fix(caching): read s-max-age first on the sync cache read like the async read

The previous commit gave s-maxage priority on the sync read, while the async
read, the batch Redis hook and the Rust cache give s-max-age priority. A
request that sent both keys could then hit the cache on one path and miss on
the other, and `{"s-maxage": None, "s-max-age": 1}` served stale audio on
this line's speech path. The sync read now uses the async read's order

litellm/caching/caching.py

@@ -663,7 +663,7 @@ class Cache:
                 if cache_key is not None:
                     cache_control_args: Final[DynamicCacheControl] = kwargs.get("cache", {})
                     max_age: Final = cache_control_args.get(
-                        "s-maxage", cache_control_args.get("s-max-age", float("inf"))
+                        "s-max-age", cache_control_args.get("s-maxage", float("inf"))
                     )
                     cache_lookup_kwargs: Final = self._get_safe_cache_lookup_kwargs(kwargs)
                     if dynamic_cache_object is not None:

tests/unit/caching/test_caching.py

@@ -1455,6 +1455,30 @@ def test_sync_completion_with_zero_s_maxage_skips_the_cached_response(
     assert open_max_age.id == zero_max_age.id
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "cache_controls",
+    [{"s-maxage": 0, "s-max-age": 60}, {"s-maxage": 60, "s-max-age": 0}, {"s-maxage": None, "s-max-age": 0}],
+)
+async def test_sync_and_async_completion_agree_on_the_cache_when_both_max_age_keys_are_sent(
+    monkeypatch: pytest.MonkeyPatch, cache_controls: dict[str, int | None]
+) -> None:
+    monkeypatch.setattr(litellm, "cache", Cache(type=LiteLLMCacheType.LOCAL))
+    sync_messages: Final = [{"role": "user", "content": f"sync both max-age keys {uuid.uuid4().hex}"}]
+    sync_first: Final = completion(model="gpt-4o-mini", messages=sync_messages, mock_response="first")
+    sync_second: Final = completion(
+        model="gpt-4o-mini", messages=sync_messages, cache=cache_controls, mock_response="second"
+    )
+    async_messages: Final = [{"role": "user", "content": f"async both max-age keys {uuid.uuid4().hex}"}]
+    async_first: Final = await acompletion(model="gpt-4o-mini", messages=async_messages, mock_response="first")
+    await asyncio.gather(*_PENDING_CACHE_WRITES)
+    async_second: Final = await acompletion(
+        model="gpt-4o-mini", messages=async_messages, cache=cache_controls, mock_response="second"
+    )
+
+    assert (sync_second.id == sync_first.id) == (async_second.id == async_first.id)
+
+
 @pytest.mark.asyncio
 async def test_embedding_caching_redis_ttl(monkeypatch: pytest.MonkeyPatch, redis_client_without_network: None) -> None:
     mock_pipeline: Final = AsyncMock()