From c348625986a89840f79f7ebd7dc6bef51806930f Mon Sep 17 00:00:00 2001
From: Tanvir Alam <tanviralamshuvo@gmail.com>
Date: Mon, 21 Sep 2026 00:28:35 -0400
Subject: [PATCH] fix(caching): tolerate missing fastapi on SDK cache hits

Cache-hit callbacks imported Anthropic response_cache, which pulls in
litellm.proxy (fastapi). Plain pip install litellm does not include the
proxy extra, so the second completion crashed with ModuleNotFoundError.

Guard that optional import so chat/text cache hits keep working (#42195).
---
 litellm/caching/caching_handler.py            | 22 +++++++++++-----
 .../caching/test_caching_handler.py           | 26 +++++++++++++++++++
 2 files changed, 42 insertions(+), 6 deletions(-)

diff --git a/litellm/caching/caching_handler.py b/litellm/caching/caching_handler.py
index 50426ea89ea..68680a76316 100644
--- a/litellm/caching/caching_handler.py
+++ b/litellm/caching/caching_handler.py
@@ -124,15 +124,25 @@ def _should_defer_streaming_cache_hit_callbacks(*, cached_result: object) -> boo
     handlers when the stream finishes; firing them here too would double-count
     spend and callback records. A plain (non-stream) replay logs here, since nothing
     else will.
+
+    The Anthropic Messages cache helper imports ``litellm.proxy`` (fastapi). That is
+    only available with the ``proxy`` extra, so a missing import must not break
+    plain SDK cache hits (#42195).
     """
-    from litellm.llms.anthropic.experimental_pass_through.messages.response_cache import (
-        CachedAnthropicMessagesStreamIterator,
-    )
     from litellm.responses.streaming_iterator import BaseResponsesAPIStreamingIterator
 
-    return isinstance(
-        cached_result, (CustomStreamWrapper, BaseResponsesAPIStreamingIterator, CachedAnthropicMessagesStreamIterator)
-    )
+    streaming_types: tuple[type, ...] = (CustomStreamWrapper, BaseResponsesAPIStreamingIterator)
+    try:
+        from litellm.llms.anthropic.experimental_pass_through.messages.response_cache import (
+            CachedAnthropicMessagesStreamIterator,
+        )
+    except ImportError:
+        # response_cache -> streaming_iterator -> litellm.proxy (fastapi / orjson).
+        pass
+    else:
+        streaming_types = (*streaming_types, CachedAnthropicMessagesStreamIterator)
+
+    return isinstance(cached_result, streaming_types)
 
 
 def _prompt_tokens_details_as_mapping(details: "PromptTokensDetailsWrapper") -> Mapping[str, object]:
diff --git a/tests/test_litellm/caching/test_caching_handler.py b/tests/test_litellm/caching/test_caching_handler.py
index 39018dca41d..cab1600a004 100644
--- a/tests/test_litellm/caching/test_caching_handler.py
+++ b/tests/test_litellm/caching/test_caching_handler.py
@@ -780,3 +780,29 @@ async def test_agentic_loop_followup_cache_hit_with_converted_stream_marker_repl
     assert hit.cached_result.choices[0].message.content == "done"
     logging_obj.handle_sync_success_callbacks_for_async_calls.assert_called_once()
     assert logging_obj.handle_sync_success_callbacks_for_async_calls.call_args.kwargs["cache_hit"] is True
+
+
+def test_should_defer_streaming_cache_hit_callbacks_without_proxy_extra():
+    """Regression for #42195: cache-hit path must not require fastapi/proxy extras.
+
+    ``response_cache`` imports proxy modules. On a plain ``pip install litellm``
+    those deps are absent; marking the module as missing must not raise, and a
+    non-stream cached result must still return False.
+    """
+    import sys
+
+    from litellm.caching.caching_handler import _should_defer_streaming_cache_hit_callbacks
+
+    mod_name = "litellm.llms.anthropic.experimental_pass_through.messages.response_cache"
+    had_module = mod_name in sys.modules
+    saved = sys.modules.get(mod_name)
+    # PEP 302: sys.modules[name] = None makes the next import raise ImportError.
+    sys.modules[mod_name] = None
+    try:
+        assert _should_defer_streaming_cache_hit_callbacks(cached_result={"choices": []}) is False
+    finally:
+        if had_module:
+            sys.modules[mod_name] = saved
+        else:
+            sys.modules.pop(mod_name, None)
+