mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-16 09:22:22 +00:00
fix(responses): drop a tool_choice naming a tool the request does not send, and run the unknown-effort drop on Azure o-series
This commit is contained in:
parent
e2e7153896
commit
db720da71f
4 changed files with 145 additions and 8 deletions
|
|
@ -8,6 +8,7 @@ Translations handled by LiteLLM:
|
|||
- Other parameters follow base Azure OpenAI Responses API behavior
|
||||
"""
|
||||
|
||||
from collections.abc import Collection, Mapping
|
||||
from typing import TYPE_CHECKING, Any, Final
|
||||
|
||||
from litellm._logging import verbose_logger
|
||||
|
|
@ -51,12 +52,14 @@ class AzureOpenAIOSeriesResponsesAPIConfig(AzureOpenAIResponsesAPIConfig):
|
|||
|
||||
return o_series_supported_params
|
||||
|
||||
def map_openai_params(
|
||||
def map_openai_params_for_deployment(
|
||||
self,
|
||||
response_api_optional_params: ResponsesAPIOptionalRequestParams,
|
||||
model: str,
|
||||
drop_params: bool,
|
||||
) -> dict:
|
||||
*,
|
||||
declared_reasoning_efforts: Collection[str] | None,
|
||||
) -> Mapping[str, object]:
|
||||
"""
|
||||
Map OpenAI parameters for Azure OpenAI O-series Responses API.
|
||||
|
||||
|
|
@ -65,13 +68,20 @@ class AzureOpenAIOSeriesResponsesAPIConfig(AzureOpenAIResponsesAPIConfig):
|
|||
"""
|
||||
if not drop_params:
|
||||
return dict(response_api_optional_params)
|
||||
kept: Final = self._well_formed_params(response_api_optional_params)
|
||||
kept: Final = self.params_dropped_for_deployment(
|
||||
response_api_optional_params, model, declared_reasoning_efforts=declared_reasoning_efforts
|
||||
)
|
||||
if "temperature" in kept:
|
||||
verbose_logger.debug(
|
||||
"Dropping unsupported parameter 'temperature' for Azure OpenAI O-series responses API model %s", model
|
||||
)
|
||||
return {name: value for name, value in kept.items() if name != "temperature"}
|
||||
|
||||
def _reasoning_refusal(
|
||||
self, reasoning: object, model: str, declared_reasoning_efforts: Collection[str] | None
|
||||
) -> None:
|
||||
return None
|
||||
|
||||
def is_o_series_model(self, model: str) -> bool:
|
||||
"""
|
||||
Check if the model is an O-series model.
|
||||
|
|
|
|||
|
|
@ -1,6 +1,7 @@
|
|||
from collections.abc import Callable, Collection, Iterable, Mapping, Sequence
|
||||
from dataclasses import dataclass
|
||||
from functools import lru_cache
|
||||
from itertools import chain
|
||||
from types import MappingProxyType
|
||||
from typing import TYPE_CHECKING, Any, Final, Protocol, TypeAlias, cast, get_args, get_type_hints
|
||||
|
||||
|
|
@ -325,13 +326,50 @@ _PROVIDERS_CHECKED_FOR_MALFORMED_VALUES: Final = (LlmProviders.OPENAI, LlmProvid
|
|||
_REASONING_EFFORTS: Final = frozenset(get_args(REASONING_EFFORT))
|
||||
|
||||
|
||||
_NON_HOSTED_TOOL_TYPES: Final = frozenset({"function", "custom", "mcp", "namespace"})
|
||||
|
||||
|
||||
def _tool_mappings(tools: object) -> tuple[Mapping[str, object], ...]:
|
||||
return tuple(mapping for mapping in map(as_str_mapping, _json_array_or_empty(tools)) if mapping is not None)
|
||||
|
||||
|
||||
def _sent_tools(tools: object) -> tuple[Mapping[str, object], ...]:
|
||||
top_level: Final = _tool_mappings(tools)
|
||||
namespaced: Final = (_tool_mappings(tool.get("tools")) for tool in top_level if tool.get("type") == "namespace")
|
||||
return top_level + tuple(chain.from_iterable(namespaced))
|
||||
|
||||
|
||||
def _names_a_sent_tool(tool_choice: object, tools: tuple[Mapping[str, object], ...]) -> bool:
|
||||
choice: Final = as_str_mapping(tool_choice)
|
||||
if choice is None:
|
||||
return True
|
||||
match choice.get("type"):
|
||||
case "allowed_tools":
|
||||
allowed: Final = choice.get("tools")
|
||||
entries: Final = _json_array_or_empty(allowed)
|
||||
return not isinstance(allowed, list | tuple) or all(_names_a_sent_tool(entry, tools) for entry in entries)
|
||||
case "function":
|
||||
return any(tool.get("type") == "function" and tool.get("name") == choice.get("name") for tool in tools)
|
||||
case "custom":
|
||||
return any(tool.get("type") == "custom" for tool in tools)
|
||||
case "mcp":
|
||||
return any(
|
||||
tool.get("type") == "mcp" and tool.get("server_label") == choice.get("server_label") for tool in tools
|
||||
)
|
||||
case str():
|
||||
return any(tool.get("type") not in _NON_HOSTED_TOOL_TYPES for tool in tools)
|
||||
case _:
|
||||
return True
|
||||
|
||||
|
||||
def _accepts_tool_choice(tool_choice: object, params: Mapping[str, object]) -> bool:
|
||||
if tool_choice is None or params.get("prompt") is not None:
|
||||
return True
|
||||
tools: Final = params.get("tools")
|
||||
if isinstance(tool_choice, str):
|
||||
return tool_choice in _TOOL_CHOICES_WITHOUT_TOOLS or (tool_choice == "required" and bool(tools))
|
||||
return bool(tools)
|
||||
listed: Final = _json_array_or_empty(tools)
|
||||
return bool(tools) and (not isinstance(tools, list | tuple) or _names_a_sent_tool(tool_choice, _sent_tools(listed)))
|
||||
|
||||
|
||||
def _is_well_formed(name: str, value: object, params: Mapping[str, object]) -> bool:
|
||||
|
|
|
|||
|
|
@ -3306,6 +3306,9 @@ class TestReasoningFollowsModelSupport:
|
|||
|
||||
|
||||
_WEATHER_TOOL: Final = {"type": "function", "name": "get_weather", "parameters": {"type": "object", "properties": {}}}
|
||||
_SQL_TOOL: Final = {"type": "custom", "name": "run_sql"}
|
||||
_DOCS_MCP_TOOL: Final = {"type": "mcp", "server_label": "docs", "server_url": "https://example.com/mcp"}
|
||||
_WEATHER_NAMESPACE: Final = {"type": "namespace", "name": "weather", "description": "Weather", "tools": [_WEATHER_TOOL]}
|
||||
|
||||
|
||||
class TestMalformedParamsFollowDropParams:
|
||||
|
|
@ -3320,6 +3323,24 @@ class TestMalformedParamsFollowDropParams:
|
|||
pytest.param({"tool_choice": "required"}, id="required-without-tools"),
|
||||
pytest.param({"tool_choice": "required", "tools": []}, id="required-with-empty-tools"),
|
||||
pytest.param({"tool_choice": {"type": "function", "name": "get_weather"}}, id="function-without-tools"),
|
||||
pytest.param(
|
||||
{"tool_choice": {"type": "function", "name": "get_time"}, "tools": [_WEATHER_TOOL]},
|
||||
id="function-naming-a-tool-not-sent",
|
||||
),
|
||||
pytest.param({"tool_choice": {"type": "custom", "name": "run_sql"}, "tools": [_WEATHER_TOOL]}, id="custom"),
|
||||
pytest.param({"tool_choice": {"type": "mcp", "server_label": "wiki"}, "tools": [_DOCS_MCP_TOOL]}, id="mcp"),
|
||||
pytest.param({"tool_choice": {"type": "file_search"}, "tools": [_WEATHER_TOOL]}, id="hosted-tool-not-sent"),
|
||||
pytest.param(
|
||||
{
|
||||
"tool_choice": {
|
||||
"type": "allowed_tools",
|
||||
"mode": "auto",
|
||||
"tools": [{"type": "function", "name": "get_time"}],
|
||||
},
|
||||
"tools": [_WEATHER_TOOL],
|
||||
},
|
||||
id="allowed-tools-naming-a-tool-not-sent",
|
||||
),
|
||||
pytest.param({"parallel_tool_calls": "true"}, id="parallel-tool-calls-string"),
|
||||
pytest.param({"previous_response_id": 12}, id="previous-response-id-int"),
|
||||
pytest.param({"previous_response_id": ""}, id="previous-response-id-empty"),
|
||||
|
|
@ -3363,11 +3384,50 @@ class TestMalformedParamsFollowDropParams:
|
|||
pytest.param({"parallel_tool_calls": None}, id="parallel-tool-calls-null"),
|
||||
pytest.param({"previous_response_id": "resp_abc"}, id="previous-response-id-string"),
|
||||
pytest.param({"tool_choice": "required", "prompt": {"id": "pmpt_abc"}}, id="required-with-stored-prompt"),
|
||||
pytest.param(
|
||||
{"tool_choice": {"type": "function", "name": "get_weather"}, "tools": [_WEATHER_NAMESPACE]},
|
||||
id="function-in-a-namespace",
|
||||
),
|
||||
pytest.param(
|
||||
{"tool_choice": {"type": "custom", "name": "other"}, "tools": [_SQL_TOOL]},
|
||||
id="custom-with-a-custom-tool",
|
||||
),
|
||||
pytest.param(
|
||||
{"tool_choice": {"type": "mcp", "server_label": "docs"}, "tools": [_DOCS_MCP_TOOL]}, id="mcp-label-sent"
|
||||
),
|
||||
pytest.param(
|
||||
{"tool_choice": {"type": "computer_use"}, "tools": [{"type": "computer"}]},
|
||||
id="hosted-alias-of-a-sent-tool",
