From 6d85aae12e2257d185f02cb361bd1370f5752878 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Tue, 20 Oct 2026 03:52:43 -0700
Subject: [PATCH] fix(vertex_ai): copy request params per key before a cache
 retry so an uncopyable param no longer fails the call

cache_rebuild deep-copied optional_params as a whole, so an SDK call carrying a param deepcopy cannot copy (an unknown kwarg holding a threading.Lock) raised APIConnectionError on any cache-marked request. Each key is now copied on its own, and a key that cannot be copied keeps its reference, as safe_deep_copy does elsewhere
---
 .../context_caching/rejected_cache_retry.py   |  3 +-
 .../test_known_cached_contents.py             | 31 +++++++++++++++++++
 2 files changed, 33 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/vertex_ai/context_caching/rejected_cache_retry.py b/litellm/llms/vertex_ai/context_caching/rejected_cache_retry.py
index 5bb13684470..43e4ee8dc84 100644
--- a/litellm/llms/vertex_ai/context_caching/rejected_cache_retry.py
+++ b/litellm/llms/vertex_ai/context_caching/rejected_cache_retry.py
@@ -5,6 +5,7 @@ from typing import Final, TypeVar
 
 import httpx
 
+from litellm.litellm_core_utils.core_helpers import independent_snapshot
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.llms.vertex_ai import RequestBody
@@ -22,7 +23,7 @@ def cache_rebuild(
     """Copies the inputs before the first transform, which pops keys from ``optional_params`` and ``messages``."""
     if not any(map(is_cached_message, messages)):
         return None
-    return partial(build, **deepcopy({"messages": messages, "optional_params": optional_params}))
+    return partial(build, messages=deepcopy(messages), optional_params=independent_snapshot(optional_params))
 
 
 def _status_and_message(error: httpx.HTTPStatusError | BaseLLMException) -> tuple[int, str]:
diff --git a/tests/unit/llms/vertex_ai/context_caching/test_known_cached_contents.py b/tests/unit/llms/vertex_ai/context_caching/test_known_cached_contents.py
index 8546d963603..0c5c3facb56 100644
--- a/tests/unit/llms/vertex_ai/context_caching/test_known_cached_contents.py
+++ b/tests/unit/llms/vertex_ai/context_caching/test_known_cached_contents.py
@@ -1,5 +1,6 @@
 import asyncio
 import json
+import threading
 from datetime import datetime, timezone
 from typing import Final
 from unittest.mock import AsyncMock, MagicMock, patch
@@ -450,6 +451,36 @@ async def _complete(sync_mode: bool, stream: bool, **kwargs) -> str:
     return response.choices[0].message.content
 
 
+@pytest.mark.parametrize("sync_mode", [True, False])
+@pytest.mark.parametrize("stream", [True, False])
+@pytest.mark.asyncio
+async def test_rejected_cache_is_replaced_when_a_request_param_cannot_be_deep_copied(
+    respx_mock: respx.MockRouter,
+    fake_provider_credentials: None,
+    monkeypatch: pytest.MonkeyPatch,
+    sync_mode: bool,
+    stream: bool,
+) -> None:
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    respx_mock.get(url__regex=r".*/cachedContents.*").mock(return_value=httpx.Response(200, json={}))
+    respx_mock.post(url__regex=r".*/cachedContents$").mock(
+        side_effect=[
+            httpx.Response(200, json=_created("cachedContents/deleted", _FAR_FUTURE)),
+            httpx.Response(200, json=_created("cachedContents/fresh", _FAR_FUTURE)),
+        ]
+    )
+    generate_route: Final = respx_mock.post(url__regex=r".*:(stream)?[gG]enerateContent.*").mock(
+        side_effect=[_generated(stream), _rejected(404, "CachedContent not found"), _generated(stream)]
+    )
+    lock: Final = threading.Lock()
+
+    assert await _complete(sync_mode, stream, lock_holder=lock) == "ok"
+    assert await _complete(sync_mode, stream, lock_holder=lock) == "ok"
+
+    sent: Final = [json.loads(call.request.content)["cachedContent"] for call in generate_route.calls]
+    assert sent == ["cachedContents/deleted", "cachedContents/deleted", "cachedContents/fresh"]
+
+
 @pytest.mark.parametrize("sync_mode", [True, False])
 @pytest.mark.parametrize("stream", [True, False])
 @pytest.mark.parametrize(