mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-18 09:45:23 +00:00
Merge 2f910080b6 into dddbaa803d
This commit is contained in:
commit
2b3a07375a
11 changed files with 250 additions and 34 deletions
|
|
@ -1192,8 +1192,9 @@ class AmazonConverseConfig(BaseConfig):
|
|||
),
|
||||
status_code=400,
|
||||
)
|
||||
verbose_logger.debug(
|
||||
"Dropping unsupported `reasoning_effort=%s` for Bedrock model=%s.",
|
||||
verbose_logger.warning(
|
||||
"Dropping reasoning_effort=%r for Bedrock model=%r (drop_params=True) because the model map "
|
||||
"marks that level unsupported.",
|
||||
value,
|
||||
model,
|
||||
)
|
||||
|
|
|
|||
|
|
@ -5,6 +5,7 @@ from collections.abc import Collection
|
|||
from typing import ClassVar, Final
|
||||
|
||||
import litellm
|
||||
from litellm._logging import verbose_logger
|
||||
from litellm.utils import (
|
||||
declared_value_factory,
|
||||
is_explicitly_disabled_factory,
|
||||
|
|
@ -15,6 +16,12 @@ from .gpt_transformation import OpenAIGPTConfig
|
|||
|
||||
_GPT_SERIES_VERSION: Final = re.compile(r"^gpt-(\d+)(?:\.(\d+))?(?=[.-]|$)")
|
||||
|
||||
DROP_REFUSED_REASONING_EFFORT_WARNING: Final = (
|
||||
"Dropping %s=%r for model=%r (drop_params=True) because the model map does not mark that level supported for "
|
||||
"%r. If the deployment supports it, list every level it supports in model_info.reasoning_effort_levels, or set "
|
||||
"model_info.base_model when the deployment name is not its underlying model."
|
||||
)
|
||||
|
||||
|
||||
def _catalogue_declares_default_effort() -> bool:
|
||||
"""Whether the loaded cost map carries default_reasoning_effort for ANY entry.
|
||||
|
|
@ -342,6 +349,13 @@ class OpenAIGPT5Config(OpenAIGPTConfig):
|
|||
and self.refuses_reasoning_effort_level(model, requested_effort)
|
||||
):
|
||||
if litellm.drop_params or drop_params:
|
||||
verbose_logger.warning(
|
||||
DROP_REFUSED_REASONING_EFFORT_WARNING,
|
||||
"reasoning_effort",
|
||||
requested_effort,
|
||||
model,
|
||||
self._model_map_lookup_name(model),
|
||||
)
|
||||
non_default_params.pop("reasoning_effort", None)
|
||||
optional_params.pop("reasoning_effort", None)
|
||||
else:
|
||||
|
|
|
|||
|
|
@ -26,6 +26,7 @@ from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
|
|||
from litellm.litellm_core_utils.url_utils import encode_url_path_segment, join_api_base
|
||||
from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig
|
||||
from litellm.llms.openai.chat.gpt_5_transformation import (
|
||||
DROP_REFUSED_REASONING_EFFORT_WARNING,
|
||||
is_gpt_reasoning_series_name,
|
||||
unsupported_reasoning_effort_message,
|
||||
)
|
||||
|
|
@ -50,6 +51,9 @@ from ..common_utils import OpenAIError
|
|||
from ..workload_identity import get_workload_identity_bearer_token, resolve_openai_workload_identity_config
|
||||
|
||||
OPENAI_RESPONSES_API_MIN_MAX_OUTPUT_TOKENS: Final = 16
|
||||
DROP_UNSUPPORTED_REASONING_WARNING: Final = (
|
||||
"Dropping reasoning=%r for model=%r (drop_params=True) because its model cost map entry lacks supports_reasoning."
