Set lower bounds for LiteLLM and OpenAI deps

pyproject.toml

@@ -12,9 +12,9 @@ license = "Apache-2.0"
 license-files = ["LICENSE"]
 dependencies = [
     "jsonpath-ng>=1.7.0",
-    "litellm==1.71.3",
+    "litellm>=1.77.2",
     "numpy>=2.1.3",
-    "openai==1.87.0",
+    "openai>=2.0.0",
     "pandas>=2.0.0",
     "pandas-stubs>=2.0.0",
     "pydantic>=2.9.2",

tests/unit/conftest.py

@@ -6,6 +6,11 @@
 import pytest
 from dotenv import load_dotenv
 import litellm
+import asyncio
+import contextlib
+from collections.abc import Coroutine, Iterator
+from importlib.metadata import version
+from unittest.mock import patch
 import vcr  # type: ignore[import-untyped]
 
 from tests.helpers.completion_template_fixtures import (
@@ -41,6 +46,54 @@ def setup_logging():
     yield
 
 
+@pytest.fixture(autouse=True, scope="session")
+def patch_litellm_logging_worker_for_race_condition() -> Iterator[None]:
+    """
+    Patch litellm's GLOBAL_LOGGING_WORKER for asyncio functionality.
+
+    SEE: https://github.com/BerriAI/litellm/issues/16518
+    SEE: https://github.com/BerriAI/litellm/issues/14521
+    """
+    try:
+        from litellm.litellm_core_utils import logging_worker
+    except ImportError:
+        if tuple(int(x) for x in version(litellm.__name__).split(".")) < (1, 76, 0):
+            # Module didn't exist before https://github.com/BerriAI/litellm/pull/13905
+            yield
+            return
+        raise
+
+    class NoOpLoggingWorker:
+        """No-op worker that executes callbacks immediately without queuing."""
+
+        def start(self) -> None:
+            pass
+
+        def enqueue(self, coroutine: Coroutine) -> None:
+            # Execute immediately in current loop instead of queueing,
+            # and do nothing if there's no current loop
+            with contextlib.suppress(RuntimeError):
+                # This logging task is fire-and-forget
+                asyncio.create_task(  # type: ignore[unused-awaitable]  # noqa: RUF006
+                    coroutine
+                )
+
+        def ensure_initialized_and_enqueue(self, async_coroutine: Coroutine) -> None:
+            self.enqueue(async_coroutine)
+
+        async def stop(self) -> None:
+            pass
+
+        async def flush(self) -> None:
+            pass
+
+        async def clear_queue(self) -> None:
+            pass
+
+    with patch.object(logging_worker, "GLOBAL_LOGGING_WORKER", NoOpLoggingWorker()):
+        yield
+
+
 @pytest.fixture(scope="function")
 async def vcr_cassette_async(request: pytest.FixtureRequest):
     """

tlm/utils/completion_utils.py

@@ -133,10 +133,11 @@ async def _generate_completion(
 
         if litellm_params.get("logprobs") and hasattr(response.choices[0], "logprobs"):
             # Convert ChoiceLogprobs to dict to avoid Pydantic validation issues
+            choice_logprobs = response.choices[0].logprobs
             logprobs = ChoiceLogprobs.model_validate(
-                response.choices[0].logprobs.model_dump()
-                if hasattr(response.choices[0].logprobs, "model_dump")
-                else response.choices[0].logprobs
+                choice_logprobs.model_dump()
+                if choice_logprobs and hasattr(choice_logprobs, "model_dump")
+                else choice_logprobs
             )
 
             if raw_message_content := _get_raw_message_content(logprobs):