fix(decisions): stop forwarding caller providerOptions to Vercel and send Jev images

Backport of #50642 to lts/1.108.x.
Cherry-picked from c8122bc6a5 (mishdotdev:vercel-ai-gateway-decisions).
Adapted to the line: decisions/call.py goes back to the line's own version, since the line never had the per-model support record main edits here, and the discount and margin test this commit touches on main is not on the line.
This commit is contained in:
mateo-berri 2026-11-09 17:29:32 -08:00
parent 0222d0bdf4
commit 9bed368bfd
6 changed files with 27 additions and 76 deletions

View file

@ -294,15 +294,12 @@ def _unsupported_params(
custom_llm_provider: str,
headers: Mapping[str, str],
supports_box_questions: bool,
supports_images: bool,
) -> tuple[str, ...]:
refuses_safety_identifier: Final = (
ir_request.safety_identifier is not None and not provider_config.supports_safety_identifier
)
refused_image_param: Final = (
None
if supports_images and provider_config.supports_images_in(ir_request.input, headers)
else image_param(ir_request.input)
None if provider_config.supports_images_in(ir_request.input, headers) else image_param(ir_request.input)
)
refuses_assets: Final = bool(ir_request.assets) and not provider_config.supports_assets(headers)
refuses_box_questions: Final = _has_box_question(ir_request.questions) and not supports_box_questions
@ -335,13 +332,12 @@ def _without_unsupported_params(
provider_config: BaseDecisionsConfig,
headers: Mapping[str, str],
supports_box_questions: bool,
supports_images: bool,
) -> DecisionsIRRequest:
return replace(
ir_request,
input=(
ir_request.input
if supports_images and provider_config.supports_images_in(ir_request.input, headers)
if provider_config.supports_images_in(ir_request.input, headers)
else with_images_where(ir_request.input, _no_image)
),
safety_identifier=ir_request.safety_identifier if provider_config.supports_safety_identifier else None,
@ -395,17 +391,14 @@ def _provider_ir_request(
supports_box_questions: Final = provider_config.supports_box_questions(
model, custom_llm_provider, headers, _deployment_supports_box_questions(kwargs)
)
supports_images: Final = provider_config.supports_image_input(model, custom_llm_provider)
unsupported: Final = _unsupported_params(
ir_request, provider_config, model, custom_llm_provider, headers, supports_box_questions, supports_images
ir_request, provider_config, model, custom_llm_provider, headers, supports_box_questions
)
if not unsupported:
return ir_request
if not _drops_params(kwargs):
return _UnsupportedParams(names=unsupported)
supported: Final = _without_unsupported_params(
ir_request, provider_config, headers, supports_box_questions, supports_images
)
supported: Final = _without_unsupported_params(ir_request, provider_config, headers, supports_box_questions)
if _leaves_no_input(supported.input):
return _NothingLeftAfterDrop(names=unsupported, missing="input to decide on")
if not supported.questions:
@ -690,7 +683,7 @@ def _request_body(call: DecisionsCall) -> Mapping[str, object]:
model=call.requested_model,
llm_provider=call.custom_llm_provider,
)
return {**body, **call.provider_config.provider_request_fields(call.litellm_params)}
return body
def _format_response(response: DecisionsIRResponse, call: DecisionsCall) -> DecisionsResponse | OpenAIDecisionResponse:

View file

@ -63,7 +63,6 @@ _SYSTEMONE_RESPONSE_ADAPTER: Final[TypeAdapter[DecisionsResponse]] = TypeAdapter
_RESERVED_HEADERS: Final[frozenset[str]] = frozenset({"authorization", "content-type"})
_REMOTE_URL_PREFIXES: Final = ("http://", "https://")
_DATA_URL_PREFIX: Final = "data:"
_NO_PROVIDER_FIELDS: Final[Mapping[str, object]] = MappingProxyType({})
def decisions_text(value: DecisionsJSON) -> str:
@ -542,12 +541,6 @@ class BaseDecisionsConfig(ABC):
) -> bool:
return self.supports_images
def supports_image_input(self, model: str, custom_llm_provider: str) -> bool:
return True
def provider_request_fields(self, litellm_params: Mapping[str, object]) -> Mapping[str, object]:
return _NO_PROVIDER_FIELDS
def supports_assets(self, headers: Mapping[str, str]) -> bool:
return False

View file

@ -1,7 +1,6 @@
from collections.abc import Mapping
from dataclasses import replace
from itertools import chain
from types import MappingProxyType
from typing import Final, Literal, TypeVar
from pydantic import BaseModel, ConfigDict, Field, ValidationError
@ -9,10 +8,8 @@ from pydantic import BaseModel, ConfigDict, Field, ValidationError
from litellm.llms.base_llm.decisions.transformation import usage_reported_cost
from litellm.llms.openai.decisions.transformation import OpenAIDecisionsConfig
from litellm.types.decisions import DecisionsIRRequest, DecisionsIRResponse, DecisionsIRUsage, OpenAIDecisionUsage
from litellm.utils import is_vision_explicitly_disabled
_Model: Final = TypeVar("_Model", bound=BaseModel)
_PROVIDER_OPTIONS: Final = "providerOptions"
class _ProviderAttempt(BaseModel):
@ -95,14 +92,6 @@ class VercelAIGatewayDecisionsConfig(OpenAIDecisionsConfig):
def resolve_api_key(self, api_key: str | None) -> str | None:
return api_key
def supports_image_input(self, model: str, custom_llm_provider: str) -> bool:
return not is_vision_explicitly_disabled(model, custom_llm_provider)
def provider_request_fields(self, litellm_params: Mapping[str, object]) -> Mapping[str, object]:
if _PROVIDER_OPTIONS not in litellm_params:
return MappingProxyType({})
return MappingProxyType({_PROVIDER_OPTIONS: litellm_params[_PROVIDER_OPTIONS]})
def parse_response(self, payload: object, request: DecisionsIRRequest) -> DecisionsIRResponse:
response: Final = super().parse_response(payload, request)
metadata: Final = _validated(_GatewayDecisionResponse, payload)

View file

@ -51269,13 +51269,17 @@
"litellm_provider": "vercel_ai_gateway",
"max_input_tokens": 32000,
"metadata": {
"notes": "Vercel AI Gateway model list entry typesafe-ai/jev, read 2026-11-09: input 0.000000042 per token, output 0, context window 32000. A direct /v1/decisions probe with an input_image answered 400 invalid request body on 2026-11-09."
"notes": "Vercel AI Gateway model list entry typesafe-ai/jev, read 2026-11-09: input 0.000000042 per token, output 0, context window 32000. A direct /v1/decisions probe with a 768x768 input_image answered 200 on 2026-11-10, while a 96x96 one answered 400 invalid request body."
},
"mode": "evaluation",
"output_cost_per_token": 0.0,
"source": "https://ai-gateway.vercel.sh/v1/models",
"supported_modalities": [
"text",
"image"
],
"supports_box_questions": true,
"supports_vision": false
"supports_vision": true
},
"vercel_ai_gateway/typesafe-ai/jev-1.13.0": {
"input_cost_per_token": 4.2e-08,

View file

@ -51269,13 +51269,17 @@
"litellm_provider": "vercel_ai_gateway",
"max_input_tokens": 32000,
"metadata": {
"notes": "Vercel AI Gateway model list entry typesafe-ai/jev, read 2026-11-09: input 0.000000042 per token, output 0, context window 32000. A direct /v1/decisions probe with an input_image answered 400 invalid request body on 2026-11-09."
"notes": "Vercel AI Gateway model list entry typesafe-ai/jev, read 2026-11-09: input 0.000000042 per token, output 0, context window 32000. A direct /v1/decisions probe with a 768x768 input_image answered 200 on 2026-11-10, while a 96x96 one answered 400 invalid request body."
},
"mode": "evaluation",
"output_cost_per_token": 0.0,
"source": "https://ai-gateway.vercel.sh/v1/models",
"supported_modalities": [
"text",
"image"
],
"supports_box_questions": true,
"supports_vision": false
"supports_vision": true
},
"vercel_ai_gateway/typesafe-ai/jev-1.13.0": {
"input_cost_per_token": 4.2e-08,

View file

@ -3493,49 +3493,18 @@ async def test_vercel_ai_gateway_decisions_price_an_unreadable_charge_from_the_c
@pytest.mark.asyncio
@pytest.mark.usefixtures("vercel_gateway_key")
@pytest.mark.parametrize(
("model", "drop_params", "sent_content"),
(
(
"typesafe-ai/jev-1.14.0",
False,
[
{"type": "input_text", "text": _SCREEN_TEXT},
{"type": "input_image", "image_url": _PNG_DATA_URL, "detail": "high"},
],
),
("typesafe-ai/jev", True, [{"type": "input_text", "text": _SCREEN_TEXT}]),
),
ids=("model_reads_images", "text_only_model_drops_them"),
)
async def test_vercel_ai_gateway_decisions_send_images_only_to_models_that_read_them(
model: str, drop_params: bool, sent_content: list[Mapping[str, object]], respx_mock: respx.MockRouter
) -> None:
@pytest.mark.parametrize("model", ("typesafe-ai/jev", "typesafe-ai/jev-1.13.0", "typesafe-ai/jev-1.14.0"))
async def test_vercel_ai_gateway_jev_decisions_send_images(model: str, respx_mock: respx.MockRouter) -> None:
route: Final = respx_mock.post(_VERCEL_DECISIONS_URL).respond(json=_vercel_response(model, {}))
await litellm.adecisions(
model=f"vercel_ai_gateway/{model}",
input=_screen_messages(_PNG_DATA_URL),
questions=_PREDICATE_QUESTIONS,
drop_params=drop_params,
model=f"vercel_ai_gateway/{model}", input=_screen_messages(_PNG_DATA_URL), questions=_PREDICATE_QUESTIONS
)
assert json.loads(route.calls.last.request.content)["input"][0]["content"] == sent_content
@pytest.mark.asyncio
@pytest.mark.usefixtures("vercel_gateway_key")
async def test_vercel_ai_gateway_text_only_decisions_model_refuses_images_before_http(
respx_mock: respx.MockRouter,
) -> None:
with pytest.raises(litellm.UnsupportedParamsError, match="input_image"):
await litellm.adecisions(
model="vercel_ai_gateway/typesafe-ai/jev",
input=_screen_messages(_PNG_DATA_URL),
questions=_PREDICATE_QUESTIONS,
)
assert len(respx_mock.calls) == 0
assert json.loads(route.calls.last.request.content)["input"][0]["content"] == [
{"type": "input_text", "text": _SCREEN_TEXT},
{"type": "input_image", "image_url": _PNG_DATA_URL, "detail": "high"},
]
@pytest.mark.asyncio
@ -3560,8 +3529,7 @@ async def test_vercel_ai_gateway_decisions_inline_remote_images(
@pytest.mark.asyncio
@pytest.mark.usefixtures("vercel_gateway_key")
async def test_vercel_ai_gateway_decisions_forward_provider_options(respx_mock: respx.MockRouter) -> None:
provider_options: Final = {"gateway": {"order": ["openai"], "byok": {"openai": [{"apiKey": "caller-openai-key"}]}}}
async def test_vercel_ai_gateway_decisions_never_send_caller_provider_options(respx_mock: respx.MockRouter) -> None:
route: Final = respx_mock.post(_VERCEL_DECISIONS_URL).respond(
json=_vercel_response("openai/gpt-6-luna-decisions", {})
)
@ -3570,10 +3538,10 @@ async def test_vercel_ai_gateway_decisions_forward_provider_options(respx_mock:
model="vercel_ai_gateway/openai/gpt-6-luna-decisions",
input="The package arrived broken.",
questions=_PREDICATE_QUESTIONS,
providerOptions=provider_options,
providerOptions={"gateway": {"models": ["openai/gpt-6"], "byok": {"openai": [{"apiKey": "caller-openai-key"}]}}},
)
assert json.loads(route.calls.last.request.content)["providerOptions"] == provider_options
assert "providerOptions" not in json.loads(route.calls.last.request.content)
@pytest.mark.asyncio