scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-11 03:38:38 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 84

Merge f8c38659d4 into a2bf67a037

Browse source
This commit is contained in:
Zach Chrystall 2026-10-04 08:04:01 -04:00 • committed by GitHub
commit faec0ba0c3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 220 additions and 26 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

55
litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
View file

@ -281,14 +281,24 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
# Fallback: delegate to parent for unknown types
return super().get_json_schema_from_pydantic_object(response_format)
@staticmethod
def _strip_regional_prefix(model: str) -> str:
return re.sub(
r"(^|/)(?:[a-z0-9_-]+\.)+(?=gem(?:ini|ma)-)",
r"\1",
model,
flags=re.IGNORECASE,
)
@staticmethod
def _is_gemini_3_or_newer(model: str) -> bool:
"""
Check if the model is Gemini 3 or newer.
"""
model_name: Final = model.split("/")[-1].lower()
normalized_model: Final = VertexGeminiConfig._strip_regional_prefix(model).lower()
model_name: Final = normalized_model.split("/")[-1]
is_vertex_fine_tuned_model: Final = model_name.isdigit() or (
model.startswith("gemini/") and not model_name.startswith("gemini-")
normalized_model.startswith("gemini/") and not model_name.startswith("gemini-")
)
if not model_name or is_vertex_fine_tuned_model or model_name.startswith("gemma-"):
return False
@ -297,6 +307,18 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
return False
return True
@staticmethod
def _is_gemini_3_flash(model: str | None) -> bool:
if not model:
return False
return VertexGeminiConfig._is_gemini_3_or_newer(model) and "flash" in model.lower()
@staticmethod
def _is_gemini_3_1_pro_or_newer(model: str | None) -> bool:
if not model:
return False
return bool(re.search(r"gemini-3\.\d+-pro", model.lower()))
@staticmethod
def _forward_gemini_function_call_id(model: str) -> bool:
"""
@ -346,7 +368,7 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
if self._supports_penalty_parameters(model):
supported_params.extend(["frequency_penalty", "presence_penalty"])
if supports_reasoning(model) or self._is_gemini_3_or_newer(model):
if supports_reasoning(self._strip_regional_prefix(model)) or self._is_gemini_3_or_newer(model):
supported_params.append("reasoning_effort")
supported_params.append("thinking")
@ -870,10 +892,9 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
@staticmethod
def _supports_minimal_thinking_level(model: str) -> bool:
lowered: Final = model.lower()
is_gemini3_or_newer_flash: Final = VertexGeminiConfig._is_gemini_3_or_newer(model) and "flash" in lowered
return is_gemini3_or_newer_flash and not is_explicitly_disabled_factory(
model=model, custom_llm_provider=None, key="supports_minimal_reasoning_effort"
normalized_model: Final = VertexGeminiConfig._strip_regional_prefix(model)
return VertexGeminiConfig._is_gemini_3_flash(normalized_model) and not is_explicitly_disabled_factory(
model=normalized_model, custom_llm_provider=None, key="supports_minimal_reasoning_effort"
)
@staticmethod
@ -966,19 +987,17 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
# For Gemini 3+ models, use thinkingLevel instead of thinkingBudget
if model and VertexGeminiConfig._is_gemini_3_or_newer(model):
if thinking_enabled:
if thinking_budget == 0:
params["includeThoughts"] = False
else:
params["includeThoughts"] = True
# Follow provider defaults unless explicitly opted into legacy behavior.
if litellm.enable_gemini_default_thinking_level_low is True:
params["thinkingLevel"] = (
"minimal" if VertexGeminiConfig._supports_minimal_thinking_level(model) else "low"
)
default_low_thinking_level: Final = (
"minimal" if VertexGeminiConfig._supports_minimal_thinking_level(model) else "low"
)
if thinking_enabled and thinking_budget != 0:
params["includeThoughts"] = True
# Follow provider defaults unless explicitly opted into legacy behavior.
if litellm.enable_gemini_default_thinking_level_low is True:
params["thinkingLevel"] = default_low_thinking_level
else:
# Thinking disabled
params["includeThoughts"] = False
params["thinkingLevel"] = default_low_thinking_level
else:
# For older Gemini models, use thinkingBudget
if thinking_enabled and not VertexGeminiConfig._is_thinking_budget_zero(thinking_budget):

10
tests/llm_translation/test_gemini.py
View file

@ -1579,10 +1579,7 @@ def test_anthropic_thinking_param_to_gemini_3_provider_defaults():
)
assert result_disabled.get("includeThoughts") is False
assert (
"thinkingLevel" not in result_disabled
or result_disabled.get("thinkingLevel") is None
)
assert result_disabled.get("thinkingLevel") == "low"
# Test 3: Budget tokens = 0 for Gemini 3
thinking_param_zero: AnthropicThinkingParam = {
@ -1596,10 +1593,7 @@ def test_anthropic_thinking_param_to_gemini_3_provider_defaults():
)
assert result_zero["includeThoughts"] is False
assert (
"thinkingLevel" not in result_zero
or result_zero.get("thinkingLevel") is None
)
assert result_zero["thinkingLevel"] == "minimal"
# Test 4: Gemini 3 flash-preview should also follow provider defaults by default
result_gemini3flashpreview = VertexGeminiConfig._map_thinking_param(

181
tests/unit/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py
View file

@ -6356,3 +6356,184 @@ def test_gemini_multi_candidate_messages_do_not_share_state():
assert resp.choices[1].message.tool_calls is None
assert getattr(resp.choices[1].message, "reasoning_content", None) is None
assert resp.choices[1].provider_specific_fields["native_finish_reason"] == "STOP"
@pytest.mark.parametrize(
("model", "expected_is_gemini_3", "expected_is_flash", "expected_is_3_1_pro_or_newer"),
[
("gemini-3.5-flash", True, True, False),
("gemini-3.5-flash-preview", True, True, False),
("gemini-3.1-flash", True, True, False),
("gemini-3.1-flash-lite-preview", True, True, False),
("gemini-3-flash", True, True, False),
("au.gemini-3.5-flash", True, True, False),
("eu.gemini-3.5-flash", True, True, False),
("us.gemini-3.5-flash", True, True, False),
("ca.gemini-3.5-flash", True, True, False),
("jp.gemini-3.5-flash", True, True, False),
("uk.gemini-3.5-flash", True, True, False),
("in.gemini-3.5-flash", True, True, False),
("sg.gemini-3.5-flash", True, True, False),
("kr.gemini-3.5-flash", True, True, False),
("global.gemini-3.5-flash", True, True, False),
("apac.gemini-3.5-flash", True, True, False),
("us-central1.gemini-3.5-flash", True, True, False),
("europe-west4.gemini-3.5-flash", True, True, False),
("australia-southeast1.gemini-3.5-flash", True, True, False),
("asia-northeast1.gemini-3.5-flash", True, True, False),
("vertex_ai/au.gemini-3.5-flash", True, True, False),
("vertex_ai/eu.gemini-3.5-flash", True, True, False),
("vertex_ai/us-central1.gemini-3.5-flash", True, True, False),
("gemini/au.gemini-3.5-flash", True, True, False),
("gemini/eu.gemini-3.5-flash", True, True, False),
("gemini-3.1-pro", True, False, True),
("gemini-3.1-pro-preview", True, False, True),
("gemini-3.5-pro", True, False, True),
("au.gemini-3.5-pro", True, False, True),
("eu.gemini-3.1-pro", True, False, True),
("us-central1.gemini-3.5-pro", True, False, True),
("gemini-3-pro-preview", True, False, False),
("au.gemini-2.5-flash", False, False, False),
("eu.gemini-2.5-pro", False, False, False),
("us.gemini-2.5-flash", False, False, False),
("us-central1.gemini-2.0-flash", False, False, False),
("vertex_ai/au.gemini-2.5-flash", False, False, False),
("gemini/eu.gemini-2.5-flash", False, False, False),
("eu.gemma-3-27b-it", False, False, False),
],
)
def test_gemini_3_helpers_with_regional_prefixes(
model: str,
expected_is_gemini_3: bool,
expected_is_flash: bool,
expected_is_3_1_pro_or_newer: bool,
):
assert VertexGeminiConfig._is_gemini_3_or_newer(model) is expected_is_gemini_3
assert VertexGeminiConfig._is_gemini_3_flash(model) is expected_is_flash
assert VertexGeminiConfig._is_gemini_3_1_pro_or_newer(model) is expected_is_3_1_pro_or_newer
@pytest.mark.parametrize(
"model",
[
"gemini-3.5-flash",
"au.gemini-3.5-flash",
"eu.gemini-3.5-flash",
"us.gemini-3.5-flash",
"global.gemini-3.5-flash",
"us-central1.gemini-3.5-flash",
"europe-west4.gemini-3.5-flash",
"australia-southeast1.gemini-3.5-flash",
"vertex_ai/au.gemini-3.5-flash",
"gemini/au.gemini-3.5-flash",
"gemini-3.5-flash-preview",
"gemini-3.1-flash",
"gemini-3-flash",
],
)
@pytest.mark.parametrize(
("reasoning_effort", "expected"),
[
("none", {"thinkingLevel": "minimal", "includeThoughts": False}),
("disable", {"thinkingLevel": "minimal", "includeThoughts": False}),
("minimal", {"thinkingLevel": "minimal", "includeThoughts": True}),
("low", {"thinkingLevel": "low", "includeThoughts": True}),
("medium", {"thinkingLevel": "medium", "includeThoughts": True}),
("high", {"thinkingLevel": "high", "includeThoughts": True}),
],
)
def test_gemini_3_flash_reasoning_effort_mapping_all_regions(
local_model_cost_map,
model: str,
reasoning_effort: str,
expected,
):
result: Final = VertexGeminiConfig._map_reasoning_effort_to_thinking_level(
reasoning_effort=reasoning_effort,
model=model,
)
assert result == expected
@pytest.mark.parametrize(
"model",
[
"gemini-3.1-pro",
"gemini-3.1-pro-preview",
"gemini-3.5-pro",
"au.gemini-3.1-pro",
"eu.gemini-3.5-pro",
"us-central1.gemini-3.5-pro",
"vertex_ai/au.gemini-3.7-flash",
"us-central1.gemini-3.8-flash",
],
)
@pytest.mark.parametrize(
("reasoning_effort", "expected"),
[
("none", {"thinkingLevel": "low", "includeThoughts": False}),
("disable", {"thinkingLevel": "low", "includeThoughts": False}),
("minimal", {"thinkingLevel": "low", "includeThoughts": True}),
("low", {"thinkingLevel": "low", "includeThoughts": True}),
("medium", {"thinkingLevel": "medium", "includeThoughts": True}),
("high", {"thinkingLevel": "high", "includeThoughts": True}),
],
)
def test_gemini_3_pro_and_37_38_flash_reasoning_effort_mapping_all_regions(
local_model_cost_map,
model: str,
reasoning_effort: str,
expected,
):
result: Final = VertexGeminiConfig._map_reasoning_effort_to_thinking_level(
reasoning_effort=reasoning_effort,
model=model,
)
assert result == expected
@pytest.mark.parametrize(
("model", "expected_level"),
[
("gemini-3.5-flash", "minimal"),
("au.gemini-3.5-flash", "minimal"),
("eu.gemini-3.5-flash", "minimal"),
("us.gemini-3.5-flash", "minimal"),
("us-central1.gemini-3.5-flash", "minimal"),
("europe-west4.gemini-3.5-flash", "minimal"),
("australia-southeast1.gemini-3.5-flash", "minimal"),
("vertex_ai/au.gemini-3.5-flash", "minimal"),
("gemini/eu.gemini-3.5-flash", "minimal"),
("gemini-3.5-flash-preview", "minimal"),
("gemini-3.1-flash", "minimal"),
("gemini-3-flash", "minimal"),
("gemini-3.1-pro", "low"),
("gemini-3.1-pro-preview", "low"),
("gemini-3.5-pro", "low"),
("au.gemini-3.5-pro", "low"),
("gemini-3.8-flash", "low"),
("au.gemini-3.8-flash", "low"),
],
)
@pytest.mark.parametrize(
"thinking_param",
[
{"type": "disabled"},
{"type": "enabled", "budget_tokens": 0},
],
)
def test_map_thinking_param_disabled_or_zero_budget_gemini_3(
local_model_cost_map,
model: str,
expected_level: str,
thinking_param,
):
result: Final = VertexGeminiConfig._map_thinking_param(
thinking_param=thinking_param,
model=model,
)
assert result == {
"thinkingLevel": expected_level,
"includeThoughts": False,
}