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

[New Model]: Support GteNewModelForSequenceClassification (#23524)

Browse Source 
Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-08-28 15:36:42 +08:00
committed by GitHub
parent 186aced5ff
commit 11a7fafaa8
13 changed files with 157 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/models/language/pooling/embed_utils.py
View File

@@ -51,6 +51,9 @@ def correctness_test_embed_models(hf_runner,
vllm_extra_kwargs = vllm_extra_kwargs or {}
vllm_extra_kwargs["dtype"] = model_info.dtype
if model_info.hf_overrides is not None:
vllm_extra_kwargs["hf_overrides"] = model_info.hf_overrides
with vllm_runner(model_info.name,
runner="pooling",
max_model_len=None,

6
tests/models/language/pooling/mteb_utils.py
View File

@@ -172,6 +172,9 @@ def mteb_test_embed_models(hf_runner,
vllm_extra_kwargs = vllm_extra_kwargs or {}
vllm_extra_kwargs["dtype"] = model_info.dtype
if model_info.hf_overrides is not None:
vllm_extra_kwargs["hf_overrides"] = model_info.hf_overrides
with vllm_runner(model_info.name,
runner="pooling",
max_model_len=None,
@@ -284,6 +287,9 @@ def mteb_test_rerank_models(hf_runner,
vllm_extra_kwargs = vllm_extra_kwargs or {}
vllm_extra_kwargs["dtype"] = model_info.dtype
if model_info.hf_overrides is not None:
vllm_extra_kwargs["hf_overrides"] = model_info.hf_overrides
with vllm_runner(model_info.name,
runner="pooling",
max_model_len=None,

24
tests/models/language/pooling/test_bge_reranker_v2_gemma.py
View File

@@ -13,7 +13,14 @@ from .mteb_utils import VllmMtebEncoder, mteb_test_rerank_models
RERANK_MODELS = [
LASTPoolingRerankModelInfo("BAAI/bge-reranker-v2-gemma",
architecture="GemmaForSequenceClassification"),
architecture="GemmaForSequenceClassification",
hf_overrides={
"architectures":
["GemmaForSequenceClassification"],
"classifier_from_token": ["Yes"],
"method":
"no_post_processing",
}),
]
PROMPT = "Given a query A and a passage B, determine whether the passage contains an answer to the query by providing a prediction of either 'Yes' or 'No'." # noqa: E501
@@ -119,22 +126,9 @@ class GemmaMtebEncoder(VllmMtebEncoder):
@pytest.mark.parametrize("model_info", RERANK_MODELS)
def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo,
monkeypatch) -> None:
monkeypatch.setenv("VLLM_USE_V1", "0")
assert model_info.architecture == "GemmaForSequenceClassification"
vllm_extra_kwargs: dict[str, Any] = {
"hf_overrides": {
"architectures": ["GemmaForSequenceClassification"],
"classifier_from_token": ["Yes"],
"method": "no_post_processing",
}
}
def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
mteb_test_rerank_models(GemmaRerankerHfRunner,
vllm_runner,
model_info,
vllm_extra_kwargs,
vllm_mteb_encoder=GemmaMtebEncoder)

24
tests/models/language/pooling/test_gte.py
View File

@@ -1,6 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
import pytest
@@ -33,12 +32,15 @@ MODELS = [
########### NewModel
CLSPoolingEmbedModelInfo("Alibaba-NLP/gte-multilingual-base",
architecture="GteNewModel",
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=True),
CLSPoolingEmbedModelInfo("Alibaba-NLP/gte-base-en-v1.5",
architecture="GteNewModel",
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=True),
CLSPoolingEmbedModelInfo("Alibaba-NLP/gte-large-en-v1.5",
architecture="GteNewModel",
hf_overrides={"architectures": ["GteNewModel"]},
enable_test=True),
########### Qwen2ForCausalLM
LASTPoolingEmbedModelInfo("Alibaba-NLP/gte-Qwen2-1.5B-instruct",
@@ -60,11 +62,16 @@ MODELS = [
]
RERANK_MODELS = [
# classifier_pooling: mean
CLSPoolingRerankModelInfo(
# classifier_pooling: mean
"Alibaba-NLP/gte-reranker-modernbert-base",
architecture="ModernBertForSequenceClassification",
enable_test=True),
CLSPoolingRerankModelInfo(
"Alibaba-NLP/gte-multilingual-reranker-base",
architecture="GteNewForSequenceClassification",
hf_overrides={"architectures": ["GteNewForSequenceClassification"]},
enable_test=True),
]
@@ -75,12 +82,7 @@ def test_embed_models_mteb(hf_runner, vllm_runner,
check_transformers_version(model_info.name,
max_transformers_version="4.53.2")
vllm_extra_kwargs: dict[str, Any] = {}
if model_info.architecture == "GteNewModel":
vllm_extra_kwargs["hf_overrides"] = {"architectures": ["GteNewModel"]}
mteb_test_embed_models(hf_runner, vllm_runner, model_info,
vllm_extra_kwargs)
mteb_test_embed_models(hf_runner, vllm_runner, model_info)
@pytest.mark.parametrize("model_info", MODELS)
@@ -91,12 +93,8 @@ def test_embed_models_correctness(hf_runner, vllm_runner,
check_transformers_version(model_info.name,
max_transformers_version="4.53.2")
vllm_extra_kwargs: dict[str, Any] = {}
if model_info.architecture == "GteNewModel":
vllm_extra_kwargs["hf_overrides"] = {"architectures": ["GteNewModel"]}
correctness_test_embed_models(hf_runner, vllm_runner, model_info,
example_prompts, vllm_extra_kwargs)
example_prompts)
@pytest.mark.parametrize("model_info", RERANK_MODELS)

