scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-07 02:59:05 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 103

chore: drop the sibling-PR changes that leaked onto this branch

Browse source 
Only router.py, its test and the shard entry belong here; the
llm_response_utils changes are #43170/#43173.
This commit is contained in:
sclfcz 2026-09-28 00:05:32 +08:00
parent 2d3b45c918
commit 75896cb1f1
2 changed files with 1 additions and 93 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
View file

@ -406,11 +406,7 @@ def _handle_invalid_parallel_tool_calls(
shift = 0
for i, replacement in replacements.items():
tool_calls[:] = tool_calls[: i + shift] + replacement + tool_calls[i + shift + 1 :]
# One entry is replaced by ``len(replacement)`` entries, so the
# offsets of everything after it move by the difference - not by
# the full length, which would skip one entry per expansion and
# leave the next ``multi_tool_use.parallel`` call in place.
shift += len(replacement) - 1
shift += len(replacement)
return tool_calls
except json.JSONDecodeError:

88
tests/test_litellm/litellm_core_utils/llm_response_utils/test_parallel_tool_call_shift.py
View file

@ -1,88 +0,0 @@
"""Every hallucinated multi_tool_use.parallel call in one message must expand.
Replacing one entry with len(expansions) entries moves everything after it by
len(expansions) - 1; advancing by the full length skipped one entry per
expansion, so with two such calls the second stayed in place (and the call
after it was overwritten).
"""
import json
from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
convert_to_model_response_object,
)
from litellm.types.utils import ModelResponse
def _parallel(call_id: str, *calls):
return {
"id": call_id,
"type": "function",
"function": {
"name": "multi_tool_use.parallel",
"arguments": json.dumps(
{
"tool_uses": [
{"recipient_name": name, "parameters": params} for name, params in calls
]
}
),
},
}
def _plain(call_id: str, name: str):
return {
"id": call_id,
"type": "function",
"function": {"name": name, "arguments": "{}"},
}
def test_every_parallel_tool_call_expands():
response_object = {
"id": "chatcmpl-parallel",
"object": "chat.completion",
"created": 1728933352,
"model": "gpt-4o-2024-08-06",
"choices": [
{
"index": 0,
"finish_reason": "tool_calls",
"message": {
"role": "assistant",
"content": None,
"tool_calls": [
_plain("0", "get_weather"),
# Two expansions here and one below, so the first splice
# changes the list length and the second has to land at
# the shifted offset.
_parallel(
"m1",
("functions.get_time", {"tz": "UTC"}),
("functions.get_date", {"tz": "UTC"}),
),
_plain("2", "get_news"),
_parallel("m2", ("functions.get_quote", {"sym": "AAPL"})),
_plain("4", "get_forecast"),
],
},
}
],
}
result = convert_to_model_response_object(
response_object=response_object,
model_response_object=ModelResponse(),
response_type="completion",
)
names = [tc.function.name for tc in result.choices[0].message.tool_calls]
assert names == [
"get_weather",
"get_time",
"get_date",
"get_news",
"get_quote",
"get_forecast",
]