xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 05:04:58 +08:00
Code Issues Packages Projects Releases Wiki Activity

[bugfix] Missing cached item in beam search (#27874)

Browse Source 
Signed-off-by: fake0fan <645327136@qq.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Chenguang Zheng 2025-11-01 01:34:27 +08:00 committed by GitHub
parent 70bfbd7b16
commit 103a468bbf
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 10 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
vllm/entrypoints/openai/serving_engine.py
View File

@ -345,22 +345,7 @@ class OpenAIServing:
if is_explicit_encoder_decoder_prompt(prompt):
raise NotImplementedError
else:
processed_inputs = processor.input_preprocessor._prompt_to_llm_inputs(
prompt
)
if processed_inputs["type"] == "embeds":
raise NotImplementedError
# This is a workaround to fix multimodal beam search; this is a
# bandaid fix for 2 small problems:
# 1. Multi_modal_data on the processed_inputs currently resolves to
# `None`.
# 2. preprocessing above expands the multimodal placeholders. However,
# this happens again in generation, so the double expansion causes
# a mismatch.
# TODO - would be ideal to handle this more gracefully.
prompt_text: str | None
prompt_token_ids: list[int]
multi_modal_data: MultiModalDataDict | None
@ -373,9 +358,16 @@ class OpenAIServing:
prompt_token_ids = prompt.get("prompt_token_ids", []) # type: ignore
multi_modal_data = prompt.get("multi_modal_data") # type: ignore
mm_processor_kwargs: dict[str, Any] | None = processed_inputs.get(
"mm_processor_kwargs"
) # type: ignore
mm_processor_kwargs: dict[str, Any] | None = None
# This is a workaround to fix multimodal beam search; this is a
# bandaid fix for 2 small problems:
# 1. Multi_modal_data on the processed_inputs currently resolves to
# `None`.
# 2. preprocessing above expands the multimodal placeholders. However,
# this happens again in generation, so the double expansion causes
# a mismatch.
# TODO - would be ideal to handle this more gracefully.
tokenized_length = len(prompt_token_ids)