[Bugfix][Frontend] Honor skip_decoder_start_token in async encoder-decoder rendering (#54799)

Signed-off-by: waizuichougou <2082431897@qq.com>
This commit is contained in:
waizuichougou
2026-09-01 19:00:28 +00:00
committed by GitHub
parent 0ad5652a52
commit d98bb2a879
+11 -8
View File
@@ -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(