diff --git a/data_juicer/ops/mapper/text_tagging_by_prompt_mapper.py b/data_juicer/ops/mapper/text_tagging_by_prompt_mapper.py
index a870d0035b..cad48061dc 100644
--- a/data_juicer/ops/mapper/text_tagging_by_prompt_mapper.py
+++ b/data_juicer/ops/mapper/text_tagging_by_prompt_mapper.py
@@ -114,7 +114,11 @@ def __init__(
         """  # noqa
 
         super().__init__(*args, **kwargs)
-        self.num_proc = 1
+        if not enable_vllm or not is_ray_mode():
+            # In Ray mode with vllm, num_proc is left to the caller so Ray can
+            # schedule one actor per GPU for data parallelism.
+            self.num_proc = 1
+            
 
         self.prompt = prompt
         self.tag_list = tag_list
@@ -129,9 +133,6 @@ def __init__(
         sampling_params = update_sampling_params(sampling_params, hf_model, self.enable_vllm)
 
         if enable_vllm:
-            if not is_ray_mode():
-                # cannot initialize vllm replicas on different GPUs
-                self.num_proc = 1
             self.model_key = prepare_model(
                 model_type="vllm",
                 pretrained_model_name_or_path=hf_model,