[Bug][CLI] Allow users to disable prefix caching explicitly (#10724)

Signed-off-by: rickyx <rickyx@anyscale.com>

tests/v1/engine/test_engine_args.py

@@ -4,6 +4,7 @@ from vllm import envs
 from vllm.config import VllmConfig
 from vllm.engine.arg_utils import EngineArgs
 from vllm.usage.usage_lib import UsageContext
+from vllm.utils import FlexibleArgumentParser
 
 if not envs.VLLM_USE_V1:
     pytest.skip(
@@ -12,6 +13,24 @@ if not envs.VLLM_USE_V1:
     )
 
 
+def test_prefix_caching_from_cli():
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+    args = parser.parse_args([])
+    engine_args = EngineArgs.from_cli_args(args=args)
+    assert (engine_args.enable_prefix_caching
+            ), "V1 turns on prefix caching by default."
+
+    # Turn it off possible with flag.
+    args = parser.parse_args(["--no-enable-prefix-caching"])
+    engine_args = EngineArgs.from_cli_args(args=args)
+    assert not engine_args.enable_prefix_caching
+
+    # Turn it on with flag.
+    args = parser.parse_args(["--enable-prefix-caching"])
+    engine_args = EngineArgs.from_cli_args(args=args)
+    assert engine_args.enable_prefix_caching
+
+
 def test_defaults():
     engine_args = EngineArgs(model="facebook/opt-125m")