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

[CI] Accelerate mteb test by setting SentenceTransformers mteb score to a constant (#24088)

Browse Source 
Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-09-03 17:23:56 +08:00
committed by GitHub
parent 9c99e4871f
commit 51383bd472
17 changed files with 83 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
tests/models/language/pooling/test_gte.py
View File

@@ -5,13 +5,14 @@ import pytest
from ...utils import (CLSPoolingEmbedModelInfo, CLSPoolingRerankModelInfo,
EmbedModelInfo, LASTPoolingEmbedModelInfo,
RerankModelInfo, check_transformers_version)
RerankModelInfo)
from .embed_utils import correctness_test_embed_models
from .mteb_utils import mteb_test_embed_models, mteb_test_rerank_models
MODELS = [
########## BertModel
CLSPoolingEmbedModelInfo("thenlper/gte-large",
mteb_score=0.76807651,
architecture="BertModel",
enable_test=True),
CLSPoolingEmbedModelInfo("thenlper/gte-base",
@@ -30,28 +31,37 @@ MODELS = [
architecture="BertModel",
enable_test=False),
########### NewModel
# These three architectures are almost the same, but not exactly the same.
# For example,
# - whether to use token_type_embeddings
# - whether to use context expansion
# So only test one (the most widely used) model
CLSPoolingEmbedModelInfo("Alibaba-NLP/gte-multilingual-base",
architecture="GteNewModel",
mteb_score=0.775074696,
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=True),
CLSPoolingEmbedModelInfo("Alibaba-NLP/gte-base-en-v1.5",
architecture="GteNewModel",
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=True),
enable_test=False),
CLSPoolingEmbedModelInfo("Alibaba-NLP/gte-large-en-v1.5",
architecture="GteNewModel",
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=True),
enable_test=False),
########### Qwen2ForCausalLM
LASTPoolingEmbedModelInfo("Alibaba-NLP/gte-Qwen2-1.5B-instruct",
mteb_score=0.758473459018872,
architecture="Qwen2ForCausalLM",
enable_test=True),
########## ModernBertModel
CLSPoolingEmbedModelInfo("Alibaba-NLP/gte-modernbert-base",
mteb_score=0.748193353,
architecture="ModernBertModel",
enable_test=True),
########## Qwen3ForCausalLM
LASTPoolingEmbedModelInfo("Qwen/Qwen3-Embedding-0.6B",
mteb_score=0.771163695,
architecture="Qwen3ForCausalLM",
dtype="float32",
enable_test=True),
@@ -65,10 +75,12 @@ RERANK_MODELS = [
CLSPoolingRerankModelInfo(
# classifier_pooling: mean
"Alibaba-NLP/gte-reranker-modernbert-base",
mteb_score=0.33386,
architecture="ModernBertForSequenceClassification",
enable_test=True),
CLSPoolingRerankModelInfo(
"Alibaba-NLP/gte-multilingual-reranker-base",
mteb_score=0.33062,
architecture="GteNewForSequenceClassification",
hf_overrides={"architectures": ["GteNewForSequenceClassification"]},
enable_test=True),
@@ -78,10 +90,6 @@ RERANK_MODELS = [
@pytest.mark.parametrize("model_info", MODELS)
def test_embed_models_mteb(hf_runner, vllm_runner,
model_info: EmbedModelInfo) -> None:
if model_info.name == "Alibaba-NLP/gte-Qwen2-1.5B-instruct":
check_transformers_version(model_info.name,
max_transformers_version="4.53.2")
mteb_test_embed_models(hf_runner, vllm_runner, model_info)
@@ -89,10 +97,6 @@ def test_embed_models_mteb(hf_runner, vllm_runner,
def test_embed_models_correctness(hf_runner, vllm_runner,
model_info: EmbedModelInfo,
example_prompts) -> None:
if model_info.name == "Alibaba-NLP/gte-Qwen2-1.5B-instruct":
check_transformers_version(model_info.name,
max_transformers_version="4.53.2")
correctness_test_embed_models(hf_runner, vllm_runner, model_info,
example_prompts)