mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-10 03:28:53 +00:00
Merge 7e4f1da216 into 7083c47998
This commit is contained in:
songkuan-zheng
GitHub
commit
a2152e3f97
3 changed files with 286 additions and 1 deletions
|
|
@ -7,6 +7,8 @@ from ...openai.chat.gpt_transformation import OpenAIGPTConfig
|
|||
|
||||
ZAI_API_BASE: Final = "https://api.z.ai/api/paas/v4"
|
||||
|
||||
_REASONING_PARAMS = ("thinking", "reasoning_effort")
|
||||
|
||||
|
||||
class ZAIChatConfig(OpenAIGPTConfig):
|
||||
@property
|
||||
|
|
@ -49,8 +51,25 @@ class ZAIChatConfig(OpenAIGPTConfig):
|
|||
|
||||
try:
|
||||
if litellm.supports_reasoning(model=model, custom_llm_provider=self.custom_llm_provider):
|
||||
base_params.append("thinking")
|
||||
base_params.extend(_REASONING_PARAMS)
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
return base_params
|
||||
|
||||
def _map_openai_params(
|
||||
self,
|
||||
non_default_params: dict,
|
||||
optional_params: dict,
|
||||
model: str,
|
||||
drop_params: bool,
|
||||
) -> dict:
|
||||
supported = self.get_supported_openai_params(model)
|
||||
for param, value in non_default_params.items():
|
||||
if param not in supported:
|
||||
continue
|
||||
if param in _REASONING_PARAMS:
|
||||
optional_params.setdefault("extra_body", {})[param] = value
|
||||
else:
|
||||
optional_params[param] = value
|
||||
return optional_params
|
||||
|
|
|
|||
|
|
@ -44637,6 +44637,7 @@
|
|||
"max_output_tokens": 32000,
|
||||
"mode": "chat",
|
||||
"supports_function_calling": true,
|
||||
"supports_reasoning": true,
|
||||
"supports_tool_choice": true,
|
||||
"source": "https://docs.z.ai/guides/overview/pricing"
|
||||
},
|
||||
|
|
@ -44648,6 +44649,7 @@
|
|||
"max_output_tokens": 32000,
|
||||
"mode": "chat",
|
||||
"supports_function_calling": true,
|
||||
"supports_reasoning": true,
|
||||
"supports_tool_choice": true,
|
||||
"supports_vision": true,
|
||||
"source": "https://docs.z.ai/guides/overview/pricing"
|
||||
|
|
@ -44660,6 +44662,7 @@
|
|||
"max_output_tokens": 32000,
|
||||
"mode": "chat",
|
||||
"supports_function_calling": true,
|
||||
"supports_reasoning": true,
|
||||
"supports_tool_choice": true,
|
||||
"source": "https://docs.z.ai/guides/overview/pricing"
|
||||
},
|
||||
|
|
@ -44671,6 +44674,7 @@
|
|||
"max_output_tokens": 32000,
|
||||
"mode": "chat",
|
||||
"supports_function_calling": true,
|
||||
"supports_reasoning": true,
|
||||
"supports_tool_choice": true,
|
||||
"source": "https://docs.z.ai/guides/overview/pricing"
|
||||
},
|
||||
|
|
@ -44682,6 +44686,7 @@
|
|||
"max_output_tokens": 32000,
|
||||
"mode": "chat",
|
||||
"supports_function_calling": true,
|
||||
"supports_reasoning": true,
|
||||
"supports_tool_choice": true,
|
||||
"source": "https://docs.z.ai/guides/overview/pricing"
|
||||
},
|
||||
|
|
@ -44704,6 +44709,7 @@
|
|||
"max_output_tokens": 32000,
|
||||
"mode": "chat",
|
||||
"supports_function_calling": true,
|
||||
"supports_reasoning": true,
|
||||
"supports_tool_choice": true,
|
||||
"source": "https://docs.z.ai/guides/overview/pricing"
|
||||
},
|
||||
|
|
|
|||
|
|
@ -172,3 +172,263 @@ def test_zai_sync_completion(respx_mock, zai_response, monkeypatch):
|
|||
|
||||
assert response.choices[0].message.content == "Hello! How can I help you today?"
|
||||
assert response.usage.total_tokens == 25
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def zai_thinking_response():
|
||||
return {
|
||||
"id": "chatcmpl-zai-thinking",
|
||||
"object": "chat.completion",
|
||||
"created": 1700000000,
|
||||
"model": "glm-4.6",
|
||||
"choices": [
|
||||
{
|
||||
"index": 0,
|
||||
"message": {"role": "assistant", "content": "hi"},
|
||||
"finish_reason": "stop",
|
||||
}
|
||||
],
|
||||
"usage": {"prompt_tokens": 5, "completion_tokens": 2, "total_tokens": 7},
|
||||
}
|
||||
|
||||
|
||||
def _captured_body(respx_mock):
|
||||
assert len(respx_mock.calls) == 1
|
||||
return json.loads(respx_mock.calls[0].request.content.decode("utf-8"))
|
||||
|
||||
|
||||
class TestZaiSupportedParamsWhitelistReasoning:
|
||||
"""`thinking` and `reasoning_effort` enter the whitelist only when the
|
||||
registry marks the model `supports_reasoning: true`. The registry
|
||||
update in this PR adds the flag to the entire GLM-4.5 family
|
||||
(previously every GLM-4.5 entry in
|
||||
`model_prices_and_context_window_backup.json` was missing the flag
|
||||
despite docs.z.ai listing GLM-4.5 as the first model with `thinking`
|
||||
support), so the gate now unlocks reasoning params for all of them.
|
||||
"""
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def _use_local_model_cost(self, monkeypatch):
|
||||
monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
|
||||
litellm.model_cost = litellm.get_model_cost_map(url="")
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"model",
|
||||
[
|
||||
"glm-4.5",
|
||||
"glm-4.5v",
|
||||
"glm-4.5-air",
|
||||
"glm-4.5-airx",
|
||||
"glm-4.5-x",
|
||||
"glm-4.5-flash",
|
||||
"glm-4.6",
|
||||
"glm-4.7",
|
||||
],
|
||||
)
|
||||
def test_reasoning_params_in_whitelist(self, model):
|
||||
from litellm.llms.zai.chat.transformation import ZAIChatConfig
|
||||
|
||||
params = ZAIChatConfig().get_supported_openai_params(model=model)
|
||||
assert "thinking" in params
|
||||
assert "reasoning_effort" in params
|
||||
|
||||
def test_reasoning_params_excluded_when_registry_flag_missing(self, monkeypatch):
|
||||
"""Regression guard for the gate. A model whose registry entry
|
||||
does NOT mark `supports_reasoning: true` must keep `thinking`
|
||||
and `reasoning_effort` OUT of the whitelist — otherwise a new
|
||||
ZAI model added to the registry without the flag silently
|
||||
accepts reasoning kwargs that the upstream API will reject.
|
||||
"""
|
||||
from litellm.llms.zai.chat.transformation import ZAIChatConfig
|
||||
|
||||
# Synthetic model that won't match any registry entry or
|
||||
# wildcard pattern.
|
||||
params = ZAIChatConfig().get_supported_openai_params(
|
||||
model="glm-no-such-future-model-xyz"
|
||||
)
|
||||
assert "thinking" not in params
|
||||
assert "reasoning_effort" not in params
|
||||
|
||||
|
||||
class TestZaiReasoningParamsLandInExtraBody:
|
||||
"""The OpenAI Python SDK rejects unknown top-level kwargs (e.g.
|
||||
`AsyncCompletions.create() got an unexpected keyword argument 'thinking'`),
|
||||
so ZAI-specific reasoning fields must travel inside `extra_body` and
|
||||
let the SDK flatten them into the HTTP body. Without this wrapping a
|
||||
chained-proxy topology (LiteLLM A -> LiteLLM B -> ZAI) drops the
|
||||
field on hop 2: hop 1's SDK flattens `extra_body` into a top-level
|
||||
`thinking` kwarg, hop 2 re-emits that as a top-level kwarg, and the
|
||||
SDK rejects it
|
||||
"""
|
||||
|
||||
def test_thinking_wraps_into_extra_body(self):
|
||||
from litellm.llms.zai.chat.transformation import ZAIChatConfig
|
||||
|
||||
result = ZAIChatConfig()._map_openai_params(
|
||||
non_default_params={"thinking": {"type": "disabled"}},
|
||||
optional_params={},
|
||||
model="glm-4.6",
|
||||
drop_params=False,
|
||||
)
|
||||
assert "thinking" not in result
|
||||
assert result["extra_body"]["thinking"] == {"type": "disabled"}
|
||||
|
||||
def test_reasoning_effort_wraps_into_extra_body(self):
|
||||
from litellm.llms.zai.chat.transformation import ZAIChatConfig
|
||||
|
||||
result = ZAIChatConfig()._map_openai_params(
|
||||
non_default_params={"reasoning_effort": "none"},
|
||||
optional_params={},
|
||||
model="glm-5",
|
||||
drop_params=False,
|
||||
)
|
||||
assert "reasoning_effort" not in result
|
||||
assert result["extra_body"]["reasoning_effort"] == "none"
|
||||
|
||||
def test_thinking_merges_into_existing_extra_body(self):
|
||||
from litellm.llms.zai.chat.transformation import ZAIChatConfig
|
||||
|
||||
result = ZAIChatConfig()._map_openai_params(
|
||||
non_default_params={"thinking": {"type": "disabled"}},
|
||||
optional_params={"extra_body": {"already_here": True}},
|
||||
model="glm-4.6",
|
||||
drop_params=False,
|
||||
)
|
||||
assert result["extra_body"]["already_here"] is True
|
||||
assert result["extra_body"]["thinking"] == {"type": "disabled"}
|
||||
|
||||
def test_standard_params_stay_top_level_alongside_thinking(self):
|
||||
from litellm.llms.zai.chat.transformation import ZAIChatConfig
|
||||
|
||||
result = ZAIChatConfig()._map_openai_params(
|
||||
non_default_params={
|
||||
"max_tokens": 100,
|
||||
"temperature": 0.7,
|
||||
"thinking": {"type": "enabled"},
|
||||
},
|
||||
optional_params={},
|
||||
model="glm-4.6",
|
||||
drop_params=False,
|
||||
)
|
||||
assert result["max_tokens"] == 100
|
||||
assert result["temperature"] == 0.7
|
||||
assert result["extra_body"]["thinking"] == {"type": "enabled"}
|
||||
assert "thinking" not in result
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_top_level_thinking_kwarg_reaches_http_body(
|
||||
self, respx_mock, zai_thinking_response, monkeypatch
|
||||
):
|
||||
"""Regression for the hop-2 SDK crash. Pre-fix this raised
|
||||
`AsyncCompletions.create() got an unexpected keyword argument
|
||||
'thinking'`. Post-fix the boundary HTTP body carries `thinking`
|
||||
as a top-level field (the OpenAI SDK flattened `extra_body`)
|
||||
"""
|
||||
monkeypatch.setenv("ZAI_API_KEY", "test-key")
|
||||
monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
|
||||
respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(
|
||||
json=zai_thinking_response
|
||||
)
|
||||
|
||||
await litellm.acompletion(
|
||||
model="zai/glm-4.6",
|
||||
messages=[{"role": "user", "content": "hi"}],
|
||||
thinking={"type": "disabled"},
|
||||
)
|
||||
|
||||
body = _captured_body(respx_mock)
|
||||
assert body["thinking"] == {"type": "disabled"}
|
||||
assert "extra_body" not in body
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_extra_body_thinking_reaches_http_body(
|
||||
self, respx_mock, zai_thinking_response, monkeypatch
|
||||
):
|
||||
monkeypatch.setenv("ZAI_API_KEY", "test-key")
|
||||
monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
|
||||
respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(
|
||||
json=zai_thinking_response
|
||||
)
|
||||
|
||||
await litellm.acompletion(
|
||||
model="zai/glm-4.6",
|
||||
messages=[{"role": "user", "content": "hi"}],
|
||||
extra_body={"thinking": {"type": "disabled"}},
|
||||
)
|
||||
|
||||
body = _captured_body(respx_mock)
|
||||
assert body["thinking"] == {"type": "disabled"}
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_top_level_reasoning_effort_reaches_http_body(
|
||||
self, respx_mock, zai_thinking_response, monkeypatch
|
||||
):
|
||||
monkeypatch.setenv("ZAI_API_KEY", "test-key")
|
||||
monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
|
||||
respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(
|
||||
json=zai_thinking_response
|
||||
)
|
||||
|
||||
await litellm.acompletion(
|
||||
model="zai/glm-5",
|
||||
messages=[{"role": "user", "content": "hi"}],
|
||||
reasoning_effort="none",
|
||||
)
|
||||
|
||||
body = _captured_body(respx_mock)
|
||||
assert body["reasoning_effort"] == "none"
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_thinking_works_on_glm_4_5_via_registry_flag(
|
||||
self, respx_mock, zai_thinking_response, monkeypatch
|
||||
):
|
||||
"""End-to-end: GLM-4.5 carries `supports_reasoning: true` in the
|
||||
registry, so the gate in `get_supported_openai_params` lets
|
||||
`thinking` through and the SDK boundary lands it in the HTTP
|
||||
body. Without the registry update this test would fail with the
|
||||
SDK rejecting `thinking` as an unsupported param.
|
||||
"""
|
||||
monkeypatch.setenv("ZAI_API_KEY", "test-key")
|
||||
monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
|
||||
litellm.model_cost = litellm.get_model_cost_map(url="")
|
||||
monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
|
||||
respx_mock.post("https://api.z.ai/api/paas/v4/chat/completions").respond(
|
||||
json=zai_thinking_response
|
||||
)
|
||||
|
||||
await litellm.acompletion(
|
||||
model="zai/glm-4.5",
|
||||
messages=[{"role": "user", "content": "hi"}],
|
||||
thinking={"type": "disabled"},
|
||||
)
|
||||
|
||||
body = _captured_body(respx_mock)
|
||||
assert body["thinking"] == {"type": "disabled"}
|
||||
|
||||
|
||||
class TestGlm45FamilyRegistrySupportsReasoning:
|
||||
"""docs.z.ai lists GLM-4.5 as the first model family that supports
|
||||
`thinking`. The registry had every GLM-4.5 entry marked
|
||||
`supports_reasoning: false`, which is the source-of-truth bug that
|
||||
let the SDK-kwarg crash hide for so long
|
||||
"""
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"model_key",
|
||||
[
|
||||
"zai/glm-4.5",
|
||||
"zai/glm-4.5v",
|
||||
"zai/glm-4.5-x",
|
||||
"zai/glm-4.5-air",
|
||||
"zai/glm-4.5-airx",
|
||||
"zai/glm-4.5-flash",
|
||||
],
|
||||
)
|
||||
def test_supports_reasoning_true(self, model_key):
|
||||
import os
|
||||
|
||||
os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
|
||||
litellm.model_cost = litellm.get_model_cost_map(url="")
|
||||
|
||||
assert model_key in litellm.model_cost
|
||||
assert litellm.model_cost[model_key].get("supports_reasoning") is True
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue