scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-07 08:26:10 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 36
litellm/tests/test_litellm/test_main.py
mateo-berri e16cf5ed3f Fall back to the GPT version rule when the cost map carries no breakpoint flag 
A proxy on the default remote cost map never produced a prompt cache
breakpoint: the published map has the gpt-5.6 entries without
supports_prompt_cache_breakpoint, so the model-map gate returned False
for every listed model and only LITELLM_LOCAL_MODEL_COST_MAP=True (the
repo .env, hence the passing unit tests) made the feature work. The hook
now honors the flag when the entry carries one, True or False, and
otherwise applies the GPT-5.6+ version rule to the model name, so a map
that lags the flag still gets the OpenAI dialect. The model-map tests
pin litellm.model_cost to the bundled backup map and a new test drives
the hook against an unflagged gpt-5.6 entry.

completion() and acompletion() take base_url as an alias for api_base
that only lands on api_base after the cache control hook ran, so a
GPT-5.6 call at a non-OpenAI gateway given through base_url still got
the dialect. Both seed calls and the unstamped request-params read now
look at base_url too.

ResponsesAPIRequestUtils.merge_prompt_management_input reshaped hook
output in place, retyping text parts to input_text on the caller's own
message objects. The merge now shapes a copy of each message as it
emits it, so the identity-based merge keeps working on the hook's
objects and nothing the hook or the client owns is mutated.
2026-08-20 05:47:19 -07:00

2756 lines
100 KiB
Python
Raw Permalink Blame History

import contextlib
import copy
import json
import os
import sys
import httpx
import pytest
import respx
from fastapi.testclient import TestClient
sys.path.insert(
0, os.path.abspath("../..")
) # Adds the parent directory to the system path
import urllib.parse
from unittest.mock import MagicMock, patch
import litellm
from litellm import main as litellm_main
async def _async_fake_bedrock_image_details(image_url):
return "ZmFrZS1pbWFnZQ==", "image/png"
@pytest.fixture(autouse=True)
def clear_client_cache():
"""
Clear the HTTP client cache before each test to ensure mocks are used.
This prevents cached real clients from being reused across tests.
"""
cache = getattr(litellm, "in_memory_llm_clients_cache", None)
if cache is not None:
cache.flush_cache()
yield
if cache is not None:
cache.flush_cache()
@pytest.fixture(autouse=True)
def add_api_keys_to_env(monkeypatch):
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-api03-1234567890")
monkeypatch.setenv("OPENAI_API_KEY", "sk-openai-api03-1234567890")
monkeypatch.setenv("AWS_ACCESS_KEY_ID", "my-fake-aws-access-key-id")
monkeypatch.setenv("AWS_SECRET_ACCESS_KEY", "my-fake-aws-secret-access-key")
monkeypatch.setenv("AWS_REGION", "us-east-1")
# Keep these transformation tests on the simple access-key path. A leaked
# session token or role/web-identity env var pushes Bedrock auth down a
# different branch and fails before the mocked HTTP client is exercised.
monkeypatch.delenv("AWS_SESSION_TOKEN", raising=False)
monkeypatch.delenv("AWS_ROLE_ARN", raising=False)
monkeypatch.delenv("AWS_WEB_IDENTITY_TOKEN_FILE", raising=False)
@pytest.fixture
def openai_api_response():
mock_response_data = {
"id": "chatcmpl-B0W3vmiM78Xkgx7kI7dr7PC949DMS",
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": None,
"message": {
"content": "",
"refusal": None,
"role": "assistant",
"audio": None,
"function_call": None,
"tool_calls": None,
},
}
],
"created": 1739462947,
"model": "gpt-4o-mini-2024-07-18",
"object": "chat.completion",
"service_tier": "default",
"system_fingerprint": "fp_bd83329f63",
"usage": {
"completion_tokens": 1,
"prompt_tokens": 121,
"total_tokens": 122,
"completion_tokens_details": {
"accepted_prediction_tokens": 0,
"audio_tokens": 0,
"reasoning_tokens": 0,
"rejected_prediction_tokens": 0,
},
"prompt_tokens_details": {"audio_tokens": 0, "cached_tokens": 0},
},
}
return mock_response_data
def test_completion_missing_role(openai_api_response):
from openai import OpenAI
from litellm.types.utils import ModelResponse
client = OpenAI(api_key="test_api_key")
mock_raw_response = MagicMock()
mock_raw_response.headers = {
"x-request-id": "123",
"openai-organization": "org-123",
"x-ratelimit-limit-requests": "100",
"x-ratelimit-remaining-requests": "99",
}
mock_raw_response.parse.return_value = ModelResponse(**openai_api_response)
print(f"openai_api_response: {openai_api_response}")
with patch.object(
client.chat.completions.with_raw_response, "create", mock_raw_response
) as mock_create:
litellm.completion(
model="gpt-4o-mini",
messages=[
{"role": "user", "content": "Hey"},
{
"content": "",
"tool_calls": [
{
"id": "call_m0vFJjQmTH1McvaHBPR2YFwY",
"function": {
"arguments": '{"input": "dksjsdkjdhskdjshdskhjkhlk"}',
"name": "tool_name",
},
"type": "function",
"index": 0,
},
{
"id": "call_Vw6RaqV2n5aaANXEdp5pYxo2",
"function": {
"arguments": '{"input": "jkljlkjlkjlkjlk"}',
"name": "tool_name",
},
"type": "function",
"index": 1,
},
{
"id": "call_hBIKwldUEGlNh6NlSXil62K4",
"function": {
"arguments": '{"input": "jkjlkjlkjlkj;lj"}',
"name": "tool_name",
},
"type": "function",
"index": 2,
},
],
},
],
client=client,
)
mock_create.assert_called_once()
@pytest.mark.parametrize(
"model",
[
"gemini/gemini-1.5-flash",
"bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0",
"bedrock/invoke/anthropic.claude-haiku-4-5-20251001-v1:0",
"anthropic/claude-3-5-sonnet",
],
)
@pytest.mark.parametrize("sync_mode", [True, False])
@pytest.mark.asyncio
async def test_url_with_format_param(model, sync_mode, monkeypatch):
from litellm import acompletion, completion
from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
from litellm.litellm_core_utils.prompt_templates import factory as prompt_factory
if sync_mode:
client = HTTPHandler()
else:
client = AsyncHTTPHandler()
# This test is about request shaping, not live image downloads. Stub the
# URL->image conversion helpers so suite-level network/client state from
# earlier tests cannot prevent the mocked provider client from being hit.
fake_base64_image = "data:image/png;base64,ZmFrZS1pbWFnZQ=="
monkeypatch.setattr(
prompt_factory, "convert_url_to_base64", lambda url: fake_base64_image
)
monkeypatch.setattr(
prompt_factory.BedrockImageProcessor,
"get_image_details",
staticmethod(lambda image_url: ("ZmFrZS1pbWFnZQ==", "image/png")),
)
monkeypatch.setattr(
prompt_factory.BedrockImageProcessor,
"get_image_details_async",
staticmethod(_async_fake_bedrock_image_details),
)
args = {
"model": model,
"messages": [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": "https://awsmp-logos.s3.amazonaws.com/seller-xw5kijmvmzasy/c233c9ade2ccb5491072ae232c814942.png",
"format": "image/png",
},
},
{"type": "text", "text": "Describe this image"},
],
}
],
}
if model.startswith("gemini/"):
args["api_key"] = "test-api-key"
with patch.object(client, "post", new=MagicMock()) as mock_client:
try:
if sync_mode:
response = completion(**args, client=client)
else:
response = await acompletion(**args, client=client)
print(response)
except Exception as e:
pass
mock_client.assert_called()
print(mock_client.call_args.kwargs)
if "data" in mock_client.call_args.kwargs:
json_str = mock_client.call_args.kwargs["data"]
else:
json_str = json.dumps(mock_client.call_args.kwargs["json"])
if isinstance(json_str, bytes):
json_str = json_str.decode("utf-8")
print(f"type of json_str: {type(json_str)}")
# Bedrock models convert URLs to base64, while direct Anthropic models support URLs
# bedrock/invoke models use Anthropic messages API which supports URLs
if model.startswith("bedrock/invoke/"):
# bedrock/invoke should convert URLs to base64 (doesn't support URL references)
# URL should NOT be in the JSON (it should be converted to base64)
assert "https://awsmp-logos.s3.amazonaws.com" not in json_str
# Should have base64 data in the source (type="base64", not type="url")
assert '"type":"base64"' in json_str or '"type": "base64"' in json_str
# Should have "data" field containing base64 content
assert '"data"' in json_str
elif model.startswith("bedrock/"):
# Regular Bedrock models should convert URLs to base64 (uses "bytes" field)
# URL should NOT be in the JSON (it should be converted to base64)
assert "https://awsmp-logos.s3.amazonaws.com" not in json_str
# Should have "bytes" field (Bedrock uses "bytes" not "base64" in the field name)
assert '"bytes"' in json_str or '"bytes":' in json_str
elif model.startswith("anthropic/"):
# Direct Anthropic models should pass HTTPS URLs directly (HTTP URLs are converted to base64)
# Since we're using HTTPS URL, it should be passed as-is
assert "https://awsmp-logos.s3.amazonaws.com" in json_str
# For Anthropic, URL references use "url" type, not base64
assert '"type":"url"' in json_str or '"type": "url"' in json_str
else:
# For other models, check format parameter is respected
assert "png" in json_str
assert "jpeg" not in json_str
@pytest.mark.parametrize("model", ["gpt-4o-mini"])
@pytest.mark.parametrize("sync_mode", [True, False])
@pytest.mark.asyncio
async def test_url_with_format_param_openai(model, sync_mode):
from openai import AsyncOpenAI, OpenAI
from litellm import acompletion, completion
if sync_mode:
client = OpenAI()
else:
client = AsyncOpenAI()
args = {
"model": model,
"messages": [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": "https://awsmp-logos.s3.amazonaws.com/seller-xw5kijmvmzasy/c233c9ade2ccb5491072ae232c814942.png",
"format": "image/png",
},
},
{"type": "text", "text": "Describe this image"},
],
}
],
}
with patch.object(
client.chat.completions.with_raw_response, "create"
) as mock_client:
try:
if sync_mode:
response = completion(**args, client=client)
else:
response = await acompletion(**args, client=client)
print(response)
except Exception as e:
print(e)
mock_client.assert_called()
print(mock_client.call_args.kwargs)
json_str = json.dumps(mock_client.call_args.kwargs)
assert "format" not in json_str
def test_bedrock_latency_optimized_inference():
from litellm.llms.custom_httpx.http_handler import HTTPHandler
client = HTTPHandler()
with patch.object(client, "post") as mock_post:
try:
response = litellm.completion(
model="bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0",
messages=[{"role": "user", "content": "Hello, how are you?"}],
performanceConfig={"latency": "optimized"},
client=client,
)
except Exception as e:
print(e)
mock_post.assert_called_once()
json_data = json.loads(mock_post.call_args.kwargs["data"])
assert json_data["performanceConfig"]["latency"] == "optimized"
def test_strip_input_examples_for_non_anthropic_providers():
tools = [
{
"type": "function",
"name": "example_tool",
"input_examples": [{"foo": "bar"}],
"function": {
"name": "example_tool",
"input_examples": [{"foo": "bar"}],
},
}
]
assert not litellm_main._should_allow_input_examples(
custom_llm_provider="openai", model="gpt-4o-mini"
)
cleaned = litellm_main._drop_input_examples_from_tools(tools=tools)
assert isinstance(cleaned, list)
assert "input_examples" not in cleaned[0]
assert "input_examples" not in cleaned[0]["function"]
def test_custom_provider_with_extra_headers():
from litellm.llms.custom_httpx.http_handler import HTTPHandler
with patch.object(
litellm.llms.custom_httpx.http_handler.HTTPHandler, "post"
) as mock_post:
response = litellm.completion(
model="custom/custom",
messages=[{"role": "user", "content": "Hello, how are you?"}],
headers={"X-Custom-Header": "custom-value"},
api_base="https://example.com/api/v1",
)
mock_post.assert_called_once()
assert mock_post.call_args[1]["headers"]["X-Custom-Header"] == "custom-value"
def test_custom_provider_with_extra_body():
from litellm.llms.custom_httpx.http_handler import HTTPHandler
with patch.object(
litellm.llms.custom_httpx.http_handler.HTTPHandler, "post"
) as mock_post:
response = litellm.completion(
model="custom/custom",
messages=[{"role": "user", "content": "Hello, how are you?"}],
extra_body={
"X-Custom-BodyValue": "custom-value",
"X-Custom-BodyValue2": "custom-value2",
},
api_base="https://example.com/api/v1",
)
mock_post.assert_called_once()
assert mock_post.call_args[1]["json"]["X-Custom-BodyValue"] == "custom-value"
assert mock_post.call_args[1]["json"] == {
"model": "custom",
"params": {
"prompt": ["Hello, how are you?"],
"max_tokens": None,
"temperature": None,
"top_p": None,
"top_k": None,
},
"X-Custom-BodyValue": "custom-value",
"X-Custom-BodyValue2": "custom-value2",
}
# test that extra_body is not passed if not provided
with patch.object(
litellm.llms.custom_httpx.http_handler.HTTPHandler, "post"
) as mock_post:
response = litellm.completion(
model="custom/custom",
messages=[{"role": "user", "content": "Hello, how are you?"}],
api_base="https://example.com/api/v1",
)
mock_post.assert_called_once()
assert mock_post.call_args[1]["json"] == {
"model": "custom",
"params": {
"prompt": ["Hello, how are you?"],
"max_tokens": None,
"temperature": None,
"top_p": None,
"top_k": None,
},
}
@pytest.fixture(autouse=True)
def set_openrouter_api_key():
original_api_key = os.environ.get("OPENROUTER_API_KEY")
os.environ["OPENROUTER_API_KEY"] = "fake-key-for-testing"
yield
if original_api_key is not None:
os.environ["OPENROUTER_API_KEY"] = original_api_key
else:
del os.environ["OPENROUTER_API_KEY"]
@pytest.mark.asyncio
async def test_extra_body_with_fallback(
respx_mock: respx.MockRouter, set_openrouter_api_key, monkeypatch
):
"""
test regression for https://github.com/BerriAI/litellm/issues/8425.
This was perhaps a wider issue with the acompletion function not passing kwargs such as extra_body correctly when fallbacks are specified.
"""
# Save original state to restore after test
original_disable_aiohttp = litellm.disable_aiohttp_transport
try:
# since this uses respx, we need to set use_aiohttp_transport to False
# Set both the global variable and environment variable to ensure it takes effect
litellm.disable_aiohttp_transport = True
monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
# Flush cache to ensure no stale aiohttp clients are used
litellm.in_memory_llm_clients_cache.flush_cache()
# Set up test parameters
model = "openrouter/deepseek/deepseek-chat"
messages = [{"role": "user", "content": "Hello, world!"}]
extra_body = {
"provider": {
"order": ["DeepSeek"],
"allow_fallbacks": False,
"require_parameters": True,
}
}
fallbacks = [{"model": "openrouter/google/gemini-flash-1.5-8b"}]
# Set up mock to respond to any POST request to the OpenRouter endpoint
# This ensures it works for both primary and fallback models
mock_route = respx_mock.post("https://openrouter.ai/api/v1/chat/completions")
mock_route.return_value = httpx.Response(
200,
json={
"id": "chatcmpl-123",
"object": "chat.completion",
"created": 1677652288,
"model": model,
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Hello from mocked response!",
},
"finish_reason": "stop",
}
],
"usage": {
"prompt_tokens": 9,
"completion_tokens": 12,
"total_tokens": 21,
},
},
)
response = await litellm.acompletion(
model=model,
messages=messages,
extra_body=extra_body,
fallbacks=fallbacks,
api_key="fake-openrouter-api-key",
)
# Verify the response
assert response is not None
assert (
len(respx_mock.calls) > 0
), "Mock was not called - check if aiohttp transport is properly disabled"
# Get the request from the mock
request: httpx.Request = respx_mock.calls[0].request
request_body = request.read()
request_body = json.loads(request_body)
# Verify basic parameters
assert request_body["model"] == "deepseek/deepseek-chat"
assert request_body["messages"] == messages
# Verify the extra_body parameters remain under the provider key
assert request_body["provider"]["order"] == ["DeepSeek"]
assert request_body["provider"]["allow_fallbacks"] is False
assert request_body["provider"]["require_parameters"] is True
finally:
# Restore original state to prevent test pollution
litellm.disable_aiohttp_transport = original_disable_aiohttp
litellm.in_memory_llm_clients_cache.flush_cache()
@pytest.mark.parametrize("env_base", ["OPENAI_BASE_URL", "OPENAI_API_BASE"])
@pytest.mark.asyncio
@pytest.mark.flaky(retries=3, delay=1)
async def test_openai_env_base(
respx_mock: respx.MockRouter, env_base, openai_api_response, monkeypatch
):
"This tests OpenAI env variables are honored, including legacy OPENAI_API_BASE"
# Ensure aiohttp transport is disabled to use httpx which respx can mock
litellm.disable_aiohttp_transport = True
expected_base_url = "http://localhost:12345/v1"
# Assign the environment variable based on env_base, and use a fake API key.
monkeypatch.setenv(env_base, expected_base_url)
monkeypatch.setenv("OPENAI_API_KEY", "fake_openai_api_key")
model = "gpt-4o"
messages = [{"role": "user", "content": "Hello, how are you?"}]
# Configure respx mock to intercept the request
mock_route = respx_mock.post(
url__regex=r"http://localhost:12345/v1/chat/completions.*"
).mock(
return_value=httpx.Response(
status_code=200,
json={
"id": "chatcmpl-123",
"object": "chat.completion",
"created": 1677652288,
"model": model,
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Hello from mocked response!",
},
"finish_reason": "stop",
}
],
"usage": {
"prompt_tokens": 9,
"completion_tokens": 12,
"total_tokens": 21,
},
},
)
)
try:
response = await litellm.acompletion(model=model, messages=messages)
# verify we had a response
assert response.choices[0].message.content == "Hello from mocked response!"
# Verify the mock was called
assert (
mock_route.called
), "Mock route was not called - request may have bypassed respx"
finally:
# Clean up to avoid affecting other tests
litellm.disable_aiohttp_transport = False
def build_database_url(username, password, host, dbname):
username_enc = urllib.parse.quote_plus(username)
password_enc = urllib.parse.quote_plus(password)
dbname_enc = urllib.parse.quote_plus(dbname)
return f"postgresql://{username_enc}:{password_enc}@{host}/{dbname_enc}"
def test_build_database_url():
url = build_database_url("user@name", "p@ss:word", "localhost", "db/name")
assert url == "postgresql://user%40name:p%40ss%3Aword@localhost/db%2Fname"
def test_bedrock_llama():
litellm._turn_on_debug()
from litellm.types.utils import CallTypes
from litellm.utils import return_raw_request
model = "bedrock/invoke/us.meta.llama4-scout-17b-instruct-v1:0"
request = return_raw_request(
endpoint=CallTypes.completion,
kwargs={
"model": model,
"messages": [
{"role": "user", "content": "hi"},
],
},
)
print(request)
assert (
request["raw_request_body"]["prompt"]
== "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nhi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
)
def _mocked_openai_chat_response(model: str) -> httpx.Response:
return httpx.Response(
status_code=200,
json={
"id": "chatcmpl-123",
"object": "chat.completion",
"created": 1677652288,
"model": model,
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Hello from mocked response!",
},
"finish_reason": "stop",
}
],
"usage": {
"prompt_tokens": 9,
"completion_tokens": 12,
"total_tokens": 21,
},
},
)
def test_completion_forwards_verbosity_in_raw_request(respx_mock: respx.MockRouter):
"""Regression test: completion() must forward the verbosity param to the provider request body."""
from litellm.types.utils import CallTypes
from litellm.utils import return_raw_request
model = "gpt-5.2"
messages = [{"role": "user", "content": "hi"}]
respx_mock.post("https://api.openai.com/v1/chat/completions").mock(
return_value=_mocked_openai_chat_response(model)
)
request = return_raw_request(
endpoint=CallTypes.completion,
kwargs={
"model": model,
"messages": messages,
"verbosity": "high",
},
)
assert request["raw_request_body"]["verbosity"] == "high"
assert request["raw_request_body"]["model"] == model
assert request["raw_request_body"]["messages"] == messages
@pytest.mark.asyncio
async def test_acompletion_forwards_verbosity_to_provider_request(
respx_mock: respx.MockRouter, monkeypatch
):
"""Regression test: acompletion() must forward the verbosity param to the provider request body."""
original_disable_aiohttp = litellm.disable_aiohttp_transport
try:
litellm.disable_aiohttp_transport = True
monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
litellm.in_memory_llm_clients_cache.flush_cache()
model = "gpt-5.2"
messages = [{"role": "user", "content": "hi"}]
mock_route = respx_mock.post("https://api.openai.com/v1/chat/completions").mock(
return_value=_mocked_openai_chat_response(model)
)
response = await litellm.acompletion(
model=model,
messages=messages,
verbosity="low",
api_key="fake-openai-api-key",
)
assert response.choices[0].message.content == "Hello from mocked response!"
assert mock_route.called
request_body = json.loads(respx_mock.calls[0].request.read())
assert request_body["verbosity"] == "low"
assert request_body["model"] == model
assert request_body["messages"] == messages
finally:
litellm.disable_aiohttp_transport = original_disable_aiohttp
litellm.in_memory_llm_clients_cache.flush_cache()
def test_responses_api_bridge_check_strips_responses_prefix():
"""Test that responses_api_bridge_check strips 'responses/' prefix and sets mode."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 4096}
model_info, model = responses_api_bridge_check(
model="responses/gpt-4-responses",
custom_llm_provider="openai",
)
assert model == "gpt-4-responses"
assert model_info["mode"] == "responses"
def test_responses_api_bridge_check_gpt_5_4_pro():
"""Test that gpt-5.4-pro routes through responses API bridge, not chat completions.
Regression test for https://github.com/BerriAI/litellm/issues/23014
gpt-5.4-pro is a responses-only model and must not be sent to /v1/chat/completions.
"""
from litellm.main import responses_api_bridge_check
for model_name in ["gpt-5.4-pro", "gpt-5.4-pro-2026-03-05"]:
model_info, model = responses_api_bridge_check(
model=model_name,
custom_llm_provider="openai",
)
assert (
model_info.get("mode") == "responses"
), f"{model_name} should have mode='responses', got '{model_info.get('mode')}'"
def test_responses_api_bridge_check_gpt_5_4_tools_plus_reasoning_routes_to_responses():
"""gpt-5.4 with both tools and reasoning_effort should route to Responses API."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort="xhigh",
)
assert model == "gpt-5.4"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_gpt_5_5_tools_plus_reasoning_routes_to_responses():
"""gpt-5.5+ with both tools and reasoning_effort should route to Responses API."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.5-pro",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort="xhigh",
)
assert model == "gpt-5.5-pro"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_azure_gpt_5_4_tools_plus_reasoning_routes_to_responses():
"""Azure gpt-5.4 with both tools and reasoning_effort should route to Responses API."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="azure",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort="high",
)
assert model == "gpt-5.4"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_azure_gpt_5_4_tools_with_default_reasoning_routes_to_responses():
"""
Azure gpt-5.4 with tools and UNSET reasoning_effort must bridge: OpenAI enables
reasoning by default for gpt-5.4+, and Chat Completions rejects function tools
whenever reasoning is on.
"""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="azure",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
)
assert model == "gpt-5.4"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_gpt_5_4_tools_with_default_reasoning_routes_to_responses():
"""
gpt-5.4 with tools and UNSET reasoning_effort must bridge: OpenAI enables reasoning
by default for gpt-5.4+, and Chat Completions rejects function tools whenever
reasoning is on ("use /v1/responses or set reasoning_effort to 'none'").
"""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
)
assert model == "gpt-5.4"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_gpt_5_4_tools_with_reasoning_none_stays_chat():
"""
Explicit reasoning_effort "none" is OpenAI's documented escape hatch that keeps
function tools servable on Chat Completions; the bridge must not fire.
"""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort="none",
)
assert model == "gpt-5.4"
assert model_info.get("mode") != "responses"
def test_responses_api_bridge_check_reasoning_none_with_summary_still_routes_to_responses():
"""A reasoning summary is Responses-only regardless of effort value."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="openai",
reasoning_effort="none",
reasoning_summary="detailed",
)
assert model == "gpt-5.4"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_gpt_5_4_custom_tools_only_stays_chat():
"""
Chat Completions serves custom (grammar) tools natively with reasoning on; only
FUNCTION tools trigger the OpenAI rejection. Custom-only requests must stay on chat
so responses keep the native custom tool_call shape instead of the bridge's
function-shaped mapping.
"""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "custom", "custom": {"name": "ApplyPatch", "description": "V4A patch"}}],
reasoning_effort=None,
)
assert model == "gpt-5.6"
assert model_info.get("mode") != "responses"
def test_responses_api_bridge_check_gpt_5_4_mixed_function_and_custom_tools_routes_to_responses():
"""One function tool in the mix is enough to make chat unservable with reasoning on."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[
{"type": "custom", "custom": {"name": "ApplyPatch"}},
{"type": "function", "function": {"name": "shell"}},
],
reasoning_effort=None,
)
assert model == "gpt-5.6"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_gpt_5_4_flat_function_tool_routes_to_responses():
"""Responses-style flat function tool defs still count as function tools."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "name": "shell", "parameters": {"type": "object"}}],
reasoning_effort=None,
)
assert model == "gpt-5.6"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_dict_effort_none_stays_chat():
"""The escape hatch must honor litellm's dict form: {"effort": "none"} means reasoning off."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort={"effort": "none"},
)
assert model == "gpt-5.6"
assert model_info.get("mode") != "responses"
def test_responses_api_bridge_check_dict_effort_active_routes_to_responses():
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort={"effort": "low"},
)
assert model == "gpt-5.6"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_dict_effort_none_with_summary_routes_to_responses():
"""A summary inside the dict form is Responses-only even when effort is none."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort={"effort": "none", "summary": "concise"},
)
assert model == "gpt-5.6"
assert model_info.get("mode") == "responses"
@pytest.mark.parametrize("blank_api_base", [None, "", " ", "\t"])
def test_responses_api_bridge_check_blank_api_base_is_default_openai(blank_api_base):
"""
A blank api_base (None, empty, or whitespace) resolves to the default OpenAI
endpoint downstream, which enforces the reasoning+tools constraint, so gpt-5.4+
function-tool requests with unset reasoning_effort must still auto-bridge.
"""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
api_base=blank_api_base,
)
assert model == "gpt-5.6"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_custom_api_base_with_unset_effort_stays_chat():
"""
Chat-only OpenAI-compatible backends registered under the openai provider with a
custom api_base and gpt-5.4+ model names serve tools-without-reasoning fine and
have no /responses route; the unset-effort arm must not reroute them.
"""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
api_base="http://vllm.internal:8000/v1",
)
assert model == "gpt-5.6"
assert model_info.get("mode") != "responses"
def test_responses_api_bridge_check_custom_api_base_via_global_with_unset_effort_stays_chat(monkeypatch):
"""
A custom base set through the litellm.api_base global (not the call arg) is resolved the
same way the chat handler resolves it, so the unset-effort arm must not reroute a chat-only
backend to a /responses route it lacks. Regression guard: the gate previously inspected only
the call-level api_base and bridged these requests.
"""
import litellm
from litellm.main import responses_api_bridge_check
monkeypatch.setattr(litellm, "api_base", "http://vllm.internal:8000/v1")
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
api_base=None,
)
assert model == "gpt-5.6"
assert model_info.get("mode") != "responses"
@pytest.mark.parametrize("env_var", ["OPENAI_BASE_URL", "OPENAI_API_BASE"])
def test_responses_api_bridge_check_custom_api_base_via_env_with_unset_effort_stays_chat(monkeypatch, env_var):
"""
A custom base set via OPENAI_BASE_URL/OPENAI_API_BASE env is resolved identically to the chat
handler, so the unset-effort arm leaves the request on chat instead of bridging it.
"""
import litellm
from litellm.main import responses_api_bridge_check
monkeypatch.setattr(litellm, "api_base", None)
monkeypatch.delenv("OPENAI_BASE_URL", raising=False)
monkeypatch.delenv("OPENAI_API_BASE", raising=False)
monkeypatch.setenv(env_var, "http://vllm.internal:8000/v1")
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
api_base=None,
)
assert model == "gpt-5.6"
assert model_info.get("mode") != "responses"
def test_responses_api_bridge_check_custom_api_base_with_explicit_effort_still_routes():
"""Explicit reasoning_effort keeps its pre-existing bridging behavior on any api_base."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.6",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort="high",
api_base="http://vllm.internal:8000/v1",
)
assert model == "gpt-5.6"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_azure_with_api_base_and_unset_effort_routes():
"""Azure OpenAI always sets api_base and does enforce the constraint; keep bridging."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="azure",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
api_base="https://myresource.openai.azure.com",
)
assert model == "gpt-5.4"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_older_gpt_5_tools_without_reasoning_stays_chat():
"""Pre-5.4 GPT-5 names keep the old boundary: tools alone never bridge."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.1",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort=None,
)
assert model == "gpt-5.1"
assert model_info.get("mode") != "responses"
def test_responses_api_bridge_check_gpt_5_4_reasoning_summary_without_tools_routes_to_responses():
"""gpt-5.4+ with reasoning_effort + reasoningSummary but no tools should bridge (AI SDK)."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5.4",
custom_llm_provider="openai",
tools=None,
reasoning_effort="medium",
reasoning_summary="auto",
)
assert model == "gpt-5.4"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_gpt_5_reasoning_summary_routes_to_responses():
"""Bare ``gpt-5`` with reasoning_effort + reasoningSummary should bridge (not 5.4+)."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5",
custom_llm_provider="openai",
tools=None,
reasoning_effort="medium",
reasoning_summary="auto",
)
assert model == "gpt-5"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_gpt_5_tools_without_summary_stays_chat():
"""gpt-5 with tools + reasoning_effort but no summary should stay on chat."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 128000}
model_info, model = responses_api_bridge_check(
model="gpt-5",
custom_llm_provider="openai",
tools=[{"type": "function", "function": {"name": "get_capital"}}],
reasoning_effort="medium",
reasoning_summary=None,
)
assert model == "gpt-5"
assert model_info.get("mode") != "responses"
@patch("litellm.completion_extras.responses_api_bridge.completion")
def test_gpt_5_4_responses_bridge_preserves_reasoning_summary_dict(
mock_responses_completion,
):
"""When routed to Responses, preserve reasoning_effort summary dict."""
mock_responses_completion.return_value = MagicMock()
import litellm
litellm.completion(
model="gpt-5.4",
messages=[{"role": "user", "content": "What is the capital of France?"}],
tools=[
{
"type": "function",
"function": {
"name": "get_capital",
"description": "Get the capital of a country",
"parameters": {
"type": "object",
"properties": {"country": {"type": "string"}},
},
},
}
],
reasoning_effort={"effort": "xhigh", "summary": "detailed"},
api_key="fake-key",
)
assert mock_responses_completion.called is True
optional_params = mock_responses_completion.call_args.kwargs["optional_params"]
assert optional_params["reasoning_effort"] == {
"effort": "xhigh",
"summary": "detailed",
}
@pytest.mark.parametrize(
"model, model_info, expected_model_param, expected_base_model_param",
[
("gemini/gemini-3.1-pro", None, "gemini-3.1-pro", None),
(
"gemini/gemini-3.1-pro",
{"base_model": "gemini-3.1-pro-preview"},
"gemini-3.1-pro",
"gemini-3.1-pro-preview",
),
],
)
def test_completion_optional_params_base_model(
model: str,
model_info: dict | None,
expected_model_param: str,
expected_base_model_param: str | None,
):
"""``model_info.base_model`` must reach ``get_optional_params`` as ``base_model``
(an additive capability hint), without overwriting ``model`` with the label.