|
||||
)
|
||||
_RAW_RESPONSE_JSON: Final = TypeAdapter(dict[str, object], config=ConfigDict(strict=True))
|
||||
|
||||
if TYPE_CHECKING:
|
||||
|
|
@ -401,7 +405,8 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
|
|||
and not self._supports_reasoning_param(model=model)
|
||||
):
|
||||
if drop_params or litellm.drop_params:
|
||||
params.pop("reasoning", None)
|
||||
dropped_reasoning: Final = params.pop("reasoning", None)
|
||||
verbose_logger.warning(DROP_UNSUPPORTED_REASONING_WARNING, dropped_reasoning, model)
|
||||
else:
|
||||
raise litellm.UnsupportedParamsError(
|
||||
message=(
|
||||
|
|
@ -427,6 +432,9 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
|
|||
message=unsupported_reasoning_effort_message("reasoning.effort", requested_effort, model),
|
||||
status_code=400,
|
||||
)
|
||||
verbose_logger.warning(
|
||||
DROP_REFUSED_REASONING_EFFORT_WARNING, "reasoning.effort", requested_effort, model, lookup_name
|
||||
)
|
||||
remaining: Final = reasoning.copy()
|
||||
remaining.pop("effort", None)
|
||||
if remaining:
|
||||
|
|
|
|||
|
|
@ -426,6 +426,27 @@ def _log_remap(
|
|||
logged_deployment_id: Final = _single_line(deployment_id)
|
||||
logged_model_group: Final = _single_line(model_group)
|
||||
logged_deployment_model: Final = _single_line(deployment_model)
|
||||
extra: Final = MappingProxyType(
|
||||
{
|
||||
"reasoning_effort_deployment_id": logged_deployment_id,
|
||||
"reasoning_effort_model_group": logged_model_group,
|
||||
"reasoning_effort_deployment_model": logged_deployment_model,
|
||||
"reasoning_effort_requested": logged_requested_level,
|
||||
"reasoning_effort_sent": logged_sent_level,
|
||||
"reasoning_effort_rounding": rounding,
|
||||
}
|
||||
)
|
||||
if sent_level is None and requested_level != _DEFAULT_EFFORT:
|
||||
verbose_logger.warning(
|
||||
"Dropping reasoning level %s (drop_params=True) for deployment %s, model group %s, model %s because its "
|
||||
"model_info.reasoning_effort_levels does not list that level. List every level the deployment supports.",
|
||||
logged_requested_level,
|
||||
logged_deployment_id,
|
||||
logged_model_group,
|
||||
logged_deployment_model,
|
||||
extra=extra,
|
||||
)
|
||||
return
|
||||
verbose_logger.info(
|
||||
"reasoning_effort %s sent as %s (rounding %s) to deployment %s, model group %s, model %s",
|
||||
logged_requested_level,
|
||||
|
|
@ -434,14 +455,7 @@ def _log_remap(
|
|||
logged_deployment_id,
|
||||
logged_model_group,
|
||||
logged_deployment_model,
|
||||
extra={
|
||||
"reasoning_effort_deployment_id": logged_deployment_id,
|
||||
"reasoning_effort_model_group": logged_model_group,
|
||||
"reasoning_effort_deployment_model": logged_deployment_model,
|
||||
"reasoning_effort_requested": logged_requested_level,
|
||||
"reasoning_effort_sent": logged_sent_level,
|
||||
"reasoning_effort_rounding": rounding,
|
||||
},
|
||||
extra=extra,
|
||||
)
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -4972,6 +4972,14 @@ def _declared_reasoning_efforts(model_info: Mapping[str, object] | None) -> tupl
|
|||
return deployment_reasoning_effort_levels(model_info) or ()
|
||||
|
||||
|
||||
_DROPPED_REASONING_PARAM_FIXES: Final = MappingProxyType(
|
||||
{
|
||||
"reasoning_effort": "If the deployment is a reasoning model, set model_info.base_model to its underlying model.",
|
||||
"thinking": "If the deployment is a reasoning model, send reasoning_effort instead.",
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
