Merge c02f11dd7e into 63a4f3f2f3

litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py

@@ -427,7 +427,11 @@ 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 :]
-            shift += len(replacement)
+            # 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
 
         return tool_calls
     except json.JSONDecodeError:

tests/test_litellm/litellm_core_utils/llm_response_utils/test_parallel_tool_call_shift.py

@@ -0,0 +1,88 @@
+"""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",
+    ]