[Bugfix][Multimodal] Avoid caching full prompts in fallback (#54439)
Signed-off-by: waizuichougou <2082431897@qq.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
waizuichougou committed
Aug 31, 2026 at 00:19 UTC
f6895a5fcb4710481bb44726af4fa57f4172330c
1 file changed
+3
-6
vllm/multimodal/processing/processor.py
+3
-6
@@ -1559,14 +1559,11 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
1559
# let BPE merge tokens across a segment boundary, silently
1560
# changing how a non-special-token placeholder is tokenized.
1561
new_token_ids = flatten_2d_lists(
1562
- [
1563
- cached_encode(tokenizer, text, add_special_tokens=False)
1564
- for text in out_texts
1565
- ]
1562
+ [tokenizer.encode(text, add_special_tokens=False) for text in out_texts]
1563
)
1564
else:
1568
- new_token_ids = cached_encode(
1569
- tokenizer, "".join(out_texts), add_special_tokens=False
1565
+ new_token_ids = tokenizer.encode(
1566
+ "".join(out_texts), add_special_tokens=False
1567
)
1568
1569
return new_token_ids, result