def get_optional_params(
|
||||
# use the openai defaults
|
||||
# https://platform.openai.com/docs/api-reference/chat/create
|
||||
|
|
@ -5059,19 +5067,27 @@ def get_optional_params(
|
|||
verbose_logger.info("\nLiteLLM completion() model= %s; provider = %s", model, custom_llm_provider)
|
||||
verbose_logger.debug("\nLiteLLM: Params passed to completion() %s", passed_params)
|
||||
verbose_logger.debug("\nLiteLLM: Non-Default params passed to completion() %s", non_default_params)
|
||||
unsupported_params: Final = {}
|
||||
for k in non_default_params:
|
||||
if k not in supported_params:
|
||||
if k in PROVIDER_UNVALIDATED_PARAMS:
|
||||
continue
|
||||
if k == "n" and n == 1: # langchain sends n=1 as a default value
|
||||
continue # skip this param
|
||||
unsupported_params[k] = non_default_params[k]
|
||||
unsupported_params: Final[Mapping[str, object]] = {
|
||||
k: non_default_params[k]
|
||||
for k in non_default_params
|
||||
if k not in supported_params and k not in PROVIDER_UNVALIDATED_PARAMS and not (k == "n" and n == 1)
|
||||
}
|
||||
|
||||
if unsupported_params:
|
||||
if litellm.drop_params is True or (drop_params is not None and drop_params is True):
|
||||
for k in unsupported_params:
|
||||
non_default_params.pop(k, None)
|
||||
for reasoning_param, fix in _DROPPED_REASONING_PARAM_FIXES.items():
|
||||
if reasoning_param in unsupported_params:
|
||||
verbose_logger.warning(
|
||||
"Dropping %s=%r for model=%r (drop_params=True) because %s does not support %s for it. %s",
|
||||
reasoning_param,
|
||||
unsupported_params[reasoning_param],
|
||||
model,
|
||||
custom_llm_provider,
|
||||
reasoning_param,
|
||||
fix,
|
||||
)
|
||||
else:
|
||||
raise UnsupportedParamsError(
|
||||
status_code=500,
|
||||
|
|
|
|||
|
|
@ -1,6 +1,7 @@
|
|||
import base64
|
||||
import copy
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
import re
|
||||
from collections.abc import Iterator, Mapping
|
||||
|
|
@ -966,7 +967,9 @@ def test_reasoning_effort_maps_to_reasoning_effort_for_openai_gpt5_converse(mode
|
|||
"us.openai.gpt-6-astra",
|
||||
],
|
||||
)
|
||||
def test_openai_gpt5_converse_rejects_effort_level_disabled_in_model_map(model, local_model_cost_map):
|
||||
def test_openai_gpt5_converse_rejects_effort_level_disabled_in_model_map(
|
||||
model: str, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
|
||||
) -> None:
|
||||
config = AmazonConverseConfig()
|
||||
assert litellm.utils.is_explicitly_disabled_factory(
|
||||
model=model, custom_llm_provider="bedrock_converse", key="supports_minimal_reasoning_effort"
|
||||
|
|
@ -980,15 +983,18 @@ def test_openai_gpt5_converse_rejects_effort_level_disabled_in_model_map(model,
|
|||
drop_params=False,
|
||||
)
|
||||
|
||||
optional_params = config.map_openai_params(
|
||||
non_default_params={"reasoning_effort": "minimal"},
|
||||
optional_params={},
|
||||
model=model,
|
||||
drop_params=True,
|
||||
)
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
optional_params = config.map_openai_params(
|
||||
non_default_params={"reasoning_effort": "minimal"},
|
||||
optional_params={},
|
||||
model=model,
|
||||
drop_params=True,
|
||||
)
|
||||
_, additional_request_params, _, _ = config._prepare_request_params(optional_params, model)
|
||||
assert "reasoning" not in additional_request_params
|
||||
assert "thinking" not in additional_request_params
|
||||
dropped = [r.getMessage() for r in caplog.records if "Dropping reasoning_effort='minimal'" in r.getMessage()]
|
||||
assert len(dropped) == 1 and model in dropped[0]
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
|
|
|
|||
|
|
@ -337,13 +337,21 @@ class TestDisabledReasoningEffort:
|
|||
)
|
||||
|
||||
@pytest.mark.parametrize("model", ["us.openai.gpt-5.6-luna", MODEL])
|
||||
def test_effort_level_disabled_in_model_map_is_dropped_with_drop_params(self, model, local_model_cost_map):
|
||||
params = _cfg().map_openai_params(
|
||||
response_api_optional_params={"reasoning": {"effort": "minimal", "summary": "auto"}, "max_output_tokens": 64},
|
||||
model=model,
|
||||
drop_params=True,
|
||||
)
|
||||
def test_effort_level_disabled_in_model_map_is_dropped_with_a_warning(
|
||||
self, model: str, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
|
||||
) -> None:
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
params = _cfg().map_openai_params(
|
||||
response_api_optional_params={
|
||||
"reasoning": {"effort": "minimal", "summary": "auto"},
|
||||
"max_output_tokens": 64,
|
||||
},
|
||||
model=model,
|
||||
drop_params=True,
|
||||
)
|
||||
assert params == {"reasoning": {"summary": "auto"}, "max_output_tokens": 64}
|
||||
dropped = [r.getMessage() for r in caplog.records if "Dropping reasoning.effort='minimal'" in r.getMessage()]
|
||||
assert len(dropped) == 1 and model in dropped[0]
|
||||
|
||||
def test_effort_only_reasoning_is_removed_when_dropped(self, local_model_cost_map):
|
||||
params = _cfg().map_openai_params(
|
||||
|
|
|
|||
|
|
@ -1,4 +1,5 @@
|
|||
import json
|
||||
import logging
|
||||
from collections.abc import Callable, Mapping
|
||||
from types import SimpleNamespace
|
||||
from typing import Final, cast
|
||||
|
|
@ -2387,6 +2388,57 @@ class TestResponsesSurfaceSharesTheEffortRule:
|
|||
)
|
||||
assert mapped["reasoning"] == {"summary": "auto"}
|
||||
|
||||
def test_dropping_a_refused_effort_logs_one_warning_naming_the_level_and_the_fix(
|
||||
self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
|
||||
) -> None:
|
||||
refused: Final = next(
|
||||
level
|
||||
for level in ("none", "minimal", "low")
|
||||
if litellm.model_cost["gpt-6.1-sol"].get(f"supports_{level}_reasoning_effort") is False
|
||||
)
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
|
||||
response_api_optional_params={"reasoning": {"effort": refused}},
|
||||
model="gpt-6.1-sol",
|
||||
drop_params=True,
|
||||
)
|
||||
|
||||
assert "reasoning" not in mapped
|
||||
drop_warnings: Final = [record.getMessage() for record in caplog.records if "Dropping" in record.getMessage()]
|
||||
assert len(drop_warnings) == 1
|
||||
assert all(
|
||||
fragment in drop_warnings[0]
|
||||
for fragment in (repr(refused), "gpt-6.1-sol", "model_info.reasoning_effort_levels")
|
||||
)
|
||||
|
||||
def test_dropping_reasoning_for_a_model_without_supports_reasoning_logs_one_warning(
|
||||
self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
|
||||
) -> None:
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
|
||||
response_api_optional_params={"reasoning": {"effort": "medium"}},
|
||||
model="gpt-4o",
|
||||
drop_params=True,
|
||||
)
|
||||
|
||||
assert "reasoning" not in mapped
|
||||
drop_warnings: Final = [record.getMessage() for record in caplog.records if "Dropping" in record.getMessage()]
|
||||
assert len(drop_warnings) == 1
|
||||
assert all(fragment in drop_warnings[0] for fragment in ("'medium'", "gpt-4o", "supports_reasoning"))
|
||||
|
||||
def test_a_forwarded_effort_logs_no_drop_warning(
|
||||
self, local_model_cost_map: None, caplog: pytest.LogCaptureFixture
|
||||
) -> None:
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
mapped: Final = OpenAIResponsesAPIConfig().map_openai_params(
|
||||
response_api_optional_params={"reasoning": {"effort": "medium"}},
|
||||
model="gpt-6.1-sol",
|
||||
drop_params=True,
|
||||
)
|
||||
|
||||
assert mapped["reasoning"] == {"effort": "medium"}
|
||||
assert not [record for record in caplog.records if "Dropping" in record.getMessage()]
|
||||
|
||||
@pytest.mark.parametrize("level", ["none", "minimal", "low", "medium", "high", "xhigh", "max"])
|
||||
def test_gpt5_3_codex_forwards_exactly_the_effort_levels_its_map_row_allows(self, local_model_cost_map, level):
|
||||
"""The gate turns a stale false into a local refusal, so the Responses-only codex row has to
|
||||
|
|
|
|||
|
|
@ -1,3 +1,4 @@
|
|||
import logging
|
||||
import re
|
||||
from collections.abc import Mapping
|
||||
from typing import Final
|
||||
|
|
@ -1589,3 +1590,44 @@ def test_gpt_6_astra_takes_the_reasoning_series_request_shape():
|
|||
assert "max_tokens" not in params
|
||||
assert params["reasoning_effort"] == "max"
|
||||
assert params["verbosity"] == "low"
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("custom_llm_provider", "level"),
|
||||
[("azure", "none"), ("azure", "xhigh"), ("openai", "xhigh")],
|
||||
)
|
||||
def test_a_reasoning_effort_the_level_gate_drops_logs_a_warning_naming_the_fix(
|
||||
caplog: pytest.LogCaptureFixture, custom_llm_provider: str, level: str
|
||||
) -> None:
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
params: Final = litellm.get_optional_params(
|
||||
model="gpt-5.6-terra-ptu",
|
||||
custom_llm_provider=custom_llm_provider,
|
||||
reasoning_effort=level,
|
||||
drop_params=True,
|
||||
)
|
||||
|
||||
assert "reasoning_effort" not in params
|
||||
drop_warnings: Final = [
|
||||
record.getMessage()
|
||||
for record in caplog.records
|
||||
if record.levelno == logging.WARNING and "reasoning_effort" in record.getMessage()
|
||||
]
|
||||
assert len(drop_warnings) == 1
|
||||
assert all(
|
||||
fragment in drop_warnings[0]
|
||||
for fragment in (level, "gpt-5.6-terra-ptu", "model_info.base_model", "model_info.reasoning_effort_levels")
|
||||
)
|
||||
|
||||
|
||||
def test_a_reasoning_effort_the_level_gate_forwards_logs_no_drop_warning(caplog: pytest.LogCaptureFixture) -> None:
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
params: Final = litellm.get_optional_params(
|
||||
model="gpt-5.6-terra-ptu",
|
||||
custom_llm_provider="azure",
|
||||
reasoning_effort="medium",
|
||||
drop_params=True,
|
||||
)
|
||||
|
||||
assert params["reasoning_effort"] == "medium"
|
||||
assert not [record for record in caplog.records if "reasoning_effort" in record.getMessage()]
|
||||
|
|
|
|||
|
|
@ -626,8 +626,9 @@ class TestResolveDeploymentReasoningEffort:
|
|||
_sent("turbo\nforged line", drop_params=True)
|
||||
|
||||
assert [record.getMessage() for record in caplog.records] == [
|
||||
"reasoning_effort unrecognized sent as None (rounding up) to deployment deployment-1, "
|
||||
"model group kimi-k3, model hosted_vllm/my-kimi"
|
||||
"Dropping reasoning level unrecognized (drop_params=True) for deployment deployment-1, model group "
|
||||
"kimi-k3, model hosted_vllm/my-kimi because its model_info.reasoning_effort_levels does not list that "
|
||||
"level. List every level the deployment supports."
