Merge pull request #50849 from BerriAI/litellm_rc109_beta_header_test_bundled_config

test: repair four failing tests on rc/1.109.0

tests/integration/mcp/test_interactions.py

@@ -1227,6 +1227,7 @@ def test_modern_subscription_delivers_changes_between_calls_and_refreshes_tools(
             "PYTHONPATH": os.pathsep.join((str(source), str(source / "tests"))),
             "LITELLM_SALT_KEY": "subscription-pagination-key",
             **({"LITELLM_MCP_TOOL_LISTING_TIMEOUT": "2"} if access == "probe-timeout" else {}),
+            **({"LITELLM_MCP_METADATA_TIMEOUT": "5"} if access == "mixed-capabilities" else {}),
             "STORE_MODEL_IN_DB": "False",
             "DISABLE_SCHEMA_UPDATE": "true",
             "LITELLM_LOCAL_MODEL_COST_MAP": "True",

tests/integration/sdk/test_dual_cache_redis.py

@@ -222,6 +222,7 @@ async def test_tier_pin_joins_pending_reads_without_reusing_request_state(mode:
     )
     client: Final = Redis(connection_pool=pool)
     client_key: Final = redis._get_async_client_cache_key()
+    litellm.in_memory_llm_clients_cache.flush_cache()
     litellm.in_memory_llm_clients_cache.set_cache(key=client_key, value=client)
     cache: Final = DualCache(redis_cache=redis)
     key: Final = redis.check_and_fix_namespace("pin")

tests/integration/spend/test_normalized_error_long_message.py

@@ -358,16 +358,17 @@ def _burst(
     *,
     rejected_model: str = CRAFTED_MODEL,
 ) -> tuple[_BurstResult, ...]:
+    nonce: Final = uuid.uuid4().hex
     crafted: Final = tuple(
         (f"crafted-{path}-{index}", path, _body(path, rejected_model, uuid.uuid4().hex, stream=index % 2 == 1), open_key)
         for path in (CHAT, MESSAGES, RESPONSES)
         for index in range(4)
     )
     happy: Final = tuple(
-        (f"happy-{index}", CHAT, _body(CHAT, happy_model, f"happy-{index}"), happy_key) for index in range(8)
+        (f"happy-{index}", CHAT, _body(CHAT, happy_model, f"happy-{index}-{nonce}"), happy_key) for index in range(8)
     )
     late: Final = tuple(
-        (f"late-{index}", CHAT, _body(CHAT, happy_model, f"late-{index}"), happy_key) for index in range(8)
+        (f"late-{index}", CHAT, _body(CHAT, happy_model, f"late-{index}-{nonce}"), happy_key) for index in range(8)
     )
     with (
         ThreadPoolExecutor(max_workers=28) as pool,
@@ -427,6 +428,7 @@ def test_killing_one_worker_mid_burst_leaves_the_other_serving_crafted_and_happy
         open_key: Final = scenario.key()
         control: Final = _burst(client, model, key, open_key, lambda: None, rejected_model=BENIGN_288KB_MODEL)
         assert all(result.status == (400 if result.label.startswith("crafted") else 200) for result in control), control
+        wire.drain()
         workers: Final = owned.worker_pids()
         assert len(workers) == 2, workers
         victim: Final = min(workers)
@@ -438,6 +440,8 @@ def test_killing_one_worker_mid_burst_leaves_the_other_serving_crafted_and_happy
             _assert_burst_rejections_fast(results, control)
             late: Final = tuple(result for result in results if result.label.startswith("late"))
             assert all(result.status == 200 for result in late), late
+            seen: Final = tuple(request.body.decode() for request in wire.drain())
+            assert all(any(f"normalized error audit {result.label}-" in body for body in seen) for result in late), seen
             _assert_rows_land_exactly_once(results, "late")
             survivor: Final = tuple(pid for pid in owned.worker_pids() if pid != victim)
             assert survivor, "no worker left serving"

tests/unit/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_partner_models_anthropic_transformation.py

@@ -640,7 +640,7 @@ async def test_vertex_ai_anthropic_prompt_caching_does_not_add_beta_header(
     assert "anthropic_beta" not in json.loads(request.content)
 
 
-def test_vertex_ai_partner_models_anthropic_remove_prompt_caching_scope_beta_header():
+def test_vertex_ai_partner_models_anthropic_remove_prompt_caching_scope_beta_header(local_beta_headers_config: None) -> None:
     """
     Test that remove_unsupported_beta correctly filters out prompt-caching-scope-2026-01-05
     from the anthropic-beta headers.