biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Frontend] Support multimodal inputs for late-interaction scoring (ColQwen3) + NewModel: nvidia/nemotron-colembed (#34574)

Browse Source 
Signed-off-by: craftsangjae <craftsangjae@gmail.com>
This commit is contained in:
Kata Coder
2026-02-21 13:01:40 +09:00
committed by GitHub
parent 11be2c74dc
commit 5719a4e4e6
10 changed files with 532 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
vllm/entrypoints/llm.py
View File

@@ -50,6 +50,7 @@ from vllm.entrypoints.pooling.score.utils import (
compress_token_type_ids,
compute_maxsim_score,
get_score_prompt,
score_data_to_prompts,
validate_score_input,
)
from vllm.entrypoints.utils import log_non_default_args
@@ -1395,25 +1396,13 @@ class LLM:
tokenizer = self.get_tokenizer()
# Extract text from ScoreData
text_1: list[str] = []
for text in data_1:
if not isinstance(text, str):
raise NotImplementedError(
"Late interaction scores currently do not support multimodal input."
)
text_1.append(text)
# Convert ScoreData to PromptType (handles both text and multimodal)
model_config = self.model_config
prompts_1 = score_data_to_prompts(data_1, "query", model_config)
prompts_2 = score_data_to_prompts(data_2, "document", model_config)
text_2: list[str] = []
for text in data_2:
if not isinstance(text, str):
raise NotImplementedError(
"Late interaction scores currently do not support multimodal input."
)
text_2.append(text)
encoded_output = self.encode(
text_1 + text_2,
encoded_output: list[PoolingRequestOutput] = self.encode(
prompts_1 + prompts_2,
use_tqdm=use_tqdm,
lora_request=lora_request,
pooling_params=pooling_params,
@@ -1421,8 +1410,8 @@ class LLM:
tokenization_kwargs=tokenization_kwargs,
)
encoded_output_1 = encoded_output[0 : len(text_1)]
encoded_output_2 = encoded_output[len(text_1) :]
encoded_output_1: list[PoolingRequestOutput] = encoded_output[: len(prompts_1)]
encoded_output_2: list[PoolingRequestOutput] = encoded_output[len(prompts_1) :]
if len(encoded_output_1) == 1:
encoded_output_1 = encoded_output_1 * len(encoded_output_2)