|
||||
]
|
||||
|
||||
def test_line_breaks_in_logged_names_cannot_forge_a_log_line(self, caplog: pytest.LogCaptureFixture) -> None:
|
||||
|
|
@ -669,6 +670,27 @@ class TestResolveDeploymentReasoningEffort:
|
|||
"reasoning_effort_rounding": "up",
|
||||
}
|
||||
|
||||
def test_only_a_dropped_level_logs_a_warning(self, caplog: pytest.LogCaptureFixture) -> None:
|
||||
verbose_logger.addHandler(caplog.handler)
|
||||
verbose_logger.propagate = False
|
||||
verbose_logger.setLevel(logging.INFO)
|
||||
declared: Final = {"reasoning_effort_levels": ["low", "high"]}
|
||||
|
||||
_sent("none", declared, drop_params=True)
|
||||
_sent("medium")
|
||||
_sent("default", declared, drop_params=True)
|
||||
_sent("default", declared)
|
||||
|
||||
assert [record.levelno for record in caplog.records] == [
|
||||
logging.WARNING,
|
||||
logging.INFO,
|
||||
logging.INFO,
|
||||
logging.INFO,
|
||||
]
|
||||
dropped: Final = caplog.records[0].getMessage()
|
||||
assert "Dropping reasoning level none" in dropped
|
||||
assert "model_info.reasoning_effort_levels" in dropped
|
||||
|
||||
|
||||
class TestAzureGpt6SolAndLunaAdvertiseTheOpenAiLevels:
|
||||
@pytest.mark.parametrize(
|
||||
|
|
|
|||
|
|
@ -1,4 +1,5 @@
|
|||
import json
|
||||
import logging
|
||||
from typing import Final
|
||||
from unittest.mock import MagicMock, patch
|
||||
|
||||
|
|
@ -2097,3 +2098,35 @@ def test_supports_system_message_merges_system_prompt_into_user_message() -> Non
|
|||
]
|
||||
|
||||
assert map_system_message_pt(messages=messages) == [{"role": "user", "content": "Listen here! Hello there!"}]
|
||||
|
||||
|
||||
def test_dropping_an_unsupported_reasoning_effort_logs_one_warning(
|
||||
local_model_cost_map: None, caplog: pytest.LogCaptureFixture
|
||||
) -> None:
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
optional_params: Final = get_optional_params(
|
||||
model="terra-ptu", custom_llm_provider="azure", reasoning_effort="none", drop_params=True
|
||||
)
|
||||
|
||||
assert "reasoning_effort" not in optional_params
|
||||
drop_warnings: Final = [
|
||||
record.getMessage() for record in caplog.records if "Dropping reasoning_effort" in record.getMessage()
|
||||
]
|
||||
assert len(drop_warnings) == 1
|
||||
assert all(fragment in drop_warnings[0] for fragment in ("'none'", "terra-ptu", "model_info.base_model"))
|
||||
|
||||
|
||||
def test_dropping_an_unsupported_thinking_logs_one_warning(
|
||||
local_model_cost_map: None, caplog: pytest.LogCaptureFixture
|
||||
) -> None:
|
||||
with caplog.at_level(logging.WARNING, logger="LiteLLM"):
|
||||
optional_params: Final = get_optional_params(
|
||||
model="terra-ptu", custom_llm_provider="azure", thinking={"type": "disabled"}, drop_params=True
|
||||
)
|
||||
|
||||
assert "thinking" not in optional_params
|
||||
drop_warnings: Final = [
|
||||
record.getMessage() for record in caplog.records if "Dropping thinking" in record.getMessage()
|
||||
]
|
||||
assert len(drop_warnings) == 1
|
||||
assert all(fragment in drop_warnings[0] for fragment in ("'disabled'", "terra-ptu", "send reasoning_effort"))
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue