ZIBEE/bi_150-vllm
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Upgrade to vllm 0.17.0 corex v4.1 overlay

Browse Source
This commit is contained in:
LZiBee
2026-04-29 19:38:22 +08:00
parent 8fac6062e4
commit 938d0854a5
430 changed files with 35969 additions and 14511 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
vllm/entrypoints/pooling/score/serving.py
View File

@@ -31,7 +31,7 @@ from vllm.entrypoints.pooling.score.utils import (
ScoreInputs,
_cosine_similarity,
compress_token_type_ids,
compute_maxsim_score,
compute_maxsim_scores,
get_score_prompt,
parse_score_data_single,
validate_score_input,
@@ -56,6 +56,7 @@ class ServingScores(OpenAIServing):
request_logger: RequestLogger | None,
score_template: str | None = None,
log_error_stack: bool = False,
use_gpu_for_pooling_score: bool = False,
) -> None:
super().__init__(
engine_client=engine_client,
@@ -64,6 +65,7 @@ class ServingScores(OpenAIServing):
log_error_stack=log_error_stack,
)
self.score_template = score_template
self.use_gpu_for_pooling_score = use_gpu_for_pooling_score
self._tokenizer_executor = ThreadPoolExecutor(max_workers=1)
@@ -311,19 +313,18 @@ class ServingScores(OpenAIServing):
# Compute MaxSim scores
from vllm.outputs import PoolingOutput
maxsim_scores = compute_maxsim_scores(
[emb.outputs.data for emb in emb_data_1],
[emb.outputs.data for emb in emb_data_2],
use_gpu_for_pooling_score=self.use_gpu_for_pooling_score,
)
scores: list[PoolingRequestOutput] = []
padding: list[int] = []
if (pad_token_id := tokenizer.pad_token_id) is not None:
padding = [pad_token_id]
for emb_1, emb_2 in zip(emb_data_1, emb_data_2):
# emb_1.outputs.data: [query_len, dim]
# emb_2.outputs.data: [doc_len, dim]
q_emb = emb_1.outputs.data
d_emb = emb_2.outputs.data
maxsim_score = compute_maxsim_score(q_emb, d_emb)
for emb_1, emb_2, maxsim_score in zip(emb_data_1, emb_data_2, maxsim_scores):
tokens = emb_1.prompt_token_ids + padding + emb_2.prompt_token_ids
scores.append(