + for chatgpt_env in ("CHATGPT_AUTH_FILE", "CHATGPT_API_BASE", "OPENAI_CHATGPT_API_BASE"): + monkeypatch.delenv(chatgpt_env, raising=False) + monkeypatch.setenv("CHATGPT_TOKEN_DIR", str(tmp_path)) + (tmp_path / "auth.json").write_text( + json.dumps({"access_token": "token", "account_id": "acct", "expires_at": 4102444800}) + ) + monkeypatch.setenv("GITHUB_COPILOT_TOKEN_DIR", str(tmp_path)) + (tmp_path / "api-key.json").write_text( + json.dumps({"token": "token", "expires_at": 4102444800, "endpoints": {"api": "https://api.githubcopilot.com"}}) + ) + litellm.register_model( + { + **{ + model_name: { + "litellm_provider": "openai", + "mode": "chat", + "supports_reasoning": True, + "reasoning_modes": ["standard"], + "reasoning_contexts": ["auto", "current_turn", "all_turns"], + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + for model_name in ("gpt-6", "gpt-6.1-standard-only") + }, + **{ + f"{provider}/{model_name}": { + "litellm_provider": provider, + "mode": "responses", + "supports_reasoning": True, + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + for provider, model_name in ( + ("chatgpt", "gpt-6.1-bridge-test"), + ("perplexity", "bridge-test-model"), + ("github_copilot", "gpt-6.1-bridge-test"), + ) + }, + }, + persist_across_reloads=False, + ) + backends: Final = { + "azure": ( + "azure/responses/gpt-6.1-standard-only", + "https://test.openai.azure.com/", + {"api_base": "https://test.openai.azure.com", "api_key": "fake-key"}, + ), + "azure_ai": ( + "azure_ai/responses/gpt-6.1-standard-only-prod", + _FOUNDRY_API_BASE, + {"api_base": _FOUNDRY_API_BASE, "api_key": "fake-key"}, + ), + "chatgpt": ("chatgpt/gpt-6.1-bridge-test", "https://chatgpt.com/backend-api/codex/responses", {}), + "github_copilot": ("github_copilot/gpt-6.1-bridge-test", "https://api.githubcopilot.com/responses", {}), + "databricks": ( + "databricks/responses/databricks-gpt-bridge-test", + "https://dbx.example/serving-endpoints/responses", + {"api_base": "https://dbx.example/serving-endpoints", "api_key": "fake-key"}, + ), + "perplexity": ( + "perplexity/bridge-test-model", + "https://api.perplexity.ai/v1/responses", + {"api_key": "fake-key"}, + ), + "xai": ("xai/responses/grok-bridge-test", "https://api.x.ai/v1/responses", {"api_key": "fake-key"}), + "fireworks_ai": ( + "fireworks_ai/responses/accounts/fireworks/models/bridge-test", + "https://api.fireworks.ai/inference/v1/responses", + {"api_key": "fake-key"}, + ), + "openrouter": ( + "openrouter/responses/openai/gpt-6.1-standard-only", + "https://openrouter.ai/api/v1/responses", + {"api_key": "fake-key"}, + ), + "litellm_proxy": ( + "litellm_proxy/responses/gpt-6.1-standard-only", + "https://proxy.example/v1/responses", + {"api_base": "https://proxy.example/v1", "api_key": "fake-key"}, + ), + } + + def route_backend(backend: str) -> tuple[str, Mapping[str, str], respx.Route]: + model, url, kwargs = backends[backend] + route: Final = respx_mock.post(url__startswith=url).respond( + content=_BRIDGED_RESPONSE_ANSWER, headers={"content-type": "application/json"} + ) + return model, kwargs, route + + return route_backend + + +@pytest.mark.parametrize( + ("backend", "reasoning_effort"), + [ + ("azure", {"effort": "low", "mode": "pro", "context": "all_turns"}), + ("chatgpt", {"effort": "low", "mode": "pro", "context": "all_turns"}), + ("github_copilot", {"effort": "low", "mode": "standard", "context": "all_turns"}), + ("databricks", {"effort": "low", "mode": "pro", "context": "all_turns"}), + ("openrouter", {"effort": "low", "mode": "pro", "context": "all_turns"}), + ("perplexity", {"effort": "low", "mode": "standard", "context": "all_turns", "generate_summary": "auto"}), + ("xai", {"effort": "low", "mode": "standard", "generate_summary": "auto"}), + ("fireworks_ai", {"effort": "low", "mode": "standard", "generate_summary": "auto"}), + ], +) +@pytest.mark.parametrize("summary", [{}, {"summary": "auto"}]) +def test_non_openai_responses_bridge_sends_the_whole_reasoning_effort_dict_with_or_without_summary( + backend: str, + reasoning_effort: dict[str, str], + summary: dict[str, str], + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + model, kwargs, route = non_openai_responses_bridge(backend) + requested: Final = {**reasoning_effort, **summary} + + litellm.completion( + model=model, messages=[{"role": "user", "content": "hello"}], reasoning_effort=requested, **kwargs + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == requested + + +@pytest.mark.parametrize( + ("backend", "reasoning_effort", "expected_reasoning"), + [ + ("azure", {"effort": "low", "mode": "pro", "context": "all_turns"}, {"effort": "low", "context": "all_turns"}), + ( + "chatgpt", + {"effort": "low", "mode": "pro", "context": "all_turns"}, + {"effort": "low", "context": "all_turns"}, + ), + ( + "chatgpt", + {"effort": "low", "mode": "standard", "context": "current_turn"}, + {"effort": "low", "mode": "standard", "context": "current_turn"}, + ), + ( + "github_copilot", + {"effort": "low", "mode": "pro", "context": "all_turns"}, + {"effort": "low", "context": "all_turns"}, + ), + ( + "openrouter", + {"effort": "low", "mode": "pro", "context": "all_turns"}, + {"effort": "low", "context": "all_turns"}, + ), + ( + "perplexity", + {"effort": "low", "mode": "pro", "context": "all_turns"}, + {"effort": "low", "context": "all_turns"}, + ), + ("perplexity", {"effort": "low", "mode": "standard", "context": "every_turn"}, {"effort": "low", "mode": "standard"}), + ("xai", {"effort": "low", "mode": "pro", "context": "all_turns"}, {"effort": "low"}), + ("fireworks_ai", {"effort": "low", "mode": "pro", "context": "all_turns"}, {"effort": "low"}), + ], +) +@pytest.mark.parametrize("summary", [{}, {"summary": "auto"}]) +def test_non_openai_responses_bridge_drops_reasoning_fields_the_backend_refuses_under_drop_params( + backend: str, + reasoning_effort: dict[str, str], + expected_reasoning: dict[str, str], + summary: dict[str, str], + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + model, kwargs, route = non_openai_responses_bridge(backend) + + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={**reasoning_effort, **summary}, + drop_params=True, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {**expected_reasoning, **summary} + + +@pytest.mark.parametrize( + ("backend", "ignored"), + [ + ("perplexity", {"mode": "pro"}), + ("perplexity", {"context": "every_turn"}), + ("xai", {"context": "all_turns"}), + ("fireworks_ai", {"context": "all_turns"}), + ("github_copilot", {"mode": "pro"}), + ], +) +@pytest.mark.parametrize("summary", [{}, {"summary": "auto"}]) +def test_responses_bridge_rejects_reasoning_fields_the_backend_ignores_without_drop_params( + backend: str, + ignored: dict[str, str], + summary: dict[str, str], + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + model, kwargs, route = non_openai_responses_bridge(backend) + + with pytest.raises(litellm.UnsupportedParamsError, match=rf"reasoning_effort\.{next(iter(ignored))}") as raised: + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", **ignored, **summary}, + **kwargs, + ) + + assert raised.value.status_code == 400 + assert route.called is False + + +@pytest.mark.parametrize("backend", ["perplexity", "xai", "fireworks_ai"]) +def test_responses_bridge_sends_the_rest_when_additional_drop_params_removes_the_refused_fields( + backend: str, + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + model, kwargs, route = non_openai_responses_bridge(backend) + + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns", "summary": "auto"}, + additional_drop_params=["reasoning_effort.mode", "reasoning_effort.context"], + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {"effort": "low", "summary": "auto"} + + +@pytest.mark.parametrize("backend", ["perplexity", "xai", "fireworks_ai"]) +def test_responses_bridge_forwards_a_null_reasoning_field_instead_of_rejecting_it( + backend: str, + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + model, kwargs, route = non_openai_responses_bridge(backend) + + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": None}, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {"effort": "low", "mode": None} + + +def test_chatgpt_bridge_keeps_a_reasoning_mode_its_catalog_row_declares_under_drop_params( + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + _, kwargs, route = non_openai_responses_bridge("chatgpt") + litellm.register_model( + { + "chatgpt/gpt-6.1-pro-bridge-test": { + "litellm_provider": "chatgpt", + "mode": "responses", + "supports_reasoning": True, + "reasoning_modes": ["pro"], + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + }, + persist_across_reloads=False, + ) + + litellm.completion( + model="chatgpt/gpt-6.1-pro-bridge-test", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro"}, + drop_params=True, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {"effort": "low", "mode": "pro"} + + +@pytest.mark.parametrize( + ("model", "model_info"), + [ + ("azure/responses/custom-deployment-name", {"base_model": "azure/gpt-6.1-standard-only"}), + ("azure/responses/custom-deployment-name", {"base_model": "azure/eu/gpt-6.1-standard-only"}), + ("azure/responses/custom-deployment-name", {"base_model": "azure/gpt-6.1-standard-only-2026-07-09"}), + ("azure/responses/gpt-6.1-standard-only-prod", {}), + ("azure/responses/eu-gpt-6.1-standard-only-prod", {}), + ("azure/responses/gpt-6-2026-07-09", {}), + ], +) +def test_azure_bridge_reads_reasoning_modes_from_the_deployment_model_family_under_drop_params( + model: str, + model_info: dict[str, str], + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + _, kwargs, route = non_openai_responses_bridge("azure") + + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns"}, + model_info=model_info, + drop_params=True, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {"effort": "low", "context": "all_turns"} + + +def test_azure_ai_bridge_judges_reasoning_fields_by_the_request_api_base_under_drop_params( + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], + monkeypatch: pytest.MonkeyPatch, +): + monkeypatch.setenv("AZURE_AI_API_BASE", "https://legacy-models.eastus2.inference.ai.azure.com") + model, kwargs, route = non_openai_responses_bridge("azure_ai") + + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns"}, + drop_params=True, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {"effort": "low", "context": "all_turns"} + + +@pytest.mark.parametrize("model", ["azure/responses/custom-gpt-6x", "azure/responses/gpt-6-1-sol"]) +@pytest.mark.parametrize("summary", [{}, {"summary": "auto"}]) +def test_azure_bridge_sends_only_effort_and_summary_to_an_unknown_gpt_series_deployment_under_drop_params( + model: str, + summary: dict[str, str], + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + _, kwargs, route = non_openai_responses_bridge("azure") + + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns", **summary}, + drop_params=True, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {"effort": "low", **summary} + + +@pytest.mark.parametrize( + ("backend", "model"), + [ + ("azure", "azure/responses/terra-prod"), + ("databricks", "databricks/responses/databricks-gpt-bridge-test"), + ("databricks", "databricks/responses/databricks-gpt-6.1-standard-only"), + ("litellm_proxy", "litellm_proxy/responses/gpt-6.1-standard-only"), + ], +) +@pytest.mark.parametrize("summary", [{}, {"summary": "auto"}]) +def test_responses_bridge_forwards_mode_and_context_it_cannot_judge_under_drop_params( + backend: str, + model: str, + summary: dict[str, str], + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + _, kwargs, route = non_openai_responses_bridge(backend) + requested: Final = {"effort": "low", "mode": "pro", "context": "all_turns", **summary} + + litellm.completion( + model=model, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=requested, + drop_params=True, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == requested + + +def test_azure_bridge_applies_a_deployment_reasoning_modes_only_to_that_deployment_under_drop_params( + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + _, kwargs, route = non_openai_responses_bridge("azure") + litellm.register_model( + { + "gpt-6.1-pro-capable": { + "litellm_provider": "openai", + "mode": "chat", + "supports_reasoning": True, + "reasoning_modes": ["standard", "pro"], + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + }, + persist_across_reloads=False, + ) + litellm_params: Final = {"model": "azure/responses/gpt-6.1-pro-capable", **kwargs} + router: Final = litellm.Router( + model_list=[ + { + "model_name": "standard-only", + "litellm_params": litellm_params, + "model_info": {"reasoning_modes": ["standard"]}, + }, + {"model_name": "pro-capable", "litellm_params": litellm_params}, + ] + ) + requested: Final = {"effort": "low", "mode": "pro"} + + for model_name in ("standard-only", "pro-capable"): + router.completion( + model=model_name, + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=requested, + drop_params=True, + ) + litellm.completion( + model="azure/responses/gpt-6.1-pro-capable", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=requested, + drop_params=True, + **kwargs, + ) + + assert [json.loads(call.request.content)["reasoning"] for call in route.calls] == [ + {"effort": "low"}, + requested, + requested, + ] + + +def test_azure_bridge_reads_reasoning_modes_from_the_deployment_cost_map_row_under_drop_params( + non_openai_responses_bridge: Callable[[str], tuple[str, Mapping[str, str], respx.Route]], +): + _, kwargs, route = non_openai_responses_bridge("azure") + litellm.register_model( + { + "azure/custom-standard-deployment": { + "litellm_provider": "azure", + "mode": "responses", + "supports_reasoning": True, + "reasoning_modes": ["standard"], + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + }, + persist_across_reloads=False, + ) + + litellm.completion( + model="azure/responses/custom-standard-deployment", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns"}, + drop_params=True, + **kwargs, + ) + + assert json.loads(route.calls[0].request.content)["reasoning"] == {"effort": "low", "context": "all_turns"} + + def test_responses_bridge_fake_stream_is_bounded_by_the_request_timeout_not_the_stream_timeout( restore_model_registry: None, ) -> None: From b8283d8b61f9d15bdc734d58a7f8a0a38cff5bd4 Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 30 Oct 2026 03:58:36 -0700 Subject: [PATCH 19/23] feat(usage): show subscription-covered providers and account fees separately (#44772) * feat(usage): show subscription-covered providers and account fees separately * fix(usage): write only the named fee columns and sync the root schema The fee update now writes just the columns the request names, so two admins changing different fields at once both land, and the dashboard edit form sends only what changed. The root schema.prisma carries the new table like the two copies, and a billing period anchored in the calendar's last month ends on December 31, 9999 instead of failing. * fix(usage): register the subscription table span and keep provider logic under llms * fix(subscription_accounts): register the fee routes as management routes and reject an empty account id * fix(usage): run the new tests in CI and close subscription card review gaps Adds the subscription test file to its unit shard, sets the migration lock timeout, rejects infinite fees, keeps the card from rewriting login files, lists each deployment once without aliases, shows the provider badge on entity views, and asks before removing a fee * fix(usage): resolve named-credential ChatGPT accounts and keep admin viewers read-only on the fee card * fix(usage): count a subscription fee only for accounts behind a current deployment Usage listed a fee row for an account no deployment uses while another ChatGPT deployment existed and dropped it when none did. It now lists only accounts behind a current deployment in both cases. Adds a real proxy and Postgres integration test for the fee API and for ChatGPT Responses rows on a deployment's own login, and names the card's multi-component test with the integration suffix. * fix(usage): keep a stored subscription fee listed after its account's deployments are removed The usage route lists every account behind a subscription deployment plus every account carrying a stored fee, and reads the fees on every proxy, so a fee stays visible, counted, and editable for past ranges after its last deployment goes. The integration test now polls every day the calls spanned, so a UTC midnight crossing no longer fails it. --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .github/workflows/test-unit.yml | 1 + backend/routes/allowlist.py | 1 + .../migration.sql | 21 + .../litellm_proxy_extras/schema.prisma | 16 + litellm/constants.py | 1 + litellm/integrations/otel/model/spans.py | 1 + litellm/llms/chatgpt/authenticator.py | 8 +- litellm/llms/subscription_accounts.py | 37 ++ litellm/models/__init__.py | 2 + litellm/models/subscription_account.py | 28 ++ litellm/proxy/_types.py | 4 + .../common_daily_activity.py | 35 +- .../subscription_accounts.py | 335 ++++++++++++++ litellm/proxy/proxy_server.py | 4 + litellm/proxy/schema.prisma | 16 + .../subscription_billing_periods.py | 54 +++ .../subscription_account_repository.py | 134 ++++++ litellm/responses/main.py | 6 +- .../subscription_accounts.py | 62 +++ schema.prisma | 16 + .../spend/test_chatgpt_subscription_usage.py | 322 +++++++++++++ tests/unit/llms/test_subscription_accounts.py | 110 +++++ .../proxy/db/test_db_spend_update_writer.py | 31 ++ .../test_common_daily_activity.py | 65 +++ .../test_subscription_accounts.py | 438 ++++++++++++++++++ .../test_subscription_billing_periods.py | 73 +++ tests/unit/repositories/test_repositories.py | 43 ++ .../EntityUsage/EntityUsage.test.tsx | 28 ++ .../EntityUsage/SpendByProvider.test.tsx | 28 ++ .../EntityUsage/SpendByProvider.tsx | 18 +- .../EntityUsage/entityUsageAggregations.ts | 6 +- .../SubscriptionCostCard.test.tsx | 185 ++++++++ .../SubscriptionCost/SubscriptionCostCard.tsx | 221 +++++++++ .../SubscriptionCosts.integration.test.tsx | 69 +++ .../SubscriptionCost/SubscriptionCosts.tsx | 38 ++ .../SubscriptionFeeModal.test.tsx | 83 ++++ .../SubscriptionCost/SubscriptionFeeModal.tsx | 114 +++++ .../components/SubscriptionCost/formatting.ts | 34 ++ .../components/SubscriptionCost/types.ts | 74 +++ .../useSubscriptionFeeActions.test.tsx | 79 ++++ .../useSubscriptionFeeActions.ts | 90 ++++ .../components/UsagePageView.test.tsx | 99 +++- .../_components/components/UsagePageView.tsx | 24 +- .../components/overview/UsageOverview.tsx | 4 + .../components/overview/overviewData.test.ts | 29 ++ .../components/overview/overviewData.ts | 10 + .../_components/hooks/useSubscriptionUsage.ts | 47 ++ .../src/components/networking.tsx | 35 ++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 328 +++++++++++++ 49 files changed, 3485 insertions(+), 22 deletions(-) create mode 100644 litellm-proxy-extras/litellm_proxy_extras/migrations/20261028000000_subscription_accounts/migration.sql create mode 100644 litellm/llms/subscription_accounts.py create mode 100644 litellm/models/subscription_account.py create mode 100644 litellm/proxy/management_endpoints/subscription_accounts.py create mode 100644 litellm/proxy/spend_tracking/subscription_billing_periods.py create mode 100644 litellm/repositories/subscription_account_repository.py create mode 100644 litellm/types/proxy/management_endpoints/subscription_accounts.py create mode 100644 tests/integration/spend/test_chatgpt_subscription_usage.py create mode 100644 tests/unit/llms/test_subscription_accounts.py create mode 100644 tests/unit/proxy/management_endpoints/test_subscription_accounts.py create mode 100644 tests/unit/proxy/spend_tracking/test_subscription_billing_periods.py create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.integration.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/formatting.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/types.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.test.tsx create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.ts create mode 100644 ui/litellm-dashboard/src/app/(dashboard)/usage/_components/hooks/useSubscriptionUsage.ts diff --git a/.github/workflows/test-unit.yml b/.github/workflows/test-unit.yml index 9570b4d8cc8..68823d99519 100644 --- a/.github/workflows/test-unit.yml +++ b/.github/workflows/test-unit.yml @@ -227,6 +227,7 @@ jobs: tests/unit/llms/test_oss_decision.py tests/unit/llms/test_polling_url_origin_match.py tests/unit/llms/test_predibase_transformation.py + tests/unit/llms/test_subscription_accounts.py tests/unit/llms/a2a tests/unit/llms/aiml tests/unit/llms/aiohttp_openai diff --git a/backend/routes/allowlist.py b/backend/routes/allowlist.py index 2a7522cf34e..2d3ccf40175 100644 --- a/backend/routes/allowlist.py +++ b/backend/routes/allowlist.py @@ -55,6 +55,7 @@ BACKEND_PATH_PREFIXES: tuple[str, ...] = ( "/cache_settings", "/coordination_redis/", "/cost_tracking", + "/subscription_accounts/", "/cost_optimization/", "/cost/", "/credentials", diff --git a/litellm-proxy-extras/litellm_proxy_extras/migrations/20261028000000_subscription_accounts/migration.sql b/litellm-proxy-extras/litellm_proxy_extras/migrations/20261028000000_subscription_accounts/migration.sql new file mode 100644 index 00000000000..e2788c85be9 --- /dev/null +++ b/litellm-proxy-extras/litellm_proxy_extras/migrations/20261028000000_subscription_accounts/migration.sql @@ -0,0 +1,21 @@ +SET LOCAL lock_timeout = '10s'; + +-- CreateTable +CREATE TABLE IF NOT EXISTS "LiteLLM_SubscriptionAccountTable" ( + "subscription_account_id" TEXT NOT NULL, + "custom_llm_provider" TEXT NOT NULL, + "account_id" TEXT NOT NULL, + "label" TEXT, + "monthly_fee" DOUBLE PRECISION NOT NULL, + "currency" TEXT NOT NULL, + "billing_period_start" TEXT NOT NULL, + "created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP, + "created_by" TEXT NOT NULL, + "updated_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP, + "updated_by" TEXT NOT NULL, + + CONSTRAINT "LiteLLM_SubscriptionAccountTable_pkey" PRIMARY KEY ("subscription_account_id") +); + +-- CreateIndex +CREATE UNIQUE INDEX IF NOT EXISTS "LiteLLM_SubscriptionAccountTable_provider_account_key" ON "LiteLLM_SubscriptionAccountTable"("custom_llm_provider", "account_id"); diff --git a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma index 2965174114b..179b741fe7c 100644 --- a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma +++ b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma @@ -2131,3 +2131,19 @@ model LiteLLM_AccountingDelivery { @@id([event_id, target, version]) @@index([target, state, next_attempt_at]) } + +model LiteLLM_SubscriptionAccountTable { + subscription_account_id String @id @default(uuid()) + custom_llm_provider String + account_id String + label String? + monthly_fee Float + currency String + billing_period_start String + created_at DateTime @default(now()) @map("created_at") + created_by String + updated_at DateTime @default(now()) @updatedAt @map("updated_at") + updated_by String + + @@unique([custom_llm_provider, account_id], map: "LiteLLM_SubscriptionAccountTable_provider_account_key") +} diff --git a/litellm/constants.py b/litellm/constants.py index 8e80c07b4ed..97e93309ceb 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -899,6 +899,7 @@ PROVIDERS_THAT_AUTHENTICATE_ON_PROVIDER_INFO: Final = frozenset( } ) +SUBSCRIPTION_BACKED_PROVIDERS: Final = frozenset({"chatgpt"}) A2A_CHAT_COMPLETION_BRIDGE_PROVIDERS: Final = frozenset({"databricks_agent"}) LITELLM_EMBEDDING_PROVIDERS_SUPPORTING_INPUT_ARRAY_OF_TOKENS: Final = [ diff --git a/litellm/integrations/otel/model/spans.py b/litellm/integrations/otel/model/spans.py index 4907bd4368d..320231430d2 100644 --- a/litellm/integrations/otel/model/spans.py +++ b/litellm/integrations/otel/model/spans.py @@ -361,6 +361,7 @@ _PRISMA_MODELS: Final[frozenset[str]] = frozenset( "LiteLLM_LensReview", "LiteLLM_LensIngestionKey", "LiteLLM_LensWorker", + "LiteLLM_SubscriptionAccountTable", "LiteLLM_LensSignalConfig", "LiteLLM_LensTraceSignal", ) diff --git a/litellm/llms/chatgpt/authenticator.py b/litellm/llms/chatgpt/authenticator.py index a5fb1c9c6b9..b237116c5f1 100644 --- a/litellm/llms/chatgpt/authenticator.py +++ b/litellm/llms/chatgpt/authenticator.py @@ -122,6 +122,12 @@ class Authenticator: return tokens["access_token"] def get_account_id(self, access_token: str | None = None) -> str | None: + return self._account_id(access_token, persist_derived=True) + + def read_account_id(self, access_token: str | None = None) -> str | None: + return self._account_id(access_token, persist_derived=False) + + def _account_id(self, access_token: str | None, persist_derived: bool) -> str | None: auth_data: Final = self._read_auth_file() if access_token and (not auth_data or auth_data.get("access_token") != access_token): return extract_chatgpt_account_id(access_token) @@ -133,7 +139,7 @@ class Authenticator: id_token: Final = auth_data.get("id_token") saved_access_token: Final = auth_data.get("access_token") derived: Final = extract_chatgpt_account_id(_optional_str(id_token or saved_access_token)) - if derived: + if derived and persist_derived: self._write_auth_file({**auth_data, "account_id": derived}) return derived diff --git a/litellm/llms/subscription_accounts.py b/litellm/llms/subscription_accounts.py new file mode 100644 index 00000000000..79c186ca124 --- /dev/null +++ b/litellm/llms/subscription_accounts.py @@ -0,0 +1,37 @@ +""" +Providers whose requests are covered by a flat subscription, each with the way to read +the account id of the login a deployment is signed into. +""" + +from collections.abc import Callable, Mapping +from types import MappingProxyType +from typing import Final + +from litellm.llms.chatgpt.authenticator import Authenticator, get_cached_authenticator, get_chatgpt_auth_file +from litellm.secret_managers.main import UnresolvableSecretReference, resolve_local_oidc_reference +from litellm.utils import with_named_credential + + +def _chatgpt_account_id(litellm_params: Mapping[str, object]) -> str | None: + params: Final = with_named_credential(litellm_params) + auth_file: Final = get_chatgpt_auth_file(params) + if auth_file: + return get_cached_authenticator(auth_file).read_account_id() + api_key: Final = params.get("api_key") + if not isinstance(api_key, str) or not api_key: + return Authenticator().read_account_id() + match resolve_local_oidc_reference(api_key): + case UnresolvableSecretReference(): + return None + case str() as access_token: + return Authenticator().read_account_id(access_token) + + +SUBSCRIPTION_ACCOUNT_ID_RESOLVERS: Final[Mapping[str, Callable[[Mapping[str, object]], str | None]]] = MappingProxyType( + {"chatgpt": _chatgpt_account_id} +) + + +def resolve_subscription_account_id(custom_llm_provider: str, litellm_params: Mapping[str, object]) -> str | None: + resolver: Final = SUBSCRIPTION_ACCOUNT_ID_RESOLVERS.get(custom_llm_provider) + return None if resolver is None else resolver(litellm_params) diff --git a/litellm/models/__init__.py b/litellm/models/__init__.py index 50eb6f8af4f..bb47bd52c2d 100644 --- a/litellm/models/__init__.py +++ b/litellm/models/__init__.py @@ -30,6 +30,7 @@ from litellm.models.organization_membership import LiteLLM_OrganizationMembershi from litellm.models.project import LiteLLM_ProjectTable from litellm.models.skills import LiteLLM_SkillsTable from litellm.models.spend_logs import LiteLLM_ErrorLogs, LiteLLM_SpendLogs +from litellm.models.subscription_account import LiteLLM_SubscriptionAccountTable from litellm.models.tag import LiteLLM_TagTable from litellm.models.team import LiteLLM_TeamTable from litellm.models.team_membership import LiteLLM_TeamMembership @@ -59,6 +60,7 @@ __all__ = [ "LiteLLM_ProxyModelTable", "LiteLLM_SkillsTable", "LiteLLM_SpendLogs", + "LiteLLM_SubscriptionAccountTable", "LiteLLM_TagTable", "LiteLLM_TeamMemberTable", "LiteLLM_TeamMembership", diff --git a/litellm/models/subscription_account.py b/litellm/models/subscription_account.py new file mode 100644 index 00000000000..36eaa0ffc37 --- /dev/null +++ b/litellm/models/subscription_account.py @@ -0,0 +1,28 @@ +""" +Subscription account table model. + +Canonical definition for ``litellm_subscriptionaccounttable``: the operator-entered +recurring fee of a provider account whose requests are covered by a flat subscription. +""" + +from datetime import datetime + +from pydantic import ConfigDict + +from litellm.types.llms.base import LiteLLMPydanticObjectBase + + +class LiteLLM_SubscriptionAccountTable(LiteLLMPydanticObjectBase): + model_config = ConfigDict(protected_namespaces=(), frozen=True) + + subscription_account_id: str + custom_llm_provider: str + account_id: str + label: str | None = None + monthly_fee: float + currency: str + billing_period_start: str + created_at: datetime + created_by: str + updated_at: datetime + updated_by: str diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 0a09fe4a855..f165355da5a 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -667,6 +667,10 @@ class LiteLLMRoutes(enum.Enum): "/team/spend/by_user", # gateway request counts (SGR); deployment-wide, admin-only "/gateway/daily/activity", + "/subscription_accounts/usage", + "/subscription_accounts/new", + "/subscription_accounts/update", + "/subscription_accounts/delete", # model "/model/new", "/model/update", diff --git a/litellm/proxy/management_endpoints/common_daily_activity.py b/litellm/proxy/management_endpoints/common_daily_activity.py index 98372b64015..6e4ea0bd0cb 100644 --- a/litellm/proxy/management_endpoints/common_daily_activity.py +++ b/litellm/proxy/management_endpoints/common_daily_activity.py @@ -12,7 +12,7 @@ from typing_extensions import ReadOnly, TypedDict, assert_never from litellm import constants from litellm._logging import verbose_proxy_logger -from litellm.constants import PTU_SENTINEL_API_KEY +from litellm.constants import PTU_SENTINEL_API_KEY, SUBSCRIPTION_BACKED_PROVIDERS from litellm.proxy._types import CommonProxyErrors from litellm.proxy.spend_tracking.key_metadata_recovery import ( attach_user_details, @@ -295,6 +295,12 @@ def _entity_metadata( return stored if stored is not None else {} +def provider_bucket_metadata(provider: str, base: Mapping[str, object]) -> Mapping[str, object]: + if provider in SUBSCRIPTION_BACKED_PROVIDERS: + return {**base, "subscription_covered": True} + return dict(base) + + def update_breakdown_metrics( breakdown: BreakdownMetrics, record: DailySpendRecord, @@ -395,7 +401,7 @@ def update_breakdown_metrics( if provider not in breakdown.providers: breakdown.providers[provider] = MetricWithMetadata( metrics=SpendMetrics(), - metadata=provider_metadata.get(provider, {}), # Add any provider-specific metadata here + metadata=provider_bucket_metadata(provider, provider_metadata.get(provider, {})), ) breakdown.providers[provider].metrics = update_metrics(breakdown.providers[provider].metrics, record) @@ -795,12 +801,23 @@ def _aggregate_grouping_sets_records_sync( grouped_data[date_str] = bucket return bucket - def assign_metric_with_metadata(target: dict[str, MetricWithMetadata], key: str, metrics: SpendMetrics) -> None: + def assign_metric_with_metadata( + target: dict[str, MetricWithMetadata], + key: str, + metrics: SpendMetrics, + metadata: Mapping[str, object] | None = None, + ) -> None: existing: Final = target.get(key) - if existing is None: - target[key] = MetricWithMetadata(metrics=metrics, metadata={}) - else: - existing.metrics = sum_spend_metrics((existing.metrics, metrics)) + target[key] = ( + MetricWithMetadata(metrics=metrics, metadata=dict(metadata or {})) + if existing is None + else existing.model_copy( + update={ + "metrics": sum_spend_metrics((existing.metrics, metrics)), + "metadata": {**existing.metadata, **(metadata or {})}, + } + ) + ) def assign_api_key_breakdown( target: dict[str, MetricWithMetadata], @@ -864,7 +881,9 @@ def _aggregate_grouping_sets_records_sync( # provider the base build reported. provider_metrics = metrics.model_copy(update={"flat_cost": 0.0}) provider = record.custom_llm_provider or "unknown" - assign_metric_with_metadata(breakdown.providers, provider, provider_metrics) + assign_metric_with_metadata( + breakdown.providers, provider, provider_metrics, provider_bucket_metadata(provider, {}) + ) elif level == _GROUP_DATE_PROVIDER_API_KEY: if not is_ptu_sentinel: provider = record.custom_llm_provider or "unknown" diff --git a/litellm/proxy/management_endpoints/subscription_accounts.py b/litellm/proxy/management_endpoints/subscription_accounts.py new file mode 100644 index 00000000000..363d6cfba5d --- /dev/null +++ b/litellm/proxy/management_endpoints/subscription_accounts.py @@ -0,0 +1,335 @@ +""" +Subscription accounts: the operator-entered recurring fee of a provider account whose +requests are covered by a flat subscription, shown beside metered spend on the Usage page. +""" + +import re +from collections.abc import Callable, Iterable, Mapping +from dataclasses import dataclass +from datetime import date +from itertools import chain +from typing import TYPE_CHECKING, Annotated, Final, TypeAlias + +from fastapi import APIRouter, Depends, HTTPException, Query, status +from pydantic import TypeAdapter + +from litellm.constants import LITELLM_PROXY_ADMIN_NAME, SUBSCRIPTION_BACKED_PROVIDERS +from litellm.litellm_core_utils.get_llm_provider_logic import declared_authenticating_provider +from litellm.llms.subscription_accounts import resolve_subscription_account_id +from litellm.models.subscription_account import LiteLLM_SubscriptionAccountTable +from litellm.proxy._types import CommonProxyErrors, LitellmUserRoles, UserAPIKeyAuth, user_api_key_has_admin_view +from litellm.proxy.auth.user_api_key_auth import user_api_key_auth +from litellm.proxy.spend_tracking.subscription_billing_periods import billing_periods_starting_within +from litellm.proxy.utils import get_prisma_client_or_throw +from litellm.repositories.subscription_account_repository import ( + DuplicateSubscriptionAccount, + KeepLabel, + SetLabel, + SubscriptionAccountFeeChanges, + SubscriptionAccountFeeFields, + SubscriptionAccountRepository, + SubscriptionAccountStore, +) +from litellm.types.proxy.management_endpoints.subscription_accounts import ( + BillingPeriod, + DeleteSubscriptionAccountRequest, + SubscriptionAccountFeeRequest, + SubscriptionAccountUsage, + SubscriptionUsageResponse, + UpdateSubscriptionAccountFeeRequest, +) +from litellm.types.router import DeploymentTypedDict + +if TYPE_CHECKING: + from litellm.router import Router + +router: Final = APIRouter(prefix="/subscription_accounts", tags=["subscription accounts"]) +Auth: TypeAlias = Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)] + + +@dataclass(frozen=True, slots=True) +class SubscriptionDeployment: + model_name: str + custom_llm_provider: str + account_id: str | None + + +AccountIdResolver: TypeAlias = Callable[[str, Mapping[str, object]], str | None] +_MODEL_INFO_ADAPTER: Final = TypeAdapter(Mapping[str, object]) + + +def _declared_provider(litellm_params: Mapping[str, object]) -> str | None: + model: Final = litellm_params.get("model") + configured_provider: Final = litellm_params.get("custom_llm_provider") + return declared_authenticating_provider( + model if isinstance(model, str) else None, + configured_provider if isinstance(configured_provider, str) else None, + ) + + +def _model_info(deployment: Mapping[str, object]) -> Mapping[str, object]: + return _MODEL_INFO_ADAPTER.validate_python(deployment.get("model_info") or {}) + + +def _public_model_name(entry: DeploymentTypedDict) -> str: + team_public_model_name: Final = _model_info(entry).get("team_public_model_name") + return team_public_model_name if isinstance(team_public_model_name, str) else entry["model_name"] + + +def _subscription_deployment( + entry: DeploymentTypedDict, account_id_resolver: AccountIdResolver +) -> SubscriptionDeployment | None: + litellm_params: Final = entry["litellm_params"] + provider: Final = _declared_provider(litellm_params) + if provider is None or provider not in SUBSCRIPTION_BACKED_PROVIDERS: + return None + return SubscriptionDeployment( + model_name=_public_model_name(entry), + custom_llm_provider=provider, + account_id=account_id_resolver(provider, litellm_params), + ) + + +def subscription_deployments_from_model_list( + model_list: Iterable[DeploymentTypedDict], account_id_resolver: AccountIdResolver +) -> tuple[SubscriptionDeployment, ...]: + candidates: Final = (_subscription_deployment(entry, account_id_resolver) for entry in model_list) + return tuple(candidate for candidate in candidates if candidate is not None) + + +def subscription_deployments_of_router( + llm_router: "Router | None", account_id_resolver: AccountIdResolver +) -> tuple[SubscriptionDeployment, ...]: + if llm_router is None: + return () + registry: Final = llm_router.registry() + indexed_names: Final = tuple(registry.index("model_name")) + deployments: Final = chain.from_iterable(registry.deployments(name, source="named") or () for name in indexed_names) + return subscription_deployments_from_model_list(deployments, account_id_resolver) + + +def get_subscription_deployments() -> tuple[SubscriptionDeployment, ...]: + from litellm.proxy.proxy_server import llm_router + + return subscription_deployments_of_router(llm_router, resolve_subscription_account_id) + + +def get_subscription_account_repository() -> SubscriptionAccountStore: + return SubscriptionAccountRepository(get_prisma_client_or_throw(CommonProxyErrors.db_not_connected_error.value)) + + +Store: TypeAlias = Annotated[SubscriptionAccountStore, Depends(get_subscription_account_repository)] +Deployments: TypeAlias = Annotated[tuple[SubscriptionDeployment, ...], Depends(get_subscription_deployments)] + + +def _forbidden(user_api_key_dict: UserAPIKeyAuth) -> HTTPException: + return HTTPException( + status_code=status.HTTP_403_FORBIDDEN, + detail={"error": f"{CommonProxyErrors.not_allowed_access.value}, your role={user_api_key_dict.user_role}"}, + ) + + +def _account_usage( + custom_llm_provider: str, + account_id: str | None, + deployments: Iterable[SubscriptionDeployment], + fee_row: LiteLLM_SubscriptionAccountTable | None, + start_date: date, + end_date: date, +) -> SubscriptionAccountUsage: + model_names: Final = tuple(sorted({deployment.model_name for deployment in deployments})) + if fee_row is None: + return SubscriptionAccountUsage( + custom_llm_provider=custom_llm_provider, + account_id=account_id, + deployments=model_names, + fee=None, + billing_periods=(), + fixed_cost=None, + ) + periods: Final = billing_periods_starting_within( + date.fromisoformat(fee_row.billing_period_start), start_date, end_date + ) + return SubscriptionAccountUsage( + custom_llm_provider=custom_llm_provider, + account_id=account_id, + deployments=model_names, + fee=fee_row, + billing_periods=tuple(BillingPeriod(start=period.start, end=period.end) for period in periods), + fixed_cost=len(periods) * fee_row.monthly_fee, + ) + + +def build_subscription_usage( + deployments: Iterable[SubscriptionDeployment], + fee_rows: Iterable[LiteLLM_SubscriptionAccountTable], + start_date: date, + end_date: date, +) -> SubscriptionUsageResponse: + deployment_tuple: Final = tuple(deployments) + fees_by_account: Final = {(row.custom_llm_provider, row.account_id): row for row in fee_rows} + deployment_keys: Final = {(d.custom_llm_provider, d.account_id) for d in deployment_tuple} + account_keys: Final = sorted(deployment_keys | set(fees_by_account), key=lambda key: (key[0], key[1] or "~")) + accounts: Final = tuple( + _account_usage( + provider, + account_id, + (d for d in deployment_tuple if (d.custom_llm_provider, d.account_id) == (provider, account_id)), + fees_by_account.get((provider, account_id)) if account_id is not None else None, + start_date, + end_date, + ) + for provider, account_id in account_keys + ) + return SubscriptionUsageResponse( + start_date=start_date, + end_date=end_date, + subscription_providers=tuple(sorted(SUBSCRIPTION_BACKED_PROVIDERS)), + accounts=accounts, + ) + + +_DAY_PATTERN: Final = re.compile(r"\d{4}-\d{2}-\d{2}") + + +def _parse_day(value: str) -> date | None: + if _DAY_PATTERN.fullmatch(value) is None: + return None + try: + return date.fromisoformat(value) + except ValueError: + return None + + +def _parse_date(value: str, name: str) -> date: + parsed: Final = _parse_day(value) + if parsed is None: + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, detail={"error": f"{name} must be YYYY-MM-DD, got {value!r}"} + ) + return parsed + + +@router.get("/usage", response_model=SubscriptionUsageResponse, dependencies=[Depends(user_api_key_auth)]) +async def get_subscription_usage( + start_date: Annotated[str, Query(description="First day of the Usage page filter, YYYY-MM-DD")], + end_date: Annotated[str, Query(description="Last day of the Usage page filter, YYYY-MM-DD")], + user_api_key_dict: Auth, + repository: Store, + deployments: Deployments, +) -> SubscriptionUsageResponse: + """ + Subscription-backed accounts behind the proxy's deployments or carrying a stored fee, each with + its configured monthly fee and the fixed cost attributed to the date range: the fee once per billing + period whose start date falls inside the range, however many deployments share the account. + """ + if not user_api_key_has_admin_view(user_api_key_dict): + raise _forbidden(user_api_key_dict) + start: Final = _parse_date(start_date, "start_date") + end: Final = _parse_date(end_date, "end_date") + if end < start: + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, detail={"error": "end_date must not be before start_date"} + ) + return build_subscription_usage(deployments, await repository.list_all(), start, end) + + +def _require_proxy_admin(user_api_key_dict: UserAPIKeyAuth) -> None: + if user_api_key_dict.user_role != LitellmUserRoles.PROXY_ADMIN: + raise _forbidden(user_api_key_dict) + + +def _actor(user_api_key_dict: UserAPIKeyAuth) -> str: + return user_api_key_dict.user_id or LITELLM_PROXY_ADMIN_NAME + + +@router.post("/new", response_model=LiteLLM_SubscriptionAccountTable, dependencies=[Depends(user_api_key_auth)]) +async def new_subscription_account( + request: SubscriptionAccountFeeRequest, + user_api_key_dict: Auth, + repository: Store, +) -> LiteLLM_SubscriptionAccountTable: + """Record the monthly fee of one subscription-backed provider account.""" + _require_proxy_admin(user_api_key_dict) + if request.custom_llm_provider not in SUBSCRIPTION_BACKED_PROVIDERS: + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail={ + "error": f"{request.custom_llm_provider!r} is not a subscription-backed provider; " + f"expected one of {sorted(SUBSCRIPTION_BACKED_PROVIDERS)}" + }, + ) + created: Final = await repository.create_account( + request.custom_llm_provider, + request.account_id, + SubscriptionAccountFeeFields( + monthly_fee=request.monthly_fee, + currency=request.currency, + billing_period_start=request.billing_period_start.isoformat(), + label=request.label, + ), + _actor(user_api_key_dict), + ) + if isinstance(created, DuplicateSubscriptionAccount): + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail={ + "error": f"A fee is already configured for {created.custom_llm_provider} account " + f"{created.account_id}; use /subscription_accounts/update to change it" + }, + ) + return created + + +def _fee_changes(request: UpdateSubscriptionAccountFeeRequest) -> SubscriptionAccountFeeChanges: + return SubscriptionAccountFeeChanges( + monthly_fee=request.monthly_fee, + currency=request.currency, + billing_period_start=( + request.billing_period_start.isoformat() if request.billing_period_start is not None else None + ), + label=SetLabel(request.label) if "label" in request.model_fields_set else KeepLabel(), + ) + + +@router.post("/update", response_model=LiteLLM_SubscriptionAccountTable, dependencies=[Depends(user_api_key_auth)]) +async def update_subscription_account( + request: UpdateSubscriptionAccountFeeRequest, + user_api_key_dict: Auth, + repository: Store, +) -> LiteLLM_SubscriptionAccountTable: + """ + Change the fee, currency, billing period start, or label of a subscription account. + Only the fields named in the request are written, so two admins changing different + fields at the same time both land. + """ + _require_proxy_admin(user_api_key_dict) + updated: Final = await repository.update_fee( + request.subscription_account_id, _fee_changes(request), _actor(user_api_key_dict) + ) + if updated is None: + raise _not_found(request.subscription_account_id) + return updated + + +@router.post("/delete", response_model=LiteLLM_SubscriptionAccountTable, dependencies=[Depends(user_api_key_auth)]) +async def delete_subscription_account( + request: DeleteSubscriptionAccountRequest, + user_api_key_dict: Auth, + repository: Store, +) -> LiteLLM_SubscriptionAccountTable: + """Stop tracking a subscription account's fee; its requests stay marked as subscription-covered.""" + _require_proxy_admin(user_api_key_dict) + if await repository.find_by_id(request.subscription_account_id) is None: + raise _not_found(request.subscription_account_id) + deleted: Final = await repository.delete_account(request.subscription_account_id) + if deleted is None: + raise _not_found(request.subscription_account_id) + return deleted + + +def _not_found(subscription_account_id: str) -> HTTPException: + return HTTPException( + status_code=status.HTTP_404_NOT_FOUND, + detail={"error": f"No subscription account with id {subscription_account_id}"}, + ) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 3e004f1ced7..ae0e7fce667 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -777,6 +777,9 @@ from litellm.proxy.management_endpoints.session_endpoints import ( router as session_management_router, ) from litellm.proxy.management_endpoints.sso.saml_sso import SAML_AUTHN_REQUEST_TTL_SECONDS +from litellm.proxy.management_endpoints.subscription_accounts import ( + router as subscription_accounts_router, +) from litellm.proxy.management_endpoints.tag_management_endpoints import ( router as tag_management_router, ) @@ -21769,6 +21772,7 @@ app.include_router(memory_router) app.include_router(lens_router) app.include_router(plugin_router) app.include_router(cost_tracking_settings_router) +app.include_router(subscription_accounts_router) app.include_router(prompt_caching_requests_router) app.include_router(router_settings_router) app.include_router(fallback_management_router) diff --git a/litellm/proxy/schema.prisma b/litellm/proxy/schema.prisma index 2965174114b..179b741fe7c 100644 --- a/litellm/proxy/schema.prisma +++ b/litellm/proxy/schema.prisma @@ -2131,3 +2131,19 @@ model LiteLLM_AccountingDelivery { @@id([event_id, target, version]) @@index([target, state, next_attempt_at]) } + +model LiteLLM_SubscriptionAccountTable { + subscription_account_id String @id @default(uuid()) + custom_llm_provider String + account_id String + label String? + monthly_fee Float + currency String + billing_period_start String + created_at DateTime @default(now()) @map("created_at") + created_by String + updated_at DateTime @default(now()) @updatedAt @map("updated_at") + updated_by String + + @@unique([custom_llm_provider, account_id], map: "LiteLLM_SubscriptionAccountTable_provider_account_key") +} diff --git a/litellm/proxy/spend_tracking/subscription_billing_periods.py b/litellm/proxy/spend_tracking/subscription_billing_periods.py new file mode 100644 index 00000000000..a7282cc942c --- /dev/null +++ b/litellm/proxy/spend_tracking/subscription_billing_periods.py @@ -0,0 +1,54 @@ +from calendar import monthrange +from dataclasses import dataclass +from datetime import date, timedelta +from typing import Final + + +@dataclass(frozen=True, slots=True) +class BillingPeriodRange: + start: date + end: date + + +def _month_index(anchor: date, months_after_anchor: int) -> int: + return anchor.year * 12 + (anchor.month - 1) + months_after_anchor + + +_LAST_MONTH_INDEX: Final = _month_index(date.max, 0) + + +def _month_start(anchor: date, months_after_anchor: int) -> date: + month_index: Final = _month_index(anchor, months_after_anchor) + year: Final = month_index // 12 + month: Final = month_index % 12 + 1 + return date(year, month, min(anchor.day, monthrange(year, month)[1])) + + +def _period_end(anchor: date, months_after_anchor: int) -> date: + if _month_index(anchor, months_after_anchor + 1) > _LAST_MONTH_INDEX: + return date.max + return _month_start(anchor, months_after_anchor + 1) - timedelta(days=1) + + +def _first_candidate_index(anchor: date, start_date: date) -> int: + months_apart: Final = (start_date.year - anchor.year) * 12 + (start_date.month - anchor.month) + return max(0, months_apart - 1) + + +def billing_periods_starting_within(anchor: date, start_date: date, end_date: date) -> tuple[BillingPeriodRange, ...]: + """Monthly periods anchored on `anchor`'s day of month whose start lies inside [start_date, end_date].""" + if end_date < start_date or end_date < anchor: + return () + first_index: Final = _first_candidate_index(anchor, start_date) + last_index: Final = min( + first_index + _month_span(start_date, end_date) + 1, _LAST_MONTH_INDEX - _month_index(anchor, 0) + ) + candidates: Final = ( + BillingPeriodRange(start=_month_start(anchor, index), end=_period_end(anchor, index)) + for index in range(first_index, last_index + 1) + ) + return tuple(period for period in candidates if start_date <= period.start <= end_date) + + +def _month_span(start_date: date, end_date: date) -> int: + return (end_date.year - start_date.year) * 12 + (end_date.month - start_date.month) diff --git a/litellm/repositories/subscription_account_repository.py b/litellm/repositories/subscription_account_repository.py new file mode 100644 index 00000000000..ddf441a9191 --- /dev/null +++ b/litellm/repositories/subscription_account_repository.py @@ -0,0 +1,134 @@ +""" +Subscription account repository for database operations on LiteLLM_SubscriptionAccountTable. +""" + +from collections.abc import Mapping +from dataclasses import dataclass +from types import MappingProxyType +from typing import TYPE_CHECKING, Final, Protocol, TypeAlias + +from litellm.models.subscription_account import LiteLLM_SubscriptionAccountTable +from litellm.repositories.base_repository import BaseRepository, is_unique_violation +from litellm.repositories.prisma_protocols import TableActions +from litellm.repositories.table_repositories import PrismaTableRepository + +if TYPE_CHECKING: + from prisma import models as prisma_models + + +@dataclass(frozen=True, slots=True) +class DuplicateSubscriptionAccount: + custom_llm_provider: str + account_id: str + + +@dataclass(frozen=True, slots=True) +class SubscriptionAccountFeeFields: + monthly_fee: float + currency: str + billing_period_start: str + label: str | None + + +@dataclass(frozen=True, slots=True) +class KeepLabel: + pass + + +@dataclass(frozen=True, slots=True) +class SetLabel: + value: str | None + + +LabelChange: TypeAlias = KeepLabel | SetLabel + + +@dataclass(frozen=True, slots=True) +class SubscriptionAccountFeeChanges: + monthly_fee: float | None = None + currency: str | None = None + billing_period_start: str | None = None + label: LabelChange = KeepLabel() + + def columns(self) -> Mapping[str, float | str | None]: + named: Final = { + "monthly_fee": self.monthly_fee, + "currency": self.currency, + "billing_period_start": self.billing_period_start, + } + present: Final = {column: value for column, value in named.items() if value is not None} + label_column: Final = {"label": self.label.value} if isinstance(self.label, SetLabel) else {} + return MappingProxyType({**present, **label_column}) + + +class SubscriptionAccountStore(Protocol): + async def list_all(self) -> tuple[LiteLLM_SubscriptionAccountTable, ...]: ... + + async def find_by_id(self, id_value: str) -> LiteLLM_SubscriptionAccountTable | None: ... + + async def create_account( + self, custom_llm_provider: str, account_id: str, fee: SubscriptionAccountFeeFields, created_by: str + ) -> LiteLLM_SubscriptionAccountTable | DuplicateSubscriptionAccount: ... + + async def update_fee( + self, subscription_account_id: str, changes: SubscriptionAccountFeeChanges, updated_by: str + ) -> LiteLLM_SubscriptionAccountTable | None: ... + + async def delete_account(self, subscription_account_id: str) -> LiteLLM_SubscriptionAccountTable | None: ... + + +class _SubscriptionAccountTable(PrismaTableRepository["prisma_models.LiteLLM_SubscriptionAccountTable"]): + table_name = "litellm_subscriptionaccounttable" + + +class SubscriptionAccountRepository(BaseRepository[LiteLLM_SubscriptionAccountTable]): + def __init__(self, prisma_client: object) -> None: + super().__init__(prisma_client) + self._table: Final = _SubscriptionAccountTable(prisma_client) + + @property + def table(self) -> TableActions["prisma_models.LiteLLM_SubscriptionAccountTable"]: + return self._table.table + + @property + def model_class(self) -> type[LiteLLM_SubscriptionAccountTable]: + return LiteLLM_SubscriptionAccountTable + + async def list_all(self) -> tuple[LiteLLM_SubscriptionAccountTable, ...]: + return tuple(await self.find_many(order={"created_at": "asc"})) + + async def find_by_id( + self, id_value: str, id_field: str = "subscription_account_id" + ) -> LiteLLM_SubscriptionAccountTable | None: + return await super().find_by_id(id_value, id_field) + + async def create_account( + self, custom_llm_provider: str, account_id: str, fee: SubscriptionAccountFeeFields, created_by: str + ) -> LiteLLM_SubscriptionAccountTable | DuplicateSubscriptionAccount: + try: + return await self.create( + { + "custom_llm_provider": custom_llm_provider, + "account_id": account_id, + "monthly_fee": fee.monthly_fee, + "currency": fee.currency, + "billing_period_start": fee.billing_period_start, + "label": fee.label, + "created_by": created_by, + "updated_by": created_by, + } + ) + except Exception as exc: + if is_unique_violation(exc): + return DuplicateSubscriptionAccount(custom_llm_provider=custom_llm_provider, account_id=account_id) + raise + + async def update_fee( + self, subscription_account_id: str, changes: SubscriptionAccountFeeChanges, updated_by: str + ) -> LiteLLM_SubscriptionAccountTable | None: + return await self.update( + subscription_account_id, {**changes.columns(), "updated_by": updated_by}, id_field="subscription_account_id" + ) + + async def delete_account(self, subscription_account_id: str) -> LiteLLM_SubscriptionAccountTable | None: + return await self.delete(subscription_account_id, id_field="subscription_account_id") diff --git a/litellm/responses/main.py b/litellm/responses/main.py index 063ef9d91ea..5c8350589b8 100644 --- a/litellm/responses/main.py +++ b/litellm/responses/main.py @@ -31,7 +31,10 @@ from litellm.integrations.anthropic_cache_control_hook import ( from litellm.litellm_core_utils.asyncify import run_async_function from litellm.litellm_core_utils.completion_timeout import CompletionTimeout, StreamKind from litellm.litellm_core_utils.core_helpers import drops_unsupported_params, normalize_drop_params -from litellm.litellm_core_utils.get_llm_provider_logic import get_provider_affinity_header +from litellm.litellm_core_utils.get_llm_provider_logic import ( + declared_authenticating_provider, + get_provider_affinity_header, +) from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup from litellm.litellm_core_utils.prompt_templates.common_utils import ( @@ -805,6 +808,7 @@ def provider_stores_the_conversation(model: str, custom_llm_provider: str | None return not _will_bridge_to_chat_completions( model, custom_llm_provider + or declared_authenticating_provider(model) or litellm.get_llm_provider(model=model, api_base=api_base, litellm_params=litellm_params)[1], bool(kwargs.get("use_chat_completions_api")), kwargs.get("model_info"), diff --git a/litellm/types/proxy/management_endpoints/subscription_accounts.py b/litellm/types/proxy/management_endpoints/subscription_accounts.py new file mode 100644 index 00000000000..a924b97034b --- /dev/null +++ b/litellm/types/proxy/management_endpoints/subscription_accounts.py @@ -0,0 +1,62 @@ +from datetime import date +from typing import Final + +from pydantic import BaseModel, ConfigDict, Field + +from litellm.models.subscription_account import LiteLLM_SubscriptionAccountTable + +CURRENCY_CODE_PATTERN: Final = r"^[A-Z]{3}$" + + +class SubscriptionAccountFeeRequest(BaseModel): + model_config = ConfigDict(frozen=True) + + custom_llm_provider: str = Field(min_length=1) + account_id: str = Field(min_length=1) + monthly_fee: float = Field(gt=0, allow_inf_nan=False) + currency: str = Field(pattern=CURRENCY_CODE_PATTERN) + billing_period_start: date + label: str | None = None + + +class UpdateSubscriptionAccountFeeRequest(BaseModel): + model_config = ConfigDict(frozen=True) + + subscription_account_id: str + monthly_fee: float | None = Field(default=None, gt=0, allow_inf_nan=False) + currency: str | None = Field(default=None, pattern=CURRENCY_CODE_PATTERN) + billing_period_start: date | None = None + label: str | None = None + + +class DeleteSubscriptionAccountRequest(BaseModel): + model_config = ConfigDict(frozen=True) + + subscription_account_id: str + + +class BillingPeriod(BaseModel): + model_config = ConfigDict(frozen=True) + + start: date + end: date + + +class SubscriptionAccountUsage(BaseModel): + model_config = ConfigDict(frozen=True) + + custom_llm_provider: str + account_id: str | None + deployments: tuple[str, ...] + fee: LiteLLM_SubscriptionAccountTable | None + billing_periods: tuple[BillingPeriod, ...] + fixed_cost: float | None + + +class SubscriptionUsageResponse(BaseModel): + model_config = ConfigDict(frozen=True) + + start_date: date + end_date: date + subscription_providers: tuple[str, ...] + accounts: tuple[SubscriptionAccountUsage, ...] diff --git a/schema.prisma b/schema.prisma index 2965174114b..179b741fe7c 100644 --- a/schema.prisma +++ b/schema.prisma @@ -2131,3 +2131,19 @@ model LiteLLM_AccountingDelivery { @@id([event_id, target, version]) @@index([target, state, next_attempt_at]) } + +model LiteLLM_SubscriptionAccountTable { + subscription_account_id String @id @default(uuid()) + custom_llm_provider String + account_id String + label String? + monthly_fee Float + currency String + billing_period_start String + created_at DateTime @default(now()) @map("created_at") + created_by String + updated_at DateTime @default(now()) @updatedAt @map("updated_at") + updated_by String + + @@unique([custom_llm_provider, account_id], map: "LiteLLM_SubscriptionAccountTable_provider_account_key") +} diff --git a/tests/integration/spend/test_chatgpt_subscription_usage.py b/tests/integration/spend/test_chatgpt_subscription_usage.py new file mode 100644 index 00000000000..8f1fb2d0f74 --- /dev/null +++ b/tests/integration/spend/test_chatgpt_subscription_usage.py @@ -0,0 +1,322 @@ +import uuid +from collections.abc import Iterator, Mapping +from dataclasses import dataclass +from datetime import UTC, datetime +from typing import Final + +import httpx +import pytest +from integration._support import codex_vendor as cv +from integration._support import responses_vendor as rv +from integration._support.client import Gateway, Scenario, eventually, gateway_from_environment, string_value +from integration._support.process import OwnedProxy, owned_proxy_process +from integration._support.wire import Wire, wire_server +from pydantic import BaseModel, ConfigDict, JsonValue, TypeAdapter + +_MODEL: Final = "chatgpt/gpt-5.5" +_PROVIDER: Final = "chatgpt" +_FEE_TABLE_QUERY: Final = ( + 'SELECT subscription_account_id FROM "LiteLLM_SubscriptionAccountTable" WHERE custom_llm_provider = %s' + " AND account_id = %s" +) +_FEE_ROW_QUERY: Final = ( + 'SELECT subscription_account_id FROM "LiteLLM_SubscriptionAccountTable" WHERE subscription_account_id = %s' +) +_MONTHLY_FEE: Final = 200.0 +_BILLING_PERIOD_START: Final = "2026-01-15" +_RANGE: Final = {"start_date": "2026-03-01", "end_date": "2026-03-31"} +_PERIOD_IN_RANGE: Final = ("2026-03-15", "2026-04-14") +_TOKENS: Final = TypeAdapter(int) + + +class _Metrics(BaseModel): + model_config = ConfigDict(frozen=True) + + spend: float + prompt_tokens: int + completion_tokens: int + api_requests: int + successful_requests: int + failed_requests: int + + +class _ProviderUsage(BaseModel): + model_config = ConfigDict(frozen=True) + + metrics: _Metrics + metadata: Mapping[str, JsonValue] + + +class _Breakdown(BaseModel): + model_config = ConfigDict(frozen=True) + + providers: Mapping[str, _ProviderUsage] + + +class _Day(BaseModel): + model_config = ConfigDict(frozen=True) + + breakdown: _Breakdown + + +class _DailyActivity(BaseModel): + model_config = ConfigDict(frozen=True) + + results: tuple[_Day, ...] + + +class _Fee(BaseModel): + model_config = ConfigDict(frozen=True) + + subscription_account_id: str + custom_llm_provider: str + account_id: str + monthly_fee: float + currency: str + billing_period_start: str + + +class _Period(BaseModel): + model_config = ConfigDict(frozen=True) + + start: str + end: str + + +class _Account(BaseModel): + model_config = ConfigDict(frozen=True) + + custom_llm_provider: str + account_id: str | None + deployments: tuple[str, ...] + fee: _Fee | None + billing_periods: tuple[_Period, ...] + fixed_cost: float | None + + +class _SubscriptionUsage(BaseModel): + model_config = ConfigDict(frozen=True) + + accounts: tuple[_Account, ...] + + +@dataclass(frozen=True, slots=True) +class _Rig: + wire: Wire + proxy: OwnedProxy + model: str + + @property + def gateway(self) -> Gateway: + return self.proxy.gateway + + +@pytest.fixture(scope="module") +def rig(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Rig]: + directory: Final = tmp_path_factory.mktemp("chatgpt-subscription-usage") + no_default_login: Final = directory / "no-default-login" + no_default_login.mkdir() + deployment_login: Final = cv.login(directory) / "auth.json" + with gateway_from_environment() as gateway, wire_server(cv.CodexVendor().respond) as wire: + overrides: Final = {"CHATGPT_TOKEN_DIR": str(no_default_login), "CHATGPT_API_BASE": wire.url} + config: Final = cv.proxy_config(directory, probe=False) + with ( + owned_proxy_process(gateway, directory, overrides, config=config, workers=2) as owned, + owned.gateway.scenario() as scenario, + ): + model: Final = scenario.model( + model=_MODEL, api_base=wire.url, api_key=None, chatgpt_auth_file=str(deployment_login) + ) + yield _Rig(wire, owned, model) + + +def _responses_body(model: str, marker: str, *, stream: bool) -> Mapping[str, JsonValue]: + return { + "model": model, + "input": [{"role": "user", "content": f"Say marker-{marker}"}], + "stream": stream, + "cache": {"no-cache": True}, + } + + +def _assert_answered(response: httpx.Response, marker: str, content_type: str) -> None: + assert response.status_code == 200, response.text + assert response.headers["content-type"].startswith(content_type), response.headers + assert rv.answer(marker) in response.text, response.text + + +def _today() -> str: + return datetime.now(UTC).date().isoformat() + + +def _chatgpt_usage_between(gateway: Gateway, first_day: str, last_day: str) -> tuple[_ProviderUsage, ...]: + response: Final = gateway.request( + "GET", "/user/daily/activity", params={"start_date": first_day, "end_date": last_day} + ) + assert response.status_code == 200, response.text + activity: Final = _DailyActivity.model_validate_json(response.content) + return tuple( + result.breakdown.providers[_PROVIDER] for result in activity.results if _PROVIDER in result.breakdown.providers + ) + + +def _summed(buckets: tuple[_ProviderUsage, ...]) -> _Metrics: + return _Metrics( + spend=sum(bucket.metrics.spend for bucket in buckets), + prompt_tokens=sum(bucket.metrics.prompt_tokens for bucket in buckets), + completion_tokens=sum(bucket.metrics.completion_tokens for bucket in buckets), + api_requests=sum(bucket.metrics.api_requests for bucket in buckets), + successful_requests=sum(bucket.metrics.successful_requests for bucket in buckets), + failed_requests=sum(bucket.metrics.failed_requests for bucket in buckets), + ) + + +def _usage(gateway: Gateway, *, key: str | None = None) -> httpx.Response: + return gateway.request("GET", "/subscription_accounts/usage", params=_RANGE, key=key) + + +def _accounts(gateway: Gateway) -> tuple[_Account, ...]: + response: Final = _usage(gateway) + assert response.status_code == 200, response.text + return _SubscriptionUsage.model_validate_json(response.content).accounts + + +def _deployment_account(rig: _Rig) -> _Account: + (account,) = _accounts(rig.gateway) + assert (account.custom_llm_provider, account.account_id, account.deployments) == ( + _PROVIDER, + cv.ACCOUNT, + (rig.model,), + ), account + return account + + +def _assert_no_fee(account: _Account) -> None: + assert (account.fee, account.billing_periods, account.fixed_cost) == (None, (), None), account + + +def _fee_body(account_id: str, monthly_fee: float = _MONTHLY_FEE) -> Mapping[str, JsonValue]: + return { + "custom_llm_provider": _PROVIDER, + "account_id": account_id, + "monthly_fee": monthly_fee, + "currency": "USD", + "billing_period_start": _BILLING_PERIOD_START, + "label": "integration", + } + + +def _delete_fee_if_present(gateway: Gateway, identity: str) -> None: + if gateway.rows(_FEE_ROW_QUERY, (identity,)): + gateway.post("/subscription_accounts/delete", {"subscription_account_id": identity}) + assert gateway.rows(_FEE_ROW_QUERY, (identity,)) == [] + + +def _set_fee(scenario: Scenario, account_id: str) -> str: + created: Final = scenario.gateway.post("/subscription_accounts/new", _fee_body(account_id)) + identity: Final = string_value(created["subscription_account_id"]) + scenario.cleanups.callback(_delete_fee_if_present, scenario.gateway, identity) + return identity + + +def _assert_fee_reported(account: _Account, identity: str) -> None: + assert account.fee is not None, account + assert ( + account.fee.subscription_account_id, + account.fee.account_id, + account.fee.monthly_fee, + account.fee.currency, + account.fee.billing_period_start, + ) == (identity, cv.ACCOUNT, _MONTHLY_FEE, "USD", _BILLING_PERIOD_START), account + assert tuple((period.start, period.end) for period in account.billing_periods) == (_PERIOD_IN_RANGE,), account + assert account.fixed_cost == _MONTHLY_FEE, account + + +def test_responses_calls_on_a_deployment_login_count_as_subscription_covered_chatgpt_usage(rig: _Rig) -> None: + first_day: Final = _today() + rig.wire.drain() + plain_marker: Final = uuid.uuid4().hex + stream_marker: Final = uuid.uuid4().hex + plain: Final = rig.gateway.request("POST", "/v1/responses", _responses_body(rig.model, plain_marker, stream=False)) + _assert_answered(plain, plain_marker, "application/json") + streamed: Final = rig.gateway.request( + "POST", "/v1/responses", _responses_body(rig.model, stream_marker, stream=True) + ) + _assert_answered(streamed, stream_marker, "text/event-stream") + forwarded: Final = rig.wire.drain() + assert len(forwarded) == 2, [request.target for request in forwarded] + for request, marker in zip(forwarded, (plain_marker, stream_marker), strict=True): + cv.forwarded(request, marker) + last_day: Final = _today() + buckets: Final = eventually( + lambda: _chatgpt_usage_between(rig.gateway, first_day, last_day), + lambda found: _summed(found).api_requests >= 2, + seconds=70, + ) + totals: Final = _summed(buckets) + assert (totals.api_requests, totals.successful_requests, totals.failed_requests, totals.spend) == ( + 2, + 2, + 0, + 0, + ), buckets + assert (totals.prompt_tokens, totals.completion_tokens) == ( + 2 * _TOKENS.validate_python(cv.USAGE["input_tokens"]), + 2 * _TOKENS.validate_python(cv.USAGE["output_tokens"]), + ), buckets + assert all(bucket.metadata.get("subscription_covered") is True for bucket in buckets), buckets + + +def test_fee_on_the_deployment_account_reports_its_fixed_cost_until_deleted(rig: _Rig) -> None: + _assert_no_fee(_deployment_account(rig)) + with rig.gateway.scenario() as scenario: + identity: Final = _set_fee(scenario, cv.ACCOUNT) + _assert_fee_reported(_deployment_account(rig), identity) + deleted: Final = rig.gateway.post("/subscription_accounts/delete", {"subscription_account_id": identity}) + assert deleted["subscription_account_id"] == identity, deleted + assert rig.gateway.rows(_FEE_ROW_QUERY, (identity,)) == [] + _assert_no_fee(_deployment_account(rig)) + + +def test_second_fee_for_the_same_account_answers_400_and_keeps_the_first(rig: _Rig) -> None: + with rig.gateway.scenario() as scenario: + identity: Final = _set_fee(scenario, cv.ACCOUNT) + duplicate: Final = rig.gateway.request( + "POST", "/subscription_accounts/new", _fee_body(cv.ACCOUNT, monthly_fee=3 * _MONTHLY_FEE) + ) + assert duplicate.status_code == 400, duplicate.text + assert rig.gateway.rows(_FEE_TABLE_QUERY, (_PROVIDER, cv.ACCOUNT)) == [{"subscription_account_id": identity}] + _assert_fee_reported(_deployment_account(rig), identity) + + +def test_update_and_delete_of_an_unknown_fee_answer_404(rig: _Rig) -> None: + unknown: Final = str(uuid.uuid4()) + updated: Final = rig.gateway.request( + "POST", "/subscription_accounts/update", {"subscription_account_id": unknown, "monthly_fee": _MONTHLY_FEE} + ) + assert updated.status_code == 404, updated.text + deleted: Final = rig.gateway.request("POST", "/subscription_accounts/delete", {"subscription_account_id": unknown}) + assert deleted.status_code == 404, deleted.text + assert rig.gateway.rows(_FEE_ROW_QUERY, (unknown,)) == [] + + +def test_internal_user_is_refused_setting_a_fee_and_reading_usage(rig: _Rig) -> None: + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(user_id=scenario.user(user_role="internal_user")) + created: Final = rig.gateway.request("POST", "/subscription_accounts/new", _fee_body(cv.ACCOUNT), key=key) + assert created.status_code == 401, created.text + assert rig.gateway.rows(_FEE_TABLE_QUERY, (_PROVIDER, cv.ACCOUNT)) == [] + usage: Final = _usage(rig.gateway, key=key) + assert usage.status_code == 401, usage.text + + +def test_fee_on_an_account_no_deployment_uses_stays_listed_and_counted(rig: _Rig) -> None: + unused: Final = f"acct-unused-{uuid.uuid4().hex}" + with rig.gateway.scenario() as scenario: + identity: Final = _set_fee(scenario, unused) + accounts: Final = _accounts(rig.gateway) + assert frozenset(account.account_id for account in accounts) == {cv.ACCOUNT, unused}, accounts + (unused_account,) = (account for account in accounts if account.account_id == unused) + assert unused_account.fee is not None, unused_account + assert (unused_account.deployments, unused_account.fee.subscription_account_id) == ((), identity) + assert unused_account.fixed_cost == _MONTHLY_FEE, unused_account diff --git a/tests/unit/llms/test_subscription_accounts.py b/tests/unit/llms/test_subscription_accounts.py new file mode 100644 index 00000000000..6995ee5d4df --- /dev/null +++ b/tests/unit/llms/test_subscription_accounts.py @@ -0,0 +1,110 @@ +import base64 +import json +from pathlib import Path +from typing import Final + +import pytest + +import litellm +from litellm.constants import SUBSCRIPTION_BACKED_PROVIDERS +from litellm.llms.chatgpt.common_utils import OPENAI_AUTH_CLAIM_KEY +from litellm.llms.subscription_accounts import SUBSCRIPTION_ACCOUNT_ID_RESOLVERS, resolve_subscription_account_id +from litellm.types.utils import CredentialItem + + +def _write_login(path: Path, account_id: str) -> Path: + path.parent.mkdir(parents=True, exist_ok=True) + path.write_text(json.dumps({"account_id": account_id})) + return path + + +def _access_token_for(account_id: str) -> str: + claims: Final = json.dumps({OPENAI_AUTH_CLAIM_KEY: {"chatgpt_account_id": account_id}}).encode() + payload: Final = base64.urlsafe_b64encode(claims).decode().rstrip("=") + return f"header.{payload}.signature" + + +@pytest.fixture +def signed_in_chatgpt_token_dir(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> Path: + monkeypatch.setenv("CHATGPT_TOKEN_DIR", str(tmp_path)) + monkeypatch.delenv("CHATGPT_AUTH_FILE", raising=False) + _write_login(tmp_path / "auth.json", "acct-from-token-file") + return tmp_path + + +def test_every_subscription_backed_provider_has_an_account_id_resolver() -> None: + assert frozenset(SUBSCRIPTION_ACCOUNT_ID_RESOLVERS) == SUBSCRIPTION_BACKED_PROVIDERS + + +def test_chatgpt_resolves_to_the_account_in_the_signed_in_token_file(signed_in_chatgpt_token_dir: Path) -> None: + assert resolve_subscription_account_id("chatgpt", {"model": "chatgpt/gpt-5.6-terra"}) == "acct-from-token-file" + + +def test_each_deployment_takes_the_account_of_its_own_auth_file(signed_in_chatgpt_token_dir: Path) -> None: + first: Final = _write_login(signed_in_chatgpt_token_dir / "team-a" / "auth.json", "acct-team-a") + second: Final = _write_login(signed_in_chatgpt_token_dir / "team-b" / "auth.json", "acct-team-b") + + assert resolve_subscription_account_id("chatgpt", {"chatgpt_auth_file": str(first)}) == "acct-team-a" + assert resolve_subscription_account_id("chatgpt", {"chatgpt_auth_file": str(second)}) == "acct-team-b" + + +def test_a_configured_access_token_resolves_to_its_own_account(signed_in_chatgpt_token_dir: Path) -> None: + litellm_params: Final = {"model": "chatgpt/gpt-5.6-terra", "api_key": _access_token_for("acct-from-api-key")} + + assert resolve_subscription_account_id("chatgpt", litellm_params) == "acct-from-api-key" + + +def test_a_metered_provider_resolves_to_no_account_even_with_a_chatgpt_login( + signed_in_chatgpt_token_dir: Path, +) -> None: + assert resolve_subscription_account_id("openai", {"model": "openai/gpt-5.6-terra"}) is None + + +def test_chatgpt_without_a_login_resolves_to_no_account(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: + empty_dir: Final = tmp_path / "no-login" + monkeypatch.setenv("CHATGPT_TOKEN_DIR", str(empty_dir)) + monkeypatch.delenv("CHATGPT_AUTH_FILE", raising=False) + assert resolve_subscription_account_id("chatgpt", {"model": "chatgpt/gpt-5.6-terra"}) is None + + +def test_resolving_an_account_leaves_the_login_file_untouched(signed_in_chatgpt_token_dir: Path) -> None: + login: Final = signed_in_chatgpt_token_dir / "no-saved-account" / "auth.json" + login.parent.mkdir(parents=True) + saved: Final = json.dumps({"access_token": _access_token_for("acct-derived"), "refresh_token": "rt-current"}) + login.write_text(saved) + + assert resolve_subscription_account_id("chatgpt", {"chatgpt_auth_file": str(login)}) == "acct-derived" + assert login.read_text() == saved + + +@pytest.mark.parametrize("credential_field", ["chatgpt_auth_file", "api_key"]) +def test_a_named_credential_resolves_to_the_account_it_signs_in_as( + credential_field: str, signed_in_chatgpt_token_dir: Path, monkeypatch: pytest.MonkeyPatch +) -> None: + login: Final = _write_login(signed_in_chatgpt_token_dir / "named" / "auth.json", "acct-named-credential") + values: Final = { + "chatgpt_auth_file": {"chatgpt_auth_file": str(login)}, + "api_key": {"api_key": _access_token_for("acct-named-credential")}, + }[credential_field] + monkeypatch.setattr( + litellm, + "credential_list", + [CredentialItem(credential_name="chatgpt-team", credential_values=values, credential_info={})], + ) + litellm_params: Final = {"model": "chatgpt/gpt-5.6-terra", "litellm_credential_name": "chatgpt-team"} + + assert resolve_subscription_account_id("chatgpt", litellm_params) == "acct-named-credential" + + +@pytest.mark.parametrize( + ("api_key", "expected_account"), + [("oidc/env/CHATGPT_TEAM_TOKEN", "acct-from-oidc-env"), ("oidc/env/CHATGPT_UNSET_TOKEN", None)], +) +def test_an_oidc_env_api_key_resolves_to_the_account_of_the_token_it_points_to( + api_key: str, expected_account: str | None, signed_in_chatgpt_token_dir: Path, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setenv("CHATGPT_TEAM_TOKEN", _access_token_for("acct-from-oidc-env")) + monkeypatch.delenv("CHATGPT_UNSET_TOKEN", raising=False) + litellm_params: Final = {"model": "chatgpt/gpt-5.6-terra", "api_key": api_key} + + assert resolve_subscription_account_id("chatgpt", litellm_params) == expected_account diff --git a/tests/unit/proxy/db/test_db_spend_update_writer.py b/tests/unit/proxy/db/test_db_spend_update_writer.py index c161aa5fe50..c3a93fd89be 100644 --- a/tests/unit/proxy/db/test_db_spend_update_writer.py +++ b/tests/unit/proxy/db/test_db_spend_update_writer.py @@ -6,6 +6,7 @@ import re from collections.abc import AsyncIterator, Callable from contextlib import AbstractAsyncContextManager, asynccontextmanager from datetime import datetime, timedelta, timezone +from pathlib import Path from types import SimpleNamespace from typing import Final, cast from unittest.mock import AsyncMock, MagicMock, call, patch @@ -4325,6 +4326,36 @@ async def test_insert_spend_log_asks_for_an_immediate_flush_only_on_rows_the_bri assert prisma.spend_log_flush_requested.is_set() is expects_flush + +@pytest.mark.asyncio +@pytest.mark.timeout(30) +async def test_a_chatgpt_responses_row_queues_without_a_default_chatgpt_login( + tmp_path: Path, monkeypatch: pytest.MonkeyPatch +): + monkeypatch.setenv("CHATGPT_TOKEN_DIR", str(tmp_path / "no-default-login")) + monkeypatch.delenv("CHATGPT_AUTH_FILE", raising=False) + own_login: Final = tmp_path / "team" / "auth.json" + own_login.parent.mkdir() + own_login.write_text(json.dumps({"access_token": "at-team", "account_id": "acct-team"})) + router: Final = litellm.Router( + model_list=[ + { + "model_name": "chatgpt-team", + "litellm_params": {"model": "chatgpt/gpt-5.6-terra", "chatgpt_auth_file": str(own_login)}, + "model_info": {"id": "dep-chatgpt"}, + } + ] + ) + prisma: Final = _tool_usage_prisma() + prisma.spend_log_flush_requested = asyncio.Event() + payload: Final = {"request_id": "req-1", "call_type": "aresponses", "model_id": "dep-chatgpt"} + + await DBSpendUpdateWriter()._insert_spend_log_to_db(payload=payload, prisma_client=prisma, llm_router=router) + + assert prisma.spend_log_transactions == [payload] + assert prisma.spend_log_flush_requested.is_set() is False + + def _batch_cost_payload() -> dict: return { **_minimal_spend_payload(), diff --git a/tests/unit/proxy/management_endpoints/test_common_daily_activity.py b/tests/unit/proxy/management_endpoints/test_common_daily_activity.py index c80fa3572fd..5fe4cd2edcd 100644 --- a/tests/unit/proxy/management_endpoints/test_common_daily_activity.py +++ b/tests/unit/proxy/management_endpoints/test_common_daily_activity.py @@ -2553,6 +2553,71 @@ async def test_get_daily_activity_rejects_non_canonical_dates_before_querying(st mock_table.find_many.assert_not_awaited() +def _provider_record(provider: str, *, model: str, spend: float): + base = vars(_spend_record("real-key", model=model, spend=spend)) + return SimpleNamespace(**{**base, "custom_llm_provider": provider, "api_requests": 1, "successful_requests": 1}) + + +def test_per_row_provider_bucket_marks_subscription_backed_providers_only(): + from litellm.proxy.management_endpoints.common_daily_activity import update_breakdown_metrics + from litellm.types.proxy.management_endpoints.common_daily_activity import BreakdownMetrics + + breakdown = BreakdownMetrics() + update_breakdown_metrics(breakdown, _provider_record("chatgpt", model="gpt-5.6-terra", spend=0.0), {}, {}, {}) + update_breakdown_metrics(breakdown, _provider_record("openai", model="gpt-5.6", spend=0.0), {}, {}, {}) + update_breakdown_metrics(breakdown, _provider_record("chatgpt", model="gpt-5.6-sol", spend=0.0), {}, {}, {}) + + assert breakdown.providers["chatgpt"].metadata == {"subscription_covered": True} + assert "subscription_covered" not in breakdown.providers["openai"].metadata + assert breakdown.providers["chatgpt"].metrics.api_requests == 2 + assert "subscription_covered" not in breakdown.models["gpt-5.6-terra"].metadata + + +def test_per_row_provider_marker_keeps_existing_provider_metadata(): + from litellm.proxy.management_endpoints.common_daily_activity import update_breakdown_metrics + from litellm.types.proxy.management_endpoints.common_daily_activity import BreakdownMetrics + + breakdown = BreakdownMetrics() + update_breakdown_metrics( + breakdown, + _provider_record("chatgpt", model="gpt-5.6-terra", spend=0.0), + {}, + {"chatgpt": {"region": "us"}}, + {}, + ) + + assert breakdown.providers["chatgpt"].metadata == {"region": "us", "subscription_covered": True} + + +@pytest.mark.parametrize("provider_row_first", [True, False]) +def test_grouping_sets_provider_bucket_marks_subscription_backed_providers_in_any_row_order(provider_row_first): + from litellm.proxy.management_endpoints.common_daily_activity import ( + _GROUP_DATE_PROVIDER, + _GROUP_DATE_PROVIDER_API_KEY, + _GROUP_GRAND_TOTAL, + _aggregate_grouping_sets_records_sync, + ) + + provider_rows = [ + _grouping_row(_GROUP_DATE_PROVIDER, custom_llm_provider="chatgpt", spend=0.0), + _grouping_row(_GROUP_DATE_PROVIDER_API_KEY, custom_llm_provider="chatgpt", api_key="real-key", spend=0.0), + ] + ordered_rows = provider_rows if provider_row_first else list(reversed(provider_rows)) + records = [ + *ordered_rows, + _grouping_row(_GROUP_DATE_PROVIDER, custom_llm_provider="openai", spend=0.0), + _grouping_row(_GROUP_DATE_PROVIDER_API_KEY, custom_llm_provider="openai", api_key="real-key", spend=0.0), + _grouping_row(_GROUP_GRAND_TOTAL, spend=0.0), + ] + + aggregated = _aggregate_grouping_sets_records_sync(records=records, api_key_metadata={}) + + providers = aggregated["results"][0].breakdown.providers + assert providers["chatgpt"].metadata == {"subscription_covered": True} + assert "real-key" in providers["chatgpt"].api_key_breakdown + assert "subscription_covered" not in providers["openai"].metadata + + @pytest.mark.parametrize("missing", [None, ""]) def test_missing_dimensions_preserve_all_paginated_metrics(missing: str | None) -> None: records: Final = ( diff --git a/tests/unit/proxy/management_endpoints/test_subscription_accounts.py b/tests/unit/proxy/management_endpoints/test_subscription_accounts.py new file mode 100644 index 00000000000..bcd31efd4b7 --- /dev/null +++ b/tests/unit/proxy/management_endpoints/test_subscription_accounts.py @@ -0,0 +1,438 @@ +from collections.abc import Iterator +from datetime import datetime, timezone +from typing import Final +from uuid import uuid4 + +from unittest.mock import MagicMock + +import pytest +from fastapi import FastAPI, HTTPException, Request +from fastapi.testclient import TestClient +from pydantic import ValidationError + +from litellm.models.subscription_account import LiteLLM_SubscriptionAccountTable +from litellm.proxy._types import LiteLLM_JWTAuth, LitellmUserRoles, UserAPIKeyAuth +from litellm.proxy.auth.auth_checks import allowed_routes_check +from litellm.proxy.auth.route_checks import RouteChecks +from litellm.proxy.auth.user_api_key_auth import user_api_key_auth +from litellm.proxy.management_endpoints.subscription_accounts import ( + SubscriptionDeployment, + get_subscription_account_repository, + get_subscription_deployments, + router, + subscription_deployments_from_model_list, + subscription_deployments_of_router, +) +from litellm.router import Router +from litellm.types.proxy.management_endpoints.subscription_accounts import ( + SubscriptionAccountFeeRequest, + UpdateSubscriptionAccountFeeRequest, +) +from litellm.repositories.subscription_account_repository import ( + DuplicateSubscriptionAccount, + SubscriptionAccountFeeChanges, + SubscriptionAccountFeeFields, +) + +_ACCOUNT: Final = "org-team-workspace" +_TEST_FEE: Final = { + "custom_llm_provider": "chatgpt", + "account_id": _ACCOUNT, + "monthly_fee": 25.0, + "currency": "USD", + "billing_period_start": "2026-10-01", + "label": "Team plan (test value)", +} +_TWO_DEPLOYMENTS_ONE_ACCOUNT: Final = ( + SubscriptionDeployment(model_name="gpt-5.6-terra", custom_llm_provider="chatgpt", account_id=_ACCOUNT), + SubscriptionDeployment(model_name="gpt-5.6-sol", custom_llm_provider="chatgpt", account_id=_ACCOUNT), +) + + +class _FakeStore: + def __init__(self) -> None: + self.rows: dict[str, LiteLLM_SubscriptionAccountTable] = {} + self.written_columns: list[dict[str, float | str | None]] = [] + + async def list_all(self) -> list[LiteLLM_SubscriptionAccountTable]: + return sorted(self.rows.values(), key=lambda row: row.created_at) + + async def find_by_id(self, id_value: str) -> LiteLLM_SubscriptionAccountTable | None: + return self.rows.get(id_value) + + async def create_account( + self, custom_llm_provider: str, account_id: str, fee: SubscriptionAccountFeeFields, created_by: str + ) -> LiteLLM_SubscriptionAccountTable | DuplicateSubscriptionAccount: + for row in self.rows.values(): + if (row.custom_llm_provider, row.account_id) == (custom_llm_provider, account_id): + return DuplicateSubscriptionAccount(custom_llm_provider=custom_llm_provider, account_id=account_id) + now = datetime.now(timezone.utc) + row = LiteLLM_SubscriptionAccountTable( + subscription_account_id=str(uuid4()), + custom_llm_provider=custom_llm_provider, + account_id=account_id, + label=fee.label, + monthly_fee=fee.monthly_fee, + currency=fee.currency, + billing_period_start=fee.billing_period_start, + created_at=now, + created_by=created_by, + updated_at=now, + updated_by=created_by, + ) + self.rows[row.subscription_account_id] = row + return row + + async def update_fee( + self, subscription_account_id: str, changes: SubscriptionAccountFeeChanges, updated_by: str + ) -> LiteLLM_SubscriptionAccountTable | None: + columns = dict(changes.columns()) + self.written_columns.append(columns) + existing = self.rows.get(subscription_account_id) + if existing is None: + return None + updated = existing.model_copy( + update={**columns, "updated_by": updated_by, "updated_at": datetime.now(timezone.utc)} + ) + self.rows[subscription_account_id] = updated + return updated + + async def delete_account(self, subscription_account_id: str) -> LiteLLM_SubscriptionAccountTable | None: + return self.rows.pop(subscription_account_id, None) + + +def _resolve_auth(request: Request) -> UserAPIKeyAuth: + role: Final = LitellmUserRoles(request.headers.get("x-user-role", LitellmUserRoles.PROXY_ADMIN.value)) + return UserAPIKeyAuth(user_id="admin", user_role=role, api_key="sk-test") + + +@pytest.fixture +def deployments() -> tuple[SubscriptionDeployment, ...]: + return _TWO_DEPLOYMENTS_ONE_ACCOUNT + + +@pytest.fixture +def client(deployments: tuple[SubscriptionDeployment, ...]) -> Iterator[tuple[TestClient, _FakeStore]]: + store: Final = _FakeStore() + app: Final = FastAPI() + app.include_router(router) + app.dependency_overrides[get_subscription_account_repository] = lambda: store + app.dependency_overrides[get_subscription_deployments] = lambda: deployments + app.dependency_overrides[user_api_key_auth] = _resolve_auth + with TestClient(app) as test_client: + yield test_client, store + + +def _usage(test_client: TestClient, start_date: str, end_date: str, **headers: str) -> dict: + response = test_client.get( + "/subscription_accounts/usage", params={"start_date": start_date, "end_date": end_date}, headers=headers + ) + assert response.status_code == 200, response.text + return response.json() + + +def test_untracked_account_lists_its_deployments_with_no_fee(client): + test_client, _ = client + + body = _usage(test_client, "2026-10-01", "2026-10-07") + + assert body["subscription_providers"] == ["chatgpt"] + assert body["accounts"] == [ + { + "custom_llm_provider": "chatgpt", + "account_id": _ACCOUNT, + "deployments": ["gpt-5.6-sol", "gpt-5.6-terra"], + "fee": None, + "billing_periods": [], + "fixed_cost": None, + } + ] + + +def test_fee_counts_once_per_billing_period_however_many_deployments_share_the_account(client): + test_client, _ = client + created = test_client.post("/subscription_accounts/new", json=_TEST_FEE) + assert created.status_code == 200, created.text + + one_period = _usage(test_client, "2026-10-01", "2026-10-07") + two_periods = _usage(test_client, "2026-09-15", "2026-11-15") + between_starts = _usage(test_client, "2026-10-02", "2026-10-31") + + (account,) = one_period["accounts"] + assert account["deployments"] == ["gpt-5.6-sol", "gpt-5.6-terra"] + assert account["fee"]["monthly_fee"] == 25.0 + assert account["fee"]["currency"] == "USD" + assert account["fee"]["label"] == "Team plan (test value)" + assert account["billing_periods"] == [{"start": "2026-10-01", "end": "2026-10-31"}] + assert account["fixed_cost"] == 25.0 + assert two_periods["accounts"][0]["fixed_cost"] == 50.0 + assert [period["start"] for period in two_periods["accounts"][0]["billing_periods"]] == [ + "2026-10-01", + "2026-11-01", + ] + assert between_starts["accounts"][0]["fixed_cost"] == 0.0 + assert between_starts["accounts"][0]["billing_periods"] == [] + + +def test_fee_of_an_account_no_deployment_uses_stays_listed_and_counted(client): + test_client, _ = client + other = {**_TEST_FEE, "account_id": "org-other", "monthly_fee": 200.0, "currency": "EUR"} + assert test_client.post("/subscription_accounts/new", json=other).status_code == 200 + assert test_client.post("/subscription_accounts/new", json=_TEST_FEE).status_code == 200 + + body = _usage(test_client, "2026-10-01", "2026-10-07") + + unused, used = body["accounts"] + assert (unused["account_id"], unused["deployments"], unused["fee"]["currency"]) == ("org-other", [], "EUR") + assert unused["fixed_cost"] == 200.0 + assert (used["account_id"], used["fixed_cost"]) == (_ACCOUNT, 25.0) + + +@pytest.mark.parametrize( + "deployments", + [(SubscriptionDeployment(model_name="gpt-5.6-terra", custom_llm_provider="chatgpt", account_id=None),)], +) +def test_account_not_signed_in_shows_no_account_and_takes_no_fee(client): + test_client, _ = client + assert test_client.post("/subscription_accounts/new", json=_TEST_FEE).status_code == 200 + + body = _usage(test_client, "2026-10-01", "2026-10-07") + + unsigned = next(account for account in body["accounts"] if account["account_id"] is None) + assert unsigned["deployments"] == ["gpt-5.6-terra"] + assert unsigned["fee"] is None + assert unsigned["fixed_cost"] is None + + +def test_update_changes_the_fee_and_delete_returns_the_account_to_untracked(client): + test_client, store = client + created = test_client.post("/subscription_accounts/new", json=_TEST_FEE).json() + account_id = created["subscription_account_id"] + + updated = test_client.post( + "/subscription_accounts/update", + json={"subscription_account_id": account_id, "monthly_fee": 30.0, "billing_period_start": "2026-10-05"}, + ) + assert updated.status_code == 200, updated.text + assert updated.json()["monthly_fee"] == 30.0 + assert updated.json()["currency"] == "USD" + assert updated.json()["billing_period_start"] == "2026-10-05" + assert updated.json()["label"] == _TEST_FEE["label"] + assert _usage(test_client, "2026-10-01", "2026-10-07")["accounts"][0]["fixed_cost"] == 30.0 + + cleared = test_client.post( + "/subscription_accounts/update", json={"subscription_account_id": account_id, "label": None} + ) + assert cleared.status_code == 200, cleared.text + assert cleared.json()["label"] is None + assert cleared.json()["monthly_fee"] == 30.0 + assert store.written_columns == [{"monthly_fee": 30.0, "billing_period_start": "2026-10-05"}, {"label": None}] + + unknown = test_client.post( + "/subscription_accounts/update", json={"subscription_account_id": "no-such-account", "monthly_fee": 1.0} + ) + assert unknown.status_code == 404 + + deleted = test_client.post("/subscription_accounts/delete", json={"subscription_account_id": account_id}) + assert deleted.status_code == 200, deleted.text + assert store.rows == {} + assert _usage(test_client, "2026-10-01", "2026-10-07")["accounts"][0]["fixed_cost"] is None + + missing = test_client.post("/subscription_accounts/delete", json={"subscription_account_id": account_id}) + assert missing.status_code == 404 + + +def test_rejects_a_second_fee_for_the_same_account(client): + test_client, _ = client + assert test_client.post("/subscription_accounts/new", json=_TEST_FEE).status_code == 200 + + duplicate = test_client.post("/subscription_accounts/new", json={**_TEST_FEE, "monthly_fee": 99.0}) + + assert duplicate.status_code == 400 + assert "/subscription_accounts/update" in duplicate.json()["detail"]["error"] + + +def test_rejects_a_fee_on_a_metered_provider(client): + test_client, store = client + + response = test_client.post("/subscription_accounts/new", json={**_TEST_FEE, "custom_llm_provider": "openai"}) + + assert response.status_code == 400 + assert store.rows == {} + + +@pytest.mark.parametrize( + "payload", + [ + {**_TEST_FEE, "monthly_fee": 0}, + {**_TEST_FEE, "account_id": ""}, + {**_TEST_FEE, "currency": "usd"}, + {**_TEST_FEE, "currency": "$"}, + {**_TEST_FEE, "billing_period_start": "10/01/2026"}, + ], +) +def test_rejects_malformed_fee_fields(client, payload): + test_client, store = client + + assert test_client.post("/subscription_accounts/new", json=payload).status_code == 422 + assert store.rows == {} + + +def test_rejects_an_infinite_monthly_fee(): + with pytest.raises(ValidationError): + SubscriptionAccountFeeRequest.model_validate({**_TEST_FEE, "monthly_fee": float("inf")}) + with pytest.raises(ValidationError): + UpdateSubscriptionAccountFeeRequest.model_validate( + {"subscription_account_id": "fee-1", "monthly_fee": float("inf")} + ) + + +@pytest.mark.parametrize("deployments", [()]) +def test_fee_stays_listed_after_the_last_subscription_deployment_is_removed(client): + test_client, _ = client + assert test_client.post("/subscription_accounts/new", json=_TEST_FEE).status_code == 200 + + body = _usage(test_client, "2026-10-01", "2026-10-07") + + (account,) = body["accounts"] + assert (account["account_id"], account["deployments"], account["fixed_cost"]) == (_ACCOUNT, [], 25.0) + + +def test_usage_rejects_a_malformed_or_inverted_range(client): + test_client, _ = client + + malformed = [ + test_client.get("/subscription_accounts/usage", params={"start_date": start, "end_date": "2026-10-07"}) + for start in ("2026-10-1", "20261001", "2026-W40-1", "2026-13-01") + ] + inverted = test_client.get( + "/subscription_accounts/usage", params={"start_date": "2026-10-07", "end_date": "2026-10-01"} + ) + + assert [response.status_code for response in malformed] == [400, 400, 400, 400] + assert inverted.status_code == 400 + + +def test_viewer_reads_usage_but_cannot_write(client): + test_client, store = client + viewer = {"x-user-role": LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY.value} + + assert _usage(test_client, "2026-10-01", "2026-10-07", **viewer)["accounts"][0]["account_id"] == _ACCOUNT + assert test_client.post("/subscription_accounts/new", json=_TEST_FEE, headers=viewer).status_code == 403 + assert store.rows == {} + + +_SUBSCRIPTION_ROUTES: Final = ( + "/subscription_accounts/usage", + "/subscription_accounts/new", + "/subscription_accounts/update", + "/subscription_accounts/delete", +) + + +def test_routes_follow_the_admin_route_policy(): + jwt_admin_defaults = LiteLLM_JWTAuth() + + for route in _SUBSCRIPTION_ROUTES: + assert allowed_routes_check( + user_role=LitellmUserRoles.PROXY_ADMIN, user_route=route, litellm_proxy_roles=jwt_admin_defaults + ) + assert RouteChecks.is_management_route(route) + + +def test_viewer_route_gate_reads_usage_and_blocks_writes(): + viewer_role: Final = LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY.value + get_usage = MagicMock(spec=Request, method="GET", query_params={}) + post_new = MagicMock(spec=Request, method="POST", query_params={}) + + RouteChecks._check_proxy_admin_viewer_access( + route="/subscription_accounts/usage", _user_role=viewer_role, request_data={}, request=get_usage + ) + with pytest.raises(HTTPException) as denied: + RouteChecks._check_proxy_admin_viewer_access( + route="/subscription_accounts/new", _user_role=viewer_role, request_data=_TEST_FEE, request=post_new + ) + assert denied.value.status_code == 403 + + +def test_internal_user_cannot_read_subscription_accounts(client): + test_client, _ = client + + response = test_client.get( + "/subscription_accounts/usage", + params={"start_date": "2026-10-01", "end_date": "2026-10-07"}, + headers={"x-user-role": LitellmUserRoles.INTERNAL_USER.value}, + ) + + assert response.status_code == 403 + + +def test_deployments_take_the_account_of_their_own_login(): + model_list = [ + {"model_name": "gpt-5.6-terra", "litellm_params": {"model": "chatgpt/gpt-5.6-terra"}}, + {"model_name": "gpt-5.6-sol", "litellm_params": {"model": "chatgpt/gpt-5.6-sol"}}, + { + "model_name": "gpt-5.6-luna", + "litellm_params": {"model": "chatgpt/gpt-5.6-luna", "chatgpt_auth_file": "/logins/other.json"}, + }, + {"model_name": "sonnet", "litellm_params": {"model": "anthropic/claude-sonnet-4-5", "api_key": "sk-x"}}, + {"model_name": "broken", "litellm_params": {"model": "no-such-provider-model"}}, + ] + lookups: list[tuple[str, object]] = [] + + def resolver(provider: str, litellm_params) -> str | None: + lookups.append((provider, litellm_params.get("model"))) + return "org-other" if litellm_params.get("chatgpt_auth_file") == "/logins/other.json" else _ACCOUNT + + deployments = subscription_deployments_from_model_list(model_list, resolver) + + assert deployments == ( + *_TWO_DEPLOYMENTS_ONE_ACCOUNT, + SubscriptionDeployment(model_name="gpt-5.6-luna", custom_llm_provider="chatgpt", account_id="org-other"), + ) + assert lookups == [ + ("chatgpt", "chatgpt/gpt-5.6-terra"), + ("chatgpt", "chatgpt/gpt-5.6-sol"), + ("chatgpt", "chatgpt/gpt-5.6-luna"), + ] + + +def test_deployments_list_each_router_deployment_once_by_the_name_callers_use(): + llm_router: Final = Router( + model_list=[ + {"model_name": "gpt-5.6-sol", "litellm_params": {"model": "chatgpt/gpt-5.6-sol"}}, + { + "model_name": "model_name_team-1_7f3a", + "litellm_params": {"model": "chatgpt/gpt-5.6-terra"}, + "model_info": {"team_id": "team-1", "team_public_model_name": "team-terra"}, + }, + ], + model_group_alias={"sol-alias": "gpt-5.6-sol"}, + ) + + deployments = subscription_deployments_of_router(llm_router, lambda provider, litellm_params: _ACCOUNT) + + assert sorted(deployment.model_name for deployment in deployments) == ["gpt-5.6-sol", "team-terra"] + + +@pytest.mark.parametrize( + "deployments", + [ + ( + *_TWO_DEPLOYMENTS_ONE_ACCOUNT, + SubscriptionDeployment(model_name="gpt-5.6-luna", custom_llm_provider="chatgpt", account_id="org-other"), + ) + ], +) +def test_two_logins_each_count_their_own_fee_once(client): + test_client, _ = client + other = {**_TEST_FEE, "account_id": "org-other", "monthly_fee": 40.0} + assert test_client.post("/subscription_accounts/new", json=_TEST_FEE).status_code == 200 + assert test_client.post("/subscription_accounts/new", json=other).status_code == 200 + + body = _usage(test_client, "2026-10-01", "2026-10-07") + + assert [(a["account_id"], a["deployments"], a["fixed_cost"]) for a in body["accounts"]] == [ + ("org-other", ["gpt-5.6-luna"], 40.0), + (_ACCOUNT, ["gpt-5.6-sol", "gpt-5.6-terra"], 25.0), + ] diff --git a/tests/unit/proxy/spend_tracking/test_subscription_billing_periods.py b/tests/unit/proxy/spend_tracking/test_subscription_billing_periods.py new file mode 100644 index 00000000000..c2c98a05b3f --- /dev/null +++ b/tests/unit/proxy/spend_tracking/test_subscription_billing_periods.py @@ -0,0 +1,73 @@ +from datetime import date + +from litellm.proxy.spend_tracking.subscription_billing_periods import ( + BillingPeriodRange, + billing_periods_starting_within, +) + + +def test_one_period_per_month_start_inside_the_filter(): + periods = billing_periods_starting_within(date(2026, 8, 5), date(2026, 9, 1), date(2026, 10, 31)) + + assert periods == ( + BillingPeriodRange(start=date(2026, 9, 5), end=date(2026, 10, 4)), + BillingPeriodRange(start=date(2026, 10, 5), end=date(2026, 11, 4)), + ) + + +def test_filter_boundaries_are_inclusive(): + anchor = date(2026, 10, 1) + + assert billing_periods_starting_within(anchor, date(2026, 10, 1), date(2026, 10, 7)) == ( + BillingPeriodRange(start=date(2026, 10, 1), end=date(2026, 10, 31)), + ) + assert billing_periods_starting_within(anchor, date(2026, 9, 25), date(2026, 10, 1)) == ( + BillingPeriodRange(start=date(2026, 10, 1), end=date(2026, 10, 31)), + ) + + +def test_filter_between_two_period_starts_attributes_nothing(): + assert billing_periods_starting_within(date(2026, 10, 15), date(2026, 10, 1), date(2026, 10, 7)) == () + + +def test_no_period_before_the_anchor(): + assert billing_periods_starting_within(date(2026, 10, 1), date(2026, 8, 1), date(2026, 9, 30)) == () + assert billing_periods_starting_within(date(2026, 10, 1), date(2026, 8, 1), date(2026, 10, 1)) == ( + BillingPeriodRange(start=date(2026, 10, 1), end=date(2026, 10, 31)), + ) + + +def test_anchor_day_clamps_to_short_months_without_drifting(): + periods = billing_periods_starting_within(date(2026, 1, 31), date(2026, 2, 1), date(2026, 5, 31)) + + assert [period.start for period in periods] == [ + date(2026, 2, 28), + date(2026, 3, 31), + date(2026, 4, 30), + date(2026, 5, 31), + ] + assert periods[0].end == date(2026, 3, 30) + assert periods[-1].end == date(2026, 6, 29) + + +def test_period_ends_the_day_before_the_next_start_across_a_year_boundary(): + periods = billing_periods_starting_within(date(2025, 12, 20), date(2025, 12, 1), date(2026, 1, 31)) + + assert periods == ( + BillingPeriodRange(start=date(2025, 12, 20), end=date(2026, 1, 19)), + BillingPeriodRange(start=date(2026, 1, 20), end=date(2026, 2, 19)), + ) + + +def test_inverted_filter_is_empty(): + assert billing_periods_starting_within(date(2026, 1, 1), date(2026, 3, 1), date(2026, 2, 1)) == () + + +def test_the_calendar_s_last_month_ends_on_its_last_day_instead_of_failing(): + assert billing_periods_starting_within(date(9999, 11, 15), date(9999, 11, 1), date(9999, 12, 31)) == ( + BillingPeriodRange(start=date(9999, 11, 15), end=date(9999, 12, 14)), + BillingPeriodRange(start=date(9999, 12, 15), end=date(9999, 12, 31)), + ) + assert billing_periods_starting_within(date(9999, 12, 1), date(9999, 12, 1), date(9999, 12, 31)) == ( + BillingPeriodRange(start=date(9999, 12, 1), end=date(9999, 12, 31)), + ) diff --git a/tests/unit/repositories/test_repositories.py b/tests/unit/repositories/test_repositories.py index cbcd0b1cd28..4e410e22b7f 100644 --- a/tests/unit/repositories/test_repositories.py +++ b/tests/unit/repositories/test_repositories.py @@ -30,6 +30,11 @@ from litellm.repositories.object_permission_repository import ( ) from litellm.repositories.organization_repository import OrganizationRepository from litellm.repositories.project_repository import ProjectRepository +from litellm.repositories.subscription_account_repository import ( + SetLabel, + SubscriptionAccountFeeChanges, + SubscriptionAccountRepository, +) from litellm.repositories.team_repository import TeamRepository from litellm.repositories.user_repository import UserRepository from litellm.repositories.verification_token_repository import ( @@ -140,6 +145,7 @@ class MockPrismaClient: self.db.litellm_projecttable = MockTable(pk_field="project_id") self.db.litellm_objectpermissiontable = MockTable(pk_field="object_permission_id") self.db.litellm_credentialstable = MockTable() + self.db.litellm_subscriptionaccounttable = MockTable(pk_field="subscription_account_id") class TestBaseRepository: @@ -2461,3 +2467,40 @@ class TestAutoRouterSessionRepository: assert AutoRouterSessionRepository(client).table is client.db.litellm_autoroutersession with pytest.raises(RuntimeError, match="No DB Connected"): _ = AutoRouterSessionRepository(None).table + + +class TestSubscriptionAccountRepository: + @pytest.mark.asyncio + async def test_update_fee_writes_only_the_named_columns(self): + client = MockPrismaClient() + stored_at = datetime(2026, 10, 1) + await client.db.litellm_subscriptionaccounttable.create( + data={ + "subscription_account_id": "sub-1", + "custom_llm_provider": "chatgpt", + "account_id": "acct_123", + "label": "Team plan (test value)", + "monthly_fee": 25.0, + "currency": "USD", + "billing_period_start": "2026-10-01", + "created_at": stored_at, + "created_by": "admin-a", + "updated_at": stored_at, + "updated_by": "admin-a", + } + ) + repo = SubscriptionAccountRepository(client) + + fee_only = await repo.update_fee("sub-1", SubscriptionAccountFeeChanges(monthly_fee=30.0), "admin-b") + label_cleared = await repo.update_fee("sub-1", SubscriptionAccountFeeChanges(label=SetLabel(None)), "admin-c") + missing = await repo.update_fee("sub-2", SubscriptionAccountFeeChanges(monthly_fee=1.0), "admin-b") + + assert fee_only is not None and label_cleared is not None + assert (fee_only.monthly_fee, fee_only.currency, fee_only.label, fee_only.updated_by) == ( + 30.0, + "USD", + "Team plan (test value)", + "admin-b", + ) + assert (label_cleared.monthly_fee, label_cleared.label, label_cleared.updated_by) == (30.0, None, "admin-c") + assert missing is None diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx index 660168fa6ab..5c8b284c72f 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/EntityUsage.test.tsx @@ -1160,6 +1160,34 @@ describe("EntityUsage", () => { expect(screen.getByTestId("provider-spend-total")).toHaveTextContent("$100.50"); }); + it("lists a subscription-covered provider on the team view as included in the subscription at zero spend", async () => { + const day = mockSpendData.results[0]; + mockTeamDailyActivityCall.mockResolvedValue({ + ...mockSpendData, + results: [ + { + ...day, + breakdown: { + ...day.breakdown, + providers: { + ...day.breakdown.providers, + chatgpt: { + metrics: { ...day.breakdown.providers.openai.metrics, spend: 0, api_requests: 4 }, + metadata: { subscription_covered: true }, + }, + }, + }, + }, + ], + }); + + render(); + + expect(await screen.findByText("Included in subscription")).toBeInTheDocument(); + expect(screen.getByText("chatgpt")).toBeInTheDocument(); + expect(screen.getAllByTestId("provider-share-segment")).toHaveLength(1); + }); + it("should label the chart with user_email metadata instead of the raw UUID (LIT-3889)", async () => { const userUuid = "c0e68be8-057e-4e2f-9d3a-000000000000"; const spendDataForUser = { diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.test.tsx index 8da636e68a7..dca5ff91fa6 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.test.tsx @@ -258,6 +258,34 @@ describe("SpendByProvider", () => { expect(screen.getByTestId("provider-spend-total")).toHaveTextContent("$200.50"); }); + it("always shows a subscription-covered provider and marks it included in the subscription", () => { + const providerSpendWithSubscription = [ + { + provider: "chatgpt", + spend: 0, + requests: 7, + successful_requests: 7, + failed_requests: 0, + tokens: 11608, + subscription_covered: true, + }, + { + provider: "google", + spend: 0, + requests: 0, + successful_requests: 0, + failed_requests: 0, + tokens: 0, + }, + ]; + render(); + expect(screen.getAllByText("chatgpt").length).toBeGreaterThan(0); + expect(screen.getByText("Included in subscription")).toBeInTheDocument(); + expect(screen.getByText("11,608")).toBeInTheDocument(); + expect(screen.queryByText("google")).not.toBeInTheDocument(); + expect(screen.getByText("Included in subscription").closest("td")).not.toHaveTextContent("$"); + }); + it("should include all providers with spend greater than zero by default", () => { const providerSpendWithMixed = [ { diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.tsx index 13d9a812a12..befb7d13b55 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/SpendByProvider.tsx @@ -4,6 +4,7 @@ import { MoneyCell } from "@/components/shared/table_cells"; import { formatNumberWithCommas } from "@/utils/dataUtils"; import { Info, Server } from "lucide-react"; import type { ColumnDef } from "@tanstack/react-table"; +import { Badge } from "@/components/ui/badge"; import { Switch } from "@/components/ui/switch"; import { Tooltip, TooltipContent, TooltipTrigger } from "@/components/ui/tooltip"; import React, { useId, useState } from "react"; @@ -11,13 +12,14 @@ import { ProviderLogo } from "@/components/molecules/models/ProviderLogo"; import { ChartLoader } from "@/components/shared/chart_loader"; import { Panel } from "../overview/Primitives"; -type ProviderSpendData = { +export type ProviderSpendData = { provider: string; spend: number; requests: number; successful_requests: number; failed_requests: number; tokens: number; + subscription_covered?: boolean; }; interface SpendByProviderProps { @@ -48,7 +50,15 @@ const columns: ColumnDef[] = [ header: "Spend", accessorKey: "spend", meta: { numeric: true, ...QUIET_HEADER }, - cell: ({ row }) => , + cell: ({ row }) => + row.original.subscription_covered ? ( +
+ {row.original.spend > 0 && } + Included in subscription +
+ ) : ( + + ), }, { header: "Successful", @@ -164,6 +174,10 @@ const SpendByProvider: React.FC = ({ loading, isDateChangi return includeUnknown; } + if (provider.subscription_covered) { + return true; + } + // If includeZeroSpend is true, include all providers (including those with 0 spend) // Otherwise, only include providers with spend > 0 if (includeZeroSpend) { diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/entityUsageAggregations.ts b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/entityUsageAggregations.ts index 54569608b0e..be9f44d0786 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/entityUsageAggregations.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/EntityUsage/entityUsageAggregations.ts @@ -1,6 +1,7 @@ import { keyActivityLabel } from "@/components/UsagePage/keyActivityLabel"; import type { TopKeyItem } from "@/components/UsagePage/components/EntityUsage/TopKeyView"; import { BreakdownMetrics, DailyData, KeyMetricWithMetadata, TagUsage } from "@/components/UsagePage/types"; +import { subscriptionCoveredProviders } from "../SubscriptionCost/types"; export type ExtendedDailyData = DailyData & { breakdown: BreakdownMetrics; @@ -15,6 +16,7 @@ export interface ProviderSpendRow extends Record { successful_requests: number; failed_requests: number; tokens: number; + subscription_covered: boolean; } export const getTopModels = ( @@ -205,6 +207,7 @@ export const getTopAPIKeys = (results: ExtendedDailyData[], topKeysLimit: number }; export const getProviderSpend = (results: ExtendedDailyData[]): ProviderSpendRow[] => { + const covered = subscriptionCoveredProviders(results); const providerSpend: Record = {}; results.forEach((day) => { Object.entries(day.breakdown.providers || {}).forEach(([provider, metrics]) => { @@ -216,6 +219,7 @@ export const getProviderSpend = (results: ExtendedDailyData[]): ProviderSpendRow successful_requests: 0, failed_requests: 0, tokens: 0, + subscription_covered: covered.has(provider), }; } try { @@ -231,6 +235,6 @@ export const getProviderSpend = (results: ExtendedDailyData[]): ProviderSpendRow }); return Object.values(providerSpend) - .filter((provider) => provider.spend > 0) + .filter((provider) => provider.spend > 0 || provider.subscription_covered) .sort((a, b) => b.spend - a.spend); }; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.test.tsx new file mode 100644 index 00000000000..37f3bf4abf4 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.test.tsx @@ -0,0 +1,185 @@ +import { fireEvent, render, screen } from "@testing-library/react"; +import { describe, expect, it, vi } from "vitest"; +import SubscriptionCostCard from "./SubscriptionCostCard"; +import type { SubscriptionAccountUsage, SubscriptionFee } from "./types"; + +vi.mock("@/components/shared/chart_loader", () => ({ + ChartLoader: () =>
Loading chart data...
