[bugfix] Fix duplicate 'load_format' argument being passed in rollout (#7312)

Author: hjh0119

swift/llm/infer/rollout.py

@@ -394,9 +394,13 @@ def get_infer_engine(args: RolloutArguments, template=None, **kwargs):
         engine_kwargs = kwargs.get('engine_kwargs', {})
         # for RL rollout model weight sync
         engine_kwargs.update({'worker_extension_cls': 'swift.llm.infer.rollout.WeightSyncWorkerExtension'})
-        # Use load_format from engine_kwargs if provided, otherwise default to 'dummy'
-        if 'load_format' not in engine_kwargs:
-            engine_kwargs['load_format'] = 'dummy'
+
+        # For RL rollout, we use 'dummy' load_format to prevent vLLM from loading weights from disk,
+        # as they will be synced from the trainer process.
+        # This will accelerate the rollout speed.
+        load_format = engine_kwargs.pop('load_format', 'dummy')
+        kwargs['load_format'] = load_format
+
         if args.vllm_use_async_engine and args.vllm_data_parallel_size > 1:
             engine_kwargs['data_parallel_size'] = args.vllm_data_parallel_size
         kwargs['engine_kwargs'] = engine_kwargs