improve the robustness of parsing vlms config in AutoRound (vllm-project#18894)

Signed-off-by: wenhuach21 <wenhua.cheng@intel.com>
Signed-off-by: amit <amit.man@gmail.com>

Author: wenhuach21

vllm/model_executor/layers/quantization/auto_round.py

@@ -116,8 +116,9 @@ def get_layer_config(self, layer, layer_name: str):
 
         quantized = True
         if self.block_name_to_quantize:
-            quantized = any(name in layer_name
-                            for name in self.block_name_to_quantize)
+            quantized = any(
+                layer_name.startswith(name)
+                for name in self.block_name_to_quantize)
         elif isinstance(layer, ParallelLMHead):
             quantized = False