Disable prefix cache by default for benchmark (#18639)

Signed-off-by: cascade812 <[email protected]>

Author: cascade812

benchmarks/benchmark_latency.py

@@ -189,5 +189,8 @@ def run_to_completion(profile_dir: Optional[str] = None):
     )
 
     parser = EngineArgs.add_cli_args(parser)
+    # V1 enables prefix caching by default which skews the latency
+    # numbers. We need to disable prefix caching by default.
+    parser.set_defaults(enable_prefix_caching=False)
     args = parser.parse_args()
     main(args)

vllm/benchmarks/latency.py

@@ -80,6 +80,9 @@ def add_cli_args(parser: argparse.ArgumentParser):
     )
 
     parser = EngineArgs.add_cli_args(parser)
+    # V1 enables prefix caching by default which skews the latency
+    # numbers. We need to disable prefix caching by default.
+    parser.set_defaults(enable_prefix_caching=True)
 
 
 def main(args: argparse.Namespace):