mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-17 09:34:48 +00:00
fix(oci): skip non-function tools and gate reasoning_effort on models that reject it
Fixes #31449. Two OCI GENERIC request-building defects that made Grok-on-OCI unusable for tool-using coding-agent clients (Codex over the Responses API, Claude Code over the Anthropic Messages API): 1. `adapt_tool_definition_to_oci_standard` raised `OCIError("OCI only supports function tools")` for any tool whose type is not "function", aborting the whole request when a client mixed a built-in/server tool in with its function tools. Non-function tools are now skipped with a warning; if none remain, `tools`/`toolChoice` are omitted from the request instead of sending an empty array. 2. The GENERIC param map forwarded `reasoning_effort` to every non-Cohere model, but OCI rejects it for all served xAI Grok models (`Model grok-4.20 does not support parameter reasoningEffort`) and for OpenAI's non-reasoning models (`Unrecognized request argument supplied: reasoning_effort` on gpt-4o / gpt-4.1); gpt-5, the o-series, gpt-oss-* and Meta/Google models accept it. `_model_supports_reasoning_effort` encodes that, `get_supported_openai_params` stops advertising the param for those models, and `map_openai_params` follows the standard unsupported-param contract (dropped under drop_params, clear error otherwise) so `drop_params` / `additional_drop_params` now work as documented instead of the 400 surfacing from the service. Behaviour verified against OCI GenAI in us-chicago-1 (2026-09-11) per model. The two existing reasoning_effort normalisation tests move from a Grok model to openai.gpt-5, which is the case they exercise. 21 new tests; the OCI suite is 406 passed.
This commit is contained in:
parent
b3882d8e43
commit
0ff303d1ea
4 changed files with 191 additions and 5 deletions
|
|
@ -13,6 +13,7 @@ from typing import Any, Final
|
|||
import httpx
|
||||
from pydantic import ValidationError
|
||||
|
||||
from litellm._logging import verbose_logger
|
||||
from litellm.llms.oci.common_utils import (
|
||||
OCIError,
|
||||
resolve_oci_schema_anyof,
|
||||
|
|
@ -218,8 +219,18 @@ def adapt_tool_definition_to_oci_standard(tools: list[dict], vendor: OCIVendors)
|
|||
"""
|
||||
new_tools: Final = []
|
||||
for tool in tools:
|
||||
if tool["type"] != "function":
|
||||
raise OCIError(status_code=400, message="OCI only supports function tools")
|
||||
tool_type = tool.get("type")
|
||||
if tool_type != "function":
|
||||
# Coding-agent clients (Codex over the Responses API, Claude Code
|
||||
# over the Anthropic Messages API) mix built-in / server tools in
|
||||
# with their function tools. OCI GENERIC only knows FUNCTION tools,
|
||||
# so translate what we can and skip the rest instead of failing
|
||||
# the whole request (see BerriAI/litellm#31449).
|
||||
verbose_logger.warning(
|
||||
"OCI GENERIC only supports function tools; skipping tool of type %r",
|
||||
tool_type,
|
||||
)
|
||||
continue
|
||||
|
||||
tool_function = tool.get("function")
|
||||
if not isinstance(tool_function, dict):
|
||||
|
|
|
|||
|
|
@ -230,6 +230,33 @@ def _normalize_response_format(selected_params: dict, vendor: OCIVendors) -> Non
|
|||
selected_params["responseFormat"] = {"type": "JSON_OBJECT" if fmt == "JSON" else fmt}
|
||||
|
||||
|
||||
def _model_supports_reasoning_effort(model: str) -> bool:
|
||||
"""Return True when the OCI-hosted model accepts ``reasoningEffort``.
|
||||
|
||||
Verified against the service (us-chicago-1, 2026-09): every served xAI
|
||||
Grok model rejects the parameter with ``400 "Model ... does not support
|
||||
parameter reasoningEffort"`` (including the ``*-reasoning`` variants), and
|
||||
OpenAI's non-reasoning commercial models (``gpt-4o``, ``gpt-4.1``) reject it
|
||||
with ``Unrecognized request argument supplied: reasoning_effort``. OpenAI
|
||||
reasoning models (gpt-5 family, o-series, driven by the catalog's
|
||||
``supports_reasoning`` flag) and the ``gpt-oss-*`` open weights accept it,
|
||||
and Meta / Google models ignore it harmlessly. Cohere is handled by its own
|
||||
param map. See BerriAI/litellm#31449.
|
||||
"""
|
||||
if not model:
|
||||
return True
|
||||
name: Final = model[4:] if model.lower().startswith("oci/") else model
|
||||
lowered: Final = name.lower()
|
||||
vendor: Final = lowered.split(".")[0]
|
||||
if vendor == "xai":
|
||||
return False
|
||||
if vendor == "openai":
|
||||
if lowered.startswith("openai.gpt-oss"):
|
||||
return True
|
||||
return bool(supports_reasoning(model=name, custom_llm_provider="oci"))
|
||||
return True
|
||||
|
||||
|
||||
def get_vendor_from_model(model: str) -> OCIVendors:
|
||||
"""Return the OCI vendor enum for a model name.
|
||||
|
||||
|
|
@ -323,7 +350,10 @@ class OCIChatConfig(BaseConfig):
|
|||
# honoured and advertising it would be misleading. Callers that gate on
|
||||
# this list strip n=1 (a no-op, matching what map_openai_params does);
|
||||
# callers that bypass it have n=1 dropped there. Both paths converge.
|
||||
return [key for key, value in param_map.items() if value]
|
||||
supported: Final = [key for key, value in param_map.items() if value]
|
||||
if "reasoning_effort" in supported and not _model_supports_reasoning_effort(model):
|
||||
supported.remove("reasoning_effort")
|
||||
return supported
|
||||
|
||||
def map_openai_params(
|
||||
self,
|
||||
|
|
@ -340,6 +370,20 @@ class OCIChatConfig(BaseConfig):
|
|||
|
||||
for key, value in {**non_default_params, **optional_params}.items():
|
||||
alias = param_map.get(key)
|
||||
if key == "reasoning_effort" and alias and not _model_supports_reasoning_effort(model):
|
||||
# The GENERIC map forwards reasoning_effort for every non-Cohere
|
||||
# model, but xAI Grok and OpenAI's non-reasoning models 400 on
|
||||
# it. Follow the standard unsupported-param contract so
|
||||
# drop_params (and additional_drop_params) work as documented.
|
||||
if drop_params or litellm.drop_params:
|
||||
continue
|
||||
raise OCIError(
|
||||
status_code=400,
|
||||
message=(
|
||||
f"param `reasoning_effort` is not supported on OCI model `{model}` "
|
||||
"(the model rejects reasoningEffort); set drop_params=True to omit it"
|
||||
),
|
||||
)
|
||||
if alias is False:
|
||||
# max_retries is a litellm-level control param (litellm applies
|
||||
# retries itself); it is never a generation param OCI accepts, so
|
||||
|
|
@ -507,6 +551,12 @@ class OCIChatConfig(BaseConfig):
|
|||
selected_params["tools"],
|
||||
vendor,
|
||||
)
|
||||
if not selected_params["tools"]:
|
||||
# Every tool was skipped (no function tools): send a plain chat
|
||||
# request rather than an empty tools array + toolChoice.
|
||||
selected_params.pop("tools")
|
||||
for key in ("toolChoice", "tool_choice"):
|
||||
selected_params.pop(key, None)
|
||||
|
||||
# Normalise tool_choice to OCI's flat uppercase dict form
|
||||
# ({"type": "AUTO"|"NONE"|"REQUIRED"} or {"type": "FUNCTION", "name": "<fn>"}).
|
||||
|
|
|
|||
|
|
@ -1178,15 +1178,19 @@ class TestOCIReasoningEffort:
|
|||
return body["chatRequest"]
|
||||
|
||||
def test_reasoning_effort_lowercase_uppercased(self):
|
||||
# gpt-5 accepts reasoningEffort; every OCI-hosted Grok model rejects it
|
||||
# (see TestReasoningEffortGating in test_oci_tool_and_reasoning_gating.py).
|
||||
chat_request = self._build_chat_request(
|
||||
"xai.grok-4-fast-reasoning",
|
||||
"openai.gpt-5",
|
||||
{"reasoning_effort": "low"},
|
||||
)
|
||||
assert chat_request.get("reasoningEffort") == "LOW"
|
||||
|
||||
def test_reasoning_effort_disable_mapped_to_none(self):
|
||||
# gpt-5 accepts reasoningEffort; every OCI-hosted Grok model rejects it
|
||||
# (see TestReasoningEffortGating in test_oci_tool_and_reasoning_gating.py).
|
||||
chat_request = self._build_chat_request(
|
||||
"xai.grok-4-fast-reasoning",
|
||||
"openai.gpt-5",
|
||||
{"reasoning_effort": "disable"},
|
||||
)
|
||||
assert chat_request.get("reasoningEffort") == "NONE"
|
||||
|
|
|
|||
|
|
@ -0,0 +1,121 @@
|
|||
"""Regression tests for BerriAI/litellm#31449.
|
||||
|
||||
1. Non-function tools must be skipped, not fail the whole request.
|
||||
2. ``reasoning_effort`` must not be forwarded to OCI models that reject it
|
||||
(xAI Grok, OpenAI non-reasoning models), so ``drop_params`` works.
