[Bugfix] Fix logic for choosing default prefix caching setting (vllm-project#29393)

Signed-off-by: Thomas Parnell <[email protected]>

Author: tdoublep

tests/engine/test_arg_utils.py

@@ -277,8 +277,9 @@ def test_prefix_cache_default():
     parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
     args = parser.parse_args([])
 
+    # should be None by default (depends on model).
     engine_args = EngineArgs.from_cli_args(args=args)
-    assert engine_args.enable_prefix_caching, "prefix caching should default to on."
+    assert engine_args.enable_prefix_caching is None
 
     # with flag to turn it on.
     args = parser.parse_args(["--enable-prefix-caching"])

vllm/engine/arg_utils.py

@@ -880,7 +880,11 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             "--num-gpu-blocks-override", **cache_kwargs["num_gpu_blocks_override"]
         )
         cache_group.add_argument(
-            "--enable-prefix-caching", **cache_kwargs["enable_prefix_caching"]
+            "--enable-prefix-caching",
+            **{
+                **cache_kwargs["enable_prefix_caching"],
+                "default": None,
+            },
         )
         cache_group.add_argument(
             "--prefix-caching-hash-algo", **cache_kwargs["prefix_caching_hash_algo"]