diff --git a/vllm/multimodal/processing/processor.py b/vllm/multimodal/processing/processor.py
index 2c3ac6d591..593a48b337 100644
--- a/vllm/multimodal/processing/processor.py
+++ b/vllm/multimodal/processing/processor.py
@@ -1559,14 +1559,11 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
             # let BPE merge tokens across a segment boundary, silently
             # changing how a non-special-token placeholder is tokenized.
             new_token_ids = flatten_2d_lists(
-                [
-                    cached_encode(tokenizer, text, add_special_tokens=False)
-                    for text in out_texts
-                ]
+                [tokenizer.encode(text, add_special_tokens=False) for text in out_texts]
             )
         else:
-            new_token_ids = cached_encode(
-                tokenizer, "".join(out_texts), add_special_tokens=False
+            new_token_ids = tokenizer.encode(
+                "".join(out_texts), add_special_tokens=False
             )
 
         return new_token_ids, result