Regression for #29618: overwriting ``model`` with a friendly ``base_model``
label made Bedrock drop ``tools``/``tool_choice`` under ``drop_params``."""
with patch("litellm.main.get_optional_params") as mock_get_optional_params:
mock_get_optional_params.return_value = MagicMock()
import litellm
kwargs = {
"model": model,
"messages": [{"role": "user", "content": "What is the capital of France?"}],
"api_key": "fake-key",
"mock_response": "Hey, how's it going?",
}
if model_info is not None:
kwargs["model_info"] = model_info
litellm.completion(**kwargs)
assert mock_get_optional_params.called is True
call_kwargs = mock_get_optional_params.call_args.kwargs
assert call_kwargs["model"] == expected_model_param
assert call_kwargs["base_model"] == expected_base_model_param
@patch("litellm.completion_extras.responses_api_bridge.completion")
def test_gpt_5_4_responses_bridge_merges_reasoning_summary_kwarg_without_tools(
mock_responses_completion,
):
"""reasoningSummary without tools should route and merge into reasoning_effort dict."""
mock_responses_completion.return_value = MagicMock()
import litellm
litellm.completion(
model="gpt-5.4",
messages=[{"role": "user", "content": "ok"}],
reasoning_effort="medium",
reasoningSummary="auto",
api_key="fake-key",
)
assert mock_responses_completion.called is True
optional_params = mock_responses_completion.call_args.kwargs["optional_params"]
assert optional_params["reasoning_effort"] == {
"effort": "medium",
"summary": "auto",
}
assert "reasoningSummary" not in optional_params
assert "reasoning_summary" not in optional_params
@patch("litellm.completion_extras.responses_api_bridge.completion")
def test_responses_bridge_preserves_reasoning_summary_without_effort(
mock_responses_completion,
):
"""Reasoning summary should survive responses routing even without effort."""
mock_responses_completion.return_value = MagicMock()
import litellm
with patch.object(litellm, "route_all_chat_openai_to_responses", True):
litellm.completion(
model="gpt-4o",
messages=[{"role": "user", "content": "ok"}],
reasoningSummary="auto",
api_key="fake-key",
)
assert mock_responses_completion.called is True
optional_params = mock_responses_completion.call_args.kwargs["optional_params"]
assert optional_params["reasoning_effort"] == {"summary": "auto"}
assert "reasoningSummary" not in optional_params
assert "reasoning_summary" not in optional_params
@patch("litellm.completion_extras.responses_api_bridge.completion")
def test_gpt_5_responses_bridge_tools_and_reasoning_summary(
mock_responses_completion,
):
"""Bare gpt-5 with tools + reasoningSummary should bridge (OpenCode-style)."""
mock_responses_completion.return_value = MagicMock()
import litellm
litellm.completion(
model="gpt-5",
messages=[{"role": "user", "content": "ok"}],
tools=[
{
"type": "function",
"function": {
"name": "apply_patch",
"parameters": {"type": "object", "properties": {}},
},
}
],
tool_choice="auto",
reasoning_effort="medium",
reasoningSummary="auto",
stream=True,
api_key="fake-key",
)
assert mock_responses_completion.called is True
optional_params = mock_responses_completion.call_args.kwargs["optional_params"]
assert optional_params.get("reasoning_effort") == {
"effort": "medium",
"summary": "auto",
}
def test_responses_api_bridge_check_handles_exception():
"""Test that responses_api_bridge_check handles exceptions and still processes responses/ models."""
from litellm.main import responses_api_bridge_check
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.side_effect = Exception("Model not found")
model_info, model = responses_api_bridge_check(
model="responses/custom-model", custom_llm_provider="custom"
)
assert model == "custom-model"
assert model_info["mode"] == "responses"
def test_responses_api_bridge_check_global_flag_routes_openai():
"""When route_all_chat_openai_to_responses is True, any OpenAI model routes to responses."""
from litellm.main import responses_api_bridge_check
with patch.object(litellm, "route_all_chat_openai_to_responses", True):
model_info, model = responses_api_bridge_check(
model="gpt-4o",
custom_llm_provider="openai",
)
assert model == "gpt-4o"
assert model_info.get("mode") == "responses"
def test_responses_api_bridge_check_global_flag_does_not_affect_azure():
"""route_all_chat_openai_to_responses should not affect Azure models."""
from litellm.main import responses_api_bridge_check
with patch.object(litellm, "route_all_chat_openai_to_responses", True):
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 4096}
model_info, model = responses_api_bridge_check(
model="gpt-4o",
custom_llm_provider="azure",
)
assert model_info.get("mode") != "responses"
def test_responses_api_bridge_check_global_flag_default_false():
"""By default, route_all_chat_openai_to_responses is False and doesn't affect routing."""
from litellm.main import responses_api_bridge_check
with patch.object(litellm, "route_all_chat_openai_to_responses", False):
with patch("litellm.main._get_model_info_helper") as mock_get_model_info:
mock_get_model_info.return_value = {"max_tokens": 4096}
model_info, model = responses_api_bridge_check(
model="gpt-4o",
custom_llm_provider="openai",
)
assert model_info.get("mode") != "responses"
@pytest.mark.asyncio
async def test_async_mock_delay():
"""Use asyncio await for mock delay on acompletion"""
import time
from litellm import acompletion
start_time = time.time()
result = await acompletion(
model="gpt-3.5-turbo",
messages=[{"role": "user", "content": "Hey, how's it going?"}],
mock_delay=0.01,
mock_response="Hello world",
)
end_time = time.time()
delay = end_time - start_time
assert delay >= 0.01
def test_stream_chunk_builder_thinking_blocks():
from litellm import stream_chunk_builder
from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
chunks = [
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content="I need to summar",
thinking_blocks=[
{
"type": "thinking",
"thinking": "I need to summar",
"signature": None,
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": "I need to summar",
"signature": None,
}
]
},
content="",
role="assistant",
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content="ize the previous agent's thinking process into a",
thinking_blocks=[
{
"type": "thinking",
"thinking": "ize the previous agent's thinking process into a",
"signature": None,
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": "ize the previous agent's thinking process into a",
"signature": None,
}
]
},
content="",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content=" short description. Based on the input data provide",
thinking_blocks=[
{
"type": "thinking",
"thinking": " short description. Based on the input data provide",
"signature": None,
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": " short description. Based on the input data provide",
"signature": None,
}
]
},
content="",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content="d, it seems the agent was planning to refine their search",
thinking_blocks=[
{
"type": "thinking",
"thinking": "d, it seems the agent was planning to refine their search",
"signature": None,
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": "d, it seems the agent was planning to refine their search",
"signature": None,
}
]
},
content="",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content=" to focus more on technical aspects of home automation and home",
thinking_blocks=[
{
"type": "thinking",
"thinking": " to focus more on technical aspects of home automation and home",
"signature": None,
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": " to focus more on technical aspects of home automation and home",
"signature": None,
}
]
},
content="",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content=" energy system management.\n\nI'll create a brief",
thinking_blocks=[
{
"type": "thinking",
"thinking": " energy system management.\n\nI'll create a brief",
"signature": None,
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": " energy system management.\n\nI'll create a brief",
"signature": None,
}
]
},
content="",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content=" summary of what the agent was doing.",
thinking_blocks=[
{
"type": "thinking",
"thinking": " summary of what the agent was doing.",
"signature": None,
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": " summary of what the agent was doing.",
"signature": None,
}
]
},
content="",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=0,
delta=Delta(
reasoning_content="",
thinking_blocks=[
{
"type": "thinking",
"thinking": "",
"signature": "ErUBCkYIBRgCIkAKBSMkB2+MBF643wiWxlERsGXVdlhbPx9lnTIbygzjFIeZ5uhTV+HNWDon9vQV4hmXvAKwQfwS8vkNFB366l05Egzt2U18IpRrZRyQn1UaDDdYvKHYP8Ps1IbWjSIw8eSYOU9gtqNcwR6D0wY7iOPx2GliDEatLI5rSs96CByoTIoADL2M5bX8KP0jEpbHKh0ccYryigdH/3J8EiFt/BmGUceVASP5l9r22dFWiBgC",
}
],
provider_specific_fields={
"thinking_blocks": [
{
"type": "thinking",
"thinking": "",
"signature": "ErUBCkYIBRgCIkAKBSMkB2+MBF643wiWxlERsGXVdlhbPx9lnTIbygzjFIeZ5uhTV+HNWDon9vQV4hmXvAKwQfwS8vkNFB366l05Egzt2U18IpRrZRyQn1UaDDdYvKHYP8Ps1IbWjSIw8eSYOU9gtqNcwR6D0wY7iOPx2GliDEatLI5rSs96CByoTIoADL2M5bX8KP0jEpbHKh0ccYryigdH/3J8EiFt/BmGUceVASP5l9r22dFWiBgC",
}
]
},
content="",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content='{"a',
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content='gent_doing"',
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content=': "Re',
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content="searching",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content=" technic",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content="al aspect",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content="s of home au",
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason=None,
index=1,
delta=Delta(
provider_specific_fields=None,
content='tomation"}',
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
citations=None,
),
ModelResponseStream(
id="chatcmpl-e8febeb7-cf7d-4947-9417-59ae5e6989f9",
created=1751934860,
model="claude-3-7-sonnet-latest",
object="chat.completion.chunk",
system_fingerprint=None,
choices=[
StreamingChoices(
finish_reason="tool_calls",
index=0,
delta=Delta(
provider_specific_fields=None,
content=None,
role=None,
function_call=None,
tool_calls=None,
audio=None,
),
logprobs=None,
)
],
provider_specific_fields=None,
),
]
response = stream_chunk_builder(chunks=chunks)
print(response)
assert response is not None
assert response.choices[0].message.content is not None
assert response.choices[0].message.thinking_blocks is not None
from litellm.llms.openai.openai import OpenAIChatCompletion
def throw_retryable_error(*_, **__):
raise RuntimeError("BOOM")
@pytest.mark.asyncio
async def test_retrying() -> None:
litellm.num_retries = 10
with (
patch.object(
OpenAIChatCompletion,
"make_openai_chat_completion_request",
side_effect=throw_retryable_error,
) as mock_request,
pytest.raises(litellm.InternalServerError, match="LiteLLM Retried: 10 times"),
):
await litellm.acompletion(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "Hello"}],
)
def test_anthropic_disable_url_suffix_env_var():
"""Test that LITELLM_ANTHROPIC_DISABLE_URL_SUFFIX prevents /v1/messages suffix."""
import os
from unittest.mock import MagicMock, patch
from litellm import completion
# Test with environment variable disabled (default behavior)
with patch.dict(os.environ, {"ANTHROPIC_API_BASE": "https://api.example.com"}):
actual_api_base = None
with patch("litellm.main.anthropic_chat_completions") as mock_anthropic:
def capture_completion(**kwargs):
nonlocal actual_api_base
actual_api_base = kwargs.get("api_base")
mock_response = MagicMock()
mock_response.choices = [MagicMock()]
return mock_response
mock_anthropic.completion = capture_completion
# This should append /v1/messages
completion(
model="anthropic/claude-3-sonnet",
messages=[{"role": "user", "content": "test"}],
api_key="test-key",
)
# Verify the api_base has /v1/messages appended
assert actual_api_base.endswith("/v1/messages")
assert actual_api_base == "https://api.example.com/v1/messages"
# Test with environment variable enabled
with patch.dict(
os.environ,
{
"ANTHROPIC_API_BASE": "https://api.example.com/custom/path",
"LITELLM_ANTHROPIC_DISABLE_URL_SUFFIX": "true",
},
):
actual_api_base = None
with patch("litellm.main.anthropic_chat_completions") as mock_anthropic:
def capture_completion(**kwargs):
nonlocal actual_api_base
actual_api_base = kwargs.get("api_base")
mock_response = MagicMock()
mock_response.choices = [MagicMock()]
return mock_response
mock_anthropic.completion = capture_completion
# This should NOT append /v1/messages
completion(
model="anthropic/claude-3-sonnet",
messages=[{"role": "user", "content": "test"}],
api_key="test-key",
)
# Verify the api_base does not have /v1/messages appended
assert actual_api_base == "https://api.example.com/custom/path"
assert not actual_api_base.endswith("/v1/messages")
def test_anthropic_text_disable_url_suffix_env_var():
"""Test that LITELLM_ANTHROPIC_DISABLE_URL_SUFFIX prevents /v1/complete suffix for anthropic_text."""
import os
from unittest.mock import MagicMock, patch
from litellm import completion
# Test with environment variable disabled (default behavior)
with patch.dict(os.environ, {"ANTHROPIC_API_BASE": "https://api.example.com"}):
actual_api_base = None
with patch("litellm.main.base_llm_http_handler") as mock_handler:
def capture_completion(**kwargs):
nonlocal actual_api_base
actual_api_base = kwargs.get("api_base")
return MagicMock()
mock_handler.completion = capture_completion
# This should append /v1/complete
completion(
model="anthropic_text/claude-instant-1",
messages=[{"role": "user", "content": "test"}],
api_key="test-key",
)
# Verify the api_base has /v1/complete appended
assert actual_api_base.endswith("/v1/complete")
assert actual_api_base == "https://api.example.com/v1/complete"
# Test with environment variable enabled
with patch.dict(
os.environ,
{
"ANTHROPIC_API_BASE": "https://api.example.com/custom/complete",
"LITELLM_ANTHROPIC_DISABLE_URL_SUFFIX": "true",
},
):
actual_api_base = None
with patch("litellm.main.base_llm_http_handler") as mock_handler:
def capture_completion(**kwargs):
nonlocal actual_api_base
actual_api_base = kwargs.get("api_base")
return MagicMock()
mock_handler.completion = capture_completion
# This should NOT append /v1/complete
completion(
model="anthropic_text/claude-instant-1",
messages=[{"role": "user", "content": "test"}],
api_key="test-key",
)
# Verify the api_base does not have /v1/complete appended
assert actual_api_base == "https://api.example.com/custom/complete"
assert not actual_api_base.endswith("/v1/complete")
def test_image_edit_merges_headers_and_extra_headers():
from litellm.images.main import base_llm_http_handler
combined_headers = {
"x-test-header-one": "value-1",
"x-test-header-two": "value-2",
}
mock_image_edit_config = MagicMock()
mock_image_edit_config.get_supported_openai_params.return_value = set()
mock_image_edit_config.map_openai_params.side_effect = lambda **kwargs: dict(
kwargs["image_edit_optional_params"]
)
with (
patch(
"litellm.images.main.ProviderConfigManager.get_provider_image_edit_config",
return_value=mock_image_edit_config,
) as mock_config,
patch.object(
base_llm_http_handler,
"image_edit_handler",
return_value="ok",
) as mock_handler,
):
response = litellm.image_edit(
image=MagicMock(name="image"),
prompt="test",
model="azure/gpt-image-1",
headers={"x-test-header-one": "value-1"},
extra_headers={
"x-test-header-two": "value-2",
},
)
assert response == "ok"
mock_config.assert_called_once()
handler_kwargs = mock_handler.call_args.kwargs
assert handler_kwargs["extra_headers"] == combined_headers
assert "extra_headers" not in handler_kwargs["image_edit_optional_request_params"]
def test_mock_completion_stream_with_model_response():
"""Test that mock_completion correctly handles stream=True with a ModelResponse as mock_response."""
from litellm import completion
from litellm.types.utils import Choices, Message, ModelResponse, Usage
# Create a ModelResponse object
mock_model_response = ModelResponse(
id="chatcmpl-test-123",
created=1234567890,
model="gpt-4o-mini",
object="chat.completion",
choices=[
Choices(
finish_reason="stop",
index=0,
message=Message(
content="This is a test response",
role="assistant",
),
)
],
usage=Usage(
prompt_tokens=10,
completion_tokens=20,
total_tokens=30,
),
)
# Call completion with stream=True and mock_response as ModelResponse
response = completion(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "Hello"}],
stream=True,
mock_response=mock_model_response,
)
# Verify that the response is a stream
assert response is not None
# Collect all chunks from the stream
chunks = []
for chunk in response:
chunks.append(chunk)
print(f"Chunk: {chunk}")
# Verify we got chunks
assert len(chunks) > 0
# Verify the content is streamed correctly
accumulated_content = ""
for chunk in chunks:
if (
hasattr(chunk.choices[0].delta, "content")
and chunk.choices[0].delta.content
):
accumulated_content += chunk.choices[0].delta.content
assert "This is a test response" in accumulated_content or len(chunks) > 0
@pytest.mark.asyncio
async def test_async_mock_completion_stream_with_model_response():
"""Test that async mock_completion correctly handles stream=True with a ModelResponse as mock_response."""
