[Model] rename use_pad_token to use_sep_token (#31784)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>

examples/pooling/score/convert_model_to_seq_cls.py

@@ -108,7 +108,7 @@ method_map = {
 
 
 def converting(
-    model_name, classifier_from_tokens, path, method, use_pad_token=False, device="cpu"
+    model_name, classifier_from_tokens, path, method, use_sep_token=False, device="cpu"
 ):
     """
     Main conversion function to transform a CausalLM model to SequenceClassification.
@@ -118,7 +118,7 @@ def converting(
         classifier_from_tokens: List of tokens used for classification
         path: Output path to save the converted model
         method: Conversion method ('from_2_way_softmax' or 'no_post_processing')
-        use_pad_token: Whether to use padding token in the sequence classification model
+        use_sep_token: Whether to use separating token in the sequence classification model
         device: Device to load the model on ('cpu' or 'cuda')
     """
     assert method in method_map, f"Unknown method: {method}"
@@ -149,10 +149,10 @@ def converting(
         causal_lm, seq_cls_model, tokenizer, classifier_from_tokens, device
     )
 
-    # Configure padding token settings
-    # Note: Reranker models typically don't use padding tokens by default
-    seq_cls_model.config.use_pad_token = use_pad_token
-    seq_cls_model.config.pad_token_id = tokenizer.pad_token_id
+    # Configure separating token settings
+    # Note: `llm as reranker` defaults to not using separating token.
+    seq_cls_model.config.use_sep_token = use_sep_token
+    seq_cls_model.config.sep_token_id = tokenizer.sep_token_id
 
     # Save the converted model and tokenizer
     seq_cls_model.save_pretrained(path)
@@ -203,6 +203,6 @@ if __name__ == "__main__":
         model_name=args.model_name,
         classifier_from_tokens=json.loads(args.classifier_from_tokens),
         method=args.method,
-        use_pad_token=args.use_pad_token,
+        use_sep_token=args.use_sep_token,
         path=args.path,
     )

tests/entrypoints/pooling/score/test_utils.py

@@ -51,9 +51,9 @@ def llm_reranker_model_config():
         CROSS_ENCODER_MODEL_ID,
         runner="pooling",
     )
-    # use_pad_token is a property that reads from hf_config,
+    # use_sep_token is a property that reads from hf_config,
     # so we set it there to override the default (True)
-    config.hf_config.use_pad_token = False
+    config.hf_config.use_sep_token = False
     return config
 
 
@@ -230,7 +230,7 @@ class TestGetScorePrompt:
             cross_encoder_tokenizer, full_prompt, engine_prompt
         )
 
-    def test_fallback_with_pad_token(
+    def test_fallback_with_sep_token(
         self,
         cross_encoder_model_config,
         cross_encoder_tokenizer,
@@ -238,7 +238,7 @@ class TestGetScorePrompt:
         mock_model_no_score_template,
     ):
         """Test fallback path when ChatTemplateResolutionError
-        and use_pad_token=True."""
+        and use_sep_token=True."""
         with (
             patch(
                 "vllm.model_executor.model_loader.get_model_cls",
@@ -250,7 +250,7 @@ class TestGetScorePrompt:
             ),
         ):
             full_prompt, engine_prompt = get_score_prompt(
-                cross_encoder_model_config,  # use_pad_token=True
+                cross_encoder_model_config,  # use_sep_token=True
                 cross_encoder_tokenizer,
                 tokenization_kwargs,
                 "query",
@@ -281,7 +281,7 @@ class TestGetScorePrompt:
             add_special_tokens=False,
         )
 
-    def test_fallback_without_pad_token(
+    def test_fallback_without_sep_token(
         self,
         llm_reranker_model_config,
         cross_encoder_tokenizer,
@@ -289,7 +289,7 @@ class TestGetScorePrompt:
         mock_model_no_score_template,
     ):
         """Test fallback path when ChatTemplateResolutionError
-        and use_pad_token=False."""
+        and use_sep_token=False."""
         with (
             patch(
                 "vllm.model_executor.model_loader.get_model_cls",
@@ -301,7 +301,7 @@ class TestGetScorePrompt:
             ),
         ):
             full_prompt, engine_prompt = get_score_prompt(
-                llm_reranker_model_config,  # use_pad_token=False
+                llm_reranker_model_config,  # use_sep_token=False
                 cross_encoder_tokenizer,
                 tokenization_kwargs,
                 "query",

vllm/config/model.py

@@ -1434,10 +1434,18 @@ class ModelConfig:
         return getattr(self.hf_config, "matryoshka_dimensions", None)
 
     @property
-    def use_pad_token(self) -> bool:
-        # cross_encoder models defaults to using pad_token.
-        # `llm as reranker` models defaults to not using pad_token.
-        return getattr(self.hf_config, "use_pad_token", True)
+    def use_sep_token(self) -> bool:
+        # cross_encoder models defaults to using separating token.
+        # `llm as reranker` defaults to not using separating token.
+
+        use_pad_token = getattr(self.hf_config, "use_pad_token", None)
+        if use_pad_token is not None:
+            logger.warning_once(
+                "use_pad_token has been deprecated; please use use_sep_token instead."
+            )
+            return use_pad_token
+
+        return getattr(self.hf_config, "use_sep_token", True)
 
     @property
     def head_dtype(self) -> torch.dtype:

vllm/entrypoints/score_utils.py

@@ -199,14 +199,14 @@ def get_score_prompt(
             full_prompt = _apply_model_score_template(model_config, prompt_1, prompt_2)
             prompt_inputs = tokenizer(full_prompt, **tokenization_kwargs)
         else:
-            if model_config.use_pad_token:
-                # cross_encoder models defaults to using pad_token.
+            if model_config.use_sep_token:
+                # cross_encoder models defaults to using separating token.
                 prompt_inputs = tokenizer(
                     text=prompt_1, text_pair=prompt_2, **tokenization_kwargs
                 )
                 full_prompt = tokenizer.decode(prompt_inputs["input_ids"])
             else:
-                # `llm as reranker` models defaults to not using pad_token.
+                # `llm as reranker` defaults to not using separating token.
                 full_prompt = prompt_1 + prompt_2
                 prompt_inputs = tokenizer(text=full_prompt, **tokenization_kwargs)
         return full_prompt, prompt_inputs

vllm/model_executor/models/adapters.py

@@ -382,9 +382,9 @@ class SequenceClassificationConfig(VerifyAndUpdateConfig):
         else:
             text_config.num_labels = len(tokens)
 
-        # `llm as reranker` defaults to not using pad_token
-        use_pad_token = getattr(text_config, "use_pad_token", False)
-        text_config.use_pad_token = use_pad_token
+        # `llm as reranker` defaults to not using separating token.
+        use_sep_token = getattr(text_config, "use_sep_token", False)
+        text_config.use_sep_token = use_sep_token
 
 
 def load_weights_using_from_2_way_softmax(