test(litellm): parametrize sync/async test pairs

[starfolkai]

Summary

Collapses 13 duplicated sync/async LiteLLM test pairs into single parametrized tests, addressing the ~15 mechanical duplicate pairs in test_litellm.py. The file shrinks from 1088 → 828 lines (−242 net) with no loss of coverage.

A shared marker plus two small helpers drive the dispatch:

sync_async = pytest.mark.parametrize("is_async", [False, True], ids=["sync", "async"])

async def _call(is_async, sync_fn, async_fn, *args, **kwargs): ...   # awaits async_fn when is_async
async def _collect_stream(stream, is_async): ...                     # drains sync or async streams

Parametrized pairs: completion, text_completion, completion_streaming, responses, embedding, moderation, image_generation, completion_with_system_prompt, transcription, speech, completion_error, responses_streaming, rerank.

Genuinely-single tests (prompt_caching, async_parallel_requests, tool_calls, async_streaming_with_break, the boolean/rerank unit tests, the patch_litellm_* subprocess tests, and auto-instrument) are left untouched.

Small coverage gains

Where an async variant historically asserted less than its sync twin, the unified body applies the stronger set to both — verified against the existing cassettes first:

• streaming variants now assert response content / "24" in the async case too
• rerank[async] now checks results[0]["index"] == 0, output-key shape, and search_units
• completion_metrics[async] now checks the litellm-auto span origin

Cassettes

Because pytest-recording names cassettes after the test node, the 24 paired cassettes were renamed to their parametrized names (e.g. test_litellm_acompletion_metrics.yaml → test_litellm_completion_metrics[async].yaml) across both the latest/ and 1.74.0/ matrix dirs — 48 pure git mv renames, no recorded content changed.

Validation

• nox -s "test_litellm(latest)" → 38 passed (matches pre-change baseline)
• nox -s "test_litellm(1.74.0)" → 38 passed
• make pylint (--errors-only) → clean; pre-commit (ruff-format, ruff check, codespell) → clean
• make check-stale-cassettes → no stale dirs

🤖 Generated with Claude Code

Created by Abhijeet Prasad (@AbhiPrasad)

Slack thread

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5ecf1a5954

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Run sync cases outside the asyncio loop

For every is_async=False parameter, sync_fn is now invoked from inside a coroutine managed by pytest-asyncio, so these tests no longer exercise LiteLLM's synchronous entry points in the no-running-event-loop context used by ordinary synchronous callers. LiteLLM and its transports can select loop-dependent execution paths, meaning this refactor can hide sync-only failures even though both parameter values pass; keep a synchronous test driver for the sync variants while sharing only the call arguments and assertions.

Useful? React with 👍 / 👎.