[Misc] Remove unnecessary detokenization in multimodal processing (#12868)

vllm/inputs/preprocess.py

@@ -260,9 +260,6 @@ class InputPreprocessor:
         mm_processor = self.mm_registry.create_processor(
             self.model_config, tokenizer)
 
-        if isinstance(prompt, list):
-            prompt = tokenizer.decode(prompt)
-
         if mm_processor_kwargs is None:
             mm_processor_kwargs = {}