from litellm import acompletion
from litellm.types.utils import Choices, Message, ModelResponse, Usage
# Create a ModelResponse object
mock_model_response = ModelResponse(
id="chatcmpl-test-456",
created=1234567890,
model="gpt-4o-mini",
object="chat.completion",
choices=[
Choices(
finish_reason="stop",
index=0,
message=Message(
content="This is an async test response",
role="assistant",
),
)
],
usage=Usage(
prompt_tokens=15,
completion_tokens=25,
total_tokens=40,
),
)
# Call acompletion with stream=True and mock_response as ModelResponse
response = await acompletion(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "Hello async"}],
stream=True,
mock_response=mock_model_response,
)
# Verify that the response is a stream
assert response is not None
# Collect all chunks from the stream
chunks = []
async for chunk in response:
chunks.append(chunk)
print(f"Async Chunk: {chunk}")
# Verify we got chunks
assert len(chunks) > 0
# Verify the content is streamed correctly
accumulated_content = ""
for chunk in chunks:
if (
hasattr(chunk.choices[0].delta, "content")
and chunk.choices[0].delta.content
):
accumulated_content += chunk.choices[0].delta.content
assert "This is an async test response" in accumulated_content or len(chunks) > 0
class TestCallTypesOCR:
"""Test that OCR call types are properly defined in CallTypes enum.
Fixes https://github.com/BerriAI/litellm/issues/17381
"""
def test_ocr_call_type_exists(self):
"""Test that CallTypes.ocr exists and has correct value."""
from litellm.types.utils import CallTypes
assert hasattr(CallTypes, "ocr")
assert CallTypes.ocr.value == "ocr"
def test_aocr_call_type_exists(self):
"""Test that CallTypes.aocr exists and has correct value."""
from litellm.types.utils import CallTypes
assert hasattr(CallTypes, "aocr")
assert CallTypes.aocr.value == "aocr"
def test_ocr_call_type_from_string(self):
"""Test that CallTypes can be constructed from 'ocr' string."""
from litellm.types.utils import CallTypes
call_type = CallTypes("ocr")
assert call_type == CallTypes.ocr
def test_aocr_call_type_from_string(self):
"""Test that CallTypes can be constructed from 'aocr' string.
This is the actual use case that was failing - the OCR endpoint
uses route_type='aocr' and guardrails try to instantiate
CallTypes('aocr').
"""
from litellm.types.utils import CallTypes
call_type = CallTypes("aocr")
assert call_type == CallTypes.aocr
def test_stream_chunk_builder_text_completion_combines_text_and_usage():
from litellm.main import stream_chunk_builder_text_completion
from litellm.types.utils import TextCompletionResponse
chunks = [
TextCompletionResponse(
id="cmpl-1",
object="text_completion",
created=1,
model="gpt-3.5-turbo-instruct",
choices=[{"text": "Hello", "index": 0, "logprobs": None, "finish_reason": None}],
),
TextCompletionResponse(
id="cmpl-1",
object="text_completion",
created=1,
model="gpt-3.5-turbo-instruct",
choices=[{"text": " world", "index": 0, "logprobs": None, "finish_reason": "stop"}],
),
]
response = stream_chunk_builder_text_completion(
chunks=chunks, messages=[{"role": "user", "content": "say hello"}]
)
assert response.choices[0].text == "Hello world"
assert response.choices[0].finish_reason == "stop"
assert response.usage.prompt_tokens > 0
assert response.usage.completion_tokens > 0
assert response.usage.total_tokens == response.usage.prompt_tokens + response.usage.completion_tokens
def test_completion_forwards_store_and_prompt_cache_key_to_openai():
"""
Regression test for https://github.com/BerriAI/litellm/issues/33184
store and prompt_cache_key are documented OpenAI chat completion params that
were accepted as supported but silently dropped before the provider request
was built, because they were not named parameters of completion() and
get_optional_params() the way safety_identifier is.
"""
from openai import OpenAI
client = OpenAI(api_key="fake-api-key")
with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
try:
litellm.completion(
model="openai/gpt-4o",
messages=[{"role": "user", "content": "Hello"}],
store=False,
prompt_cache_key="test-cache-key",
client=client,
)
except Exception as e:
print(e)
mock_client.assert_called_once()
request_body = mock_client.call_args.kwargs
assert request_body["store"] is False
assert request_body["prompt_cache_key"] == "test-cache-key"
@pytest.mark.asyncio
async def test_acompletion_forwards_store_and_prompt_cache_key_to_openai():
"""
Async variant of the store/prompt_cache_key forwarding regression test for
https://github.com/BerriAI/litellm/issues/33184
"""
from openai import AsyncOpenAI
client = AsyncOpenAI(api_key="fake-api-key")
with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
try:
await litellm.acompletion(
model="openai/gpt-4o",
messages=[{"role": "user", "content": "Hello"}],
store=False,
prompt_cache_key="test-cache-key",
client=client,
)
except Exception as e:
print(e)
mock_client.assert_called_once()
request_body = mock_client.call_args.kwargs
assert request_body["store"] is False
assert request_body["prompt_cache_key"] == "test-cache-key"
def test_completion_omits_store_and_prompt_cache_key_when_not_passed():
"""
When store and prompt_cache_key are not passed, they must not appear in the
outbound request body (guards against always forwarding None defaults).
"""
from openai import OpenAI
client = OpenAI(api_key="fake-api-key")
with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
try:
litellm.completion(
model="openai/gpt-4o",
messages=[{"role": "user", "content": "Hello"}],
client=client,
)
except Exception as e:
print(e)
mock_client.assert_called_once()
request_body = mock_client.call_args.kwargs
assert "store" not in request_body
assert "prompt_cache_key" not in request_body
def test_completion_forwards_store_and_prompt_cache_key_to_mcp_gateway():
"""
Regression test for the MCP gateway early-return in completion(): store and
prompt_cache_key are named params, so they no longer travel via **kwargs and
must be forwarded explicitly like safety_identifier and service_tier.
"""
with patch(
"litellm.responses.mcp.chat_completions_handler.acompletion_with_mcp"
) as mock_mcp:
result = litellm.completion(
model="openai/gpt-4o",
messages=[{"role": "user", "content": "Hello"}],
tools=[{"type": "mcp", "server_url": "litellm_proxy"}],
store=False,
prompt_cache_key="test-cache-key",
)
result.close()
mock_mcp.assert_called_once()
call_kwargs = mock_mcp.call_args.kwargs
assert call_kwargs["store"] is False
assert call_kwargs["prompt_cache_key"] == "test-cache-key"
@pytest.mark.asyncio
@pytest.mark.parametrize(
"aws_credential_kwargs",
[
{
"aws_session_name": "litellm-gcp",
"aws_role_name": "arn:aws:iam::123456789012:role/litellm-bedrock-role",
"aws_web_identity_token": "oidc/google/108963886734710037768",
},
{
"aws_access_key_id": "AKIASTATICKEYFORTEST",
"aws_secret_access_key": "static-secret-key",
"aws_session_token": "static-session-token",
},
],
ids=["web_identity", "static_keys"],
)
async def test_acompletion_forwards_aws_credentials_through_responses_bridge(
respx_mock: respx.MockRouter, monkeypatch, aws_credential_kwargs: dict
):
from botocore.credentials import Credentials
from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
original_disable_aiohttp = litellm.disable_aiohttp_transport
try:
litellm.disable_aiohttp_transport = True
monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
litellm.in_memory_llm_clients_cache.flush_cache()
monkeypatch.delenv("AWS_BEARER_TOKEN_BEDROCK", raising=False)
monkeypatch.delenv("BEDROCK_MANTLE_API_KEY", raising=False)
get_credentials_mock = MagicMock(return_value=Credentials("fake-key", "fake-secret"))
monkeypatch.setattr(BaseAWSLLM, "get_credentials", get_credentials_mock)
respx_mock.post("https://bedrock-mantle.us-east-2.api.aws/openai/v1/responses").respond(
json={
"id": "resp_123",
"object": "response",
"created_at": 1760144904,
"status": "completed",
"model": "openai.gpt-5.4",
"output": [
{
"type": "message",
"id": "msg_1",
"role": "assistant",
"status": "completed",
"content": [{"type": "output_text", "text": "ok", "annotations": []}],
}
],
}
)
response = await litellm.acompletion(
model="bedrock_mantle/openai.gpt-5.4",
messages=[{"role": "user", "content": "hi"}],
api_base="https://bedrock-mantle.us-east-2.api.aws/v1",
aws_region_name="us-east-2",
num_retries=0,
**aws_credential_kwargs,
)
assert response.choices[0].message.content == "ok"
credential_kwargs = get_credentials_mock.call_args.kwargs
assert credential_kwargs["aws_region_name"] == "us-east-2"
for key, value in aws_credential_kwargs.items():
assert credential_kwargs[key] == value
authorization = respx_mock.calls.last.request.headers["Authorization"]
assert authorization.startswith("AWS4-HMAC-SHA256")
assert "fake-key" in authorization
finally:
litellm.disable_aiohttp_transport = original_disable_aiohttp
litellm.in_memory_llm_clients_cache.flush_cache()
_GEMINI_RESPONSE_BODY = {
"candidates": [{"content": {"parts": [{"text": "hello"}], "role": "model"}, "finishReason": "STOP"}],
"usageMetadata": {"promptTokenCount": 2, "candidatesTokenCount": 1, "totalTokenCount": 3},
}
def _gemini_client_returning_a_reply():
"""An injected HTTP client whose post() answers like generativelanguage does."""
