mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-11 08:38:24 +00:00
d489c04554
* fix(anthropic): drop unknown legacy thinking keys under drop_params when rebuilding adaptive thinking * fix(bedrock): drop unknown thinking keys under drop_params on display updates rebuilds Bedrock Invoke copied every caller thinking key when adding display updates, so an unknown key still reached the provider with drop_params on. Filter that copy the same way, and cover each chat call site plus the global litellm.drop_params path in tests * fix(anthropic): drop budget_tokens from rebuilt adaptive thinking under drop_params * test(anthropic): cover drop_params sources for rebuilt adaptive thinking on /v1/messages and Bedrock Invoke
107 lines
4.3 KiB
Python
107 lines
4.3 KiB
Python
import json
|
|
from typing import Final
|
|
|
|
import pytest
|
|
from integration._support.client import Gateway
|
|
from integration._support.wire import Reply, Request, wire_server
|
|
from pydantic import JsonValue
|
|
|
|
MODEL: Final = "bedrock/invoke/us.anthropic.claude-opus-4-8"
|
|
TOKEN: Final = "synthetic-bedrock-bearer"
|
|
RESPONSE: Final = json.dumps(
|
|
{
|
|
"id": "msg_adaptive_control",
|
|
"type": "message",
|
|
"role": "assistant",
|
|
"model": "us.anthropic.claude-opus-4-8",
|
|
"content": [{"type": "text", "text": "adaptive thinking control"}],
|
|
"stop_reason": "end_turn",
|
|
"stop_sequence": None,
|
|
"usage": {"input_tokens": 12, "output_tokens": 5},
|
|
}
|
|
).encode()
|
|
|
|
|
|
def adaptive_peer(request: Request) -> Reply:
|
|
assert request.method == "POST" and request.target == "/model/us.anthropic.claude-opus-4-8/invoke"
|
|
assert request.headers["authorization"] == f"Bearer {TOKEN}"
|
|
body: Final = json.loads(request.body)
|
|
assert body["messages"] == [{"role": "user", "content": [{"type": "text", "text": "synthetic effort request"}]}]
|
|
assert body["thinking"]["type"] == "adaptive", body
|
|
assert body["output_config"] == {"effort": "high"}, body
|
|
assert "budget_tokens" not in json.dumps(body), body
|
|
return Reply(body=RESPONSE)
|
|
|
|
|
|
@pytest.mark.covers("other.provider_wire.bedrock.prefixed_opus_4_8_reasoning_effort_sends_adaptive_thinking")
|
|
def test_prefixed_opus_4_8_reasoning_effort_reaches_bedrock_as_adaptive_thinking_not_budget_tokens(
|
|
gateway: Gateway,
|
|
) -> None:
|
|
with wire_server(adaptive_peer) as wire, gateway.scenario() as scenario:
|
|
model: Final = scenario.model(
|
|
model=MODEL,
|
|
api_key=TOKEN,
|
|
aws_region_name="us-east-1",
|
|
api_base=wire.url,
|
|
aws_bedrock_runtime_endpoint=wire.url,
|
|
)
|
|
response: Final = gateway.request(
|
|
"POST",
|
|
"/v1/chat/completions",
|
|
{
|
|
"model": model,
|
|
"messages": [{"role": "user", "content": "synthetic effort request"}],
|
|
"max_tokens": 4096,
|
|
"reasoning_effort": "high",
|
|
},
|
|
)
|
|
assert response.status_code == 200, response.text
|
|
assert response.json()["choices"][0]["message"]["content"] == "adaptive thinking control"
|
|
assert response.json()["usage"]["prompt_tokens"] == 12 and response.json()["usage"]["completion_tokens"] == 5
|
|
assert len(wire.drain()) == 1
|
|
|
|
|
|
BLOCK_BINDING: Final = {"prefix_mismatch_behavior": "error"}
|
|
DISPLAY_UPDATES_THINKING: Final = {"type": "adaptive", "display": "updates", "block_binding": BLOCK_BINDING}
|
|
NON_ADAPTIVE_THINKING_KEYS: Final = {"budget_tokens": 2048, "not_a_thinking_key": True}
|
|
|
|
|
|
def recorded_reply(_request: Request) -> Reply:
|
|
return Reply(body=RESPONSE)
|
|
|
|
|
|
@pytest.mark.parametrize("route", ("/v1/chat/completions", "/v1/messages"))
|
|
@pytest.mark.parametrize(
|
|
("client_drop_params", "carried_keys"),
|
|
(
|
|
pytest.param({"drop_params": True}, {}, id="drop-params-sends-only-adaptive-fields"),
|
|
pytest.param({}, NON_ADAPTIVE_THINKING_KEYS, id="without-drop-params-forwards-every-key"),
|
|
),
|
|
)
|
|
def test_bedrock_display_updates_rebuild_drops_non_adaptive_thinking_keys_only_under_drop_params(
|
|
gateway: Gateway, route: str, client_drop_params: dict[str, JsonValue], carried_keys: dict[str, JsonValue]
|
|
) -> None:
|
|
with wire_server(recorded_reply) as wire, gateway.scenario() as scenario:
|
|
model: Final = scenario.model(
|
|
model=MODEL,
|
|
api_key=TOKEN,
|
|
aws_region_name="us-east-1",
|
|
api_base=wire.url,
|
|
aws_bedrock_runtime_endpoint=wire.url,
|
|
)
|
|
response: Final = gateway.request(
|
|
"POST",
|
|
route,
|
|
{
|
|
"model": model,
|
|
"messages": [{"role": "user", "content": "synthetic display updates request"}],
|
|
"max_tokens": 4096,
|
|
"thinking": {**DISPLAY_UPDATES_THINKING, **NON_ADAPTIVE_THINKING_KEYS},
|
|
**client_drop_params,
|
|
},
|
|
)
|
|
assert response.status_code == 200, response.text
|
|
received: Final = wire.drain()
|
|
assert len(received) == 1, received
|
|
assert received[0].target == "/model/us.anthropic.claude-opus-4-8/invoke", received[0].target
|
|
assert json.loads(received[0].body)["thinking"] == {**DISPLAY_UPDATES_THINKING, **carried_keys}
|