, +})); + +vi.mock("@/components/molecules/models/ProviderLogo", () => ({ + ProviderLogo: ({ provider }: { provider: string }) =>
, +})); + +const fee = (overrides: Partial = {}): SubscriptionFee => ({ + subscription_account_id: "fee-1", + custom_llm_provider: "chatgpt", + account_id: "acct_123", + label: null, + monthly_fee: 25, + currency: "USD", + billing_period_start: "2026-10-01", + created_at: "2026-10-06T00:00:00Z", + updated_at: "2026-10-06T00:00:00Z", + created_by: "admin", + updated_by: "admin", + ...overrides, +}); + +const account = (overrides: Partial = {}): SubscriptionAccountUsage => ({ + custom_llm_provider: "chatgpt", + account_id: "acct_123", + deployments: ["gpt-5.6-terra", "gpt-5.6-sol"], + fee: null, + billing_periods: [], + fixed_cost: null, + ...overrides, +}); + +const handlers = () => ({ onSetFee: vi.fn(), onEditFee: vi.fn(), onRemoveFee: vi.fn() }); + +const renderCard = (accounts: SubscriptionAccountUsage[], props: { canEdit?: boolean; meteredSpend?: number } = {}) => { + const actions = handlers(); + const view = render( + , + ); + return { ...view, ...actions }; +}; + +describe("SubscriptionCostCard", () => { + it("renders nothing when there are no subscription accounts", () => { + const { container } = render( + , + ); + expect(container).toBeEmptyDOMElement(); + expect(screen.queryByText("Subscription fixed costs")).not.toBeInTheDocument(); + }); + + it("calls out an untracked fixed cost and offers to set the fee only to an editor", () => { + const { onSetFee, unmount } = renderCard([account()]); + expect(screen.getByText("Fixed cost not tracked")).toBeInTheDocument(); + expect(screen.getByText("gpt-5.6-terra, gpt-5.6-sol")).toBeInTheDocument(); + expect(screen.getByText("None in range")).toBeInTheDocument(); + expect(screen.getByTestId("fixed-cost-total")).toHaveTextContent("Fixed cost in range: none"); + fireEvent.click(screen.getByRole("button", { name: "Set monthly fee" })); + expect(onSetFee).toHaveBeenCalledWith(account()); + unmount(); + + renderCard([account()], { canEdit: false }); + expect(screen.getByText("Fixed cost not tracked")).toBeInTheDocument(); + expect(screen.queryByRole("button", { name: "Set monthly fee" })).not.toBeInTheDocument(); + }); + + it("shows a configured fee once for two deployments sharing the account and the fixed cost in range", () => { + const covered = account({ + fee: fee({ label: "Team workspace" }), + billing_periods: [{ start: "2026-10-01", end: "2026-10-31" }], + fixed_cost: 25, + }); + const { onEditFee, onRemoveFee } = renderCard([covered], { meteredSpend: 0.0838 }); + + expect(screen.getAllByText("25.00 USD")).toHaveLength(2); + expect(screen.getByText("Team workspace")).toBeInTheDocument(); + expect(screen.getByText("Oct 1 - Oct 31, 2026")).toBeInTheDocument(); + expect(screen.getByTestId("fixed-cost-total")).toHaveTextContent("Fixed cost in range: 25.00 USD"); + expect(screen.getByTestId("metered-spend")).toHaveTextContent("Usage-billed spend in range, all providers: $0.08"); + expect(screen.queryByText(/25\.08/)).not.toBeInTheDocument(); + + fireEvent.click(screen.getByRole("button", { name: "Edit" })); + expect(onEditFee).toHaveBeenCalledWith(covered); + }); + + it("removes a fee only after the admin confirms it", async () => { + const covered = account({ fee: fee(), billing_periods: [], fixed_cost: 0 }); + const { onRemoveFee } = renderCard([covered]); + + fireEvent.click(screen.getByRole("button", { name: "Remove" })); + expect(onRemoveFee).not.toHaveBeenCalled(); + fireEvent.click(await screen.findByRole("button", { name: "Cancel" })); + expect(onRemoveFee).not.toHaveBeenCalled(); + + fireEvent.click(screen.getByRole("button", { name: "Remove" })); + fireEvent.click(await screen.findByRole("button", { name: "Remove fee" })); + expect(onRemoveFee).toHaveBeenCalledWith(covered); + }); + + it("sums fixed costs per currency and keeps a second currency separate", () => { + const october = { start: "2026-10-01", end: "2026-10-31" }; + const twoPeriods = { + fee: fee(), + billing_periods: [{ start: "2026-09-01", end: "2026-09-30" }, october], + fixed_cost: 50, + }; + const secondUsd = { subscription_account_id: "fee-2", account_id: "acct_456", monthly_fee: 10 }; + const noDeployments = { + account_id: "acct_456", + deployments: [], + fee: fee(secondUsd), + billing_periods: [october], + fixed_cost: 10, + }; + const euro = { subscription_account_id: "fee-3", account_id: "acct_789", monthly_fee: 20, currency: "EUR" }; + const inEuros = { account_id: "acct_789", fee: fee(euro), billing_periods: [october], fixed_cost: 20 }; + renderCard([account(twoPeriods), account(noDeployments), account(inEuros)]); + + const totals = screen.getAllByTestId("fixed-cost-total").map((node) => node.textContent); + expect(totals).toEqual(["Fixed cost in range: 60.00 USD", "Fixed cost in range: 20.00 EUR"]); + expect(screen.getByText("None")).toBeInTheDocument(); + expect(screen.getByText("Sep 1 - Sep 30, 2026")).toBeInTheDocument(); + }); + + it("explains an unresolved account and offers no action for it", () => { + renderCard([account({ account_id: null })]); + expect(screen.getByText("Account not signed in on the proxy")).toBeInTheDocument(); + expect(screen.getByText("Fixed cost not tracked")).toBeInTheDocument(); + expect(screen.queryByRole("button")).not.toBeInTheDocument(); + }); + + it("shows the loader instead of the table while a range is loading", () => { + render( + , + ); + expect(screen.getByTestId("chart-loader")).toBeInTheDocument(); + expect(screen.queryByText("25.00 USD")).not.toBeInTheDocument(); + }); + + it("says so instead of disappearing when the usage request fails", () => { + render( + , + ); + expect(screen.getByText("Could not load subscription accounts for this range.")).toBeInTheDocument(); + expect(screen.queryByTestId("chart-loader")).not.toBeInTheDocument(); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.tsx new file mode 100644 index 00000000000..deee30a2db5 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCostCard.tsx @@ -0,0 +1,221 @@ +import type { ColumnDef } from "@tanstack/react-table"; +import { Info } from "lucide-react"; +import React from "react"; +import { ProviderLogo } from "@/components/molecules/models/ProviderLogo"; +import { ChartLoader } from "@/components/shared/chart_loader"; +import { DataTable } from "@/components/shared/DataTable"; +import { + AlertDialog, + AlertDialogAction, + AlertDialogCancel, + AlertDialogContent, + AlertDialogDescription, + AlertDialogFooter, + AlertDialogHeader, + AlertDialogTitle, + AlertDialogTrigger, +} from "@/components/ui/alert-dialog"; +import { Badge } from "@/components/ui/badge"; +import { Button } from "@/components/ui/button"; +import { Card, CardContent, CardHeader, CardTitle } from "@/components/ui/card"; +import { Tooltip, TooltipContent, TooltipTrigger } from "@/components/ui/tooltip"; +import { formatNumberWithCommas } from "@/utils/dataUtils"; +import { ATTRIBUTION_RULE, fixedCostByCurrency, formatBillingPeriod, formatFee } from "./formatting"; +import type { SubscriptionAccountUsage, SubscriptionUsageResponse } from "./types"; + +interface SubscriptionCostCardProps { + loading: boolean; + failed: boolean; + isDateChanging: boolean; + usage: SubscriptionUsageResponse | null; + meteredSpend: number; + canEdit: boolean; + onSetFee: (account: SubscriptionAccountUsage) => void; + onEditFee: (account: SubscriptionAccountUsage) => void; + onRemoveFee: (account: SubscriptionAccountUsage) => void; +} + +type FeeActions = Pick; + +const accountRowId = (account: SubscriptionAccountUsage): string => + `${account.custom_llm_provider}:${account.account_id ?? "unresolved"}`; + +const AccountCell = ({ account }: { account: SubscriptionAccountUsage }) => { + if (account.account_id === null) { + return Account not signed in on the proxy; + } + return ( +
+ {account.account_id} + {account.fee?.label && {account.fee.label}} +
+ ); +}; + +const ActionsCell = ({ + account, + onSetFee, + onEditFee, + onRemoveFee, +}: { account: SubscriptionAccountUsage } & FeeActions) => { + if (account.account_id === null) return null; + if (account.fee === null) { + return ( + + ); + } + return ( +
+ + + }>Remove + + + Remove this monthly fee? + + {account.account_id} goes back to "Fixed cost not tracked". Its requests stay included in the + subscription + + + + Cancel + onRemoveFee(account)}> + Remove fee + + + + +
+ ); +}; + +const baseColumns: ColumnDef[] = [ + { + header: "Provider", + accessorKey: "custom_llm_provider", + cell: ({ row }) => ( +
+ + {row.original.custom_llm_provider} +
+ ), + }, + { + header: "Account", + accessorKey: "account_id", + cell: ({ row }) => , + }, + { + header: "Deployments", + id: "deployments", + cell: ({ row }) => (row.original.deployments.length === 0 ? "None" : row.original.deployments.join(", ")), + }, + { + header: "Monthly fee", + id: "monthly_fee", + meta: { numeric: true }, + cell: ({ row }) => + row.original.fee === null ? ( + Fixed cost not tracked + ) : ( + formatFee(row.original.fee.monthly_fee, row.original.fee.currency) + ), + }, + { + header: "Billing periods in range", + id: "billing_periods", + cell: ({ row }) => + row.original.billing_periods.length === 0 ? ( + None in range + ) : ( +
+ {row.original.billing_periods.map((period) => ( + {formatBillingPeriod(period)} + ))} +
+ ), + }, + { + header: "Fixed cost in range", + id: "fixed_cost", + meta: { numeric: true }, + cell: ({ row }) => + row.original.fee === null || row.original.fixed_cost === null + ? "-" + : formatFee(row.original.fixed_cost, row.original.fee.currency), + }, +]; + +const actionsColumn = (actions: FeeActions): ColumnDef => ({ + header: "", + id: "actions", + meta: { numeric: true }, + cell: ({ row }) => , +}); + +const SubscriptionCostCard: React.FC = ({ + loading, + failed, + isDateChanging, + usage, + meteredSpend, + canEdit, + onSetFee, + onEditFee, + onRemoveFee, +}) => { + if (!failed && (usage === null || usage.accounts.length === 0)) return null; + const columns = canEdit ? [...baseColumns, actionsColumn({ onSetFee, onEditFee, onRemoveFee })] : baseColumns; + const accounts = usage?.accounts ?? []; + const totals = fixedCostByCurrency(accounts); + + return ( + + + + Subscription fixed costs + + } /> + {ATTRIBUTION_RULE} + + + + + {failed &&

Could not load subscription accounts for this range.

} + {!failed && loading && } + {!failed && !loading && ( +
+ +
+ {totals.length === 0 ? ( + Fixed cost in range: none + ) : ( + totals.map(({ currency, total }) => ( + + Fixed cost in range: {formatFee(total, currency)} + + )) + )} + + Usage-billed spend in range, all providers: ${formatNumberWithCommas(meteredSpend, 2)} + +
+

{ATTRIBUTION_RULE}

+
+ )} + + + ); +}; + +export default SubscriptionCostCard; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.integration.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.integration.test.tsx new file mode 100644 index 00000000000..f057e98dd70 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.integration.test.tsx @@ -0,0 +1,69 @@ +import { fireEvent, screen, waitFor } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; +import { beforeEach, describe, expect, it, vi } from "vitest"; +import { renderWithProviders, testQueryClient } from "@/../tests/test-utils"; +import { subscriptionAccountCreateCall, subscriptionUsageCall } from "@/components/networking"; +import SubscriptionCosts from "./SubscriptionCosts"; +import type { SubscriptionUsageResponse } from "./types"; + +vi.mock("@/components/networking", () => ({ + formatDate: (date: Date) => date.toISOString().slice(0, 10), + subscriptionUsageCall: vi.fn(), + subscriptionAccountCreateCall: vi.fn().mockResolvedValue({}), + subscriptionAccountUpdateCall: vi.fn().mockResolvedValue({}), + subscriptionAccountDeleteCall: vi.fn().mockResolvedValue({}), +})); +vi.mock("@/lib/toast", () => ({ toast: { success: vi.fn(), fromError: vi.fn() } })); + +const untracked: SubscriptionUsageResponse = { + start_date: "2026-10-01", + end_date: "2026-10-07", + subscription_providers: ["chatgpt"], + accounts: [ + { + custom_llm_provider: "chatgpt", + account_id: "acct_123", + deployments: ["gpt-5.6-sol", "gpt-5.6-terra"], + fee: null, + billing_periods: [], + fixed_cost: null, + }, + ], +}; + +const fill = (label: string, value: string) => fireEvent.change(screen.getByLabelText(label), { target: { value } }); + +describe("SubscriptionCosts", () => { + beforeEach(() => { + testQueryClient.clear(); + vi.clearAllMocks(); + vi.mocked(subscriptionUsageCall).mockResolvedValue(untracked); + }); + + it("saves a fee set from the card on that card's account and reloads the card", async () => { + const user = userEvent.setup(); + renderWithProviders( + , + ); + + await user.click(await screen.findByRole("button", { name: "Set monthly fee" })); + expect(screen.getByText("chatgpt account acct_123")).toBeInTheDocument(); + fill("Monthly fee", "25"); + fill("Currency", "USD"); + fill("Billing period start", "2026-10-01"); + await user.click(screen.getByRole("button", { name: "Save" })); + + await waitFor(() => expect(subscriptionUsageCall).toHaveBeenCalledTimes(2)); + expect(subscriptionAccountCreateCall).toHaveBeenCalledWith( + "sk-admin", + expect.objectContaining({ custom_llm_provider: "chatgpt", account_id: "acct_123", monthly_fee: 25 }), + ); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.tsx new file mode 100644 index 00000000000..e326fa1ff01 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionCosts.tsx @@ -0,0 +1,38 @@ +import React from "react"; +import { useSubscriptionUsage, type SubscriptionUsageOptions } from "../../hooks/useSubscriptionUsage"; +import SubscriptionCostCard from "./SubscriptionCostCard"; +import SubscriptionFeeModal from "./SubscriptionFeeModal"; +import { useSubscriptionFeeActions } from "./useSubscriptionFeeActions"; + +interface SubscriptionCostsProps extends SubscriptionUsageOptions { + meteredSpend: number; + canEdit: boolean; +} + +const SubscriptionCosts: React.FC = ({ meteredSpend, canEdit, ...usageOptions }) => { + const usage = useSubscriptionUsage(usageOptions); + const feeActions = useSubscriptionFeeActions({ accessToken: usageOptions.accessToken, refetch: usage.refetch }); + return ( + <> + + + + ); +}; + +export default SubscriptionCosts; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.test.tsx new file mode 100644 index 00000000000..a920f6ec883 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.test.tsx @@ -0,0 +1,83 @@ +import { fireEvent, render, screen, waitFor } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; +import { describe, expect, it, vi } from "vitest"; +import SubscriptionFeeModal from "./SubscriptionFeeModal"; +import type { SubscriptionAccountUsage } from "./types"; + +const account: SubscriptionAccountUsage = { + custom_llm_provider: "chatgpt", + account_id: "acct_123", + deployments: ["gpt-5.6-terra"], + fee: null, + billing_periods: [], + fixed_cost: null, +}; + +const fill = (label: string, value: string) => fireEvent.change(screen.getByLabelText(label), { target: { value } }); + +describe("SubscriptionFeeModal", () => { + it("submits normalized values for a new fee", async () => { + const onSubmit = vi.fn().mockResolvedValue(undefined); + render(); + + expect(screen.getByText("Set monthly fee")).toBeInTheDocument(); + expect(screen.getByText("chatgpt account acct_123")).toBeInTheDocument(); + fill("Monthly fee", "25"); + fill("Currency", "usd"); + fill("Billing period start", "2026-10-01"); + fill("Label (optional)", "Team workspace"); + await userEvent.setup().click(screen.getByRole("button", { name: "Save" })); + + await waitFor(() => expect(onSubmit).toHaveBeenCalledTimes(1)); + const normalized = { + monthly_fee: 25, + currency: "USD", + billing_period_start: "2026-10-01", + label: "Team workspace", + }; + expect(onSubmit).toHaveBeenCalledWith(normalized); + }); + + it("rejects a zero fee and a two-letter currency", async () => { + const onSubmit = vi.fn().mockResolvedValue(undefined); + const user = userEvent.setup(); + render(); + + fill("Monthly fee", "0"); + fill("Billing period start", "2026-10-01"); + await user.click(screen.getByRole("button", { name: "Save" })); + expect(await screen.findByText("Monthly fee must be greater than 0")).toBeInTheDocument(); + + fill("Monthly fee", "25"); + fill("Currency", "US"); + await user.click(screen.getByRole("button", { name: "Save" })); + expect(await screen.findByText("Currency must be a three-letter code")).toBeInTheDocument(); + expect(onSubmit).not.toHaveBeenCalled(); + }); + + it("prefills the existing fee when editing", () => { + const existing: SubscriptionAccountUsage = { + ...account, + fee: { + subscription_account_id: "fee-1", + custom_llm_provider: "chatgpt", + account_id: "acct_123", + label: "Team workspace", + monthly_fee: 25, + currency: "EUR", + billing_period_start: "2026-09-15", + created_at: "2026-10-06T00:00:00Z", + updated_at: "2026-10-06T00:00:00Z", + created_by: "admin", + updated_by: "admin", + }, + }; + render(); + + expect(screen.getByText("Edit monthly fee")).toBeInTheDocument(); + expect(screen.getByLabelText("Monthly fee")).toHaveValue(25); + expect(screen.getByLabelText("Currency")).toHaveValue("EUR"); + expect(screen.getByLabelText("Billing period start")).toHaveValue("2026-09-15"); + expect(screen.getByLabelText("Label (optional)")).toHaveValue("Team workspace"); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.tsx new file mode 100644 index 00000000000..1c211a5280c --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/SubscriptionFeeModal.tsx @@ -0,0 +1,114 @@ +"use client"; + +import React, { useEffect } from "react"; +import { z } from "zod"; +import { formatDate } from "@/components/networking"; +import { FormField } from "@/components/shared/form/FormField"; +import { Button } from "@/components/ui/button"; +import { + Dialog, + DialogContent, + DialogDescription, + DialogFooter, + DialogHeader, + DialogTitle, +} from "@/components/ui/dialog"; +import { FieldGroup } from "@/components/ui/field"; +import { Input } from "@/components/ui/input"; +import { useZodForm } from "@/lib/forms/useZodForm"; +import type { SubscriptionAccountUsage, SubscriptionFeeFormValues } from "./types"; + +const feeShape = { + monthly_fee: z + .string() + .min(1, "Enter a monthly fee") + .refine((value) => Number(value) > 0, "Monthly fee must be greater than 0"), + currency: z + .string() + .trim() + .toUpperCase() + .regex(/^[A-Z]{3}$/, "Currency must be a three-letter code"), + billing_period_start: z.string().regex(/^\d{4}-\d{2}-\d{2}$/, "Pick a billing period start date"), + label: z.string().trim(), +}; + +const feeSchema = z.object(feeShape); + +type FeeFormInput = z.input; + +const formInputFor = (account: SubscriptionAccountUsage | null): FeeFormInput => ({ + monthly_fee: account?.fee ? String(account.fee.monthly_fee) : "", + currency: account?.fee?.currency ?? "USD", + billing_period_start: account?.fee?.billing_period_start ?? formatDate(new Date()), + label: account?.fee?.label ?? "", +}); + +interface SubscriptionFeeModalProps { + visible: boolean; + account: SubscriptionAccountUsage | null; + onCancel: () => void; + onSubmit: (values: SubscriptionFeeFormValues) => Promise; +} + +const SubscriptionFeeModal: React.FC = ({ visible, account, onCancel, onSubmit }) => { + const form = useZodForm(feeSchema, { defaultValues: formInputFor(account) }); + const { reset } = form; + + useEffect(() => { + if (visible) reset(formInputFor(account)); + }, [visible, account, reset]); + + const handleFinish = async (values: z.output) => { + const submitted: SubscriptionFeeFormValues = { + monthly_fee: Number(values.monthly_fee), + currency: values.currency, + billing_period_start: values.billing_period_start, + label: values.label === "" ? null : values.label, + }; + await onSubmit(submitted); + }; + + const handleCancel = () => { + reset(formInputFor(null)); + onCancel(); + }; + + return ( + !open && handleCancel()}> + + + {account?.fee ? "Edit monthly fee" : "Set monthly fee"} + + {account ? `${account.custom_llm_provider} account ${account.account_id ?? ""}` : ""} + + +
+ + + {({ ref, ...field }) => } + + + {({ ref, ...field }) => } + + + {({ ref, ...field }) => } + + + {({ ref, ...field }) => } + + + + + + +
+ + + ); +}; + +export default SubscriptionFeeModal; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/formatting.ts b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/formatting.ts new file mode 100644 index 00000000000..b1d4e0a48ef --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/formatting.ts @@ -0,0 +1,34 @@ +import type { BillingPeriod } from "./types"; + +export const ATTRIBUTION_RULE = + "A subscription fee is counted once for each monthly billing period whose start date falls inside the selected dates. Fixed costs sit beside metered spend and are never added to it, to key or team spend, or to budgets."; + +export const formatFee = (amount: number, currency: string): string => `${amount.toFixed(2)} ${currency}`; + +const parseLocalDay = (day: string): Date => { + const [year, month, date] = day.split("-").map(Number); + return new Date(year, month - 1, date); +}; + +export const formatBillingPeriod = ({ start, end }: BillingPeriod): string => { + const startDate = parseLocalDay(start); + const endDate = parseLocalDay(end); + const startLabel = startDate.toLocaleDateString("en-US", { + month: "short", + day: "numeric", + year: startDate.getFullYear() !== endDate.getFullYear() ? "numeric" : undefined, + }); + const endLabel = endDate.toLocaleDateString("en-US", { month: "short", day: "numeric", year: "numeric" }); + return `${startLabel} - ${endLabel}`; +}; + +export const fixedCostByCurrency = ( + accounts: readonly { fee: { currency: string } | null; fixed_cost: number | null }[], +): readonly { currency: string; total: number }[] => { + const totals = new Map(); + accounts.forEach((account) => { + if (account.fee === null || account.fixed_cost === null) return; + totals.set(account.fee.currency, (totals.get(account.fee.currency) ?? 0) + account.fixed_cost); + }); + return Array.from(totals, ([currency, total]) => ({ currency, total })); +}; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/types.ts b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/types.ts new file mode 100644 index 00000000000..6af5e0cdf3b --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/types.ts @@ -0,0 +1,74 @@ +import type { DailyData } from "@/components/UsagePage/types"; + +export interface SubscriptionFee { + subscription_account_id: string; + custom_llm_provider: string; + account_id: string; + label: string | null; + monthly_fee: number; + currency: string; + billing_period_start: string; + created_at: string; + updated_at: string; + created_by: string; + updated_by: string; +} + +export interface BillingPeriod { + start: string; + end: string; +} + +export interface SubscriptionAccountUsage { + custom_llm_provider: string; + account_id: string | null; + deployments: string[]; + fee: SubscriptionFee | null; + billing_periods: BillingPeriod[]; + fixed_cost: number | null; +} + +export interface SubscriptionUsageResponse { + subscription_providers: string[]; + accounts: SubscriptionAccountUsage[]; +} + +export interface SubscriptionAccountCreateRequest { + custom_llm_provider: string; + account_id: string; + monthly_fee: number; + currency: string; + billing_period_start: string; + label: string | null; +} + +export interface SubscriptionAccountUpdateRequest { + subscription_account_id: string; + monthly_fee?: number; + currency?: string; + billing_period_start?: string; + label?: string | null; +} + +export interface SubscriptionAccountDeleteRequest { + subscription_account_id: string; +} + +export interface SubscriptionFeeFormValues { + monthly_fee: number; + currency: string; + billing_period_start: string; + label: string | null; +} + +export const isSubscriptionCovered = (metadata: object | undefined): boolean => + metadata !== undefined && "subscription_covered" in metadata && metadata.subscription_covered === true; + +export const subscriptionCoveredProviders = (results: readonly DailyData[]): ReadonlySet => + new Set( + results.flatMap((day) => + Object.entries(day.breakdown.providers ?? {}) + .filter(([, entry]) => isSubscriptionCovered(entry.metadata)) + .map(([provider]) => provider), + ), + ); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.test.tsx new file mode 100644 index 00000000000..9406e07b497 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.test.tsx @@ -0,0 +1,79 @@ +import { act, renderHook } from "@testing-library/react"; +import { beforeEach, describe, expect, it, vi } from "vitest"; +import { subscriptionAccountCreateCall, subscriptionAccountUpdateCall } from "@/components/networking"; +import type { SubscriptionAccountUsage, SubscriptionFeeFormValues } from "./types"; +import { useSubscriptionFeeActions } from "./useSubscriptionFeeActions"; + +vi.mock("@/components/networking", () => ({ + subscriptionAccountCreateCall: vi.fn().mockResolvedValue({}), + subscriptionAccountUpdateCall: vi.fn().mockResolvedValue({}), + subscriptionAccountDeleteCall: vi.fn(), +})); +vi.mock("@/lib/toast", () => ({ toast: { success: vi.fn(), fromError: vi.fn() } })); + +const stored: SubscriptionFeeFormValues = { + monthly_fee: 25, + currency: "USD", + billing_period_start: "2026-10-01", + label: "Team plan (test value)", +}; + +const tracked: SubscriptionAccountUsage = { + custom_llm_provider: "chatgpt", + account_id: "acct_123", + deployments: ["gpt-5.6-terra"], + fee: { + subscription_account_id: "sub-1", + custom_llm_provider: "chatgpt", + account_id: "acct_123", + ...stored, + created_at: "2026-10-01T00:00:00Z", + updated_at: "2026-10-01T00:00:00Z", + created_by: "admin", + updated_by: "admin", + }, + billing_periods: [{ start: "2026-10-01", end: "2026-10-31" }], + fixed_cost: 25, +}; + +const submit = async (account: SubscriptionAccountUsage, values: SubscriptionFeeFormValues) => { + const refetch = vi.fn().mockResolvedValue(undefined); + const { result } = renderHook(() => useSubscriptionFeeActions({ accessToken: "sk-admin", refetch })); + act(() => result.current.openFeeModal(account)); + await act(() => result.current.submitFee(values)); + return { refetch, result }; +}; + +describe("useSubscriptionFeeActions", () => { + beforeEach(() => vi.clearAllMocks()); + + it("sends only the fields the admin changed when editing a fee", async () => { + const { refetch, result } = await submit(tracked, { ...stored, monthly_fee: 30 }); + + expect(subscriptionAccountUpdateCall).toHaveBeenCalledWith("sk-admin", { + subscription_account_id: "sub-1", + monthly_fee: 30, + }); + expect(refetch).toHaveBeenCalledTimes(1); + expect(result.current.feeModalAccount).toBeNull(); + }); + + it("clears a label with an explicit null and nothing else", async () => { + await submit(tracked, { ...stored, label: null }); + + expect(subscriptionAccountUpdateCall).toHaveBeenCalledWith("sk-admin", { + subscription_account_id: "sub-1", + label: null, + }); + }); + + it("creates a fee with every field for an untracked account", async () => { + const untracked: SubscriptionAccountUsage = { ...tracked, fee: null, billing_periods: [], fixed_cost: null }; + const createRequest = { custom_llm_provider: "chatgpt", account_id: "acct_123", ...stored }; + + await submit(untracked, stored); + + expect(subscriptionAccountCreateCall).toHaveBeenCalledWith("sk-admin", createRequest); + expect(subscriptionAccountUpdateCall).not.toHaveBeenCalled(); + }); +}); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.ts b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.ts new file mode 100644 index 00000000000..dd2b2f01a30 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/SubscriptionCost/useSubscriptionFeeActions.ts @@ -0,0 +1,90 @@ +import { useCallback, useState } from "react"; +import { + subscriptionAccountCreateCall, + subscriptionAccountDeleteCall, + subscriptionAccountUpdateCall, +} from "@/components/networking"; +import { toast } from "@/lib/toast"; +import type { + SubscriptionAccountUpdateRequest, + SubscriptionAccountUsage, + SubscriptionFee, + SubscriptionFeeFormValues, +} from "./types"; + +interface Options { + accessToken: string | null; + refetch: () => Promise; +} + +interface Result { + feeModalAccount: SubscriptionAccountUsage | null; + openFeeModal: (account: SubscriptionAccountUsage) => void; + closeFeeModal: () => void; + submitFee: (values: SubscriptionFeeFormValues) => Promise; + removeFee: (account: SubscriptionAccountUsage) => Promise; +} + +const changedFields = (fee: SubscriptionFee, values: SubscriptionFeeFormValues): SubscriptionAccountUpdateRequest => ({ + subscription_account_id: fee.subscription_account_id, + ...(values.monthly_fee !== fee.monthly_fee && { monthly_fee: values.monthly_fee }), + ...(values.currency !== fee.currency && { currency: values.currency }), + ...(values.billing_period_start !== fee.billing_period_start && { + billing_period_start: values.billing_period_start, + }), + ...(values.label !== fee.label && { label: values.label }), +}); + +const saveFee = ( + accessToken: string, + account: SubscriptionAccountUsage, + values: SubscriptionFeeFormValues, +): Promise | null => { + if (account.fee) { + return subscriptionAccountUpdateCall(accessToken, changedFields(account.fee, values)); + } + if (account.account_id === null) return null; + return subscriptionAccountCreateCall(accessToken, { + custom_llm_provider: account.custom_llm_provider, + account_id: account.account_id, + ...values, + }); +}; + +export function useSubscriptionFeeActions({ accessToken, refetch }: Options): Result { + const [feeModalAccount, setFeeModalAccount] = useState(null); + const closeFeeModal = useCallback(() => setFeeModalAccount(null), []); + + const submitFee = useCallback( + async (values: SubscriptionFeeFormValues) => { + if (!accessToken || feeModalAccount === null) return; + const request = saveFee(accessToken, feeModalAccount, values); + if (request === null) return; + try { + await request; + toast.success("Monthly fee saved"); + setFeeModalAccount(null); + await refetch(); + } catch (error) { + toast.fromError(error); + } + }, + [accessToken, feeModalAccount, refetch], + ); + + const removeFee = useCallback( + async (account: SubscriptionAccountUsage) => { + if (!accessToken || account.fee === null) return; + try { + await subscriptionAccountDeleteCall(accessToken, account.fee.subscription_account_id); + toast.success("Monthly fee removed"); + await refetch(); + } catch (error) { + toast.fromError(error); + } + }, + [accessToken, refetch], + ); + + return { feeModalAccount, openFeeModal: setFeeModalAccount, closeFeeModal, submitFee, removeFee }; +} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.test.tsx index f46a4569369..9d39e9c080c 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.test.tsx @@ -8,7 +8,7 @@ import type { ComponentProps } from "react"; import userEvent from "@testing-library/user-event"; import { beforeAll, beforeEach, describe, expect, it, vi } from "vitest"; import type { UrlUpdateEvent } from "nuqs/adapters/testing"; -import { chooseSelectOption, renderWithProviders } from "@/../tests/test-utils"; +import { chooseSelectOption, renderWithProviders, testQueryClient } from "@/../tests/test-utils"; import { toast } from "@/lib/toast"; import { processActivityData } from "@/components/activity_metrics"; import type { Organization } from "@/components/networking"; @@ -74,6 +74,11 @@ beforeAll(() => { // Mock the networking module vi.mock("@/components/networking", () => ({ + formatDate: (date: Date) => date.toISOString().slice(0, 10), + subscriptionUsageCall: vi.fn(), + subscriptionAccountCreateCall: vi.fn(), + subscriptionAccountUpdateCall: vi.fn(), + subscriptionAccountDeleteCall: vi.fn(), dailyActivityAggregatedCall: vi.fn(), dailyActivityKeyPageCall: vi.fn(), userDailyActivityUserPageCall: vi.fn(), @@ -142,6 +147,14 @@ vi.mock("./EntityUsage/SpendByProvider", () => ({ default: () =>
Spend By Provider
, })); +vi.mock("./SubscriptionCost/SubscriptionCostCard", () => ({ + default: ({ usage, canEdit }: { usage: unknown; canEdit: boolean }) => ( +
+ Subscription Cost +
+ ), +})); + vi.mock("./EndpointUsage/EndpointUsage", () => ({ default: () =>
Endpoint Usage
, })); @@ -278,6 +291,7 @@ describe("UsagePage", () => { const mockTagListCall = vi.mocked(networking.tagListCall); const mockGatewayDailyActivityCall = vi.mocked(networking.gatewayDailyActivityCall); const mockRequestErrorActivityCall = vi.mocked(networking.requestErrorActivityCall); + const mockSubscriptionUsageCall = vi.mocked(networking.subscriptionUsageCall); const mockUseAgents = vi.mocked(useAgents); const mockUseAuthorized = vi.mocked(useAuthorized); const mockUseCurrentUser = vi.mocked(useCurrentUser); @@ -496,7 +510,24 @@ describe("UsagePage", () => { organizations: [], }; + const mockSubscriptionUsage = { + subscription_providers: ["chatgpt"], + accounts: [ + { + custom_llm_provider: "chatgpt", + account_id: "acct_123", + deployments: ["gpt-5.6-terra", "gpt-5.6-sol"], + fee: null, + billing_periods: [], + fixed_cost: null, + }, + ], + }; + beforeEach(() => { + testQueryClient.clear(); + mockSubscriptionUsageCall.mockReset(); + mockSubscriptionUsageCall.mockResolvedValue(mockSubscriptionUsage); mockClientUsagePanel.mockClear(); vi.mocked(processActivityData).mockReset(); vi.mocked(processActivityData).mockImplementation((_data, key) => ({ @@ -578,6 +609,72 @@ describe("UsagePage", () => { } as any); }); + it("hands the subscription usage to the fixed cost card for an admin and never fetches it otherwise", async () => { + const { unmount } = renderWithProviders(); + await waitFor(() => { + expect(mockSubscriptionUsageCall).toHaveBeenCalledWith("test-token", expect.any(Date), expect.any(Date)); + }); + await waitFor(() => { + expect(screen.getByTestId("subscription-cost-card")).toHaveAttribute( + "data-usage", + JSON.stringify(mockSubscriptionUsage), + ); + }); + unmount(); + + for (const session of [nonAdminSession, { ...nonAdminSession, userRole: "org_admin" }]) { + mockSubscriptionUsageCall.mockClear(); + mockUserDailyActivityAggregatedCall.mockClear(); + mockUseAuthorized.mockReturnValue(session); + const rendered = renderWithProviders(); + await waitFor(() => { + expect(mockUserDailyActivityAggregatedCall).toHaveBeenCalled(); + }); + expect(mockSubscriptionUsageCall).not.toHaveBeenCalled(); + expect(screen.getByTestId("subscription-cost-card")).toHaveAttribute("data-usage", "null"); + rendered.unmount(); + } + }); + + it("lets a proxy admin edit subscription fees but shows an admin viewer the same card read-only", async () => { + const adminRender = renderWithProviders(); + await waitFor(() => { + expect(screen.getByTestId("subscription-cost-card")).toHaveAttribute( + "data-usage", + JSON.stringify(mockSubscriptionUsage), + ); + }); + expect(screen.getByTestId("subscription-cost-card")).toHaveAttribute("data-can-edit", "true"); + adminRender.unmount(); + + const adminViewerSession = { + ...nonAdminSession, + userRole: "Admin", + userRoleLabel: "Admin Viewer", + isViewOnly: true, + }; + mockUseAuthorized.mockReturnValue(adminViewerSession); + renderWithProviders(); + await waitFor(() => { + expect(screen.getByTestId("subscription-cost-card")).toHaveAttribute( + "data-usage", + JSON.stringify(mockSubscriptionUsage), + ); + }); + expect(screen.getByTestId("subscription-cost-card")).toHaveAttribute("data-can-edit", "false"); + }); + + it("never sets an account-wide subscription fee beside one user's spend when an admin filters by user", async () => { + renderWithProviders(, { searchParams: "?user=user-001" }); + await waitFor(() => { + expect(mockUserDailyActivityAggregatedCall).toHaveBeenCalledWith( + expect.objectContaining({ entityIds: ["user-001"] }), + ); + }); + expect(mockSubscriptionUsageCall).not.toHaveBeenCalled(); + expect(screen.getByTestId("subscription-cost-card")).toHaveAttribute("data-usage", "null"); + }); + it("should render and fetch usage data on mount", async () => { renderWithProviders(); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx index d5ec3ef7675..7c891754195 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/UsagePageView.tsx @@ -26,7 +26,7 @@ import { useOrganizations } from "@/app/(dashboard)/hooks/organizations/useOrgan import { toast } from "@/lib/toast"; import { hasCapability } from "@/utils/capabilities"; import { copyToClipboard } from "@/utils/dataUtils"; -import { hasProxyWideSpendView, internalUserRoles } from "@/utils/roles"; +import { hasProxyWideSpendView, internalUserRoles, isProxyAdminRole } from "@/utils/roles"; import { ActivityMetrics, processActivityData } from "@/components/activity_metrics"; import CloudZeroExportModal from "@/components/cloudzero_export_modal"; import UserDropdown from "@/components/common_components/UserDropdown"; @@ -92,7 +92,7 @@ import { usageViewFromParams, usageViewPatch, } from "../usageUrlState"; -import { overviewTotals, rollUpBreakdown } from "./overview/overviewData"; +import { overviewTotals, providerSpendRows } from "./overview/overviewData"; import TopAgents from "./overview/TopAgents"; import { parseUsageClient, topAgents, type AgentRow } from "./overview/agentCatalog"; import ClientUsagePanel from "./ClientUsagePanel"; @@ -100,6 +100,7 @@ import { resolveClientUsageScope } from "../hooks/clientUsageScope"; import { useUsagePanelContext } from "../hooks/useUsagePanelContext"; import ClientUsageSelect from "./ClientUsageSelect"; import SpendByProvider from "./EntityUsage/SpendByProvider"; +import SubscriptionCosts from "./SubscriptionCost/SubscriptionCosts"; import { useTagSummary } from "@/app/(dashboard)/hooks/tags/useTagSummary"; import { Panel } from "./overview/Primitives"; import UsageOverview from "./overview/UsageOverview"; @@ -112,7 +113,7 @@ interface UsagePageProps { } const UsagePage: React.FC = ({ teams, organizations, teamsLoading = false, defaultTimeRange }) => { - const { accessToken, userRole, userId: userID, premiumUser, isLoading: authorizationLoading } = useAuthorized(); + const { accessToken, userRole, userId: userID, premiumUser, isViewOnly, isLoading: authLoading } = useAuthorized(); const [gatewayActivityData, setGatewayActivityData] = useState(null); const defaultRange = useMemo(() => defaultUsageDateRange(defaultTimeRange), [defaultTimeRange]); @@ -187,7 +188,7 @@ const UsagePage: React.FC = ({ teams, organizations, teamsLoadin }; const { isLoading: organizationsLoading } = useOrganizations(); - const authorizationSettled = !authorizationLoading && userRole !== null; + const authorizationSettled = !authLoading && userRole !== null; const accessSettled = authorizationSettled && !organizationsLoading && !teamsLoading; useEffect(() => { if (!accessSettled) return; @@ -374,10 +375,7 @@ const UsagePage: React.FC = ({ teams, organizations, teamsLoadin () => overviewTotals(userSpendData.metadata, gatewayActivity), [userSpendData.metadata, gatewayActivity], ); - const providerSpend = useMemo( - () => rollUpBreakdown(userSpendData.results, "providers").map(({ key, ...row }) => ({ provider: key, ...row })), - [userSpendData.results], - ); + const providerSpend = useMemo(() => providerSpendRows(userSpendData.results), [userSpendData.results]); const { data: tagSummary, isLoading: tagSummaryLoading, @@ -546,6 +544,16 @@ const UsagePage: React.FC = ({ teams, organizations, teamsLoadin loading={loading} requestCountsPending={requestCountsPending} budget={currentUser?.max_budget ?? null} + subscriptionCosts={ + + } topKeys={ + {subscriptionCosts} + {topAgents} diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.test.ts b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.test.ts index f00d53ea1bc..1dfdfd1da78 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.test.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.test.ts @@ -7,6 +7,7 @@ import { labelForDate, modelColorMap, OTHER_COLOR, + providerSpendRows, rollUpBreakdown, seriesBy, } from "./overviewData"; @@ -88,6 +89,34 @@ describe("bucketSeries", () => { }); }); +const providerDay = (date: string, providers: Record>): DailyData => + ({ + ...day(date, 0, {}), + breakdown: { + ...day(date, 0, {}).breakdown, + providers: Object.fromEntries( + Object.entries(providers).map(([name, metadata]) => [ + name, + { metrics: metrics(0), metadata, api_key_breakdown: {} }, + ]), + ), + }, + }) as unknown as DailyData; + +describe("providerSpendRows", () => { + it("marks a provider covered by a subscription on any day of the range and leaves metered ones unmarked", () => { + const rows = providerSpendRows([ + providerDay("2026-10-01", { chatgpt: {}, openai: {} }), + providerDay("2026-10-02", { chatgpt: { subscription_covered: true }, openai: { subscription_covered: false } }), + ]); + expect(Object.fromEntries(rows.map((row) => [row.provider, row.subscription_covered]))).toEqual({ + chatgpt: true, + openai: false, + }); + expect(rows.find((row) => row.provider === "chatgpt")).toEqual(expect.objectContaining({ requests: 2 })); + }); +}); + describe("rollUpBreakdown", () => { it("sums a dimension across days without changing the input", () => { const days = [day("2026-10-01", 3, { a: 3 }), day("2026-10-02", 2, { a: 2 })]; diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.ts b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.ts index c103c2d7c60..9d9e8f9fa32 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.ts +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/components/overview/overviewData.ts @@ -2,6 +2,7 @@ import type { DailyData, SpendMetrics } from "@/components/UsagePage/types"; import { stackedUsageColor } from "@/components/shared/charts"; import { EMPTY_DAILY_ACTIVITY_METADATA, type DailyActivityMetadata } from "@/components/UsagePage/dailyActivityApi"; import type { GatewayActivity } from "../gatewayActivity"; +import { subscriptionCoveredProviders } from "../SubscriptionCost/types"; export type BreakdownDimension = "models" | "model_groups" | "providers"; @@ -40,6 +41,15 @@ export const rollUpBreakdown = (results: readonly DailyData[], dimension: Breakd return [...totals.values()].sort((a, b) => b.spend - a.spend); }; +export const providerSpendRows = (results: readonly DailyData[]) => { + const covered = subscriptionCoveredProviders(results); + return rollUpBreakdown(results, "providers").map(({ key, ...row }) => ({ + provider: key, + ...row, + subscription_covered: covered.has(key), + })); +}; + export const modelColorMap = (results: readonly DailyData[]): ReadonlyMap => { const groups = rollUpBreakdown(results, "model_groups"); const models = rollUpBreakdown(results, "models"); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/hooks/useSubscriptionUsage.ts b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/hooks/useSubscriptionUsage.ts new file mode 100644 index 00000000000..afda5ec57b8 --- /dev/null +++ b/ui/litellm-dashboard/src/app/(dashboard)/usage/_components/hooks/useSubscriptionUsage.ts @@ -0,0 +1,47 @@ +import { keepPreviousData, queryOptions, skipToken, useQuery } from "@tanstack/react-query"; +import { formatDate, subscriptionUsageCall } from "@/components/networking"; +import type { SubscriptionUsageResponse } from "../components/SubscriptionCost/types"; + +export interface SubscriptionUsageOptions { + accessToken: string | null; + startTime: Date | null; + endTime: Date | null; + enabled: boolean; +} + +interface Result { + data: SubscriptionUsageResponse | null; + loading: boolean; + failed: boolean; + refetch: () => Promise; +} + +interface UsageRequest { + accessToken: string; + startTime: Date; + endTime: Date; +} + +const subscriptionUsageQuery = (request: UsageRequest | null) => { + const queryKey = [ + "subscriptionUsage", + request ? formatDate(request.startTime) : null, + request ? formatDate(request.endTime) : null, + ]; + const queryFn = request + ? () => subscriptionUsageCall(request.accessToken, request.startTime, request.endTime) + : skipToken; + return queryOptions({ queryKey, queryFn, placeholderData: keepPreviousData }); +}; + +export function useSubscriptionUsage({ accessToken, startTime, endTime, enabled }: SubscriptionUsageOptions): Result { + const range = startTime && endTime ? { startTime, endTime } : null; + const request = enabled && accessToken && range ? { accessToken, ...range } : null; + const query = useQuery(subscriptionUsageQuery(request)); + return { + data: request && query.data ? query.data : null, + loading: request !== null && (query.isPending || query.isPlaceholderData), + failed: query.isError, + refetch: query.refetch, + }; +} diff --git a/ui/litellm-dashboard/src/components/networking.tsx b/ui/litellm-dashboard/src/components/networking.tsx index 99a9dc7a707..c13052be245 100644 --- a/ui/litellm-dashboard/src/components/networking.tsx +++ b/ui/litellm-dashboard/src/components/networking.tsx @@ -119,6 +119,13 @@ import { MCP_TOOLS_PREVIEW_FORBIDDEN_MESSAGE } from "./mcp_tools/constants"; import type { ComplexityRouterConfigPayload } from "./add_model/build_complexity_router_config"; import type { AutoRouterPresetsResponse } from "@/lib/autorouter_presets"; import type { VectorStoreIndex } from "@/app/(dashboard)/vector-stores/_components/IndexesTab"; +import type { + SubscriptionAccountCreateRequest, + SubscriptionAccountDeleteRequest, + SubscriptionAccountUpdateRequest, + SubscriptionFee, + SubscriptionUsageResponse, +} from "@/app/(dashboard)/usage/_components/components/SubscriptionCost/types"; import type { RoutingDecision } from "./logs/detail/RoutingDecisionCard"; import type { SpanDetail, @@ -1395,6 +1402,34 @@ export const dailyActivityAggregatedCall = ( query: dailyActivityQuery(entity, req, req.apiKeyLimit === undefined ? {} : { api_key_limit: req.apiKeyLimit }), }); +export const subscriptionUsageCall = ( + accessToken: string, + startTime: Date, + endTime: Date, +): Promise => + apiClient.get("/subscription_accounts/usage", { + accessToken, + query: { start_date: formatDate(startTime), end_date: formatDate(endTime) }, + }); + +export const subscriptionAccountCreateCall = ( + accessToken: string, + body: SubscriptionAccountCreateRequest, +): Promise => apiClient.post("/subscription_accounts/new", { accessToken, body }); + +export const subscriptionAccountUpdateCall = ( + accessToken: string, + body: SubscriptionAccountUpdateRequest, +): Promise => apiClient.post("/subscription_accounts/update", { accessToken, body }); + +export const subscriptionAccountDeleteCall = ( + accessToken: string, + subscriptionAccountId: string, +): Promise => { + const body: SubscriptionAccountDeleteRequest = { subscription_account_id: subscriptionAccountId }; + return apiClient.post("/subscription_accounts/delete", { accessToken, body }); +}; + export const dailyActivityKeyPageCall = ( entity: DailyActivityEntity, req: DailyActivityRequest, diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 7d9e9839a5f..f53e57c2f18 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -16080,6 +16080,90 @@ export interface paths { patch?: never; trace?: never; }; + "/subscription_accounts/delete": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + get?: never; + put?: never; + /** + * Delete Subscription Account + * @description Stop tracking a subscription account's fee; its requests stay marked as subscription-covered. + */ + post: operations["delete_subscription_account_subscription_accounts_delete_post"]; + delete?: never; + options?: never; + head?: never; + patch?: never; + trace?: never; + }; + "/subscription_accounts/new": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + get?: never; + put?: never; + /** + * New Subscription Account + * @description Record the monthly fee of one subscription-backed provider account. + */ + post: operations["new_subscription_account_subscription_accounts_new_post"]; + delete?: never; + options?: never; + head?: never; + patch?: never; + trace?: never; + }; + "/subscription_accounts/update": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + get?: never; + put?: never; + /** + * Update Subscription Account + * @description Change the fee, currency, billing period start, or label of a subscription account. + * Only the fields named in the request are written, so two admins changing different + * fields at the same time both land. + */ + post: operations["update_subscription_account_subscription_accounts_update_post"]; + delete?: never; + options?: never; + head?: never; + patch?: never; + trace?: never; + }; + "/subscription_accounts/usage": { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + /** + * Get Subscription Usage + * @description Subscription-backed accounts behind the proxy's deployments or carrying a stored fee, each with + * its configured monthly fee and the fixed cost attributed to the date range: the fee once per billing + * period whose start date falls inside the range, however many deployments share the account. + */ + get: operations["get_subscription_usage_subscription_accounts_usage_get"]; + put?: never; + post?: never; + delete?: never; + options?: never; + head?: never; + patch?: never; + trace?: never; + }; "/systemone": { parameters: { query?: never; @@ -27402,6 +27486,19 @@ export interface components { /** Entities */ entities: components["schemas"]["BedrockChecksSensitiveInformationEntityItem"][]; }; + /** BillingPeriod */ + BillingPeriod: { + /** + * End + * Format: date + */ + end: string; + /** + * Start + * Format: date + */ + start: string; + }; /** BlockKeyRequest */ BlockKeyRequest: { /** Key */ @@ -31882,6 +31979,11 @@ export interface components { */ type: string; }; + /** DeleteSubscriptionAccountRequest */ + DeleteSubscriptionAccountRequest: { + /** Subscription Account Id */ + subscription_account_id: string; + }; /** DeleteTeamRequest */ DeleteTeamRequest: { /** Team Ids */ @@ -36017,6 +36119,37 @@ export interface components { */ user: string | null; }; + /** LiteLLM_SubscriptionAccountTable */ + LiteLLM_SubscriptionAccountTable: { + /** Account Id */ + account_id: string; + /** Billing Period Start */ + billing_period_start: string; + /** + * Created At + * Format: date-time + */ + created_at: string; + /** Created By */ + created_by: string; + /** Currency */ + currency: string; + /** Custom Llm Provider */ + custom_llm_provider: string; + /** Label */ + label?: string | null; + /** Monthly Fee */ + monthly_fee: number; + /** Subscription Account Id */ + subscription_account_id: string; + /** + * Updated At + * Format: date-time + */ + updated_at: string; + /** Updated By */ + updated_by: string; + }; /** LiteLLM_TeamMembership */ LiteLLM_TeamMembership: { /** Budget Id */ @@ -46225,6 +46358,55 @@ export interface components { /** Window Start */ window_start: number; }; + /** SubscriptionAccountFeeRequest */ + SubscriptionAccountFeeRequest: { + /** Account Id */ + account_id: string; + /** + * Billing Period Start + * Format: date + */ + billing_period_start: string; + /** Currency */ + currency: string; + /** Custom Llm Provider */ + custom_llm_provider: string; + /** Label */ + label?: string | null; + /** Monthly Fee */ + monthly_fee: number; + }; + /** SubscriptionAccountUsage */ + SubscriptionAccountUsage: { + /** Account Id */ + account_id: string | null; + /** Billing Periods */ + billing_periods: components["schemas"]["BillingPeriod"][]; + /** Custom Llm Provider */ + custom_llm_provider: string; + /** Deployments */ + deployments: string[]; + fee: components["schemas"]["LiteLLM_SubscriptionAccountTable"] | null; + /** Fixed Cost */ + fixed_cost: number | null; + }; + /** SubscriptionUsageResponse */ + SubscriptionUsageResponse: { + /** Accounts */ + accounts: components["schemas"]["SubscriptionAccountUsage"][]; + /** + * End Date + * Format: date + */ + end_date: string; + /** + * Start Date + * Format: date + */ + start_date: string; + /** Subscription Providers */ + subscription_providers: string[]; + }; /** * SuccessfulKeyUpdate * @description Successfully updated key with its updated information @@ -48898,6 +49080,19 @@ export interface components { UpdateSearchToolRequest: { search_tool: components["schemas"]["SearchTool"]; }; + /** UpdateSubscriptionAccountFeeRequest */ + UpdateSubscriptionAccountFeeRequest: { + /** Billing Period Start */ + billing_period_start?: string | null; + /** Currency */ + currency?: string | null; + /** Label */ + label?: string | null; + /** Monthly Fee */ + monthly_fee?: number | null; + /** Subscription Account Id */ + subscription_account_id: string; + }; /** * UpdateTeamMemberPermissionsRequest * @description Request to update the team member permissions for a team @@ -72988,6 +73183,139 @@ export interface operations { }; }; }; + delete_subscription_account_subscription_accounts_delete_post: { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + requestBody: { + content: { + "application/json": components["schemas"]["DeleteSubscriptionAccountRequest"]; + }; + }; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["LiteLLM_SubscriptionAccountTable"]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; + new_subscription_account_subscription_accounts_new_post: { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + requestBody: { + content: { + "application/json": components["schemas"]["SubscriptionAccountFeeRequest"]; + }; + }; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["LiteLLM_SubscriptionAccountTable"]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; + update_subscription_account_subscription_accounts_update_post: { + parameters: { + query?: never; + header?: never; + path?: never; + cookie?: never; + }; + requestBody: { + content: { + "application/json": components["schemas"]["UpdateSubscriptionAccountFeeRequest"]; + }; + }; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["LiteLLM_SubscriptionAccountTable"]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; + get_subscription_usage_subscription_accounts_usage_get: { + parameters: { + query: { + /** @description First day of the Usage page filter, YYYY-MM-DD */ + start_date: string; + /** @description Last day of the Usage page filter, YYYY-MM-DD */ + end_date: string; + }; + header?: never; + path?: never; + cookie?: never; + }; + requestBody?: never; + responses: { + /** @description Successful Response */ + 200: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["SubscriptionUsageResponse"]; + }; + }; + /** @description Validation Error */ + 422: { + headers: { + [name: string]: unknown; + }; + content: { + "application/json": components["schemas"]["HTTPValidationError"]; + }; + }; + }; + }; systemone_systemone_post: { parameters: { query?: never; From 0070ce23b311375c0e2b34a51f193e90ec5f6c65 Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 30 Oct 2026 04:27:11 -0700 Subject: [PATCH 20/23] fix(router)!: apply content_policy_fallbacks when router settings update at runtime (#45428) * fix(router): apply content_policy_fallbacks when router settings update at runtime * test(router): annotate the fallback regression test's locals as Final * fix(proxy): validate router_settings.content_policy_fallbacks on /config/update like its siblings * test(proxy): annotate the content_policy_fallbacks config update test locals as Final * fix(proxy): keep litellm_settings content policy fallbacks, accept dict targets, skip invalid stored rules A runtime content_policy_fallbacks rule now layers over the litellm_settings rules instead of replacing them, so models the runtime rules do not name keep their file fallbacks. Fallback targets may be deployment dicts as well as model names. A stored rule that fails validation is logged and skipped at sync, so the other router settings still apply. Per-user token discovery now reads every fallback key from key and team router_settings. * fix(proxy): validate content policy fallback rules one by one and keep a config-owned list alone An invalid stored rule is now skipped on its own instead of dropping every runtime rule, a router_settings list from the config file stays the router's only rules across DB syncs, and GET /fallback names the model of a dict target. * fix(proxy): answer an error for a string-valued fallback rule on GET /fallback instead of its characters * fix(proxy): keep config-file content-policy rules as Router init accepted them and log a skipped rule once per change * fix(proxy): log a skipped content-policy rule whenever the saved rules change * fix(proxy): keep general fallbacks when no content-policy rule is saved, name a model-less target's model, and take only model names in key rules * fix(proxy): reject an empty dict target in a key's content-policy rule --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- litellm/constants.py | 1 + litellm/proxy/litellm_pre_call_utils.py | 17 +- .../fallback_management_endpoints.py | 95 ++++++- litellm/proxy/proxy_server.py | 33 ++- .../management_endpoints/key_requests.py | 24 +- litellm/types/router.py | 6 + .../test_key_management_endpoints.py | 16 +- .../test_fallback_management_endpoints.py | 100 ++++++- .../unit/proxy/test_litellm_pre_call_utils.py | 11 +- ...test_proxy_server_endpoints_and_startup.py | 243 +++++++++++++++++- tests/unit/test_router/test_router.py | 33 +++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 6 + 12 files changed, 553 insertions(+), 32 deletions(-) diff --git a/litellm/constants.py b/litellm/constants.py index 97e93309ceb..c586a5f767a 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -68,6 +68,7 @@ RUNTIME_UPDATABLE_ROUTER_SETTINGS: Final[frozenset[str]] = frozenset( "retry_after", "fallbacks", "context_window_fallbacks", + "content_policy_fallbacks", "retry_policy", "model_group_retry_policy", "model_group_alias", diff --git a/litellm/proxy/litellm_pre_call_utils.py b/litellm/proxy/litellm_pre_call_utils.py index c1e541d20a9..9b52f2f8a8d 100644 --- a/litellm/proxy/litellm_pre_call_utils.py +++ b/litellm/proxy/litellm_pre_call_utils.py @@ -3188,9 +3188,9 @@ def _fallback_lists( ) -> tuple[Sequence[object], ...]: """Every fallback list the router can route this request through: the three router-level lists, the same three keys from the request body (which replace - the router's when present, so the union is the safe superset), and the - key/team ``router_settings.fallbacks`` the proxy prefers over router-level - fallbacks in ``_configured_fallbacks``.""" + the router's when present, so the union is the safe superset), and the same + three keys from key/team ``router_settings``, which the proxy forwards into + the request after this discovery runs.""" lists: Final[list[Sequence[object]]] = [] # mutable-ok: accumulates each fallback list found configured: Final = llm_router.settings() for router_list in ( @@ -3205,11 +3205,12 @@ def _fallback_lists( if isinstance(entries, list): lists.append(cast("Sequence[object]", entries)) # cast-ok: request fallback lists hold mixed entry shapes for settings in (router_settings, team_router_settings): - key_fallbacks: object = settings.get("fallbacks") if settings is not None else None - if isinstance(key_fallbacks, list): - lists.append( - cast("Sequence[object]", key_fallbacks) # cast-ok: settings fallback lists hold mixed entry shapes - ) + for key in _REQUEST_FALLBACK_KEYS: + key_fallbacks: object = settings.get(key) if settings is not None else None + if isinstance(key_fallbacks, list): + lists.append( + cast("Sequence[object]", key_fallbacks) # cast-ok: settings fallback lists hold mixed entry shapes + ) return tuple(lists) diff --git a/litellm/proxy/management_endpoints/fallback_management_endpoints.py b/litellm/proxy/management_endpoints/fallback_management_endpoints.py index d688aac0ed8..477cf44073f 100644 --- a/litellm/proxy/management_endpoints/fallback_management_endpoints.py +++ b/litellm/proxy/management_endpoints/fallback_management_endpoints.py @@ -14,10 +14,12 @@ import json from collections.abc import Mapping, Sequence from typing import TYPE_CHECKING, Annotated, Final, Literal -from pydantic import Field, JsonValue, TypeAdapter +from pydantic import Field, JsonValue, TypeAdapter, ValidationError +import litellm from litellm._logging import verbose_proxy_logger from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy.auth.auth_utils import fallback_target_model_name from litellm.proxy.auth.model_checks import get_all_fallbacks from litellm.proxy.auth.user_api_key_auth import user_api_key_auth from litellm.proxy.utils import PrismaClient, invalidate_config_param @@ -26,6 +28,7 @@ if TYPE_CHECKING: from fastapi import APIRouter, Depends, HTTPException, status from litellm.proxy.proxy_server import ProxyConfig + from litellm.router import Router else: try: from fastapi import APIRouter, Depends, HTTPException, status @@ -41,6 +44,7 @@ from litellm.types.management_endpoints.router_settings_endpoints import ( FallbackGetResponse, FallbackResponse, ) +from litellm.types.router import ContentPolicyFallbackRule, ContentPolicyFallbackRules, FallbackTargets router: Final = APIRouter() @@ -50,12 +54,83 @@ StoredFallback = Annotated[FallbackRule | dict[str, object] | str, Field(union_m _STORED_FALLBACKS: Final = TypeAdapter(list[StoredFallback]) _JSON_RULES: Final = TypeAdapter(list[JsonValue]) _STORED_ROUTER_SETTINGS: Final = TypeAdapter(dict[str, JsonValue]) +_OPTIONAL_JSON_RULES: Final[TypeAdapter[Sequence[JsonValue] | None]] = TypeAdapter(Sequence[JsonValue] | None) +_CONTENT_POLICY_FALLBACK_RULE: Final[TypeAdapter[ContentPolicyFallbackRule]] = TypeAdapter(ContentPolicyFallbackRule) +_FALLBACK_TARGETS: Final[TypeAdapter[FallbackTargets]] = TypeAdapter(FallbackTargets) def _rule_covers(entry: StoredFallback, model: str) -> bool: return isinstance(entry, dict) and model in entry +def _layered_over_config_rules( + runtime_rules: ContentPolicyFallbackRules, config_rules: Sequence[Mapping[str, JsonValue]] +) -> Sequence[Mapping[str, object]]: + covered: Final = frozenset[str]().union(*runtime_rules) + return [*runtime_rules, *(rule for rule in config_rules if covered.isdisjoint(rule))] + + +def _config_file_rules(rules: object) -> Sequence[Mapping[str, JsonValue]]: + try: + candidates: Final = _OPTIONAL_JSON_RULES.validate_python(rules) or () + except ValidationError: + return () + return tuple(rule for rule in candidates if isinstance(rule, Mapping)) + + +def _log_ignored_content_policy_rules(rules: object, error: ValidationError) -> None: + verbose_proxy_logger.error( + "Ignoring invalid content_policy_fallbacks %r from the DB, every other rule still applies. " + "Fix it through POST /fallback or /config/update: %s", + rules, + error, + ) + + +def _checked_content_policy_rule(rule: JsonValue) -> ContentPolicyFallbackRule | ValidationError: + try: + return _CONTENT_POLICY_FALLBACK_RULE.validate_python(rule) + except ValidationError as invalid_rule: + return invalid_rule + + +def _checked_content_policy_rules( + rules: object, +) -> Sequence[tuple[object, ContentPolicyFallbackRule | ValidationError]]: + try: + candidates: Final = _OPTIONAL_JSON_RULES.validate_python(rules) or () + except ValidationError as not_a_rule_list: + return ((rules, not_a_rule_list),) + return tuple((candidate, _checked_content_policy_rule(candidate)) for candidate in candidates) + + +def apply_content_policy_fallbacks(llm_router: "Router", rules: object, *, log_skipped_rules: bool) -> None: + checked_rules: Final = _checked_content_policy_rules(rules) + runtime_rules: Final = tuple(result for _, result in checked_rules if not isinstance(result, ValidationError)) + layered_rules: Final = ( + _layered_over_config_rules( + runtime_rules, + _config_file_rules( + litellm.content_policy_fallbacks # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType] # untyped module list + ), + ) + or None + ) + for rule, result in checked_rules: + if log_skipped_rules and isinstance(result, ValidationError): + _log_ignored_content_policy_rules(rule, result) + if layered_rules == llm_router.settings().content_policy_fallbacks: # pyright: ignore[reportUnknownMemberType] # untyped RouterSettings list + return + llm_router.apply(ChangeSet(settings={"content_policy_fallbacks": layered_rules})) + + +def _apply_fallback_rules(llm_router: "Router", fallback_key: str, rules: Sequence[StoredFallback]) -> None: + if fallback_key == "content_policy_fallbacks": + apply_content_policy_fallbacks(llm_router, rules, log_skipped_rules=True) + return + llm_router.apply(ChangeSet(assign={fallback_key: rules})) + + async def _reload_config_file_ownership(proxy_config: "ProxyConfig") -> None: await proxy_config.get_config() @@ -83,6 +158,7 @@ def _editable_rules( async def _write_rules( prisma_client: PrismaClient, proxy_config: "ProxyConfig", + llm_router: "Router", stored_settings: Mapping[str, JsonValue], fallback_key: str, rules: Sequence[StoredFallback], @@ -91,8 +167,10 @@ async def _write_rules( section_name=ROUTER_SETTINGS_PARAM, changed_keys={fallback_key: _JSON_RULES.validate_python(rules)} ) if proxy_config.router_settings.owned_by_config(fallback_key): + llm_router.apply(ChangeSet(assign={fallback_key: rules})) return await _persist_router_settings(prisma_client, {**stored_settings, fallback_key: rules}) + _apply_fallback_rules(llm_router, fallback_key, rules) async def _persist_router_settings(prisma_client: PrismaClient, router_settings: Mapping[str, object]) -> None: @@ -218,10 +296,7 @@ async def create_fallback( # Add new fallback existing_fallbacks.append({data.model: data.fallback_models}) - await _write_rules(prisma_client, proxy_config, router_settings, fallback_key, existing_fallbacks) - - # Update the in-memory router configuration - llm_router.apply(ChangeSet(assign={fallback_key: existing_fallbacks})) + await _write_rules(prisma_client, proxy_config, llm_router, router_settings, fallback_key, existing_fallbacks) verbose_proxy_logger.info( "Fallback configured: %s -> %s (type: %s)", data.model, data.fallback_models, data.fallback_type @@ -277,7 +352,10 @@ async def get_fallback( ) # Get fallbacks using the existing utility function - fallback_models: Final = get_all_fallbacks(model=model, llm_router=llm_router, fallback_type=fallback_type) + fallback_targets: Final = _FALLBACK_TARGETS.validate_python( + get_all_fallbacks(model=model, llm_router=llm_router, fallback_type=fallback_type) + ) + fallback_models: Final = [fallback_target_model_name(target) or model for target in fallback_targets] if not fallback_models: raise HTTPException( @@ -367,10 +445,7 @@ async def delete_fallback( detail={"error": f"No {fallback_type} fallbacks configured for model '{model}'"}, ) - await _write_rules(prisma_client, proxy_config, router_settings, fallback_key, updated_fallbacks) - - # Update the in-memory router configuration - llm_router.apply(ChangeSet(assign={fallback_key: updated_fallbacks})) + await _write_rules(prisma_client, proxy_config, llm_router, router_settings, fallback_key, updated_fallbacks) verbose_proxy_logger.info("Fallback deleted: %s (type: %s)", model, fallback_type) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index ae0e7fce667..f078267c6b2 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -718,6 +718,9 @@ from litellm.proxy.management_endpoints.customer_endpoints import ( router as customer_router, ) from litellm.proxy.management_endpoints.daily_activity_routes import router as daily_activity_router +from litellm.proxy.management_endpoints.fallback_management_endpoints import ( + apply_content_policy_fallbacks, +) from litellm.proxy.management_endpoints.fallback_management_endpoints import ( router as fallback_management_router, ) @@ -8163,14 +8166,38 @@ class ProxyConfig: if db_router_settings is not None and db_router_settings.param_value is not None else _EMPTY_SETTINGS_MAPPING ) + previous_content_policy_rules: Final = self.router_settings.resolved().get("content_policy_fallbacks") self.router_settings.apply_db_row("router_settings", db_values) combined_router_settings: Final = self.router_settings.resolved() if combined_router_settings: - self._apply_router_settings(llm_router, combined_router_settings) + self._apply_router_settings( + llm_router, + combined_router_settings, + content_policy_owned_by_config=self.router_settings.owned_by_config("content_policy_fallbacks"), + content_policy_rules_changed=( + combined_router_settings.get("content_policy_fallbacks") != previous_content_policy_rules + ), + ) @staticmethod - def _apply_router_settings(llm_router: Router, router_settings: Mapping[str, object]) -> None: - llm_router.apply(ChangeSet(settings={k: v for k, v in router_settings.items() if k != "routing_groups"})) + def _apply_router_settings( + llm_router: Router, + router_settings: Mapping[str, object], + *, + content_policy_owned_by_config: bool, + content_policy_rules_changed: bool, + ) -> None: + llm_router.apply( + ChangeSet( + settings={ + k: v for k, v in router_settings.items() if k not in ("routing_groups", "content_policy_fallbacks") + } + ) + ) + if "content_policy_fallbacks" in router_settings and not content_policy_owned_by_config: + apply_content_policy_fallbacks( + llm_router, router_settings["content_policy_fallbacks"], log_skipped_rules=content_policy_rules_changed + ) if "routing_groups" not in router_settings: return try: diff --git a/litellm/types/proxy/management_endpoints/key_requests.py b/litellm/types/proxy/management_endpoints/key_requests.py index 172357d8ac4..b49cb63c8e0 100644 --- a/litellm/types/proxy/management_endpoints/key_requests.py +++ b/litellm/types/proxy/management_endpoints/key_requests.py @@ -1,14 +1,14 @@ import enum -from collections.abc import Mapping +from collections.abc import Iterator, Mapping from datetime import datetime from types import MappingProxyType -from typing import Any, Literal # noqa: TID251 # allowed_routes and metadata stay Any for existing callers +from typing import Any, Final, Literal # noqa: TID251 # allowed_routes and metadata stay Any for existing callers from pydantic import Field, PositiveInt, field_validator, model_validator from typing_extensions import TypeIs # noqa: TID251 # narrows the raw payload a before-validator receives from litellm.types.proxy.management_endpoints.request_base import GenerateRequestBase -from litellm.types.router import UpdateRouterConfig +from litellm.types.router import ContentPolicyFallbackRules, UpdateRouterConfig from litellm.types.utils import LiteLLMPydanticObjectBase @@ -17,6 +17,12 @@ class AllowedVectorStoreIndexItem(LiteLLMPydanticObjectBase): index_permissions: list[Literal["read", "write"]] +def _fallback_targets(rules: ContentPolicyFallbackRules) -> Iterator[object]: + for rule in rules: + for targets in rule.values(): + yield from targets + + class KeyRequestBase(GenerateRequestBase): key: str | None = None tpd_limit: int | None = None @@ -42,6 +48,18 @@ class KeyRequestBase(GenerateRequestBase): router_settings: UpdateRouterConfig | None = None access_group_ids: list[str] | None = None + @field_validator("router_settings") + @classmethod + def content_policy_fallbacks_name_models( + cls, router_settings: UpdateRouterConfig | None + ) -> UpdateRouterConfig | None: + rules: Final = None if router_settings is None else router_settings.content_policy_fallbacks + if rules is not None and not all(isinstance(target, str) for target in _fallback_targets(rules)): + raise ValueError( + 'A key\'s content_policy_fallbacks targets must be model names, e.g. [{"gpt-4": ["gpt-4o"]}]' + ) + return router_settings + class LiteLLMKeyType(str, enum.Enum): """ diff --git a/litellm/types/router.py b/litellm/types/router.py index f927ae0475b..726a328e913 100644 --- a/litellm/types/router.py +++ b/litellm/types/router.py @@ -192,6 +192,11 @@ OptionalPreCallChecks = list[ ] +FallbackTargets: TypeAlias = Sequence[str | Mapping[str, JsonValue]] +ContentPolicyFallbackRule: TypeAlias = Annotated[Mapping[str, FallbackTargets], Field(min_length=1, max_length=1)] +ContentPolicyFallbackRules: TypeAlias = Sequence[ContentPolicyFallbackRule] + + class UpdateRouterConfig(LiteLLMBaseModel): """ Set of params that you can modify via `router.update_settings()`. @@ -211,6 +216,7 @@ class UpdateRouterConfig(LiteLLMBaseModel): retry_after: float | None = None fallbacks: list[dict] | None = None context_window_fallbacks: list[dict] | None = None + content_policy_fallbacks: ContentPolicyFallbackRules | None = None model_group_alias: dict[str, str | dict] | None = Field(default={}) enable_tag_filtering: bool | None = None weights: RouterWeights | None = None diff --git a/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py index b0ef7e64e60..a0c32f2235f 100644 --- a/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py @@ -13,7 +13,7 @@ import pytest import yaml from fastapi import HTTPException from fastapi.testclient import TestClient -from pydantic import JsonValue +from pydantic import JsonValue, ValidationError import litellm from litellm.constants import PROXY_DB_LOOKUP_STALL_WINDOW_SECONDS @@ -7597,6 +7597,7 @@ async def test_generate_key_with_router_settings(monkeypatch): "num_retries": 3, "model_group_retry_policy": {"gpt-4": {"RateLimitErrorRetries": 5}}, "weights": {"gpt-4": {"weighted-id": 1}}, + "content_policy_fallbacks": [{"gpt-4": ["gpt-4o", "gpt-4o-mini"]}], } request_data = GenerateKeyRequest( @@ -7736,6 +7737,19 @@ async def test_update_key_with_router_settings( assert json.loads(cleared["router_settings"]) == {} +@pytest.mark.parametrize("request_type", [GenerateKeyRequest, UpdateKeyRequest]) +@pytest.mark.parametrize("dict_target", [{"model": "gpt-4o-mini"}, {}]) +def test_a_key_content_policy_rule_can_only_name_fallback_models( + request_type: type[GenerateKeyRequest] | type[UpdateKeyRequest], + dict_target: dict[str, str], +): + with pytest.raises(ValidationError, match="must be model names"): + request_type( + key="sk-key", + router_settings={"content_policy_fallbacks": [{"gpt-4": ["gpt-4o", dict_target]}]}, + ) + + @pytest.mark.asyncio async def test_validate_max_budget(): """ diff --git a/tests/unit/proxy/test_fallback_management_endpoints.py b/tests/unit/proxy/test_fallback_management_endpoints.py index d28d61af72b..e3d0d072e0a 100644 --- a/tests/unit/proxy/test_fallback_management_endpoints.py +++ b/tests/unit/proxy/test_fallback_management_endpoints.py @@ -12,7 +12,7 @@ import copy import json from collections.abc import AsyncIterator, Mapping from types import MappingProxyType, SimpleNamespace -from typing import Final, get_args +from typing import Final, Literal, get_args from unittest.mock import AsyncMock, MagicMock, patch import pytest @@ -20,6 +20,7 @@ from fastapi import HTTPException from pydantic import JsonValue, TypeAdapter from redis.exceptions import NoPermissionError +import litellm from litellm import Router from litellm.proxy.config_resolvers.settings_rules import Section from litellm.proxy.management_endpoints.fallback_management_endpoints import ( @@ -388,6 +389,88 @@ class TestFallbackWritesSeeTheLatestStoredRules: assert stored.stored_fallbacks() == [GATEWAY_RULE, TEAM_RULE] + async def test_content_policy_writes_keep_the_litellm_settings_rules_for_other_models( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + monkeypatch.setattr(litellm, "content_policy_fallbacks", [TEAM_RULE]) + router: Final = _team_router() + stored: Final = _StoredRouterSettings(None) + with ( + patch("litellm.proxy.proxy_server.llm_router", router), + patch("litellm.proxy.proxy_server.prisma_client", stored.prisma_client), + patch("litellm.proxy.proxy_server.proxy_config", stored.proxy_config), + patch("litellm.proxy.proxy_server.store_model_in_db", True), + ): + await create_fallback( + FallbackCreateRequest( + model="gpt-5.4-mini", fallback_models=["team-fallback"], fallback_type="content_policy" + ), + MagicMock(), + ) + after_create: Final = await get_fallback("team-primary", "content_policy", MagicMock()) + await delete_fallback("gpt-5.4-mini", "content_policy", MagicMock()) + after_delete: Final = await get_fallback("team-primary", "content_policy", MagicMock()) + + assert after_create.fallback_models == ["team-fallback"] + assert after_delete.fallback_models == ["team-fallback"] + + async def test_a_content_policy_write_applies_beside_an_invalid_stored_rule(self) -> None: + router: Final = _team_router() + stored: Final = _StoredRouterSettings({"content_policy_fallbacks": [MALFORMED_GATEWAY_RULE]}) + with ( + patch("litellm.proxy.proxy_server.llm_router", router), + patch("litellm.proxy.proxy_server.prisma_client", stored.prisma_client), + patch("litellm.proxy.proxy_server.proxy_config", stored.proxy_config), + patch("litellm.proxy.proxy_server.store_model_in_db", True), + ): + await create_fallback( + FallbackCreateRequest( + model="team-primary", fallback_models=["team-fallback"], fallback_type="content_policy" + ), + MagicMock(), + ) + after_create: Final = await get_fallback("team-primary", "content_policy", MagicMock()) + + assert after_create.fallback_models == ["team-fallback"] + + @pytest.mark.parametrize( + ("fallback_type", "router_setting"), + [("general", "fallbacks"), ("content_policy", "content_policy_fallbacks")], + ) + async def test_get_names_the_model_of_a_dict_target( + self, fallback_type: Literal["general", "content_policy"], router_setting: str + ) -> None: + router: Final = _team_router() + router.update_settings( + **{router_setting: [{"team-primary": [{"model": "gpt-5.4-mini", "api_key": "sk-test"}, "team-fallback"]}]} + ) + with patch("litellm.proxy.proxy_server.llm_router", router): + response: Final = await get_fallback("team-primary", fallback_type, MagicMock()) + + assert response.fallback_models == ["gpt-5.4-mini", "team-fallback"] + + @pytest.mark.parametrize( + ("fallback_type", "router_setting"), + [("general", "fallbacks"), ("content_policy", "content_policy_fallbacks")], + ) + async def test_get_names_the_rules_own_model_for_a_dict_target_that_names_none( + self, fallback_type: Literal["general", "content_policy"], router_setting: str + ) -> None: + router: Final = _team_router() + router.update_settings(**{router_setting: [{"team-primary": [{"temperature": 0}, "team-fallback"]}]}) + with patch("litellm.proxy.proxy_server.llm_router", router): + response: Final = await get_fallback("team-primary", fallback_type, MagicMock()) + + assert response.fallback_models == ["team-primary", "team-fallback"] + + async def test_get_answers_an_error_for_a_string_valued_rule_instead_of_its_characters(self) -> None: + router: Final = _team_router() + router.update_settings(fallbacks=[{"team-primary": "team-fallback"}]) + with patch("litellm.proxy.proxy_server.llm_router", router), pytest.raises(HTTPException) as raised: + await get_fallback("team-primary", "general", MagicMock()) + + assert raised.value.status_code == 500 + async def test_a_create_keeps_a_rule_another_instance_stored_since_this_one_last_read(self) -> None: stored: Final = _StoredRouterSettings({}) await get_config_param(stored.prisma_client, "router_settings") @@ -568,6 +651,21 @@ class TestFallbackWritesHonourTheConfigFile: assert stored.row is None assert router.fallbacks == [CONFIG_RULE, CONFIG_WINDOW_RULE] + async def test_a_content_policy_restate_keeps_the_config_list_without_the_litellm_settings_rules( + self, router: Router, monkeypatch: pytest.MonkeyPatch + ) -> None: + monkeypatch.setattr(litellm, "content_policy_fallbacks", [TEAM_RULE]) + stored: Final = _StoredRouterSettings(None, config_file={"content_policy_fallbacks": [CONFIG_RULE]}) + request: Final = FallbackCreateRequest( + model="gpt-5.4-mini", fallback_models=["team-fallback"], fallback_type="content_policy" + ) + + response: Final = await self._create(request, router, stored) + + assert response.fallback_models == ["team-fallback"] + assert stored.row is None + assert router.content_policy_fallbacks == [CONFIG_RULE] + async def test_a_create_that_restates_a_config_rule_next_to_an_env_reference_writes_nothing( self, router: Router, monkeypatch: pytest.MonkeyPatch ) -> None: diff --git a/tests/unit/proxy/test_litellm_pre_call_utils.py b/tests/unit/proxy/test_litellm_pre_call_utils.py index 2edb032b048..c5136ff6e4a 100644 --- a/tests/unit/proxy/test_litellm_pre_call_utils.py +++ b/tests/unit/proxy/test_litellm_pre_call_utils.py @@ -10143,9 +10143,12 @@ class TestResolveUserProviderCredentials: assert data["secret_fields"]["user_provider_credentials_user_id"] == "user-a" @pytest.mark.asyncio - async def test_key_router_settings_fallback_loads_per_user_tokens(self, monkeypatch): - """The proxy prefers key/team router_settings.fallbacks over router-level - fallbacks (_configured_fallbacks), so discovery must read them too.""" + @pytest.mark.parametrize("fallback_key", ["fallbacks", "context_window_fallbacks", "content_policy_fallbacks"]) + async def test_key_router_settings_fallback_loads_per_user_tokens( + self, monkeypatch: pytest.MonkeyPatch, fallback_key: str + ): + """The proxy forwards key/team router_settings fallback lists into the + request after discovery, so discovery must read every one of them.""" from litellm.proxy.litellm_pre_call_utils import _resolve_user_provider_credentials_for_request table = self._env(monkeypatch) @@ -10175,7 +10178,7 @@ class TestResolveUserProviderCredentials: authenticated_user_id="user-a", team_id=None, llm_router=router, - router_settings={"fallbacks": [{"gpt-4o": [{"model": "copilot-chat"}]}]}, + router_settings={fallback_key: [{"gpt-4o": [{"model": "copilot-chat"}]}]}, ) table.find_many.assert_awaited() assert data["secret_fields"]["user_provider_credentials_user_id"] == "user-a" diff --git a/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py b/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py index 62a90109d26..bc3a228321a 100644 --- a/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py +++ b/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py @@ -10,6 +10,7 @@ import subprocess import sys import time import types +from collections.abc import Callable from datetime import datetime, timedelta, timezone from pathlib import Path from typing import Final @@ -5583,6 +5584,215 @@ async def test_valid_db_routing_groups_still_replace_router_groups(): assert router._get_routing_context("m2", None)[0] == "least-busy" +def _content_policy_router(**router_kwargs: object) -> litellm.Router: + return litellm.Router( + model_list=[ + { + "model_name": "primary", + "litellm_params": {"model": "openai/gpt-5.4-mini", "api_key": "fake", "mock_response": "primary answered"}, + }, + { + "model_name": "secondary", + "litellm_params": {"model": "openai/gpt-5.4-mini", "api_key": "fake", "mock_response": "secondary answered"}, + }, + { + "model_name": "backup", + "litellm_params": {"model": "openai/gpt-5.4-nano", "api_key": "fake", "mock_response": "backup answered"}, + }, + ], + num_retries=0, + **router_kwargs, + ) + + +def _router_settings_row(content_policy_fallbacks: object) -> MagicMock: + row: Final = MagicMock() + row.param_value = {"allowed_fails": 7, "content_policy_fallbacks": content_policy_fallbacks} + prisma: Final = MagicMock() + prisma.db.litellm_config.find_first = AsyncMock(return_value=row) + return prisma + + +async def _answer_after_content_policy_refusal(router: litellm.Router, model: str = "primary") -> str | None: + response: Final = await router.acompletion( + model=model, messages=[{"role": "user", "content": "hi"}], mock_testing_content_policy_fallbacks=True + ) + return response.choices[0].message.content + + +@pytest.mark.asyncio +@pytest.mark.parametrize("stored_target", ["backup", {"model": "backup"}]) +async def test_db_content_policy_fallbacks_reach_the_router(stored_target: str | dict[str, str]): + router: Final = _content_policy_router() + + await proxy_server_module.ProxyConfig()._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row([{"primary": [stored_target]}]) + ) + + assert await _answer_after_content_policy_refusal(router) == "backup answered" + + +@pytest.mark.asyncio +@pytest.mark.parametrize("stored_rules", [[], [{"primary": "backup"}]]) +async def test_a_saved_content_policy_list_with_no_valid_rule_leaves_refusals_to_the_general_fallbacks( + monkeypatch: pytest.MonkeyPatch, stored_rules: list[object] +): + monkeypatch.setattr(litellm, "content_policy_fallbacks", None) + router: Final = _content_policy_router(fallbacks=[{"primary": ["backup"]}]) + + await proxy_server_module.ProxyConfig()._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row(stored_rules) + ) + + assert await _answer_after_content_policy_refusal(router) == "backup answered" + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("litellm_settings_rules", "stored_rules", "refused_model"), + [ + ([{"secondary": ["backup"]}], [{"primary": ["backup"]}], "secondary"), + ([{"*": ["secondary"]}], [{"*": ["backup"]}], "primary"), + ([{"primary": ["backup"]}], [], "primary"), + ], +) +async def test_db_content_policy_fallbacks_layer_over_the_litellm_settings_rules( + monkeypatch: pytest.MonkeyPatch, + litellm_settings_rules: list[dict[str, list[str]]], + stored_rules: list[dict[str, list[str]]], + refused_model: str, +): + monkeypatch.setattr(litellm, "content_policy_fallbacks", litellm_settings_rules) + router: Final = _content_policy_router() + + await proxy_server_module.ProxyConfig()._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row(stored_rules) + ) + + assert await _answer_after_content_policy_refusal(router, refused_model) == "backup answered" + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "invalid_rule", + [{"secondary": "backup"}, ["backup"], dict[str, list[str]](), {"secondary": ["backup"], "primary": ["secondary"]}], +) +async def test_an_invalid_db_content_policy_rule_skips_only_itself(monkeypatch: pytest.MonkeyPatch, invalid_rule: object): + monkeypatch.setattr(litellm, "content_policy_fallbacks", [{"primary": ["backup"]}]) + router: Final = _content_policy_router() + + await proxy_server_module.ProxyConfig()._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row([invalid_rule, {"secondary": ["backup"]}]) + ) + + assert router.allowed_fails == 7 + assert await _answer_after_content_policy_refusal(router) == "backup answered" + assert await _answer_after_content_policy_refusal(router, "secondary") == "backup answered" + + +@pytest.mark.asyncio +async def test_a_db_content_policy_value_that_is_not_a_list_keeps_the_litellm_settings_rules( + monkeypatch: pytest.MonkeyPatch, +): + monkeypatch.setattr(litellm, "content_policy_fallbacks", [{"primary": ["backup"]}]) + router: Final = _content_policy_router() + + await proxy_server_module.ProxyConfig()._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row("oops") + ) + + assert router.allowed_fails == 7 + assert await _answer_after_content_policy_refusal(router) == "backup answered" + + +@pytest.mark.asyncio +async def test_a_db_sync_keeps_a_litellm_settings_rule_with_a_target_router_init_accepted( + monkeypatch: pytest.MonkeyPatch, +): + monkeypatch.setattr(litellm, "content_policy_fallbacks", [{"primary": ["backup", 5]}]) + router: Final = _content_policy_router() + + await proxy_server_module.ProxyConfig()._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row([{"secondary": ["backup"]}]) + ) + + assert await _answer_after_content_policy_refusal(router) == "backup answered" + assert await _answer_after_content_policy_refusal(router, "secondary") == "backup answered" + + +@pytest.mark.asyncio +async def test_an_invalid_db_content_policy_rule_is_logged_once_per_change_of_the_rules( + monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture +): + monkeypatch.setattr(litellm, "content_policy_fallbacks", []) + router: Final = _content_policy_router() + proxy_config: Final = proxy_server_module.ProxyConfig() + stored_rules: Final = [{"secondary": "backup"}, {"primary": ["backup"]}] + changed_rules: Final = [{"secondary": "backup"}, {"secondary": ["backup"]}] + + with caplog.at_level(logging.ERROR, logger="LiteLLM Proxy"): + await proxy_config._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row(stored_rules) + ) + await proxy_config._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row(stored_rules) + ) + await proxy_config._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row(changed_rules) + ) + + notices: Final = [ + record for record in caplog.records if "Ignoring invalid content_policy_fallbacks" in record.getMessage() + ] + assert len(notices) == 2 + assert await _answer_after_content_policy_refusal(router, "secondary") == "backup answered" + + +@pytest.mark.asyncio +async def test_a_db_sync_logs_a_new_invalid_content_policy_rule_when_no_valid_rule_changes( + monkeypatch: pytest.MonkeyPatch, caplog: pytest.LogCaptureFixture +): + monkeypatch.setattr(litellm, "content_policy_fallbacks", []) + router: Final = _content_policy_router() + proxy_config: Final = proxy_server_module.ProxyConfig() + + with caplog.at_level(logging.ERROR, logger="LiteLLM Proxy"): + await proxy_config._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row([{"secondary": "backup"}]) + ) + await proxy_config._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row([{"secondary": "backup"}, {"primary": "backup"}]) + ) + + notices: Final = [ + record.getMessage() + for record in caplog.records + if "Ignoring invalid content_policy_fallbacks" in record.getMessage() + ] + assert len(notices) == 3 + assert "{'primary': 'backup'}" in notices[-1] + + +@pytest.mark.asyncio +async def test_config_file_content_policy_fallbacks_stay_the_only_rules_after_a_db_sync( + monkeypatch: pytest.MonkeyPatch, +): + monkeypatch.setattr(litellm, "content_policy_fallbacks", [{"secondary": ["backup"]}]) + file_rules: Final = [{"primary": ["backup"]}] + router: Final = _content_policy_router(content_policy_fallbacks=file_rules) + proxy_config: Final = proxy_server_module.ProxyConfig() + proxy_config.router_settings.load_yaml({"content_policy_fallbacks": file_rules}) + + await proxy_config._add_router_settings_from_db_config( + llm_router=router, prisma_client=_router_settings_row([{"secondary": ["backup"]}]) + ) + + assert router.allowed_fails == 7 + assert await _answer_after_content_policy_refusal(router) == "backup answered" + with pytest.raises(litellm.ContentPolicyViolationError): + await _answer_after_content_policy_refusal(router, "secondary") + + @pytest.mark.asyncio async def test_add_router_settings_from_db_config_empty_db_lists_do_not_clobber_config_fallbacks(): """ @@ -5591,6 +5801,7 @@ async def test_add_router_settings_from_db_config_empty_db_lists_do_not_clobber_ empty lists are "no value" and must not clobber config.yaml fallbacks, matching _deep_merge_dicts semantics. Non-empty DB lists still win. """ + from collections import ChainMap from unittest.mock import AsyncMock, MagicMock from litellm.proxy.proxy_server import ProxyConfig @@ -5625,10 +5836,10 @@ async def test_add_router_settings_from_db_config_empty_db_lists_do_not_clobber_ prisma_client=mock_prisma_client, ) - combined_settings = mock_router.update_settings.call_args.kwargs + combined_settings = ChainMap(*(call.kwargs for call in reversed(mock_router.update_settings.call_args_list))) assert combined_settings["fallbacks"] == [{"gpt-oss-120b": ["granite-4-h-small"]}] assert combined_settings["context_window_fallbacks"] == [{"gpt-oss-120b": ["granite-4-h-small"]}] - assert combined_settings["content_policy_fallbacks"] == [{"gpt-oss-120b": ["granite-4-h-small"]}] + assert combined_settings.get("content_policy_fallbacks") != [{"gpt-oss-120b": ["other-model"]}] assert combined_settings["num_retries"] == 3 @@ -9997,6 +10208,34 @@ def test_update_config_accepts_disjoint_routing_groups(_update_config_setup): restore() +def test_update_config_validates_content_policy_fallbacks_like_fallbacks( + _update_config_setup: Callable[..., tuple[TestClient, _FakePrismaClient, Callable[[], None]]], +): + client, prisma, restore = _update_config_setup(initial_rows={"router_settings": {"num_retries": 2}}) + rule: Final = [{"gpt-5.4-mini": ["gpt-5.4-nano", {"model": "gpt-5.4-nano", "temperature": 1}]}] + malformed: Final = ( + "oops", + 5, + [1, 2], + [{"gpt-5.4-mini": "gpt-5.4-nano"}], + [{"gpt-5.4-mini": ["gpt-5.4-nano", 7]}], + [dict[str, list[str]]()], + [{"gpt-5.4-mini": ["gpt-5.4-nano"], "gpt-5.4": ["gpt-5.4-nano"]}], + ) + try: + rejected: Final = tuple( + client.post("/config/update", json={"router_settings": {"content_policy_fallbacks": value}}).status_code + for value in malformed + ) + assert rejected == (422,) * len(malformed) + assert prisma.db.litellm_config.rows["router_settings"] == {"num_retries": 2} + accepted: Final = client.post("/config/update", json={"router_settings": {"content_policy_fallbacks": rule}}) + assert accepted.status_code == 200 + assert prisma.db.litellm_config.rows["router_settings"] == {"num_retries": 2, "content_policy_fallbacks": rule} + finally: + restore() + + def test_update_config_env_var_round_trip_not_double_encrypted(_update_config_setup, monkeypatch): """Endpoint-level regression for the /config/update double-encryption bug. diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index 74959155380..1ac44b5976d 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -2731,6 +2731,39 @@ def test_update_settings_model_group_alias_drops_cached_group_info(): assert after.input_cost_per_token is not None and after.input_cost_per_token > 0 +@pytest.mark.parametrize( + "fallback_kind, mock_testing_param", + [ + ("fallbacks", "mock_testing_fallbacks"), + ("context_window_fallbacks", "mock_testing_context_fallbacks"), + ("content_policy_fallbacks", "mock_testing_content_policy_fallbacks"), + ], +) +async def test_update_settings_fallback_rule_of_every_kind_is_used_on_its_error( + fallback_kind: str, mock_testing_param: str +): + router: Final = Router( + model_list=[ + { + "model_name": "primary", + "litellm_params": {"model": "gpt-5.4-mini", "api_key": "fake", "mock_response": "primary answered"}, + }, + { + "model_name": "backup", + "litellm_params": {"model": "gpt-5.4-nano", "api_key": "fake", "mock_response": "backup answered"}, + }, + ], + num_retries=0, + ) + + router.update_settings(**{fallback_kind: [{"primary": ["backup"]}]}) + + response: Final = await router.acompletion( + model="primary", messages=[{"role": "user", "content": "hi"}], **{mock_testing_param: True} + ) + assert response.choices[0].message.content == "backup answered" + + _PAID_INPUT_COST_PER_TOKEN: Final = 3e-06 _PAID_OUTPUT_COST_PER_TOKEN: Final = 1.5e-05 diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index f53e57c2f18..1aec7d5b761 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -49017,6 +49017,12 @@ export interface components { UpdateRouterConfig: { /** Allowed Fails */ allowed_fails?: number | null; + /** Content Policy Fallbacks */ + content_policy_fallbacks?: { + [key: string]: (string | { + [key: string]: components["schemas"]["JsonValue"]; + })[]; + }[] | null; /** Context Window Fallbacks */ context_window_fallbacks?: { [key: string]: unknown; From 02ef44c3885b2b0500cf267abf57064fbd206dcd Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 30 Oct 2026 04:41:02 -0700 Subject: [PATCH 21/23] fix(streaming): convert upstream logprobs so streams with logprobs end with their finish chunk (#33549) * fix(streaming): normalize provider logprobs to prevent MockValSer serialization crash * fix(streaming): convert upstream chunk logprobs to litellm ChoiceLogprobs so the finish chunk serializes A streaming chat completion with logprobs on the OpenAI SDK path attached the SDK's own ChoiceLogprobs object to the outgoing chunk when the chunk had no content to rebuild from, which is the provider's finish chunk (empty delta, logprobs for the end token, finish_reason). The SDK parses stream chunks without building that nested class, so its pydantic serializer is still a MockValSer and model_dump_json fails. The proxy then sent an error frame in place of the finish chunk, and when the same chunk also carried usage the SDK wrapper itself raised MidStreamFallbackError. The attach site now validates upstream logprobs through a TypeAdapter that dumps any non-litellm pydantic model before validating into litellm's ChoiceLogprobs, and passes litellm instances through unchanged. * fix(streaming): keep upstream logprobs that don't fit ChoiceLogprobs as plain JSON so the stream still finishes * fix(streaming): build ChoiceLogprobs' serializer up front and keep nonconforming logprobs as JSON on chunk rebuilds ChoiceLogprobs inherits the openai SDK's deferred build, so a fresh worker whose first logprobs arrive on the finish chunk still had no serializer. Foreign logprobs models now dump in full, so a missing top_logprobs becomes [] and SDK callers keep attribute access, and chunk rebuilds for content and usage keep logprobs that don't fit as plain JSON instead of raising. * fix(streaming): keep only the logprobs keys upstream sent and type refusal logprobs * fix(types): pass nonconforming non-streaming logprobs through instead of failing the response --------- Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../litellm_core_utils/streaming_handler.py | 8 +- litellm/types/utils.py | 30 +- .../test_convert_dict_to_response.py | 39 ++ .../test_streaming_handler.py | 354 ++++++++++++++++++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 4 +- 5 files changed, 428 insertions(+), 7 deletions(-) diff --git a/litellm/litellm_core_utils/streaming_handler.py b/litellm/litellm_core_utils/streaming_handler.py index 5542249d2c7..682b2a2f538 100644 --- a/litellm/litellm_core_utils/streaming_handler.py +++ b/litellm/litellm_core_utils/streaming_handler.py @@ -40,6 +40,7 @@ from litellm.llms.ollama.common_utils import OllamaResponseIdNormalizer from litellm.types.llms.base import LiteLLMBaseModel from litellm.types.router import GenericLiteLLMParams from litellm.types.utils import ( + CHOICE_LOGPROBS_OR_JSON, CacheCreationTokenDetails, CompletionTokensDetailsWrapper, Delta, @@ -1201,7 +1202,10 @@ class CustomStreamWrapper: choice_json.pop( "finish_reason", None ) # for mistral etc. which return a value in their last chunk (not-openai compatible). - choices.append(StreamingChoices(**choice_json)) + choice_json.pop("logprobs", None) + choices.append( + StreamingChoices(**choice_json, logprobs=getattr(choice, "logprobs", None)) + ) except Exception: choices.append(StreamingChoices()) setattr(model_response, "choices", choices) @@ -1605,7 +1609,7 @@ class CustomStreamWrapper: model_response.system_fingerprint = response_obj["original_chunk"].system_fingerprint self.system_fingerprint = response_obj["original_chunk"].system_fingerprint if response_obj["logprobs"] is not None: - model_response.choices[0].logprobs = response_obj["logprobs"] + model_response.choices[0].logprobs = CHOICE_LOGPROBS_OR_JSON.validate_python(response_obj["logprobs"]) if response_obj["usage"] is not None: if isinstance(response_obj["usage"], dict): diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 6a758961c7b..780c9ef6824 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -7,6 +7,7 @@ from enum import Enum from types import MappingProxyType from typing import ( TYPE_CHECKING, + Annotated, Any, Final, Literal, @@ -38,6 +39,7 @@ from openai.types.moderation_create_response import ( ) from pydantic import ( BaseModel, + BeforeValidator, ConfigDict, Field, FieldSerializationInfo, @@ -45,6 +47,7 @@ from pydantic import ( JsonValue, PrivateAttr, SkipValidation, + TypeAdapter, ValidationError, field_serializer, field_validator, @@ -1199,7 +1202,7 @@ class ChatCompletionTokenLogprob(OpenAIObject): unlikely. """ - top_logprobs: list[TopLogprob] + top_logprobs: list[TopLogprob] = Field(default_factory=list) """List of the most likely tokens and their log probability, at this token position. @@ -1236,9 +1239,13 @@ class ChatCompletionTokenLogprob(OpenAIObject): class ChoiceLogprobs(OpenAIObject): + model_config = ConfigDict(defer_build=False) + content: list[ChatCompletionTokenLogprob] | None = None """A list of message content tokens with log probability information.""" + refusal: Sequence[ChatCompletionTokenLogprob] | None = None + def __contains__(self, key) -> bool: # Define custom behavior for the 'in' operator return hasattr(self, key) @@ -1252,6 +1259,21 @@ class ChoiceLogprobs(OpenAIObject): return getattr(self, key) +def _sdk_logprobs_as_dict(logprobs: object) -> object: + if isinstance(logprobs, ChoiceLogprobs) or not isinstance(logprobs, BaseModel): + return logprobs + return logprobs.model_dump(exclude_unset=True) + + +CHOICE_LOGPROBS_OR_JSON: Final[TypeAdapter[ChoiceLogprobs | JsonValue]] = TypeAdapter( + Annotated[ + ChoiceLogprobs | JsonValue, + Field(union_mode="left_to_right"), + BeforeValidator(_sdk_logprobs_as_dict), + ] +) + + class FunctionCall(OpenAIObject): arguments: str name: str | None = None @@ -1765,7 +1787,7 @@ class Choices(SafeAttributeModel, OpenAIObject): params["message"] = Message(**dump) if logprobs is not None: if isinstance(logprobs, dict): - params["logprobs"] = ChoiceLogprobs(**logprobs) + params["logprobs"] = CHOICE_LOGPROBS_OR_JSON.validate_python(logprobs) else: params["logprobs"] = logprobs else: @@ -2131,8 +2153,8 @@ class StreamingChoices(OpenAIObject): if enhancements is not None: self.enhancements = enhancements - if logprobs is not None and isinstance(logprobs, dict): - self.logprobs = ChoiceLogprobs(**logprobs) + if isinstance(logprobs, (dict, BaseModel)): + self.logprobs = CHOICE_LOGPROBS_OR_JSON.validate_python(logprobs) else: self.logprobs = logprobs diff --git a/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py b/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py index faecb8d8ef6..de8d6a7a8c5 100644 --- a/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py +++ b/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py @@ -1,6 +1,7 @@ import asyncio import importlib import json +from collections.abc import Mapping from datetime import datetime, timedelta from typing import Final @@ -1111,6 +1112,44 @@ def test_convert_to_model_response_object_with_logprobs(): assert result.system_fingerprint is None assert result._hidden_params is not None + +@pytest.mark.parametrize( + "upstream_logprobs", + [ + pytest.param({"content": [{"token": "Hi", "logprob": None}]}, id="null-content-logprob"), + pytest.param( + { + "content": [{"token": "Hi", "logprob": -0.1, "top_logprobs": []}], + "refusal": [{"token": "I", "logprob": None}], + }, + id="null-refusal-logprob", + ), + ], +) +def test_convert_to_model_response_object_passes_nonconforming_logprobs_through( + upstream_logprobs: Mapping[str, object], +) -> None: + result: Final = convert_to_model_response_object( + response_object={ + "id": "chatcmpl-123", + "object": "chat.completion", + "created": 1702685778, + "model": "test-model", + "choices": [ + { + "index": 0, + "message": {"role": "assistant", "content": "Hi"}, + "logprobs": upstream_logprobs, + "finish_reason": "stop", + } + ], + }, + model_response_object=ModelResponse(), + ) + + assert json.loads(result.model_dump_json())["choices"][0]["logprobs"] == upstream_logprobs + + def test_convert_to_model_response_object_error(): """Test error handling for None response object.""" with pytest.raises(Exception, match="Error in response object format"): diff --git a/tests/unit/litellm_core_utils/test_streaming_handler.py b/tests/unit/litellm_core_utils/test_streaming_handler.py index 41e8c50fa85..b9c8feb223d 100644 --- a/tests/unit/litellm_core_utils/test_streaming_handler.py +++ b/tests/unit/litellm_core_utils/test_streaming_handler.py @@ -2,12 +2,15 @@ import asyncio import datetime import itertools import json +import os import queue +import subprocess import sys import threading import time from collections.abc import Mapping, Sequence from dataclasses import dataclass +from pathlib import Path from types import CodeType, FrameType, MappingProxyType from typing import AsyncIterator, Callable, Final, Iterator, Literal, NoReturn, Optional, cast from unittest.mock import AsyncMock, MagicMock, Mock, patch @@ -33,6 +36,7 @@ from litellm.llms.custom_httpx.http_handler import HTTPHandler from litellm.llms.openai.chat.gpt_transformation import OpenAIChatCompletionStreamingHandler from litellm.responses.mcp.chat_completions_handler import _add_mcp_metadata_to_response from litellm.types.utils import ( + ChoiceLogprobs, CompletionTokensDetailsWrapper, Delta, ModelResponse, @@ -1624,6 +1628,356 @@ async def test_stream_with_null_choices_chunk_finishes_normally( assert [choice.finish_reason for choice in streamed_choices if choice.finish_reason] == ["stop"] +def _token_logprobs(token: str) -> dict[str, object]: + return { + "content": [{"token": token, "logprob": -0.5, "bytes": list(token.encode()), "top_logprobs": []}], + "refusal": None, + } + + +@pytest.mark.parametrize("async_stream", [False, True]) +@pytest.mark.parametrize("transport", ["sdk", "http_handler"]) +@pytest.mark.parametrize( + ("finish_chunk_extra", "stream_options", "expected_usage_totals"), + [ + pytest.param({}, None, [], id="finish-chunk-with-logprobs"), + pytest.param( + {"usage": {"prompt_tokens": 3, "completion_tokens": 3, "total_tokens": 6}}, + {"include_usage": True}, + [6], + id="finish-chunk-with-logprobs-and-usage", + ), + ], +) +async def test_stream_finish_chunk_with_logprobs_reaches_caller_as_litellm_logprobs( + async_stream: bool, + transport: str, + finish_chunk_extra: dict[str, object], + stream_options: dict[str, bool] | None, + expected_usage_totals: list[int], + logging_obj: Logging, +) -> None: + upstream_tokens: Final = ("Hel", "lo", "<|return|>") + upstream_chunks: Final = ( + { + "choices": [ + { + "index": 0, + "delta": {"role": "assistant", "content": "Hel"}, + "logprobs": _token_logprobs("Hel"), + "finish_reason": None, + } + ] + }, + { + "choices": [ + {"index": 0, "delta": {"content": "lo"}, "logprobs": _token_logprobs("lo"), "finish_reason": None} + ] + }, + { + "choices": [{"index": 0, "delta": {}, "logprobs": _token_logprobs("<|return|>"), "finish_reason": "stop"}], + **finish_chunk_extra, + }, + ) + body: Final = ( + "".join( + "data: " + + json.dumps( + { + "id": "chatcmpl-logprobs", + "object": "chat.completion.chunk", + "created": 1, + "model": "test-model", + **chunk, + } + ) + + "\n\n" + for chunk in upstream_chunks + ) + + "data: [DONE]\n\n" + ) + raw_response: Final = httpx.Response( + 200, + content=body, + headers={"content-type": "text/event-stream"}, + request=httpx.Request("POST", "https://provider.example/v1/chat/completions"), + ) + client: Final = AsyncOpenAI(api_key="unused") if async_stream else OpenAI(api_key="unused") + stream: Final = ( + ( + AsyncStream(cast_to=ChatCompletionChunk, response=raw_response, client=client) + if isinstance(client, AsyncOpenAI) + else Stream(cast_to=ChatCompletionChunk, response=raw_response, client=client) + ) + if transport == "sdk" + else OpenAIChatCompletionStreamingHandler( + streaming_response=raw_response.aiter_lines() if async_stream else raw_response.iter_lines(), + sync_stream=not async_stream, + ) + ) + response: Final = CustomStreamWrapper( + completion_stream=stream, + model="test-model", + logging_obj=logging_obj, + custom_llm_provider="openai" if transport == "sdk" else "hosted_vllm", + stream_options=stream_options, + ) + + try: + chunks: Final = [chunk async for chunk in response] if async_stream else list(response) + finally: + if isinstance(client, AsyncOpenAI): + await client.close() + else: + client.close() + + wire_choices: Final = tuple( + itertools.chain.from_iterable( + json.loads(chunk.model_dump_json(exclude_none=True, exclude_unset=True))["choices"] for chunk in chunks + ) + ) + finish_choices: Final = [choice for choice in wire_choices if choice.get("finish_reason")] + streamed_logprobs: Final = tuple( + choice.logprobs + for choice in itertools.chain.from_iterable(chunk.choices for chunk in chunks) + if choice.logprobs is not None + ) + streamed_tokens: Final = tuple( + entry.token for entry in itertools.chain.from_iterable(logprobs.content or () for logprobs in streamed_logprobs) + ) + + assert [choice["finish_reason"] for choice in finish_choices] == ["stop"] + assert [entry["token"] for entry in finish_choices[0]["logprobs"]["content"]] == [upstream_tokens[-1]] + assert all(isinstance(logprobs, ChoiceLogprobs) for logprobs in streamed_logprobs) + assert streamed_tokens == upstream_tokens + assert [chunk.usage.total_tokens for chunk in chunks if getattr(chunk, "usage", None)] == expected_usage_totals + + +async def _sdk_stream_chunks( + upstream_chunks: Sequence[Mapping[str, object]], + async_stream: bool, + logging_obj: Logging, + stream_options: dict[str, bool] | None = None, +) -> list[ModelResponseStream]: + body: Final = ( + "".join( + "data: " + + json.dumps( + { + "id": "chatcmpl-logprobs", + "object": "chat.completion.chunk", + "created": 1, + "model": "test-model", + **chunk, + } + ) + + "\n\n" + for chunk in upstream_chunks + ) + + "data: [DONE]\n\n" + ) + raw_response: Final = httpx.Response( + 200, + content=body, + headers={"content-type": "text/event-stream"}, + request=httpx.Request("POST", "https://provider.example/v1/chat/completions"), + ) + client: Final = AsyncOpenAI(api_key="unused") if async_stream else OpenAI(api_key="unused") + stream: Final = ( + AsyncStream(cast_to=ChatCompletionChunk, response=raw_response, client=client) + if isinstance(client, AsyncOpenAI) + else Stream(cast_to=ChatCompletionChunk, response=raw_response, client=client) + ) + response: Final = CustomStreamWrapper( + completion_stream=stream, + model="test-model", + logging_obj=logging_obj, + custom_llm_provider="openai", + stream_options=stream_options, + ) + try: + return [chunk async for chunk in response] if async_stream else list(response) + finally: + if isinstance(client, AsyncOpenAI): + await client.close() + else: + client.close() + + +@pytest.mark.parametrize("async_stream", [False, True]) +@pytest.mark.parametrize( + ("upstream_logprobs", "finish_chunk_extra", "stream_options", "expected_usage_totals"), + [ + pytest.param( + {"content": [{"token": "<|return|>", "bytes": None, "logprob": None, "top_logprobs": None}], "refusal": None}, + {}, + None, + [], + id="null-logprob", + ), + pytest.param( + {"content": [{"token": "<|return|>", "logprob": None}]}, + {}, + None, + [], + id="null-logprob-keeps-only-the-keys-upstream-sent", + ), + pytest.param(-0.25, {}, None, [], id="scalar-logprobs"), + pytest.param( + {"content": [{"token": "<|return|>", "bytes": None, "logprob": None, "top_logprobs": None}], "refusal": None}, + {"usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5}}, + {"include_usage": True}, + [5], + id="null-logprob-and-usage", + ), + ], +) +async def test_stream_with_nonstandard_logprobs_passes_them_through_and_finishes( + async_stream: bool, + upstream_logprobs: object, + finish_chunk_extra: dict[str, object], + stream_options: dict[str, bool] | None, + expected_usage_totals: list[int], + logging_obj: Logging, +) -> None: + chunks: Final = await _sdk_stream_chunks( + ( + { + "choices": [ + { + "index": 0, + "delta": {"role": "assistant", "content": "hi"}, + "logprobs": upstream_logprobs, + "finish_reason": None, + } + ] + }, + { + "choices": [{"index": 0, "delta": {}, "logprobs": upstream_logprobs, "finish_reason": "stop"}], + **finish_chunk_extra, + }, + ), + async_stream, + logging_obj, + stream_options, + ) + + wire_choices: Final = tuple( + itertools.chain.from_iterable( + json.loads(chunk.model_dump_json(exclude_none=True, exclude_unset=True))["choices"] for chunk in chunks + ) + ) + logprobs_choices: Final = [choice for choice in wire_choices if "logprobs" in choice] + + assert [choice.get("finish_reason") for choice in logprobs_choices] == [None, "stop"] + assert [choice["logprobs"] for choice in logprobs_choices] == [upstream_logprobs, upstream_logprobs] + assert [chunk.usage.total_tokens for chunk in chunks if getattr(chunk, "usage", None)] == expected_usage_totals + + +@pytest.mark.parametrize("async_stream", [False, True]) +@pytest.mark.parametrize("logprobs_field", ["content", "refusal"]) +async def test_stream_logprobs_without_top_logprobs_reach_caller_as_litellm_logprobs( + async_stream: bool, logprobs_field: Literal["content", "refusal"], logging_obj: Logging +) -> None: + chunks: Final = await _sdk_stream_chunks( + tuple( + { + "choices": [ + { + "index": 0, + "delta": delta, + "logprobs": {logprobs_field: [{"token": token, "logprob": -0.5}]}, + "finish_reason": finish_reason, + } + ] + } + for token, delta, finish_reason in ( + ("hi", {"role": "assistant", "content": "hi"}, None), + ("<|return|>", {}, "stop"), + ) + ), + async_stream, + logging_obj, + ) + + streamed_logprobs: Final = tuple( + choice.logprobs + for choice in itertools.chain.from_iterable(chunk.choices for chunk in chunks) + if choice.logprobs is not None + ) + streamed_entries: Final = tuple( + itertools.chain.from_iterable( + (logprobs.content if logprobs_field == "content" else logprobs.refusal) or () + for logprobs in streamed_logprobs + ) + ) + + assert [(entry.token, entry.logprob, entry.top_logprobs) for entry in streamed_entries] == [ + ("hi", -0.5, []), + ("<|return|>", -0.5, []), + ] + + +FINISH_ONLY_LOGPROBS_STREAM_SCRIPT: Final = """ +import json +import sys + +sys.path.insert(0, sys.argv[1]) + +import httpx +from openai import OpenAI + +import litellm + +upstream_choices = ( + {"index": 0, "delta": {"role": "assistant", "content": "hi"}, "finish_reason": None}, + { + "index": 0, + "delta": {}, + "logprobs": {"content": [{"token": "<|return|>", "logprob": -0.5, "top_logprobs": []}]}, + "finish_reason": "stop", + }, +) +body = "".join( + "data: " + + json.dumps({"id": "c", "object": "chat.completion.chunk", "created": 1, "model": "m", "choices": [choice]}) + + "\\n\\n" + for choice in upstream_choices +) + "data: [DONE]\\n\\n" +client = OpenAI( + api_key="unused", + base_url="https://provider.example/v1", + http_client=httpx.Client( + transport=httpx.MockTransport( + lambda request: httpx.Response(200, content=body, headers={"content-type": "text/event-stream"}) + ) + ), +) +stream = litellm.completion( + model="openai/m", messages=[{"role": "user", "content": "hi"}], stream=True, logprobs=True, client=client +) +print(json.dumps([json.loads(chunk.model_dump_json(exclude_none=True)) for chunk in stream])) +""" + + +def test_stream_with_logprobs_only_on_finish_chunk_serializes_in_a_fresh_process() -> None: + result: Final = subprocess.run( + [sys.executable, "-I", "-c", FINISH_ONLY_LOGPROBS_STREAM_SCRIPT, str(Path(litellm.__file__).parent.parent)], + capture_output=True, + text=True, + timeout=120, + env={**os.environ, "LITELLM_LOCAL_MODEL_COST_MAP": "True"}, + ) + + assert result.returncode == 0, result.stdout + result.stderr + wire_choices: Final = tuple( + itertools.chain.from_iterable(chunk["choices"] for chunk in json.loads(result.stdout.strip().splitlines()[-1])) + ) + finish_choices: Final = [choice for choice in wire_choices if choice.get("finish_reason")] + assert [choice["finish_reason"] for choice in finish_choices] == ["stop"] + assert [entry["token"] for entry in finish_choices[0]["logprobs"]["content"]] == ["<|return|>"] + + def _hosted_vllm_stream_wrapper(logging_obj: Logging, error_payload: dict) -> CustomStreamWrapper: """A CustomStreamWrapper over the real OpenAI-compatible line iterator, fed an HTTP 200 SSE body that carries an in-body error payload the way diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 1aec7d5b761..960616e2631 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -29161,7 +29161,7 @@ export interface components { /** Token */ token: string; /** Top Logprobs */ - top_logprobs: components["schemas"]["TopLogprob"][]; + top_logprobs?: components["schemas"]["TopLogprob"][]; } & { [key: string]: unknown; }; @@ -29292,6 +29292,8 @@ export interface components { ChoiceLogprobs: { /** Content */ content?: components["schemas"]["ChatCompletionTokenLogprob"][] | null; + /** Refusal */ + refusal?: components["schemas"]["ChatCompletionTokenLogprob"][] | null; } & { [key: string]: unknown; }; From 87fec5467fb7a965a2cf3998fb542c3beab17234 Mon Sep 17 00:00:00 2001 From: zoveki <349028329+zoveki@users.noreply.github.com> Date: Fri, 30 Oct 2026 05:07:00 -0700 Subject: [PATCH 22/23] fix(sso): keep CLI SSO and Claude Code gateway team lists in the user record's order (#49033) * fix(sso): bind CLI SSO and Claude Code gateway logins to the first team on the user record * fix(proxy): keep SCIM and JWT team sync in record order, and look up SSO groups once SCIM PATCH and JWT team sync saved a user's teams from a set, so the first team on the record, the one a Claude Code gateway sign-in binds, could change on any sync. Both now keep the record's order and append new teams in the order the IdP sent them. SSO group sync now drops claim groups that are not LiteLLM teams with one lookup before adding the missing teams one at a time, so a claim full of non-team groups no longer costs one failed add per group on every sign-in. * test(sso): expect the gateway team dropdown in the user record's order * fix(sso): read team existence from the writer, keep new-user and SCIM replace team order * fix(scim): keep the teams a user already has in record order on PATCH and PUT * fix(scim): keep record team order when a SCIM POST matches an existing email * fix(scim): join new teams in request order so a failed roster write keeps it * fix(scim): let a groups replace win over later group ops in a PATCH again Applying group adds and removes after a replace made a PATCH that deactivates a user and adds a team that does not exist answer 404 and leave the user active, where it answered 200 before. Keep the record's team order, but ignore group ops after the last replace as the base did. * test(sso): type the CLI team detail row helper --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Co-authored-by: rj-cairn <349020452+rj-cairn@users.noreply.github.com> --- litellm/proxy/auth/handle_jwt.py | 23 +- .../internal_user_endpoints.py | 36 +-- .../management_endpoints/scim/scim_v2.py | 83 +++-- litellm/proxy/management_endpoints/ui_sso.py | 55 +++- .../test_claude_code_gateway_oauth.py | 52 +++ tests/unit/proxy/auth/test_handle_jwt.py | 25 ++ .../scim/test_scim_patch_user.py | 73 ++++- .../scim/test_scim_v2_endpoints.py | 146 +++++++++ .../test_internal_user_endpoints.py | 59 ++++ .../proxy/management_endpoints/test_ui_sso.py | 300 +++++++++++------- 10 files changed, 661 insertions(+), 191 deletions(-) diff --git a/litellm/proxy/auth/handle_jwt.py b/litellm/proxy/auth/handle_jwt.py index 7bd7b1226b1..9e11179c686 100644 --- a/litellm/proxy/auth/handle_jwt.py +++ b/litellm/proxy/auth/handle_jwt.py @@ -2199,15 +2199,19 @@ class JWTAuthManager: # not mistaken for claimless and left with stale DB memberships the fallback # could later attribute. With the flag off, keep the upstream plural-only # reconciliation so existing deployments are unchanged. - jwt_team_ids: Final = set( - jwt_handler.get_all_jwt_team_ids(jwt_valid_token) - if jwt_handler.litellm_jwtauth.fallback_to_db_teams - else jwt_handler.get_team_ids_from_jwt(jwt_valid_token) + claim_team_ids: Final = tuple( + dict.fromkeys( + jwt_handler.get_all_jwt_team_ids(jwt_valid_token) + if jwt_handler.litellm_jwtauth.fallback_to_db_teams + else jwt_handler.get_team_ids_from_jwt(jwt_valid_token) + ) ) - existing_teams: Final = set(user_object.teams or []) - teams_to_add: Final = jwt_team_ids - existing_teams + jwt_team_ids: Final = frozenset(claim_team_ids) + existing_team_order: Final = tuple(dict.fromkeys(user_object.teams or [])) + existing_teams: Final = frozenset(existing_team_order) + teams_to_add: Final = tuple(team_id for team_id in claim_team_ids if team_id not in existing_teams) preserve_db_teams_without_claims: Final = jwt_handler.litellm_jwtauth.fallback_to_db_teams and not jwt_team_ids - teams_to_remove: Final = set() if preserve_db_teams_without_claims else existing_teams - jwt_team_ids + teams_to_remove: Final = frozenset() if preserve_db_teams_without_claims else existing_teams - jwt_team_ids if teams_to_add or teams_to_remove: from litellm.proxy.management_endpoints.scim.scim_v2 import ( patch_team_membership, @@ -2218,7 +2222,10 @@ class JWTAuthManager: teams_ids_to_add_user_to=list(teams_to_add), teams_ids_to_remove_user_from=list(teams_to_remove), ) - user_object.teams = list(jwt_team_ids) + user_object.teams = [ + *(team_id for team_id in existing_team_order if team_id not in teams_to_remove), + *teams_to_add, + ] if user_api_key_cache is not None: await user_api_key_cache.async_set_cache( key=user_object.user_id, diff --git a/litellm/proxy/management_endpoints/internal_user_endpoints.py b/litellm/proxy/management_endpoints/internal_user_endpoints.py index 6ceab27befa..6950cf08a25 100644 --- a/litellm/proxy/management_endpoints/internal_user_endpoints.py +++ b/litellm/proxy/management_endpoints/internal_user_endpoints.py @@ -13,6 +13,7 @@ These are members of a Team on LiteLLM """ import asyncio +import contextlib import json import traceback from collections.abc import Awaitable, Mapping, Sequence @@ -443,30 +444,25 @@ async def add_new_user_to_default_team( teams: Sequence[str | NewUserRequestTeam], prisma_client: "PrismaClient", ): - tasks: Final[list[Awaitable[object]]] = [] - for team in teams: - user_role: Literal["user", "admin"] = "user" - max_budget_in_team: float | None = None - if isinstance(team, str): - team_id = team - elif isinstance(team, NewUserRequestTeam): - team_id = team.team_id - user_role = team.user_role - max_budget_in_team = team.max_budget_in_team - else: - raise ValueError(f"Invalid team type: {type(team)}") - - tasks.append( - _add_user_to_team( + memberships: Final = tuple(_default_team_membership(team) for team in teams) + for membership in memberships: + with contextlib.suppress(Exception): + await _add_user_to_team( user_id=user_id, - team_id=team_id, + team_id=membership.team_id, user_email=user_email, user_api_key_dict=user_api_key_dict, - max_budget_in_team=max_budget_in_team, - user_role=user_role, + max_budget_in_team=membership.max_budget_in_team, + user_role=membership.user_role, ) - ) - await asyncio.gather(*tasks, return_exceptions=True) + + +def _default_team_membership(team: str | NewUserRequestTeam) -> NewUserRequestTeam: + if isinstance(team, str): + return NewUserRequestTeam(team_id=team) + if isinstance(team, NewUserRequestTeam): + return team + raise ValueError(f"Invalid team type: {type(team)}") async def _fetch_user_team_ids(user_id: str, prisma_client: "PrismaClient") -> tuple[str, ...]: diff --git a/litellm/proxy/management_endpoints/scim/scim_v2.py b/litellm/proxy/management_endpoints/scim/scim_v2.py index 668bd64227b..e48f68a2d0a 100644 --- a/litellm/proxy/management_endpoints/scim/scim_v2.py +++ b/litellm/proxy/management_endpoints/scim/scim_v2.py @@ -229,7 +229,11 @@ class UserProvisionerHelpers: return None requested_teams: Final = list(dict.fromkeys(new_user_request.teams or [])) - new_teams: Final = requested_teams if requested_teams else list(existing_user.teams or []) + new_teams: Final = ( + list(_keep_record_team_order(existing_user.teams or [], requested_teams)) + if requested_teams + else list(existing_user.teams or []) + ) if new_user_request.user_id != existing_user.user_id: verbose_proxy_logger.info( @@ -1165,7 +1169,7 @@ async def _handle_team_membership_changes( existing_teams_set: Final = set(existing_teams) new_teams_set: Final = set(new_teams) - teams_to_add: Final = new_teams_set - existing_teams_set + teams_to_add: Final = tuple(team for team in dict.fromkeys(new_teams) if team not in existing_teams_set) teams_to_remove: Final = existing_teams_set - new_teams_set if teams_to_add or teams_to_remove: @@ -1948,7 +1952,11 @@ async def update_user( # SCIM User.groups is readOnly (RFC 7643 4.1.2): IdPs sync membership via /Groups and send # no groups or `groups: []` on profile PUTs, so empty means unspecified, not "remove from every team" - target_teams: Final = user_data["teams"] or existing_user.teams + target_teams: Final = ( + list(_keep_record_team_order(existing_user.teams, user_data["teams"])) + if user_data["teams"] + else existing_user.teams + ) await _handle_team_membership_changes( user_id=user_id, existing_teams=existing_user.teams, @@ -2241,19 +2249,29 @@ def _handle_name_update(path: str, op_type: str, value: object, scim_metadata: d scim_metadata["familyName"] = str(value) -def _handle_group_operations(op_type: str, value: object, teams_set: set[str], path: str | None) -> set[str] | None: - """Handle group/team membership operations.""" - group_values = _extract_group_values(value) - if not group_values and value is None: - group_values = _extract_ids_from_path_filter(path, "groups") - if op_type == "replace": - return set(group_values) - elif op_type == "add": - teams_set.update(group_values) - elif op_type == "remove": - for gid in group_values: - teams_set.discard(gid) - return None +def _is_group_path(path: str | None) -> bool: + return isinstance(path, str) and path.lower().startswith("groups") + + +def _apply_group_operation(teams: tuple[str, ...], op: SCIMPatchOperation) -> tuple[str, ...]: + listed: Final = _extract_group_values(op.value) + group_values: Final = listed if listed or op.value is not None else _extract_ids_from_path_filter(op.path, "groups") + if op.op == "replace": + return tuple(dict.fromkeys(group_values)) + if op.op == "add": + return tuple(dict.fromkeys((*teams, *group_values))) + removed: Final = frozenset(group_values) + return tuple(team for team in teams if team not in removed) + + +def _keep_record_team_order(record_teams: Sequence[str], target_teams: Sequence[str]) -> tuple[str, ...]: + target: Final = frozenset(target_teams) + record: Final = tuple(dict.fromkeys(record_teams)) + known: Final = frozenset(record) + return ( + *(team for team in record if team in target), + *(team for team in dict.fromkeys(target_teams) if team not in known), + ) def _multi_valued_attribute_base(path: str) -> str: @@ -2520,8 +2538,8 @@ def _handle_generic_metadata(path: str, op_type: str, value: object, metadata: d def _apply_patch_ops( existing_user: LiteLLM_UserTable, patch_ops: SCIMPatchOp, -) -> tuple[dict[str, object], set[str]]: - """Apply patch operations and return update data and final team set.""" +) -> tuple[dict[str, object], tuple[str, ...]]: + """Apply patch operations and return update data and the final teams in order.""" update_data: Final[dict[str, object]] = {} stored_metadata: Final[object] = TypeAdapter(object).validate_python(existing_user.model_dump().get("metadata")) metadata: Final = _JSON_OBJECT_ADAPTER.validate_python(stored_metadata) if isinstance(stored_metadata, dict) else {} @@ -2541,9 +2559,6 @@ def _apply_patch_ops( if isinstance(enterprise, _InvalidEnterpriseValue): _raise_invalid_enterprise_value(enterprise) - teams_set: Final[set[str]] = set(existing_user.teams or []) - replace_team_set: set[str] | None = None - for op in patch_ops.Operations: path = op.path.lower() if isinstance(op.path, str) else "" value = op.value @@ -2585,16 +2600,20 @@ def _apply_patch_ops( _handle_name_update(path, op_type, value, scim_metadata) elif _multi_valued_attribute_base(path) in SCIM_MULTI_VALUED_ATTRIBUTE_METADATA_KEYS: _handle_multi_valued_attribute_update(path, op_type, value, metadata) - elif path.startswith("groups"): - new_replace_set = _handle_group_operations(op_type, value, teams_set, op.path) - if new_replace_set is not None: - replace_team_set = new_replace_set - elif _enterprise_extension_sub_path(path) is not None: + elif _is_group_path(path) or _enterprise_extension_sub_path(path) is not None: continue else: _handle_generic_metadata(path, op_type, value, metadata) - final_team_set: Final = replace_team_set if replace_team_set is not None else teams_set + record_teams: Final = tuple(existing_user.teams or []) + group_ops: Final = tuple(op for op in patch_ops.Operations if _is_group_path(op.path)) + last_replace: Final = next((op for op in reversed(group_ops) if op.op == "replace"), None) + final_teams: Final = _keep_record_team_order( + record_teams, + _apply_group_operation(record_teams, last_replace) + if last_replace is not None + else reduce(_apply_group_operation, group_ops, record_teams), + ) final_metadata: Final = ( { **{key: metadata_value for key, metadata_value in metadata.items() if key != SCIM_ENTERPRISE_METADATA_KEY}, @@ -2605,7 +2624,7 @@ def _apply_patch_ops( else {**metadata, "scim_metadata": scim_metadata} ) update_data["metadata"] = final_metadata - return update_data, final_team_set + return update_data, final_teams def _is_user_not_in_team_error(exc: HTTPException) -> bool: @@ -2758,7 +2777,7 @@ async def patch_user( prev_active: Final = _user_scim_active(existing_user) - update_data, final_team_set = _apply_patch_ops( + update_data, final_teams = _apply_patch_ops( existing_user=existing_user, patch_ops=patch_ops, ) @@ -2770,15 +2789,15 @@ async def patch_user( await _handle_team_membership_changes( user_id=user_id, existing_teams=existing_user.teams or [], - new_teams=list(final_team_set), + new_teams=list(final_teams), ) - update_data["teams"] = list(final_team_set) + update_data["teams"] = list(final_teams) admin_group: Final = await _get_scim_admin_group() if admin_group is not None: update_data["user_role"] = _resolve_scim_user_role( - await _scim_groups_from_team_ids(prisma_client, list(final_team_set)), + await _scim_groups_from_team_ids(prisma_client, list(final_teams)), admin_group, _default_scim_user_role(), ) diff --git a/litellm/proxy/management_endpoints/ui_sso.py b/litellm/proxy/management_endpoints/ui_sso.py index a980b9c33ff..59c2df91d77 100644 --- a/litellm/proxy/management_endpoints/ui_sso.py +++ b/litellm/proxy/management_endpoints/ui_sso.py @@ -8,7 +8,6 @@ Has all /sso/* routes /sso/debug/callback - returns the OpenID object returned by the SSO provider """ -import asyncio import base64 import hashlib import inspect @@ -151,6 +150,7 @@ from litellm.proxy.utils import ( get_custom_url, get_server_root_path, ) +from litellm.repositories.chunked_in import find_many_in from litellm.repositories.prisma_protocols import TableActions from litellm.repositories.table_repositories import SSOConfigRepository from litellm.repositories.team_repository import TeamRepository @@ -1855,7 +1855,7 @@ async def warn_if_id_jag_capture_gap(*, retention_enabled: RetentionCheck | None verbose_proxy_logger.warning("SSO debug callback ran with an oauth2_id_jag capture gap: %s", gap) -async def create_team_member_add_task(team_id, user_info): +async def create_team_member_add_task(team_id: str, user_info: NewUserResponse | LiteLLM_UserTable): """Create a task for adding a member to a team.""" try: member: Final = Member(user_id=user_info.user_id, role="user") @@ -1871,22 +1871,43 @@ async def create_team_member_add_task(team_id, user_info): verbose_proxy_logger.debug("[Non-Blocking] Error trying to add sso user to db: %s", e) -async def add_missing_team_member(user_info: NewUserResponse | LiteLLM_UserTable, sso_teams: list[str]): +async def find_existing_team_ids(team_ids: Sequence[str]) -> frozenset[str] | None: + from litellm.proxy.proxy_server import prisma_client + + if prisma_client is None: + return None + try: + rows: Final = await find_many_in(TeamRepository(prisma_client, use_writer=True).table, "team_id", team_ids) + except Exception as e: + verbose_proxy_logger.debug("[Non-Blocking] Error looking up SSO teams: %s", e) + return None + return frozenset(row.team_id for row in rows) + + +async def add_missing_team_member( + user_info: NewUserResponse | LiteLLM_UserTable, + sso_teams: Sequence[str], + *, + add_team_member: Callable[ + [str, NewUserResponse | LiteLLM_UserTable], Awaitable[object] + ] = create_team_member_add_task, + existing_team_ids: Callable[[Sequence[str]], Awaitable[frozenset[str] | None]] = find_existing_team_ids, +) -> None: """ - Get missing teams (diff b/w user_info.team_ids and sso_teams) - - Add missing user to missing teams + - Drop the IdP groups that are not teams with one lookup, since group claims often carry many of them + - Add missing user to missing teams, one at a time in sso_teams order, so user.teams keeps the claim's order """ - # Handle None as empty list for new users - user_teams: Final = user_info.teams if user_info.teams is not None else [] - missing_teams: Final = set(sso_teams) - set(user_teams) - missing_teams_list: Final = list(missing_teams) - tasks = [] - tasks = [create_team_member_add_task(team_id, user_info) for team_id in missing_teams_list] - - try: - await asyncio.gather(*tasks) - except Exception as e: - verbose_proxy_logger.debug("[Non-Blocking] Error trying to add sso user to db: %s", e) + held_teams: Final = frozenset(user_info.teams or ()) + missing_teams: Final = tuple(team_id for team_id in dict.fromkeys(sso_teams) if team_id not in held_teams) + if not missing_teams: + return + existing: Final = await existing_team_ids(missing_teams) + teams_to_add: Final = ( + missing_teams if existing is None else tuple(team_id for team_id in missing_teams if team_id in existing) + ) + for team_id in teams_to_add: + await add_team_member(team_id, user_info) def get_disabled_non_admin_personal_key_creation(): @@ -2451,7 +2472,9 @@ async def fetch_cli_sso_team_details( except Exception as e: verbose_proxy_logger.error("Error fetching team details for CLI SSO session: %s", e) return None - return tuple(_cli_sso_team_detail(team_row.model_dump()) for team_row in prisma_teams) + fetched_details: Final = tuple(_cli_sso_team_detail(team_row.model_dump()) for team_row in prisma_teams) + details_by_team_id: Final = MappingProxyType({detail.team_id: detail for detail in fetched_details}) + return tuple(details_by_team_id[team_id] for team_id in dict.fromkeys(teams) if team_id in details_by_team_id) def _cli_sso_session_teams(team_details: Sequence[CliSsoTeamDetail]) -> list[str]: diff --git a/tests/integration/authorization/test_claude_code_gateway_oauth.py b/tests/integration/authorization/test_claude_code_gateway_oauth.py index 4dedbdcd8e4..23476c439de 100644 --- a/tests/integration/authorization/test_claude_code_gateway_oauth.py +++ b/tests/integration/authorization/test_claude_code_gateway_oauth.py @@ -636,3 +636,55 @@ def test_device_flow_mints_the_session_for_the_team_the_developer_picks_in_the_b served_models: Final = {model.id for model in sdk.models.list()} assert served_models == {team_models[chosen_team]} assert len(identity_provider.drain()) == 3 + + +@pytest.mark.timeout(300) +def test_gateway_login_lists_and_defaults_to_the_first_team_on_the_user_record_after_a_write_to_its_row( + enabled_gateway: Gateway, identity_provider: Wire +) -> None: + run_id: Final = uuid.uuid4().hex + with enabled_gateway.scenario() as scenario: + user_id: Final = scenario.user(user_id=_USER_ID, user_email=_USER_EMAIL, user_role="internal_user") + first_team: Final = scenario.team(team_id=f"zeta-{run_id}") + second_team: Final = scenario.team(team_id=f"alpha-{run_id}") + for team_id in (first_team, second_team): + enabled_gateway.post( + "/team/member_add", {"team_id": team_id, "member": {"role": "user", "user_id": user_id}} + ) + enabled_gateway.post("/team/update", {"team_id": first_team, "metadata": {"updated_by_run": run_id}}) + assert enabled_gateway.rows('SELECT teams FROM "LiteLLM_UserTable" WHERE user_id = %s', (user_id,)) == [ + {"teams": [first_team, second_team]} + ] + + authorization_response: Final = enabled_gateway.client.post( + "/claude_code_gateway/oauth/device_authorization", data={"client_id": _CLIENT_ID} + ) + assert authorization_response.status_code == 200, authorization_response.text + authorization: Final = _DeviceAuthorization.model_validate_json(authorization_response.content) + verification: Final = enabled_gateway.client.get(authorization.verification_uri, follow_redirects=False) + assert verification.status_code == 303, verification.text + identity_provider_redirect: Final = enabled_gateway.client.get( + verification.headers["location"], follow_redirects=False + ) + assert identity_provider_redirect.status_code == 302, identity_provider_redirect.text + callback: Final = enabled_gateway.client.get( + identity_provider_redirect.headers["location"], follow_redirects=False + ) + assert callback.status_code == 200, callback.text + assert tuple(_TEAM_OPTION.findall(callback.text)) == (first_team, second_team), callback.text + browser_token: Final = _BROWSER_TOKEN.search(callback.text) + assert browser_token is not None, callback.text + completion_without_a_pick: Final = enabled_gateway.client.post( + f"/sso/cli/complete/{authorization.device_code.partition('.')[0]}", + data={ + "user_code": authorization.user_code, + "browser_complete_token": html.unescape(browser_token.group(1)), + }, + ) + assert completion_without_a_pick.status_code == 200, completion_without_a_pick.text + token_response: Final = _device_code_grant(enabled_gateway, authorization.device_code) + assert token_response.status_code == 200, token_response.text + token: Final = _AccessToken.model_validate_json(token_response.content) + identity_provider.drain() + + assert (token.user_id, token.team_id) == (user_id, first_team) diff --git a/tests/unit/proxy/auth/test_handle_jwt.py b/tests/unit/proxy/auth/test_handle_jwt.py index 51cf15996cd..4c223edc20d 100644 --- a/tests/unit/proxy/auth/test_handle_jwt.py +++ b/tests/unit/proxy/auth/test_handle_jwt.py @@ -913,6 +913,31 @@ async def test_sync_user_role_and_teams_cache_invalidation_on_team_change(): assert call_kwargs.kwargs["model_type"] == LiteLLM_UserTable +@pytest.mark.asyncio +async def test_sync_user_role_and_teams_keeps_record_order_and_adds_claim_teams_in_claim_order(): + jwt_handler: Final = JWTHandler() + jwt_handler.update_environment( + prisma_client=None, + user_api_key_cache=AsyncMock(), + litellm_jwtauth=LiteLLM_JWTAuth(team_ids_jwt_field="groups", sync_user_role_and_teams=True), + ) + cache: Final = AsyncMock() + user: Final = LiteLLM_UserTable(user_id="u1", teams=["team-zeta", "team-kappa", "team-alpha"]) + token: Final = {"groups": ["team-mu", "team-alpha", "team-beta", "team-mu", "team-zeta", "team-omega"]} + + with patch( + "litellm.proxy.management_endpoints.scim.scim_v2.patch_team_membership", + new_callable=AsyncMock, + ) as mock_patch: + await JWTAuthManager.sync_user_role_and_teams(jwt_handler, token, user, AsyncMock(), user_api_key_cache=cache) + + assert mock_patch.await_args.kwargs["teams_ids_to_add_user_to"] == ["team-mu", "team-beta", "team-omega"] + assert mock_patch.await_args.kwargs["teams_ids_to_remove_user_from"] == ["team-kappa"] + expected_teams: Final = ["team-zeta", "team-alpha", "team-mu", "team-beta", "team-omega"] + assert user.teams == expected_teams + assert cache.async_set_cache.await_args.kwargs["value"].teams == expected_teams + + @pytest.mark.asyncio async def test_sync_user_role_and_teams_no_cache_write_when_nothing_changes(): """Test that cache is NOT written when role and teams already match.""" diff --git a/tests/unit/proxy/management_endpoints/scim/test_scim_patch_user.py b/tests/unit/proxy/management_endpoints/scim/test_scim_patch_user.py index 31c5a54a753..1901e0efd4b 100644 --- a/tests/unit/proxy/management_endpoints/scim/test_scim_patch_user.py +++ b/tests/unit/proxy/management_endpoints/scim/test_scim_patch_user.py @@ -502,9 +502,9 @@ def test_apply_patch_ops_remove_group_filtered_path_without_value(): Operations=[SCIMPatchOperation(op="remove", path='groups[value eq "team-1"]')] ) - _, final_team_set = _apply_patch_ops(existing_user=user, patch_ops=patch_ops) + _, final_teams = _apply_patch_ops(existing_user=user, patch_ops=patch_ops) - assert final_team_set == {"team-2"} + assert final_teams == ("team-2",) def test_apply_patch_ops_add_group_filtered_path_without_value(): @@ -519,9 +519,9 @@ def test_apply_patch_ops_add_group_filtered_path_without_value(): Operations=[SCIMPatchOperation(op="add", path="groups[value eq 'team-3']")] ) - _, final_team_set = _apply_patch_ops(existing_user=user, patch_ops=patch_ops) + _, final_teams = _apply_patch_ops(existing_user=user, patch_ops=patch_ops) - assert final_team_set == {"team-1", "team-3"} + assert final_teams == ("team-1", "team-3") def test_apply_patch_ops_replace_groups_empty_value_does_not_use_path_filter(): @@ -539,9 +539,70 @@ def test_apply_patch_ops_replace_groups_empty_value_does_not_use_path_filter(): ] ) - _, final_team_set = _apply_patch_ops(existing_user=user, patch_ops=patch_ops) + _, final_teams = _apply_patch_ops(existing_user=user, patch_ops=patch_ops) - assert final_team_set == set() + assert final_teams == () + + +@pytest.mark.parametrize( + "operations,expected_teams", + [ + pytest.param( + [ + SCIMPatchOperation(op="add", path="groups", value=[{"value": "team-mu"}, {"value": "team-beta"}]), + SCIMPatchOperation(op="remove", path='groups[value eq "team-kappa"]'), + SCIMPatchOperation(op="add", path="groups", value=[{"value": "team-alpha"}, {"value": "team-omega"}]), + ], + ("team-zeta", "team-alpha", "team-delta", "team-mu", "team-beta", "team-omega"), + id="adds_follow_the_record_in_op_order", + ), + pytest.param( + [ + SCIMPatchOperation( + op="replace", + path="groups", + value=[{"value": "team-omega"}, {"value": "team-delta"}, {"value": "team-beta"}], + ) + ], + ("team-delta", "team-omega", "team-beta"), + id="replace_keeps_teams_the_user_has_in_record_order_and_adds_new_ones_in_request_order", + ), + pytest.param( + [ + SCIMPatchOperation( + op="replace", + path="groups", + value=[{"value": "team-omega"}, {"value": "team-delta"}, {"value": "team-beta"}], + ), + SCIMPatchOperation(op="add", path="groups", value=[{"value": "team-mu"}]), + SCIMPatchOperation(op="remove", path='groups[value eq "team-delta"]'), + ], + ("team-delta", "team-omega", "team-beta"), + id="a_replace_wins_over_later_group_ops", + ), + pytest.param( + [ + SCIMPatchOperation(op="remove", path='groups[value eq "team-zeta"]'), + SCIMPatchOperation(op="add", path="groups", value=[{"value": "team-zeta"}]), + ], + ("team-zeta", "team-alpha", "team-kappa", "team-delta"), + id="removing_and_re_adding_a_team_keeps_its_place", + ), + ], +) +def test_apply_patch_ops_keeps_team_order_so_the_first_team_stays_first( + operations: list[SCIMPatchOperation], expected_teams: tuple[str, ...] +): + user: Final = LiteLLM_UserTable( + user_id="user-order", + user_email="order@example.com", + teams=["team-zeta", "team-alpha", "team-kappa", "team-delta"], + metadata={}, + ) + + _, final_teams = _apply_patch_ops(existing_user=user, patch_ops=SCIMPatchOp(Operations=operations)) + + assert final_teams == expected_teams @pytest.mark.asyncio diff --git a/tests/unit/proxy/management_endpoints/scim/test_scim_v2_endpoints.py b/tests/unit/proxy/management_endpoints/scim/test_scim_v2_endpoints.py index 6014c9229f6..180362f45f1 100644 --- a/tests/unit/proxy/management_endpoints/scim/test_scim_v2_endpoints.py +++ b/tests/unit/proxy/management_endpoints/scim/test_scim_v2_endpoints.py @@ -25,6 +25,7 @@ from litellm.proxy._types import ( NewUserResponse, ProxyErrorTypes, ProxyException, + TeamMemberAddRequest, UserAPIKeyAuth, ) from litellm.proxy.management_endpoints.scim.scim_v2 import ( @@ -941,6 +942,57 @@ async def test_handle_existing_user_by_email_syncs_roster_and_dedups_teams(mocke assert update_calls[0].kwargs["data"]["teams"] == ["team-a", "team-b"] +class _EmailMatchedUserRow: + def __init__(self, user: LiteLLM_UserTable) -> None: + self.user = user + + async def find_first(self, *, where: dict[str, object]) -> LiteLLM_UserTable: + return self.user + + async def update(self, *, where: dict[str, object], data: dict[str, object]) -> LiteLLM_UserTable: + self.user = self.user.model_copy(update=data) + return self.user + + +@pytest.mark.asyncio +async def test_handle_existing_user_by_email_keeps_record_team_order(mocker: MockerFixture): + user_rows: Final = _EmailMatchedUserRow( + LiteLLM_UserTable( + user_id="same-id", + user_email="member@example.com", + user_alias="Member", + teams=["team-zeta", "team-alpha", "team-kappa"], + metadata={}, + ) + ) + mock_prisma_client: Final = mocker.MagicMock() + mock_prisma_client.db = mocker.MagicMock() + mock_prisma_client.db.litellm_usertable = user_rows + + mocker.patch( # test-quality-ok: roster helpers are module-level, not injectable into the helper + "litellm.proxy.management_endpoints.scim.scim_v2.ScimTransformations.transform_litellm_user_to_scim_user", + AsyncMock(return_value=None), + ) + mocker.patch( # test-quality-ok: roster helpers are module-level, not injectable into the helper + "litellm.proxy.management_endpoints.scim.scim_v2._handle_team_membership_changes", + AsyncMock(), + ) + + await UserProvisionerHelpers.handle_existing_user_by_email( + prisma_client=mock_prisma_client, + new_user_request=NewUserRequest( + user_id="same-id", + user_email="member@example.com", + user_alias="Member", + teams=["team-mu", "team-kappa", "team-beta", "team-zeta"], + metadata={}, + auto_create_key=False, + ), + ) + + assert user_rows.user.teams == ["team-zeta", "team-kappa", "team-mu", "team-beta"] + + @pytest.mark.asyncio async def test_handle_existing_user_by_email_without_teams_preserves_memberships(mocker): """Adoption via POST /Users without ``groups`` must keep the user's existing teams. @@ -1290,6 +1342,49 @@ async def test_handle_team_membership_changes_add_and_remove(mocker): assert call_args[1]["teams_ids_to_remove_user_from"] == ["team1"] +class _TeamRosterRecord: + def __init__(self, missing_team: str) -> None: + self.missing_team = missing_team + self.teams: tuple[str, ...] = () + + async def team_member_add(self, *, data: TeamMemberAddRequest, user_api_key_dict: UserAPIKeyAuth) -> None: + if data.team_id == self.missing_team: + raise HTTPException(status_code=404, detail={"error": "Team not found"}) + self.teams = (*self.teams, data.team_id) + + +@pytest.mark.asyncio +async def test_handle_team_membership_changes_adds_teams_in_request_order_when_one_fails(mocker: MockerFixture): + requested_teams: Final = ( + "team-zeta", + "team-alpha", + "team-kappa", + "team-kept", + "team-mu", + "team-missing", + "team-beta", + "team-omega", + "team-delta", + "team-sigma", + "team-gamma", + "team-tau", + ) + roster: Final = _TeamRosterRecord(missing_team="team-missing") + mocker.patch( # test-quality-ok: team_member_add is module-level, not injectable into the helper + "litellm.proxy.management_endpoints.scim.scim_v2.team_member_add", + roster.team_member_add, + ) + + with pytest.raises(SCIMRosterSyncError): + await _handle_team_membership_changes( + user_id="test-user", + existing_teams=["team-kept"], + new_teams=list(requested_teams), + ) + + assert roster.teams == tuple(team for team in requested_teams if team not in {"team-kept", "team-missing"}) + + @pytest.mark.asyncio async def test_update_user_success(mocker): """Should successfully update user with PUT request""" @@ -1492,6 +1587,57 @@ async def test_update_user_without_groups_preserves_memberships_and_role(mocker, assert "user_role" not in update_data +@pytest.mark.asyncio +async def test_update_user_put_keeps_the_record_order_of_teams_the_user_already_has( + mocker: MockerFixture, monkeypatch: pytest.MonkeyPatch +) -> None: + from litellm.proxy.proxy_server import proxy_config + + async def no_scim_admin_group() -> dict[str, object]: + return {} + + monkeypatch.setattr(proxy_config, "get_config", no_scim_admin_group) + monkeypatch.setattr("litellm.default_internal_user_params", None, raising=False) + + existing_user: Final = mocker.MagicMock() + existing_user.teams = ["team-zeta", "team-alpha", "team-kappa"] + existing_user.metadata = {} + scim_user: Final = SCIMUser( + schemas=["urn:ietf:params:scim:schemas:core:2.0:User"], + userName="order-user", + emails=[SCIMUserEmail(value="order@example.com")], + groups=[ + SCIMUserGroup(value="team-mu"), + SCIMUserGroup(value="team-kappa"), + SCIMUserGroup(value="team-beta"), + SCIMUserGroup(value="team-zeta"), + ], + ) + mock_prisma_client: Final = mocker.MagicMock() + mock_prisma_client.db.litellm_usertable.update = AsyncMock(return_value={"user_id": "order-user"}) + mocker.patch( # test-quality-ok: update_user's collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2._get_prisma_client_or_raise_exception", + AsyncMock(return_value=mock_prisma_client), + ) + mocker.patch( # test-quality-ok: update_user's collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2._check_user_exists", + AsyncMock(return_value=existing_user), + ) + mocker.patch( # test-quality-ok: update_user's collaborators are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2.ScimTransformations.transform_litellm_user_to_scim_user", + AsyncMock(return_value=scim_user), + ) + mocker.patch( # test-quality-ok: roster writes are module-level, not injectable + "litellm.proxy.management_endpoints.scim.scim_v2.patch_team_membership", + AsyncMock(), + ) + + await update_user(user_id="order-user", user=scim_user) + + update_data: Final = mock_prisma_client.db.litellm_usertable.update.call_args.kwargs["data"] + assert update_data["teams"] == ["team-zeta", "team-kappa", "team-mu", "team-beta"] + + @pytest.mark.asyncio async def test_update_user_not_found(mocker): """Should raise 404 when user doesn't exist""" diff --git a/tests/unit/proxy/management_endpoints/test_internal_user_endpoints.py b/tests/unit/proxy/management_endpoints/test_internal_user_endpoints.py index 42162087c9b..39586a1a204 100644 --- a/tests/unit/proxy/management_endpoints/test_internal_user_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_internal_user_endpoints.py @@ -4195,6 +4195,65 @@ async def test_add_new_user_to_default_team_string_teams_have_no_member_budget(m assert mock_add.call_args.kwargs["team_id"] == "string-team" +@pytest.mark.asyncio +async def test_add_new_user_to_default_team_joins_teams_in_the_order_given_and_past_a_failing_team( + mocker: MockerFixture, +) -> None: + from litellm.proxy.management_endpoints.internal_user_endpoints import ( + add_new_user_to_default_team, + ) + + requested: Final = ("team-zeta", "team-broken", "team-mu", "team-alpha") + seconds_per_team: Final = {"team-zeta": 0.03, "team-broken": 0.0, "team-mu": 0.02, "team-alpha": 0.0} + finished: Final = mocker.MagicMock() + + async def join_team(*, team_id: str, **_: object) -> None: + await asyncio.sleep(seconds_per_team[team_id]) + if team_id == "team-broken": + raise RuntimeError("team row locked") + finished(team_id) + + mocker.patch( + "litellm.proxy.management_endpoints.internal_user_endpoints._add_user_to_team", + side_effect=join_team, + ) + + await add_new_user_to_default_team( + user_id="scim-user", + user_email=None, + user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN), + teams=list(requested), + prisma_client=mocker.MagicMock(), + ) + + assert [call.args[0] for call in finished.call_args_list] == ["team-zeta", "team-mu", "team-alpha"] + + +@pytest.mark.asyncio +async def test_add_new_user_to_default_team_rejects_an_invalid_team_before_joining_any( + mocker: MockerFixture, +) -> None: + from litellm.proxy.management_endpoints.internal_user_endpoints import ( + add_new_user_to_default_team, + ) + + join_team: Final = mocker.patch( + "litellm.proxy.management_endpoints.internal_user_endpoints._add_user_to_team", + new_callable=mocker.AsyncMock, + ) + + with pytest.raises(ValueError, match="Invalid team type"): + await add_new_user_to_default_team( + user_id="new-user", + user_email=None, + user_api_key_dict=UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN), + teams=["team-zeta", {"team_id": "team-mu"}], + prisma_client=mocker.MagicMock(), + ) + + join_team.assert_not_awaited() + + @pytest.mark.asyncio async def test_add_user_to_team_logs_unknown_team_at_error(mocker, caplog): """A default team that no longer exists makes every membership write 404. diff --git a/tests/unit/proxy/management_endpoints/test_ui_sso.py b/tests/unit/proxy/management_endpoints/test_ui_sso.py index cfb64a85359..136cc86e67a 100644 --- a/tests/unit/proxy/management_endpoints/test_ui_sso.py +++ b/tests/unit/proxy/management_endpoints/test_ui_sso.py @@ -4049,8 +4049,8 @@ class TestCLIKeyRegenerationFlow: assert ( '" ) in page @@ -4530,6 +4530,39 @@ class TestCLIKeyRegenerationFlow: == () ) + @pytest.mark.asyncio + async def test_fetch_cli_sso_team_details_follows_the_user_record_order(self): + """The CLI login binds to the first detail, so it must be the first team on the user record. + + Postgres returns the rows in whatever order the table holds them, which every write to a + team row changes, so the bound team used to flip between sign-ins. + """ + from litellm.proxy.management_endpoints.ui_sso import ( + fetch_cli_sso_team_details, + ) + + def team_row(team_id: str) -> MagicMock: + row: Final = MagicMock() + row.model_dump.return_value = { + "team_id": team_id, + "team_alias": None, + "models": [], + "litellm_model_table": None, + } + return row + + prisma_client: Final = MagicMock() + prisma_client.db.litellm_teamtable.find_many = AsyncMock( + return_value=[team_row("team-alpha"), team_row("team-zeta")] + ) + + details: Final = await fetch_cli_sso_team_details( + prisma_client=prisma_client, + teams=["team-zeta", "team-gone", "team-alpha", "team-zeta"], + ) + + assert [detail.team_id for detail in details] == ["team-zeta", "team-alpha"] + @pytest.mark.asyncio async def test_cli_poll_key_mints_jwt_with_selected_team_grants(self): """The selected team's grants must reach the mint, not just its alias.""" @@ -6692,72 +6725,57 @@ class TestPKCEFunctionality: class TestAddMissingTeamMember: """Tests for the add_missing_team_member function""" + @staticmethod + def _recording_add() -> tuple[ + list[tuple[str, str | None]], + Callable[[str, NewUserResponse | LiteLLM_UserTable], Awaitable[None]], + ]: + added: Final[list[tuple[str, str | None]]] = [] + + async def add(team_id: str, user_info: NewUserResponse | LiteLLM_UserTable) -> None: + added.append((team_id, user_info.user_id)) + + return added, add + @pytest.mark.asyncio async def test_add_missing_team_member_with_new_user_response_teams_none(self): - """ - Bug reproduction: When a NewUserResponse has teams=None (new SSO user), - add_missing_team_member() should still add the user to the SSO teams. - - Currently FAILS: The function returns early when teams is None. - """ + """A new SSO user's NewUserResponse has teams=None, and the user must still be added to the SSO teams.""" from litellm.proxy._types import NewUserResponse from litellm.proxy.management_endpoints.ui_sso import add_missing_team_member - # Simulate a new SSO user - NewUserResponse has teams=None by default - new_user = NewUserResponse( - user_id="new-sso-user-123", - key="sk-xxxxx", - teams=None, # This is the default for NewUserResponse + new_user = NewUserResponse(user_id="new-sso-user-123", key="sk-xxxxx", teams=None) + added, add = self._recording_add() + + await add_missing_team_member( + user_info=new_user, + sso_teams=["team-from-entra-1", "team-from-entra-2"], + add_team_member=add, ) - sso_teams = ["team-from-entra-1", "team-from-entra-2"] - - with patch( - "litellm.proxy.management_endpoints.ui_sso.create_team_member_add_task" - ) as mock_add_task: - mock_add_task.return_value = AsyncMock() - - await add_missing_team_member(user_info=new_user, sso_teams=sso_teams) - - # Bug: This assertion currently FAILS - no teams are added - # because function returns early when teams is None - assert ( - mock_add_task.call_count == 2 - ), f"Expected 2 calls to add user to teams, but got {mock_add_task.call_count}" - called_team_ids = [call.args[0] for call in mock_add_task.call_args_list] - assert set(called_team_ids) == { - "team-from-entra-1", - "team-from-entra-2", - } + assert added == [ + ("team-from-entra-1", "new-sso-user-123"), + ("team-from-entra-2", "new-sso-user-123"), + ] @pytest.mark.asyncio async def test_add_missing_team_member_with_litellm_user_table_empty_teams(self): - """ - Control test: When a LiteLLM_UserTable has teams=[] (existing user, no teams), - add_missing_team_member() should add the user to SSO teams. - - This test PASSES because LiteLLM_UserTable defaults teams to [] not None. - """ + """An existing user with teams=[] is added to every SSO team.""" from litellm.proxy._types import LiteLLM_UserTable from litellm.proxy.management_endpoints.ui_sso import add_missing_team_member - # Existing user has teams=[] by default (not None) - existing_user = LiteLLM_UserTable( - user_id="existing-user-456", - teams=[], # Empty list, not None + existing_user = LiteLLM_UserTable(user_id="existing-user-456", teams=[]) + added, add = self._recording_add() + + await add_missing_team_member( + user_info=existing_user, + sso_teams=["team-from-entra-1", "team-from-entra-2"], + add_team_member=add, ) - sso_teams = ["team-from-entra-1", "team-from-entra-2"] - - with patch( - "litellm.proxy.management_endpoints.ui_sso.create_team_member_add_task" - ) as mock_add_task: - mock_add_task.return_value = AsyncMock() - - await add_missing_team_member(user_info=existing_user, sso_teams=sso_teams) - - # This PASSES - teams are added because teams=[] not None - assert mock_add_task.call_count == 2 + assert added == [ + ("team-from-entra-1", "existing-user-456"), + ("team-from-entra-2", "existing-user-456"), + ] @pytest.mark.asyncio async def test_add_user_to_teams_from_sso_response_new_user(self): @@ -6800,96 +6818,160 @@ class TestAddMissingTeamMember: @pytest.mark.asyncio async def test_sso_first_login_full_flow_adds_user_to_teams(self): - """ - End-to-end test: Simulates complete first-time SSO login with Entra groups. - Verifies teams are created AND user is added as a member. - """ + """A first-time SSO login with Entra groups adds the new user as a member of every group's team.""" from litellm.proxy._types import NewUserResponse from litellm.proxy.management_endpoints.ui_sso import add_missing_team_member - team_member_calls = [] + new_user = NewUserResponse(user_id="first-time-sso-user", key="sk-xxxxx", teams=None) + added, add = self._recording_add() - async def track_team_member_add(team_id, user_info): - team_member_calls.append({"team_id": team_id, "user_id": user_info.user_id}) - - # New SSO user with Entra groups - new_user = NewUserResponse( - user_id="first-time-sso-user", - key="sk-xxxxx", - teams=None, # The problematic default + await add_missing_team_member( + user_info=new_user, + sso_teams=["entra-team-alpha", "entra-team-beta"], + add_team_member=add, ) - sso_teams = ["entra-team-alpha", "entra-team-beta"] - - with patch( - "litellm.proxy.management_endpoints.ui_sso.create_team_member_add_task", - side_effect=track_team_member_add, - ): - await add_missing_team_member(user_info=new_user, sso_teams=sso_teams) - - # Bug: With current code, team_member_calls will be empty - # After fix: Should have 2 entries - assert ( - len(team_member_calls) == 2 - ), f"Expected 2 teams to be added, but got {len(team_member_calls)}" - assert {c["team_id"] for c in team_member_calls} == { - "entra-team-alpha", - "entra-team-beta", - } - assert all(c["user_id"] == "first-time-sso-user" for c in team_member_calls) + assert added == [ + ("entra-team-alpha", "first-time-sso-user"), + ("entra-team-beta", "first-time-sso-user"), + ] @pytest.mark.asyncio @pytest.mark.parametrize( - "user_info_factory,teams_value,expected_teams_added", + "user_info_factory,expected_teams_added", [ - # Bug case: NewUserResponse with teams=None pytest.param( lambda uid: NewUserResponse(user_id=uid, key="sk-xxx", teams=None), - None, - ["team-1", "team-2"], # Should still add teams + ["team-1", "team-2"], id="new_user_teams_none", ), - # Working case: LiteLLM_UserTable with teams=[] pytest.param( lambda uid: LiteLLM_UserTable(user_id=uid, teams=[]), - [], ["team-1", "team-2"], id="existing_user_empty_teams", ), - # Existing user with some teams already pytest.param( lambda uid: LiteLLM_UserTable(user_id=uid, teams=["team-1"]), - ["team-1"], - ["team-2"], # Only missing team should be added + ["team-2"], id="existing_user_partial_teams", ), + pytest.param( + lambda uid: LiteLLM_UserTable(user_id=uid, teams=["team-2", "team-1"]), + [], + id="existing_user_all_teams", + ), ], ) async def test_add_missing_team_member_handles_all_user_types( - self, user_info_factory, teams_value, expected_teams_added + self, user_info_factory, expected_teams_added ): - """ - Parametrized test ensuring add_missing_team_member works for all user types. - """ + """Only the SSO teams the user is not already on are added.""" from litellm.proxy.management_endpoints.ui_sso import add_missing_team_member - user_info = user_info_factory("test-user-id") - sso_teams = ["team-1", "team-2"] + added, add = self._recording_add() - added_teams = [] + await add_missing_team_member( + user_info=user_info_factory("test-user-id"), + sso_teams=["team-1", "team-2"], + add_team_member=add, + ) - async def mock_create_task(team_id, user): - added_teams.append(team_id) + assert added == [(team_id, "test-user-id") for team_id in expected_teams_added] - with patch( - "litellm.proxy.management_endpoints.ui_sso.create_team_member_add_task", - side_effect=mock_create_task, - ): - await add_missing_team_member(user_info=user_info, sso_teams=sso_teams) + @pytest.mark.asyncio + async def test_add_missing_team_member_lands_teams_in_claim_order(self): + """user.teams is appended to as each add finishes, and the first team on it is the one a CLI + login binds to, so the adds must finish in the claim's order even when an earlier add is slower. - assert set(added_teams) == set( - expected_teams_added - ), f"Expected teams {expected_teams_added}, but got {added_teams}" + The adds used to run concurrently from a set, so they finished in an arbitrary order. + """ + from litellm.proxy._types import LiteLLM_UserTable + from litellm.proxy.management_endpoints.ui_sso import add_missing_team_member + + claim: Final = ["team-zeta", "team-alpha", "team-held", "team-zeta", "team-mu"] + yields_before_finishing: Final = {"team-zeta": 3, "team-alpha": 2, "team-mu": 1} + finished: Final[list[str]] = [] + + async def slower_when_earlier(team_id: str, user_info: NewUserResponse | LiteLLM_UserTable) -> None: + for _ in range(yields_before_finishing[team_id]): + await asyncio.sleep(0) + finished.append(team_id) + + await add_missing_team_member( + user_info=LiteLLM_UserTable(user_id="group-sync-user", teams=["team-held"]), + sso_teams=claim, + add_team_member=slower_when_earlier, + ) + + assert finished == ["team-zeta", "team-alpha", "team-mu"] + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "existing,expected_teams_added", + [ + pytest.param(frozenset({"team-zeta", "team-alpha"}), ["team-zeta", "team-alpha"], id="only_teams"), + pytest.param(None, ["grp-everyone", "team-zeta", "grp-vpn", "team-alpha"], id="lookup_failed"), + ], + ) + async def test_add_missing_team_member_looks_up_missing_groups_once_and_adds_only_teams( + self, existing: frozenset[str] | None, expected_teams_added: list[str] + ): + from litellm.proxy.management_endpoints.ui_sso import add_missing_team_member + + added, add = self._recording_add() + lookups: Final[list[tuple[str, ...]]] = [] + + async def existing_team_ids(team_ids: Sequence[str]) -> frozenset[str] | None: + lookups.append(tuple(team_ids)) + return existing + + await add_missing_team_member( + user_info=LiteLLM_UserTable(user_id="group-sync-user", teams=["team-held"]), + sso_teams=["grp-everyone", "team-zeta", "team-held", "grp-vpn", "team-alpha"], + add_team_member=add, + existing_team_ids=existing_team_ids, + ) + + assert lookups == [("grp-everyone", "team-zeta", "grp-vpn", "team-alpha")] + assert added == [(team_id, "group-sync-user") for team_id in expected_teams_added] + + @pytest.mark.asyncio + async def test_add_missing_team_member_skips_the_lookup_when_every_group_is_held(self): + from litellm.proxy.management_endpoints.ui_sso import add_missing_team_member + + added, add = self._recording_add() + lookup: Final = AsyncMock(return_value=frozenset()) + + await add_missing_team_member( + user_info=LiteLLM_UserTable(user_id="group-sync-user", teams=["team-alpha", "team-zeta"]), + sso_teams=["team-zeta", "team-alpha"], + add_team_member=add, + existing_team_ids=lookup, + ) + + lookup.assert_not_awaited() + assert added == [] + + @pytest.mark.asyncio + async def test_find_existing_team_ids_reads_the_writer_and_returns_none_when_the_db_cannot_answer(self): + from litellm.proxy.management_endpoints.ui_sso import find_existing_team_ids + + stored_on_writer: Final = ("team-zeta", "team-alpha") + + async def find_many(*, where: Mapping[str, Mapping[str, Sequence[str]]]) -> list[SimpleNamespace]: + return [ + SimpleNamespace(team_id=team_id) for team_id in stored_on_writer if team_id in where["team_id"]["in"] + ] + + prisma: Final = MagicMock() + prisma.writer_db.litellm_teamtable.find_many = AsyncMock(side_effect=find_many) + prisma.db.litellm_teamtable.find_many = AsyncMock(return_value=[]) + with patch("litellm.proxy.proxy_server.prisma_client", prisma): + assert await find_existing_team_ids(("grp-vpn", "team-alpha")) == frozenset({"team-alpha"}) + prisma.writer_db.litellm_teamtable.find_many.side_effect = RuntimeError("db unavailable") + assert await find_existing_team_ids(("team-alpha",)) is None + with patch("litellm.proxy.proxy_server.prisma_client", None): + assert await find_existing_team_ids(("team-alpha",)) is None @pytest.mark.asyncio From a44e44b98059a9f740fd721fc475d139b721f89f Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Fri, 30 Oct 2026 05:11:42 -0700 Subject: [PATCH 23/23] fix(cache_control): pick the chat-path dialect from the deployment's provider prefix (#45295) * fix(cache_control): resolve the chat-path dialect with the deployment api_base * test(cache_control): type the chat-path resolver helpers * fix(cache_control): pick the chat-path dialect from the deployment's provider prefix Resolving with the deployment api_base followed the dispatch's azure_ai to azure flip, so chat on a non-Foundry Azure host sent Anthropic cache_control while /v1/messages, told azure_ai, sent the OpenAI breakpoint. Use the model's provider prefix when the caller does not name a provider, so chat matches the told path whatever AZURE_AI_API_BASE holds, and keep name-only resolution everywhere else. * fix(cache_control): keep litellm_proxy routing when the chat path reads the provider prefix With USE_LITELLM_PROXY or litellm.use_litellm_proxy on, every call goes to the upstream proxy, so the provider prefix no longer decides the cache format there and resolution matches main. The plain OpenAI check reads the same provider value as the hosted check, so one function no longer mixes two lookups * fix(cache_control): resolve the chat-path dialect with the deployment api_base, as the call is routed The hosted cache-format check now resolves the provider with the deployment's own api_base when the caller names none, the same provider the call dispatches to and the Responses path passes, so AZURE_AI_API_BASE no longer decides the chat format. The plain OpenAI check keeps main's name-only lookup. * fix(cache_control): send acompletion fallbacks the caller's request, not the primary's hooked kwargs acompletion handed async_completion_with_fallbacks the kwargs the cache seed and prompt hook had already rewritten for the first model, so a Claude fallback behind a Foundry GPT model received prompt_cache_options and got a 400. Fallbacks now get a copy of the caller's model, messages, tools, and kwargs taken before the hooks, as completion already does with its args. Both cache seeds also read base_url before api_base, since completion sends the call to base_url when both are set. * fix(cache_control): keep the caller's tools list intact for acompletion fallbacks The vector-store hook pops recognized file_search tools from the request's tools list in place, so the fallback request built from the pre-hook snapshot lost the retrieved context. Snapshot a copy of the list instead. * fix(cache_control): skip acompletion's prompt hooks when fallbacks will run them Each fallback attempt runs the prompt hooks for its own model, so running them first in the outer call only cost a vector-store search whose result the hand-off threw away. * fix(cache_control): run acompletion's prompt hooks for fallbacks and leave only the cache hook to each attempt --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../anthropic_cache_control_hook.py | 10 +- litellm/main.py | 40 +++- .../test_anthropic_cache_control_hook.py | 71 +++++++ .../litellm_core_utils/test_fallback_utils.py | 187 ++++++++++++++++++ 4 files changed, 294 insertions(+), 14 deletions(-) diff --git a/litellm/integrations/anthropic_cache_control_hook.py b/litellm/integrations/anthropic_cache_control_hook.py index 45c82589d2d..d816bc661f3 100644 --- a/litellm/integrations/anthropic_cache_control_hook.py +++ b/litellm/integrations/anthropic_cache_control_hook.py @@ -116,12 +116,12 @@ def configured_injection_points(value: object) -> Sequence[CacheControlInjection return tuple(cast(CacheControlInjectionPoint, entry) for entry in value if isinstance(entry, dict)) -def _served_provider_for_model(model: str) -> str | None: +def _served_provider_for_model(model: str, api_base: object = None) -> str | None: from litellm.exceptions import BadRequestError from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider try: - _, provider, _, _ = get_llm_provider(model=model) + _, provider, _, _ = get_llm_provider(model=model, api_base=api_base if isinstance(api_base, str) else None) except BadRequestError: return None return provider @@ -536,7 +536,7 @@ class AnthropicCacheControlHook(CustomPromptManagement): if model is None: return False hosted_flag: Final = _hosted_openai_dialect_flag( - model, custom_llm_provider, AnthropicCacheControlHook._resolve_provider + model, custom_llm_provider, lambda name: AnthropicCacheControlHook._resolve_provider(name, api_base) ) if hosted_flag is not None: return hosted_flag @@ -559,8 +559,8 @@ class AnthropicCacheControlHook(CustomPromptManagement): ) @staticmethod - def _resolve_provider(model: str) -> str | None: - return _served_provider_for_model(model) + def _resolve_provider(model: str, api_base: object = None) -> str | None: + return _served_provider_for_model(model, api_base) @staticmethod def count_request_cache_breakpoints( diff --git a/litellm/main.py b/litellm/main.py index 54afe654cec..5346741e4d1 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -681,7 +681,7 @@ async def acompletion( - The `completion` function is called using `run_in_executor` to execute synchronously in the event loop. - If `stream` is True, the function returns an async generator that yields completion lines. """ - fallbacks = kwargs.get("fallbacks", None) + fallbacks: Final = kwargs.get("fallbacks", None) or litellm.model_fallbacks mock_timeout = kwargs.get("mock_timeout", None) if mock_timeout is True: @@ -697,6 +697,7 @@ async def acompletion( loop: Final = asyncio.get_event_loop() custom_llm_provider = kwargs.get("custom_llm_provider", None) _ = _resolve_control_options(kwargs, model) + caller_kwargs: Final = cast("Mapping[str, object]", {**kwargs}) # cast-ok: acompletion takes untyped kwargs ## PROMPT MANAGEMENT HOOKS ## ######################################################### @@ -715,15 +716,34 @@ async def acompletion( custom_llm_provider=cast(str | None, custom_llm_provider), # cast-ok: read from untyped kwargs tools=tools, enable_prompt_caching=cast(bool | None, kwargs.get("enable_prompt_caching")), # cast-ok: untyped kwargs - api_base=kwargs.get("api_base") or base_url, + api_base=base_url or kwargs.get("api_base"), ) - if isinstance(litellm_logging_obj, LiteLLMLoggingObj) and ( - litellm_logging_obj.should_run_prompt_management_hooks( - prompt_id=kwargs.get("prompt_id", None), + prompt_id: Final = cast(str | None, kwargs.get("prompt_id", None)) # cast-ok: read from untyped kwargs + prompt_management_logger: Final = ( + litellm_logging_obj.get_custom_logger_for_prompt_management( + model=model, + non_default_params=kwargs, + tools=tools, + prompt_id=prompt_id, + dynamic_callback_params=litellm_logging_obj.standard_callback_dynamic_params, + ) + if isinstance(litellm_logging_obj, LiteLLMLoggingObj) + and litellm_logging_obj.should_run_prompt_management_hooks( + prompt_id=prompt_id, non_default_params=kwargs, tools=tools, ) + else None + ) + each_fallback_writes_its_own_cache_dialect: Final = fallbacks is not None and isinstance( + prompt_management_logger, AnthropicCacheControlHook + ) + + if ( + isinstance(litellm_logging_obj, LiteLLMLoggingObj) + and prompt_management_logger is not None + and not each_fallback_writes_its_own_cache_dialect ): ( model, @@ -733,8 +753,9 @@ async def acompletion( model=model, messages=messages, non_default_params=kwargs, - prompt_id=kwargs.get("prompt_id", None), + prompt_id=prompt_id, prompt_variables=kwargs.get("prompt_variables", None), + prompt_management_logger=prompt_management_logger, tools=tools, prompt_label=kwargs.get("prompt_label", None), prompt_version=kwargs.get("prompt_version", None), @@ -826,9 +847,10 @@ async def acompletion( litellm_params=(GenericLiteLLMParams.model_validate(_supplemental) if _supplemental else None), ) - fallbacks = fallbacks or litellm.model_fallbacks if fallbacks is not None: - response = await async_completion_with_fallbacks(**completion_kwargs, kwargs={"fallbacks": fallbacks, **kwargs}) + response = await async_completion_with_fallbacks( + **completion_kwargs, kwargs={"fallbacks": fallbacks, **caller_kwargs} + ) if response is None: raise Exception( "No response from fallbacks. Got none. Turn on `litellm.set_verbose=True` to see more details." @@ -5765,7 +5787,7 @@ def completion( custom_llm_provider=cast(str | None, kwargs.get("custom_llm_provider")), # cast-ok: untyped kwargs tools=tools, enable_prompt_caching=cast(bool | None, kwargs.get("enable_prompt_caching")), # cast-ok: untyped kwargs - api_base=kwargs.get("api_base") or base_url, + api_base=base_url or kwargs.get("api_base"), ) if isinstance(litellm_logging_obj, LiteLLMLoggingObj) and ( diff --git a/tests/unit/integrations/test_anthropic_cache_control_hook.py b/tests/unit/integrations/test_anthropic_cache_control_hook.py index c70c7080ab8..fa03879c05a 100644 --- a/tests/unit/integrations/test_anthropic_cache_control_hook.py +++ b/tests/unit/integrations/test_anthropic_cache_control_hook.py @@ -3998,6 +3998,77 @@ class TestChatPathProviderStamp: assert params["prompt_cache_options"] == {"mode": "explicit"} +class TestChatPathResolvesTheDeploymentProvider: + MODEL = "azure_ai/gpt-6-astra" + AMBIENT_API_BASE = "https://other-deployment.openai.azure.com" + FOUNDRY_API_BASE = "https://foundry.services.ai.azure.com" + UPSTREAM_PROXY_API_BASE = "https://upstream-proxy.example.com" + POINTS = [{"location": "message", "role": "system"}] + MESSAGES = [{"role": "system", "content": "sys"}, {"role": "user", "content": "hi"}] + + def _pin_ambient_api_base(self, monkeypatch: pytest.MonkeyPatch, pinned: bool) -> None: + if pinned: + monkeypatch.setenv("AZURE_AI_API_BASE", self.AMBIENT_API_BASE) + else: + monkeypatch.delenv("AZURE_AI_API_BASE", raising=False) + + def _chat( + self, custom_llm_provider: str | None, api_base: str, model: str = MODEL + ) -> tuple[list[AllMessageValues], dict[str, object]]: + seeded: Final[dict[str, object]] = {"cache_control_injection_points": copy.deepcopy(self.POINTS)} + AnthropicCacheControlHook.maybe_seed_default_injection_points( + non_default_params=seeded, + messages=copy.deepcopy(self.MESSAGES), + model=model, + custom_llm_provider=custom_llm_provider, + api_base=api_base, + ) + _, out, params = AnthropicCacheControlHook().get_chat_completion_prompt( + model=model, + messages=copy.deepcopy(self.MESSAGES), + non_default_params=seeded, + prompt_id=None, + prompt_variables=None, + dynamic_callback_params={}, + ) + return out, params + + def test_chat_path_stamps_the_breakpoint_on_a_foundry_deployment_while_the_ambient_base_is_azure_openai( + self, monkeypatch: pytest.MonkeyPatch + ) -> None: + self._pin_ambient_api_base(monkeypatch, True) + messages, params = self._chat(None, self.FOUNDRY_API_BASE) + assert (messages, params) == self._chat("azure_ai", self.FOUNDRY_API_BASE) + assert AnthropicCacheControlHook.count_request_cache_breakpoints(messages) == 1 + assert ("prompt_cache_options" in params) == supports_openai_prompt_cache_breakpoint(self.MODEL, "azure_ai") + + @pytest.mark.parametrize("pinned", [True, False]) + @pytest.mark.parametrize( + "api_base", + [FOUNDRY_API_BASE, "https://resource.cognitiveservices.azure.com", "https://resource.openai.azure.com"], + ) + def test_chat_path_marks_the_deployment_like_the_provider_its_api_base_dispatches_to( + self, monkeypatch: pytest.MonkeyPatch, api_base: str, pinned: bool + ) -> None: + self._pin_ambient_api_base(monkeypatch, pinned) + dispatched_provider: Final = litellm.get_llm_provider(model=self.MODEL, api_base=api_base)[1] + assert self._chat(None, api_base) == self._chat(dispatched_provider, api_base) + + @pytest.mark.parametrize("model", [MODEL, "bedrock_mantle/openai.gpt-5.6-sol"]) + @pytest.mark.parametrize("set_globally", ["env", "module"]) + def test_chat_path_marks_the_deployment_like_litellm_proxy_when_calls_route_to_one( + self, monkeypatch: pytest.MonkeyPatch, model: str, set_globally: str + ) -> None: + self._pin_ambient_api_base(monkeypatch, False) + if set_globally == "env": + monkeypatch.setenv("USE_LITELLM_PROXY", "True") + else: + monkeypatch.setattr(litellm, "use_litellm_proxy", True) + assert self._chat(None, self.UPSTREAM_PROXY_API_BASE, model) == self._chat( + "litellm_proxy", self.UPSTREAM_PROXY_API_BASE, model + ) + + class TestClientBreakpointsCountedOnce: def test_client_message_breakpoints_are_not_double_counted(self): messages = [ diff --git a/tests/unit/litellm_core_utils/test_fallback_utils.py b/tests/unit/litellm_core_utils/test_fallback_utils.py index 1e3ca0ba90d..c0a344cc6c3 100644 --- a/tests/unit/litellm_core_utils/test_fallback_utils.py +++ b/tests/unit/litellm_core_utils/test_fallback_utils.py @@ -1,6 +1,7 @@ """Tests for litellm.litellm_core_utils.fallback_utils.""" import json +import pathlib from typing import Final import httpx @@ -8,12 +9,14 @@ import pytest import respx import litellm +from litellm.litellm_core_utils import litellm_logging from litellm.litellm_core_utils.core_helpers import process_response_headers from litellm.litellm_core_utils.fallback_utils import ( async_completion_with_fallbacks, ) from litellm.main import acompletion as python_acompletion from litellm.main import completion as python_completion +from litellm.vector_stores.vector_store_registry import LiteLLM_ManagedVectorStore, VectorStoreRegistry @pytest.mark.asyncio @@ -444,3 +447,187 @@ def test_completion_fallbacks_sync( assert response.choices[0].message.content == "sync fallback answer" assert tuple(json.loads(call.request.content)["model"] for call in route.calls) == ("first", "second") + + +CACHE_POINTS: Final = ({"location": "message", "role": "system"},) +CACHED_MESSAGES: Final = ( + {"role": "system", "content": "You are a careful assistant."}, + {"role": "user", "content": "hello"}, +) +OPENAI_ANSWER: Final = { + "id": "chatcmpl-1", + "object": "chat.completion", + "created": 1, + "model": "gpt-6.1-sol", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, +} +ANTHROPIC_ANSWER: Final = { + "id": "msg_1", + "type": "message", + "role": "assistant", + "model": "claude-haiku-4-5", + "content": [{"type": "text", "text": "ok"}], + "stop_reason": "end_turn", + "stop_sequence": None, + "usage": {"input_tokens": 1, "output_tokens": 1}, +} + + +def _cached_request(**kwargs: object) -> dict[str, object]: + return { + "messages": [dict(message) for message in CACHED_MESSAGES], + "cache_control_injection_points": [dict(point) for point in CACHE_POINTS], + **kwargs, + } + + +@pytest.mark.asyncio +async def test_acompletion_fallback_sends_the_fallback_model_the_request_a_direct_call_sends( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + openai_route: Final = respx_mock.post("https://api.openai.com/v1/chat/completions").mock( + return_value=httpx.Response( + 401, + json={"error": {"message": "invalid key", "type": "authentication_error", "param": None, "code": None}}, + ) + ) + anthropic_route: Final = respx_mock.post("https://api.anthropic.com/v1/messages").mock( + return_value=httpx.Response(200, json=ANTHROPIC_ANSWER) + ) + + await python_acompletion(**_cached_request(model="anthropic/claude-haiku-4-5", api_key="anthropic-key")) + await python_acompletion( + **_cached_request( + model="openai/gpt-6.1-sol", + api_key="openai-key", + fallbacks=[{"model": "anthropic/claude-haiku-4-5", "api_key": "anthropic-key"}], + ) + ) + + assert "prompt_cache_options" in json.loads(openai_route.calls[0].request.content) + direct_body, fallback_body = (json.loads(call.request.content) for call in anthropic_route.calls) + assert fallback_body == direct_body + + +@pytest.mark.parametrize("is_async", [True, False]) +@pytest.mark.asyncio +async def test_cache_dialect_follows_base_url_over_an_api_base_the_call_does_not_use( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, is_async: bool +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + route: Final = respx_mock.post("https://api.openai.com/v1/chat/completions").mock( + return_value=httpx.Response(200, json=OPENAI_ANSWER) + ) + + async def send(**kwargs: object) -> None: + request: Final = _cached_request( + model="openai/gpt-6.1-sol", api_key="openai-key", base_url="https://api.openai.com/v1", **kwargs + ) + if is_async: + await python_acompletion(**request) + else: + python_completion(**request) + + await send() + await send(api_base="https://unused-gateway.example.com/v1") + + base_url_only, with_unused_api_base = (json.loads(call.request.content) for call in route.calls) + assert with_unused_api_base == base_url_only + + +@pytest.mark.asyncio +async def test_acompletion_fallback_sends_the_context_a_vector_store_tool_retrieves( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr( + litellm, + "vector_store_registry", + VectorStoreRegistry( + vector_stores=[ + LiteLLM_ManagedVectorStore( + vector_store_id="vs_docs", custom_llm_provider="openai", litellm_params={"api_key": "openai-key"} + ) + ] + ), + ) + search_route: Final = respx_mock.post("https://api.openai.com/v1/vector_stores/vs_docs/search").mock( + return_value=httpx.Response( + 200, + json={ + "object": "vector_store.search_results.page", + "search_query": "what is litellm?", + "data": [{"score": 1.0, "content": [{"type": "text", "text": "litellm is an LLM gateway"}]}], + }, + ) + ) + openai_route: Final = respx_mock.post("https://api.openai.com/v1/chat/completions").mock( + return_value=httpx.Response( + 401, + json={"error": {"message": "invalid key", "type": "authentication_error", "param": None, "code": None}}, + ) + ) + anthropic_route: Final = respx_mock.post("https://api.anthropic.com/v1/messages").mock( + return_value=httpx.Response(200, json=ANTHROPIC_ANSWER) + ) + + async def send(**kwargs: object) -> None: + await python_acompletion( + messages=[{"role": "user", "content": "what is litellm?"}], + tools=[{"type": "file_search", "vector_store_ids": ["vs_docs"]}], + **kwargs, + ) + + await send(model="anthropic/claude-haiku-4-5", api_key="anthropic-key") + await send( + model="openai/gpt-6.1-sol", + api_key="openai-key", + fallbacks=[{"model": "anthropic/claude-haiku-4-5", "api_key": "anthropic-key"}], + ) + + direct_body, fallback_body = (json.loads(call.request.content) for call in anthropic_route.calls) + assert "litellm is an LLM gateway" in json.dumps(direct_body) + assert fallback_body == direct_body + assert search_route.call_count <= openai_route.call_count + anthropic_route.call_count + + +@pytest.mark.asyncio +async def test_acompletion_fallback_sends_the_prompt_the_primary_model_prefix_template_renders( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, tmp_path: pathlib.Path +) -> None: + _ = (tmp_path / "brief.prompt").write_text( + "---\ninput:\n schema:\n question: string\n---\nSystem: Answer in one word.\n\nUser: {{question}}\n" + ) + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.setattr(litellm, "global_prompt_directory", str(tmp_path)) + monkeypatch.setattr(litellm_logging, "_in_memory_loggers", []) + _ = respx_mock.post("https://api.openai.com/v1/chat/completions").mock( + return_value=httpx.Response( + 401, + json={"error": {"message": "invalid key", "type": "authentication_error", "param": None, "code": None}}, + ) + ) + anthropic_route: Final = respx_mock.post("https://api.anthropic.com/v1/messages").mock( + return_value=httpx.Response(200, json=ANTHROPIC_ANSWER) + ) + + async def send(**kwargs: object) -> None: + await python_acompletion( + prompt_id="brief", + prompt_variables={"question": "What color is the sky?"}, + messages=[{"role": "user", "content": "placeholder"}], + **kwargs, + ) + + await send(model="dotprompt/anthropic/claude-haiku-4-5", api_key="anthropic-key") + await send( + model="dotprompt/openai/gpt-6.1-sol", + api_key="openai-key", + fallbacks=[{"model": "anthropic/claude-haiku-4-5", "api_key": "anthropic-key"}], + ) + + direct_body, fallback_body = (json.loads(call.request.content) for call in anthropic_route.calls) + assert "What color is the sky?" in json.dumps(direct_body) + assert fallback_body == direct_body