from litellm.llms.custom_httpx.http_handler import HTTPHandler
client = HTTPHandler()
request = httpx.Request("POST", "https://generativelanguage.googleapis.com/")
post = MagicMock(return_value=httpx.Response(200, json=_GEMINI_RESPONSE_BODY, request=request))
return client, post
@pytest.fixture
def restore_model_registry():
"""litellm.model_cost and the provider name sets are module-global.
register_model merges into the existing entry in place, hence the deep copy.
"""
model_cost = copy.deepcopy(litellm.model_cost)
openai_models = set(litellm.open_ai_chat_completion_models)
yield
litellm.model_cost.clear()
litellm.model_cost.update(model_cost)
litellm.open_ai_chat_completion_models.clear()
litellm.open_ai_chat_completion_models.update(openai_models)
def test_openai_model_name_does_not_outrank_explicit_provider():
"""`gemini/gpt-4o` goes to Google, not to litellm's OpenAI handler.
completion() checks `model in litellm.open_ai_chat_completion_models` ahead of
the gemini branch, so the call used to reach the OpenAI handler carrying
VertexGeminiConfig, whose transform_request raises NotImplementedError.
"""
assert "gpt-4o" in litellm.open_ai_chat_completion_models
client, post = _gemini_client_returning_a_reply()
with patch.object(client, "post", new=post):
response = litellm.completion(
model="gemini/gpt-4o",
messages=[{"role": "user", "content": "hello"}],
api_key="test-api-key",
client=client,
)
assert "generativelanguage.googleapis.com" in post.call_args.kwargs["url"]
assert "models/gpt-4o" in post.call_args.kwargs["url"]
assert response.choices[0].message.content == "hello"
def test_mislabelled_pricing_entry_does_not_reroute_provider(restore_model_registry):
"""register_model is the other way into the same failure.
An entry claiming litellm_provider "openai" adds its name to
open_ai_chat_completion_models, so one mislabelled price reroutes every later
call to that model in the process.
"""
litellm.register_model(
{
"gemini-2.5-pro": {
"litellm_provider": "openai",
"mode": "chat",
"input_cost_per_token": 1e-06,
"output_cost_per_token": 4e-06,
}
}
)
assert "gemini-2.5-pro" in litellm.open_ai_chat_completion_models
client, post = _gemini_client_returning_a_reply()
with patch.object(client, "post", new=post):
response = litellm.completion(
model="gemini/gemini-2.5-pro",
messages=[{"role": "user", "content": "hello"}],
api_key="test-api-key",
client=client,
)
assert "generativelanguage.googleapis.com" in post.call_args.kwargs["url"]
assert response.choices[0].message.content == "hello"
def test_openai_model_without_a_provider_still_routes_to_openai():
from openai import OpenAI
client = OpenAI(api_key="fake-key")
raw_response = client.chat.completions.with_raw_response
with patch.object(raw_response, "create") as mock_create, contextlib.suppress(Exception):
litellm.completion(
model="gpt-4o",
messages=[{"role": "user", "content": "hello"}],
client=client,
)
mock_create.assert_called()
def _openai_chat_create_kwargs(client, **completion_kwargs):
with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
with contextlib.suppress(Exception):
litellm.completion(
messages=[{"role": "system", "content": "sys"}, {"role": "user", "content": "hi"}],
cache_control_injection_points=[{"location": "message", "role": "system"}],
client=client,
**completion_kwargs,
)
mock_client.assert_called_once()
return mock_client.call_args.kwargs
@pytest.fixture
def _no_openai_api_base_override(monkeypatch):
monkeypatch.delenv("OPENAI_BASE_URL", raising=False)
monkeypatch.delenv("OPENAI_API_BASE", raising=False)
monkeypatch.setattr(litellm, "api_base", None)
@pytest.mark.usefixtures("_no_openai_api_base_override")
def test_completion_custom_api_base_sends_no_prompt_cache_breakpoint_for_gpt_5_6():
from openai import OpenAI
client = OpenAI(api_key="fake-api-key", base_url="http://127.0.0.1:9/v1")
request_body = _openai_chat_create_kwargs(client, model="gpt-5.6", api_base="http://127.0.0.1:9/v1")
assert request_body["messages"][0] == {"role": "system", "content": "sys", "cache_control": {"type": "ephemeral"}}
assert "prompt_cache_breakpoint" not in json.dumps(request_body["messages"])
assert "prompt_cache_options" not in json.dumps(request_body)
@pytest.mark.usefixtures("_no_openai_api_base_override")
def test_completion_custom_base_url_sends_no_prompt_cache_breakpoint_for_gpt_5_6():
from openai import OpenAI
client = OpenAI(api_key="fake-api-key", base_url="http://127.0.0.1:9/v1")
request_body = _openai_chat_create_kwargs(client, model="gpt-5.6", base_url="http://127.0.0.1:9/v1")
assert request_body["messages"][0] == {"role": "system", "content": "sys", "cache_control": {"type": "ephemeral"}}
assert "prompt_cache_breakpoint" not in json.dumps(request_body["messages"])
assert "prompt_cache_options" not in json.dumps(request_body)
@pytest.mark.asyncio
@pytest.mark.usefixtures("_no_openai_api_base_override")
async def test_acompletion_custom_base_url_sends_no_prompt_cache_breakpoint_for_gpt_5_6():
from openai import AsyncOpenAI
client = AsyncOpenAI(api_key="fake-api-key", base_url="http://127.0.0.1:9/v1")
with patch.object(client.chat.completions.with_raw_response, "create") as mock_create:
with contextlib.suppress(Exception):
await litellm.acompletion(
model="gpt-5.6",
messages=[{"role": "system", "content": "sys"}, {"role": "user", "content": "hi"}],
cache_control_injection_points=[{"location": "message", "role": "system"}],
client=client,
base_url="http://127.0.0.1:9/v1",
)
mock_create.assert_called_once()
request_body = mock_create.call_args.kwargs
assert request_body["messages"][0] == {"role": "system", "content": "sys", "cache_control": {"type": "ephemeral"}}
assert "prompt_cache_breakpoint" not in json.dumps(request_body["messages"])
assert "prompt_cache_options" not in json.dumps(request_body)
@pytest.mark.usefixtures("_no_openai_api_base_override")
def test_completion_default_api_base_sends_prompt_cache_breakpoint_for_gpt_5_6():
from openai import OpenAI
client = OpenAI(api_key="fake-api-key")
request_body = _openai_chat_create_kwargs(client, model="gpt-5.6")
assert request_body["messages"][0]["content"] == [
{"type": "text", "text": "sys", "prompt_cache_breakpoint": {"mode": "explicit"}}
]
assert request_body["extra_body"]["prompt_cache_options"] == {"mode": "explicit"}
Reference in a new issue View git blame Copy permalink