|
||||
"""
|
||||
|
||||
import pytest
|
||||
|
||||
import litellm
|
||||
from litellm.llms.oci.chat.generic import adapt_tool_definition_to_oci_standard
|
||||
from litellm.llms.oci.chat.transformation import (
|
||||
OCIChatConfig,
|
||||
_model_supports_reasoning_effort,
|
||||
)
|
||||
from litellm.llms.oci.common_utils import OCIError
|
||||
from litellm.types.llms.oci import OCIVendors
|
||||
|
||||
FUNCTION_TOOL = {
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "get_weather",
|
||||
"description": "Weather for a city",
|
||||
"parameters": {"type": "object", "properties": {"city": {"type": "string"}}},
|
||||
},
|
||||
}
|
||||
BUILTIN_TOOL = {"type": "web_search_preview"}
|
||||
SERVER_TOOL = {"type": "computer_use_preview", "display_width": 1024}
|
||||
|
||||
|
||||
class TestNonFunctionToolsAreSkipped:
|
||||
def test_mixed_tools_keep_function_tools_only(self):
|
||||
result = adapt_tool_definition_to_oci_standard([BUILTIN_TOOL, FUNCTION_TOOL, SERVER_TOOL], OCIVendors.GENERIC)
|
||||
assert [tool.name for tool in result] == ["get_weather"]
|
||||
assert result[0].type == "FUNCTION"
|
||||
|
||||
def test_only_non_function_tools_yield_empty_list(self):
|
||||
assert adapt_tool_definition_to_oci_standard([BUILTIN_TOOL], OCIVendors.GENERIC) == []
|
||||
|
||||
def test_tool_without_type_is_skipped_not_keyerror(self):
|
||||
assert adapt_tool_definition_to_oci_standard([{"name": "x"}], OCIVendors.GENERIC) == []
|
||||
|
||||
def test_request_drops_tools_and_tool_choice_when_none_remain(self):
|
||||
config = OCIChatConfig()
|
||||
params = config._get_optional_params(
|
||||
OCIVendors.GENERIC,
|
||||
{"tools": [BUILTIN_TOOL], "tool_choice": "auto"},
|
||||
model="xai.grok-4.20",
|
||||
)
|
||||
assert "tools" not in params
|
||||
assert "toolChoice" not in params
|
||||
assert "tool_choice" not in params
|
||||
|
||||
def test_request_keeps_function_tools_and_tool_choice(self):
|
||||
config = OCIChatConfig()
|
||||
params = config._get_optional_params(
|
||||
OCIVendors.GENERIC,
|
||||
{"tools": [BUILTIN_TOOL, FUNCTION_TOOL], "tool_choice": "auto"},
|
||||
model="xai.grok-4.20",
|
||||
)
|
||||
assert [tool.name for tool in params["tools"]] == ["get_weather"]
|
||||
assert params["toolChoice"] == {"type": "AUTO"}
|
||||
|
||||
|
||||
class TestReasoningEffortGating:
|
||||
@pytest.mark.parametrize(
|
||||
"model,expected",
|
||||
[
|
||||
("xai.grok-4.20", False),
|
||||
("oci/xai.grok-4.20-0309-reasoning", False),
|
||||
("xai.grok-code-fast-1", False),
|
||||
("openai.gpt-4o", False),
|
||||
("openai.gpt-4.1", False),
|
||||
("openai.gpt-5", True),
|
||||
("oci/openai.gpt-5-mini", True),
|
||||
("openai.gpt-oss-120b", True),
|
||||
("meta.llama-3.3-70b-instruct", True),
|
||||
("google.gemini-2.5-flash", True),
|
||||
],
|
||||
)
|
||||
def test_model_supports_reasoning_effort(self, model, expected):
|
||||
assert _model_supports_reasoning_effort(model) is expected
|
||||
|
||||
def test_supported_params_exclude_reasoning_effort_for_grok(self):
|
||||
supported = OCIChatConfig().get_supported_openai_params("xai.grok-4.20")
|
||||
assert "reasoning_effort" not in supported
|
||||
assert "tools" in supported
|
||||
|
||||
def test_supported_params_include_reasoning_effort_for_gpt5(self):
|
||||
assert "reasoning_effort" in OCIChatConfig().get_supported_openai_params("openai.gpt-5")
|
||||
|
||||
@pytest.mark.parametrize("model", ["xai.grok-4.20", "openai.gpt-4o"])
|
||||
def test_reasoning_effort_dropped_under_drop_params(self, model):
|
||||
result = OCIChatConfig().map_openai_params(
|
||||
non_default_params={"reasoning_effort": "low", "temperature": 0.1},
|
||||
optional_params={},
|
||||
model=model,
|
||||
drop_params=True,
|
||||
)
|
||||
assert "reasoningEffort" not in result
|
||||
assert "reasoning_effort" not in result
|
||||
assert result["temperature"] == 0.1
|
||||
|
||||
def test_reasoning_effort_raises_without_drop_params_for_grok(self, monkeypatch):
|
||||
monkeypatch.setattr(litellm, "drop_params", False)
|
||||
with pytest.raises(OCIError, match=r"reasoning_effort.*xai\.grok-4\.20"):
|
||||
OCIChatConfig().map_openai_params(
|
||||
non_default_params={"reasoning_effort": "low"},
|
||||
optional_params={},
|
||||
model="xai.grok-4.20",
|
||||
drop_params=False,
|
||||
)
|
||||
|
||||
def test_reasoning_effort_still_forwarded_for_reasoning_models(self):
|
||||
result = OCIChatConfig().map_openai_params(
|
||||
non_default_params={"reasoning_effort": "low"},
|
||||
optional_params={},
|
||||
model="openai.gpt-5",
|
||||
drop_params=False,
|
||||
)
|
||||
assert result["reasoningEffort"] == "low"
|
||||
Loading…
Add table
Reference in a new issue