|
||||
),
|
||||
pytest.param(
|
||||
{
|
||||
"tool_choice": {
|
||||
"type": "allowed_tools",
|
||||
"mode": "required",
|
||||
"tools": [{"type": "function", "name": "get_weather"}, {"type": "mcp", "server_label": "docs"}],
|
||||
},
|
||||
"tools": [_WEATHER_TOOL, _DOCS_MCP_TOOL],
|
||||
},
|
||||
id="allowed-tools-all-sent",
|
||||
),
|
||||
],
|
||||
)
|
||||
def test_a_well_formed_value_is_kept_under_drop_params(self, local_model_cost_map, params):
|
||||
assert OpenAIResponsesAPIConfig().map_openai_params(dict(params), "gpt-6.1-sol", True) == params
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"params",
|
||||
[
|
||||
pytest.param({"tool_choice": {"name": "get_time"}, "tools": [_WEATHER_TOOL]}, id="choice-without-a-type"),
|
||||
pytest.param({"tool_choice": ["get_time"], "tools": [_WEATHER_TOOL]}, id="choice-not-an-object"),
|
||||
pytest.param(
|
||||
{"tool_choice": {"type": "function", "name": "get_time"}, "tools": _WEATHER_TOOL}, id="tools-not-a-list"
|
||||
),
|
||||
],
|
||||
)
|
||||
def test_a_tool_choice_the_check_cannot_read_is_kept_under_drop_params(self, local_model_cost_map, params):
|
||||
assert OpenAIResponsesAPIConfig().map_openai_params(dict(params), "gpt-6.1-sol", True) == params
|
||||
|
||||
def test_providers_outside_openai_and_azure_keep_these_values_as_sent(self, local_model_cost_map):
|
||||
params: Final = {"tool_choice": "required", "parallel_tool_calls": "true", "previous_response_id": 12}
|
||||
|
||||
|
|
@ -3439,8 +3499,12 @@ class TestMalformedParamsFollowDropParams:
|
|||
|
||||
@pytest.mark.parametrize(
|
||||
("config", "model"),
|
||||
[(OpenAIResponsesAPIConfig(), "gpt-6.1-sol"), (AzureOpenAIResponsesAPIConfig(), "gpt-6.1-sol")],
|
||||
ids=["openai", "azure"],
|
||||
[
|
||||
(OpenAIResponsesAPIConfig(), "gpt-6.1-sol"),
|
||||
(AzureOpenAIResponsesAPIConfig(), "gpt-6.1-sol"),
|
||||
(AzureOpenAIOSeriesResponsesAPIConfig(), "o_series/gpt-6.1-sol"),
|
||||
],
|
||||
ids=["openai", "azure", "azure-o-series"],
|
||||
)
|
||||
def test_an_effort_outside_the_known_levels_is_dropped_unless_the_deployment_declares_it(
|
||||
self, local_model_cost_map, monkeypatch, config, model
|
||||
|
|
@ -3450,8 +3514,12 @@ class TestMalformedParamsFollowDropParams:
|
|||
|
||||
dropped: Final = config.map_openai_params({"reasoning": dict(reasoning)}, model, True)
|
||||
sent: Final = config.map_openai_params({"reasoning": dict(reasoning)}, model, False)
|
||||
declared: Final = config.map_openai_params_for_deployment(
|
||||
{"reasoning": dict(reasoning)}, model, True, declared_reasoning_efforts=("not-an-effort",)
|
||||
declared: Final = ResponsesAPIRequestUtils.get_optional_params_responses_api(
|
||||
model=model,
|
||||
responses_api_provider_config=config,
|
||||
response_api_optional_params={"reasoning": dict(reasoning)},
|
||||
drop_params=True,
|
||||
declared_reasoning_efforts=("not-an-effort",),
|
||||
)
|
||||
|
||||
assert dropped["reasoning"] == {"summary": "auto"}
|
||||
|
|
|
|||
|
|
@ -1038,3 +1038,24 @@ def test_the_chat_bridge_drops_a_reasoning_value_no_openai_row_declares_under_dr
|
|||
|
||||
assert bridged(True) == {"effort": "high", "summary": "auto"}
|
||||
assert bridged(False) == requested
|
||||
|
||||
def test_the_chat_bridge_keeps_only_effort_for_a_catalog_model_that_does_not_reason(local_model_cost_map):
|
||||
non_reasoning_model: Final = next(
|
||||
name
|
||||
for name, row in litellm.model_cost.items()
|
||||
if row.get("litellm_provider") == "openai"
|
||||
and row.get("mode") == "chat"
|
||||
and row.get("supports_reasoning") is False
|
||||
and OpenAIGPTConfig.is_openai_catalog_model(name)
|
||||
)
|
||||
|
||||
bridged: Final = openai_bridge_reasoning(
|
||||
requested={"effort": "low", "summary": "auto", "mode": "standard"},
|
||||
effort_kept=True,
|
||||
summary_alias=None,
|
||||
model=non_reasoning_model,
|
||||
drop_params=True,
|
||||
additional_drop_params=None,
|
||||
)
|
||||
|
||||
assert bridged == {"effort": "low"}
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue