Merge 24fd51b582 into e0982d6248

litellm/caching/caching_handler.py

@@ -182,14 +182,19 @@ def _should_defer_streaming_cache_hit_callbacks(*, cached_result: object) -> boo
     spend and callback records. A plain (non-stream) replay logs here, since nothing
     else will.
     """
-    from litellm.llms.anthropic.pass_through.messages.response_cache import (
-        CachedAnthropicMessagesStreamIterator,
-    )
     from litellm.responses.streaming_iterator import BaseResponsesAPIStreamingIterator
 
-    return isinstance(
-        cached_result, (CustomStreamWrapper, BaseResponsesAPIStreamingIterator, CachedAnthropicMessagesStreamIterator)
-    )
+    base_replay_classes: Final = (CustomStreamWrapper, BaseResponsesAPIStreamingIterator)
+    try:
+        from litellm.llms.anthropic.pass_through.messages.response_cache import (
+            CachedAnthropicMessagesStreamIterator,
+        )
+    except ModuleNotFoundError:
+        # The Anthropic Messages replay is only reachable through proxy-only
+        # extras (fastapi). A base SDK install can never hold one, so the
+        # isinstance check can safely go ahead without it.
+        return isinstance(cached_result, base_replay_classes)
+    return isinstance(cached_result, (*base_replay_classes, CachedAnthropicMessagesStreamIterator))
 
 
 def _prompt_tokens_details_as_mapping(details: "PromptTokensDetailsWrapper") -> Mapping[str, object]:

tests/unit/caching/test_caching_handler.py

@@ -3231,3 +3231,21 @@ async def test_a_cache_hit_inside_an_internal_sub_call_is_free_and_fires_no_call
     assert async_logging.async_success_handler.called is not internal
     assert async_logging.handle_sync_success_callbacks_for_async_calls.called is not internal
     assert sync_logging.handle_sync_cache_hit_success_callbacks.called is not internal
+
+def test_should_defer_streaming_cache_hit_callbacks_without_proxy_extras(monkeypatch):
+    """Base SDK installs lack the proxy-only extras (fastapi); the defer check must not require them."""
+    import sys
+
+    for mod in list(sys.modules):
+        if mod.startswith(("fastapi", "litellm.llms.anthropic.pass_through", "litellm.proxy")):
+            monkeypatch.delitem(sys.modules, mod, raising=False)
+    monkeypatch.setitem(sys.modules, "fastapi", None)
+
+    logging_obj = MagicMock()
+    logging_obj.model_call_details = {}
+    stream_replay = CustomStreamWrapper(
+        completion_stream=iter(()), model="gpt-4o", logging_obj=logging_obj
+    )
+    assert _should_defer_streaming_cache_hit_callbacks(cached_result=stream_replay) is True
+    assert _should_defer_streaming_cache_hit_callbacks(cached_result=ModelResponse()) is False
+    assert _should_defer_streaming_cache_hit_callbacks(cached_result={"id": "msg_1"}) is False