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

[CI] Reorganization pooling_mteb_test (#31265)

Browse Source 
Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-12-24 23:36:20 +08:00
committed by GitHub
parent 7cd288a4b3
commit 1ff67df182
18 changed files with 480 additions and 347 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
tests/models/language/pooling_mteb_test/test_gte.py
View File

@@ -5,36 +5,32 @@ import pytest
from tests.models.language.pooling.embed_utils import correctness_test_embed_models
from tests.models.utils import (
CLSPoolingEmbedModelInfo,
CLSPoolingRerankModelInfo,
EmbedModelInfo,
LASTPoolingEmbedModelInfo,
RerankModelInfo,
)
from .mteb_utils import mteb_test_embed_models, mteb_test_rerank_models
from .mteb_embed_utils import mteb_test_embed_models
from .mteb_score_utils import mteb_test_rerank_models
MODELS = [
########## BertModel
CLSPoolingEmbedModelInfo(
EmbedModelInfo(
"thenlper/gte-large",
mteb_score=0.76807651,
architecture="BertModel",
pooling_type="MEAN",
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
),
CLSPoolingEmbedModelInfo(
"thenlper/gte-base", architecture="BertModel", enable_test=False
),
CLSPoolingEmbedModelInfo(
"thenlper/gte-small", architecture="BertModel", enable_test=False
),
CLSPoolingEmbedModelInfo(
EmbedModelInfo("thenlper/gte-base", architecture="BertModel", enable_test=False),
EmbedModelInfo("thenlper/gte-small", architecture="BertModel", enable_test=False),
EmbedModelInfo(
"thenlper/gte-large-zh", architecture="BertModel", enable_test=False
),
CLSPoolingEmbedModelInfo(
"thenlper/gte-base-zh", architecture="BertModel", enable_test=False
),
CLSPoolingEmbedModelInfo(
EmbedModelInfo("thenlper/gte-base-zh", architecture="BertModel", enable_test=False),
EmbedModelInfo(
"thenlper/gte-small-zh", architecture="BertModel", enable_test=False
),
########### NewModel
@@ -43,48 +39,64 @@ MODELS = [
# - whether to use token_type_embeddings
# - whether to use context expansion
# So only test one (the most widely used) model
CLSPoolingEmbedModelInfo(
EmbedModelInfo(
"Alibaba-NLP/gte-multilingual-base",
architecture="GteNewModel",
mteb_score=0.775074696,
hf_overrides={"architectures": ["GteNewModel"]},
pooling_type="CLS",
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
),
CLSPoolingEmbedModelInfo(
EmbedModelInfo(
"Alibaba-NLP/gte-base-en-v1.5",
architecture="GteNewModel",
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=False,
),
CLSPoolingEmbedModelInfo(
EmbedModelInfo(
"Alibaba-NLP/gte-large-en-v1.5",
architecture="GteNewModel",
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=False,
),
########### Qwen2ForCausalLM
LASTPoolingEmbedModelInfo(
EmbedModelInfo(
"Alibaba-NLP/gte-Qwen2-1.5B-instruct",
mteb_score=0.758473459018872,
architecture="Qwen2ForCausalLM",
pooling_type="LAST",
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
),
########## ModernBertModel
CLSPoolingEmbedModelInfo(
EmbedModelInfo(
"Alibaba-NLP/gte-modernbert-base",
mteb_score=0.748193353,
architecture="ModernBertModel",
pooling_type="CLS",
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
),
########## Qwen3ForCausalLM
LASTPoolingEmbedModelInfo(
EmbedModelInfo(
"Qwen/Qwen3-Embedding-0.6B",
mteb_score=0.771163695,
architecture="Qwen3ForCausalLM",
dtype="float32",
pooling_type="LAST",
attn_type="decoder",
is_prefix_caching_supported=True,
is_chunked_prefill_supported=True,
enable_test=True,
),
LASTPoolingEmbedModelInfo(
EmbedModelInfo(
"Qwen/Qwen3-Embedding-4B",
architecture="Qwen3ForCausalLM",
dtype="float32",
@@ -93,18 +105,26 @@ MODELS = [
]
RERANK_MODELS = [
CLSPoolingRerankModelInfo(
RerankModelInfo(
# classifier_pooling: mean
"Alibaba-NLP/gte-reranker-modernbert-base",
mteb_score=0.33386,
architecture="ModernBertForSequenceClassification",
pooling_type="CLS",
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
),
CLSPoolingRerankModelInfo(
RerankModelInfo(
"Alibaba-NLP/gte-multilingual-reranker-base",
mteb_score=0.33062,
architecture="GteNewForSequenceClassification",
hf_overrides={"architectures": ["GteNewForSequenceClassification"]},
pooling_type="CLS",
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
),
]