[Inference] Support qwen2_5_vl (#9924)

* support qwen2_5_vl

* update support qwen2_vl & qwen2_5_vl & support multi models mapping

Author: chang-wenbin

paddlenlp/experimental/transformers/qwen2/modeling.py

@@ -1554,4 +1554,17 @@ class Qwen2VLForConditionalGenerationBlockInferenceModel(Qwen2ForCausalLMBlockIn
 
     # NOTE: (changwenbin) This function corresponds to QWen2-VL's second part, only used for QWen2-VL.
     def __init__(self, config: Qwen2Config):
-        super().__init__(config, base_model_prefix="model")
+        super().__init__(config)
+        self.qwen2.base_model_prefix = "model"
+
+
+class Qwen2_5_VLForConditionalGenerationBlockInferenceModel(Qwen2ForCausalLMBlockInferenceModel):
+    """
+    NOTE: (changwenbin) This class inherits from Qwen2ForCausalLMBlockInferenceModel.
+    Used only for QWen2-5-VL's second part.
+    """
+
+    # NOTE: (changwenbin) This function corresponds to QWen2-5-VL's second part, only used for QWen2-5-VL.
+    def __init__(self, config: Qwen2Config):
+        super().__init__(config)
+        self.qwen2.base_model_prefix = "model"

paddlenlp/transformers/auto/configuration.py

@@ -213,6 +213,14 @@
     ]
 )
 
+MULTI_MODELS_MAPPING = OrderedDict(
+    # multi models mapping
+    [
+        ("qwen2_vl", "qwen2"),
+        ("qwen2_5_vl", "qwen2"),
+    ]
+)
+
 
 def config_class_to_model_type(config):
     """Converts a config class name to the corresponding model type"""
@@ -238,8 +246,9 @@ def __init__(self, mapping):
 
     def __getitem__(self, key):
         # NOTE: (changwenbin) This is to enable the qwen2_vl language model to use qwen2 reasoning optimization
-        if key == "qwen2_vl":
-            key = "qwen2"
+        for model_type, model_key in MULTI_MODELS_MAPPING.items():
+            if key == model_type:
+                key = model_key
         if key in self._extra_content:
             return self._extra_content[key]
         if key not in self._mapping: