fix(deepseek): forward output_config.effort and thinking as sent on /v1/messages

This commit is contained in:
mateo-berri 2026-10-30 01:59:57 -07:00
parent 73aaf9e3da
commit df53f7a615
2 changed files with 50 additions and 0 deletions

View file

@ -42,6 +42,10 @@ class DeepSeekAnthropicMessagesConfig(AnthropicMessagesConfig):
It accepts the native Anthropic Messages conversation shape, including
thinking blocks in assistant history, but rejects Anthropic's explicit
custom-tool discriminator (`{"type": "custom"}`).
It reads `thinking` and `output_config.effort` natively (ignoring
`budget_tokens`) and accepts any `temperature`, so the Claude-specific
effort and temperature reshaping is skipped.
"""
@property
@ -115,6 +119,18 @@ class DeepSeekAnthropicMessagesConfig(AnthropicMessagesConfig):
return join_api_base(base_url, "v1/messages")
@staticmethod
def _translate_adaptive_effort_for_non_adaptive_model(
model: str, optional_params: Mapping[str, object], max_tokens: int | None, custom_llm_provider: str
) -> None:
return None
@staticmethod
def _drop_incompatible_temperature_for_thinking(
model: str, optional_params: Mapping[str, object], custom_llm_provider: str
) -> None:
return None
@staticmethod
def _sanitize_tools_for_deepseek(tools: object) -> object:
if not isinstance(tools, list):

View file

@ -1,3 +1,5 @@
import pytest
import litellm
from litellm.llms.anthropic.messages.native.transformation import (
AnthropicMessagesConfig,
@ -187,3 +189,35 @@ def test_deepseek_anthropic_messages_preserves_thinking_and_sanitizes_custom_too
"input_schema": {"type": "object"},
}
assert request["tools"][1]["type"] == "web_search_20260209"
def _deepseek_messages_request(optional_params: dict) -> dict:
return DeepSeekAnthropicMessagesConfig().transform_anthropic_messages_request(
model="deepseek-v4-pro",
messages=[{"role": "user", "content": "Hi"}],
anthropic_messages_optional_request_params={"max_tokens": 32000, **optional_params},
litellm_params=GenericLiteLLMParams(),
headers={},
)
@pytest.mark.parametrize("effort", ["max", "low", "none", "ultra"])
def test_deepseek_anthropic_messages_forwards_effort_as_sent(effort):
request = _deepseek_messages_request({"output_config": {"effort": effort}})
assert request["output_config"] == {"effort": effort}
assert "thinking" not in request
def test_deepseek_anthropic_messages_forwards_adaptive_thinking_with_effort():
request = _deepseek_messages_request({"thinking": {"type": "adaptive"}, "output_config": {"effort": "max"}})
assert request["thinking"] == {"type": "adaptive"}
assert request["output_config"] == {"effort": "max"}
def test_deepseek_anthropic_messages_keeps_temperature_with_effort_none():
request = _deepseek_messages_request({"output_config": {"effort": "none"}, "temperature": 0})
assert request["output_config"] == {"effort": "none"}
assert request["temperature"] == 0