19
tests/models/language/pooling/test_mxbai_rerank.py
View File

@@ -10,12 +10,20 @@ from tests.conftest import HfRunner
from ...utils import LASTPoolingRerankModelInfo, RerankModelInfo
from .mteb_utils import mteb_test_rerank_models
mxbai_rerank_hf_overrides = {
"architectures": ["Qwen2ForSequenceClassification"],
"classifier_from_token": ["0", "1"],
"method": "from_2_way_softmax",
}
RERANK_MODELS = [
LASTPoolingRerankModelInfo("mixedbread-ai/mxbai-rerank-base-v2",
architecture="Qwen2ForSequenceClassification",
hf_overrides=mxbai_rerank_hf_overrides,
enable_test=True),
LASTPoolingRerankModelInfo("mixedbread-ai/mxbai-rerank-large-v2",
architecture="Qwen2ForSequenceClassification",
hf_overrides=mxbai_rerank_hf_overrides,
enable_test=False)
]
@@ -71,13 +79,4 @@ class MxbaiRerankerHfRunner(HfRunner):
@pytest.mark.parametrize("model_info", RERANK_MODELS)
def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
vllm_extra_kwargs: dict[str, Any] = {}
if model_info.architecture == "Qwen2ForSequenceClassification":
vllm_extra_kwargs["hf_overrides"] = {
"architectures": ["Qwen2ForSequenceClassification"],
"classifier_from_token": ["0", "1"],
"method": "from_2_way_softmax",
}
mteb_test_rerank_models(MxbaiRerankerHfRunner, vllm_runner, model_info,
vllm_extra_kwargs)
mteb_test_rerank_models(MxbaiRerankerHfRunner, vllm_runner, model_info)

26
tests/models/language/pooling/test_qwen3_reranker.py
View File

@@ -11,12 +11,20 @@ from tests.utils import multi_gpu_test
from ...utils import LASTPoolingRerankModelInfo, RerankModelInfo
from .mteb_utils import mteb_test_rerank_models
qwen3_reranker_hf_overrides = {
"architectures": ["Qwen3ForSequenceClassification"],
"classifier_from_token": ["no", "yes"],
"is_original_qwen3_reranker": True,
}
RERANK_MODELS = [
LASTPoolingRerankModelInfo("Qwen/Qwen3-Reranker-0.6B",
architecture="Qwen3ForSequenceClassification",
hf_overrides=qwen3_reranker_hf_overrides,
enable_test=True),
LASTPoolingRerankModelInfo("Qwen/Qwen3-Reranker-4B",
architecture="Qwen3ForSequenceClassification",
hf_overrides=qwen3_reranker_hf_overrides,
enable_test=False)
]
@@ -74,18 +82,7 @@ class Qwen3RerankerHfRunner(HfRunner):
@pytest.mark.parametrize("model_info", RERANK_MODELS)
def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
assert model_info.architecture == "Qwen3ForSequenceClassification"
vllm_extra_kwargs: dict[str, Any] = {
"hf_overrides": {
"architectures": ["Qwen3ForSequenceClassification"],
"classifier_from_token": ["no", "yes"],
"is_original_qwen3_reranker": True,
}
}
mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info,
vllm_extra_kwargs)
mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info)
@pytest.mark.parametrize("model_info", RERANK_MODELS)
@@ -96,11 +93,6 @@ def test_rerank_models_mteb_tp(vllm_runner,
assert model_info.architecture == "Qwen3ForSequenceClassification"
vllm_extra_kwargs: dict[str, Any] = {
"hf_overrides": {
"architectures": ["Qwen3ForSequenceClassification"],
"classifier_from_token": ["no", "yes"],
"is_original_qwen3_reranker": True,
},
"tensor_parallel_size": 2,
}