update arg_utils.py

vllm/engine/arg_utils.py

@@ -184,6 +184,7 @@ class EngineArgs:
     logits_processor_pattern: Optional[str] = None
 
     speculative_config: Optional[Dict[str, Any]] = None
+    num_speculative_heads: Optional[int] = None
 
     qlora_adapter_name_or_path: Optional[str] = None
     show_hidden_metrics_for_version: Optional[str] = None
@@ -793,6 +794,12 @@ class EngineArgs:
                             default=None,
                             help='The configurations for speculative decoding.'
                             ' Should be a JSON string.')
+        parser.add_argument(
+            '--num-speculative-heads',
+            type=int,
+            default=EngineArgs.num_speculative_heads,
+            help='The number of speculative heads to sample from '
+                 'the draft model in speculative decoding.')
         parser.add_argument('--model-loader-extra-config',
                             type=nullable_str,
                             default=EngineArgs.model_loader_extra_config,
@@ -1745,4 +1752,4 @@ def _engine_args_parser():
 
 def _async_engine_args_parser():
     return AsyncEngineArgs.add_cli_args(FlexibleArgumentParser(),
-                                        async_args_only=True)
+                                        async_args_only=True)
\ No newline at end of file