diff --git a/vllm/renderers/base.py b/vllm/renderers/base.py index 540a400c7f..6f374147e7 100644 --- a/vllm/renderers/base.py +++ b/vllm/renderers/base.py @@ -914,6 +914,15 @@ class BaseRenderer(ABC, Generic[_T]): return await self._process_tokens_async(prompt, skip_mm_cache=skip_mm_cache) # type: ignore[arg-type] + def _get_skip_decoder_start_token(self) -> bool: + """Whether the multimodal processor supplies a complete decoder prefix.""" + if self.mm_processor is not None: + from vllm.multimodal.processing import EncDecMultiModalProcessor + + if isinstance(self.mm_processor, EncDecMultiModalProcessor): + return self.mm_processor.skip_decoder_start_token + return False + def _process_enc_dec( self, prompt: EncoderDecoderTokPrompt, @@ -923,13 +932,6 @@ class BaseRenderer(ABC, Generic[_T]): enc_prompt = prompt["encoder_prompt"] dec_prompt = prompt["decoder_prompt"] - skip_decoder_start_token = False - if self.mm_processor is not None: - from vllm.multimodal.processing import EncDecMultiModalProcessor - - if isinstance(self.mm_processor, EncDecMultiModalProcessor): - skip_decoder_start_token = self.mm_processor.skip_decoder_start_token - return build_enc_dec_input( encoder_input=self._process_singleton( enc_prompt, skip_mm_cache=skip_mm_cache @@ -940,7 +942,7 @@ class BaseRenderer(ABC, Generic[_T]): else self._process_singleton(dec_prompt, skip_mm_cache=skip_mm_cache) ), decoder_start_token_id=self.get_dec_start_token_id(), - skip_decoder_start_token=skip_decoder_start_token, + skip_decoder_start_token=self._get_skip_decoder_start_token(), ) async def _process_enc_dec_async( @@ -967,6 +969,7 @@ class BaseRenderer(ABC, Generic[_T]): encoder_input=encoder_input, decoder_input=decoder_input, decoder_start_token_id=self.get_dec_start_token_id(), + skip_decoder_start_token=self._get_skip_decoder_start_token(), ) def process_for_engine(