[V0 Deprecation] Remove pooling model support in V0 (#23434)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Maximilien de Bayser
2025-08-29 00:04:02 -07:00
committed by GitHub
co-authored by Woosuk Kwon
parent 934bebf192
commit 2554b27baa
38 changed files with 99 additions and 808 deletions
@@ -10,14 +10,6 @@ from vllm.platforms import current_platform
from ...utils import check_embeddings_close, check_transformers_version
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
# Simple autouse wrapper to run both engines for each test
# This can be promoted up to conftest.py to run for every
# test in a package
pass
@pytest.mark.parametrize(
"model",
[
@@ -32,21 +24,15 @@ def v1(run_with_both_engines):
"intfloat/e5-mistral-7b-instruct",
# CPU v1 doesn't support sliding window
marks=[pytest.mark.core_model]),
# the qwen models interfere with each other (see PR
# https://github.com/vllm-project/vllm/pull/18720).
# To avoid this problem, for now we skip v0 since it will be
# deprecated anyway.
pytest.param("ssmits/Qwen2-7B-Instruct-embed-base",
marks=[pytest.mark.skip_v0, pytest.mark.cpu_model]),
marks=[pytest.mark.cpu_model]),
# [Encoder-only]
pytest.param("BAAI/bge-base-en-v1.5", marks=[pytest.mark.core_model]),
pytest.param("sentence-transformers/all-MiniLM-L12-v2"),
pytest.param("intfloat/multilingual-e5-small"),
pytest.param("Alibaba-NLP/gte-Qwen2-1.5B-instruct",
marks=[pytest.mark.skip_v1]),
pytest.param("Alibaba-NLP/gte-Qwen2-1.5B-instruct"),
# [Cross-Encoder]
pytest.param("sentence-transformers/stsb-roberta-base-v2",
marks=[pytest.mark.skip_v1]),
pytest.param("sentence-transformers/stsb-roberta-base-v2"),
],
)
def test_models(
@@ -13,14 +13,6 @@ from ....conftest import HfRunner
from ...utils import check_transformers_version
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
# Simple autouse wrapper to run both engines for each test
# This can be promoted up to conftest.py to run for every
# test in a package
pass
@pytest.fixture
def math_step_prompts():
# ruff: noqa: E501
@@ -23,15 +23,6 @@ TEXTS_2 = [
"The capital of Germany is Berlin.",
]
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
# Simple autouse wrapper to run both engines for each test
# This can be promoted up to conftest.py to run for every
# test in a package
pass
DTYPE = "half"
+11 -12
View File
@@ -323,8 +323,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
_EMBEDDING_EXAMPLE_MODELS = {
# [Text-only]
"BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5", v0_only=True),
"Gemma2Model": _HfExamplesInfo("BAAI/bge-multilingual-gemma2", v0_only=True), # noqa: E501
"BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5"),
"Gemma2Model": _HfExamplesInfo("BAAI/bge-multilingual-gemma2"), # noqa: E501
"GritLM": _HfExamplesInfo("parasail-ai/GritLM-7B-vllm"),
"GteModel": _HfExamplesInfo("Snowflake/snowflake-arctic-embed-m-v2.0",
trust_remote_code=True),
@@ -337,9 +337,9 @@ _EMBEDDING_EXAMPLE_MODELS = {
"LlamaModel": _HfExamplesInfo("llama", is_available_online=False),
"MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"),
"ModernBertModel": _HfExamplesInfo("Alibaba-NLP/gte-modernbert-base",
trust_remote_code=True, v0_only=True),
trust_remote_code=True),
"NomicBertModel": _HfExamplesInfo("nomic-ai/nomic-embed-text-v2-moe",
trust_remote_code=True, v0_only=True), # noqa: E501
trust_remote_code=True), # noqa: E501
"Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
"Qwen2ForRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-RM-72B",
max_transformers_version="4.53",
@@ -347,9 +347,9 @@ _EMBEDDING_EXAMPLE_MODELS = {
"Qwen2ForProcessRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-PRM-7B",
max_transformers_version="4.53",
transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers"), # noqa: E501
"RobertaModel": _HfExamplesInfo("sentence-transformers/stsb-roberta-base-v2", v0_only=True), # noqa: E501
"RobertaForMaskedLM": _HfExamplesInfo("sentence-transformers/all-roberta-large-v1", v0_only=True), # noqa: E501
"XLMRobertaModel": _HfExamplesInfo("intfloat/multilingual-e5-small", v0_only=True), # noqa: E501
"RobertaModel": _HfExamplesInfo("sentence-transformers/stsb-roberta-base-v2"), # noqa: E501
"RobertaForMaskedLM": _HfExamplesInfo("sentence-transformers/all-roberta-large-v1"), # noqa: E501
"XLMRobertaModel": _HfExamplesInfo("intfloat/multilingual-e5-small"), # noqa: E501
# [Multimodal]
"LlavaNextForConditionalGeneration": _HfExamplesInfo("royokong/e5-v"),
"Phi3VForCausalLM": _HfExamplesInfo("TIGER-Lab/VLM2Vec-Full",
@@ -364,20 +364,19 @@ _SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
"GPT2ForSequenceClassification": _HfExamplesInfo("nie3e/sentiment-polish-gpt2-small"), # noqa: E501
# [Cross-encoder]
"BertForSequenceClassification": _HfExamplesInfo("cross-encoder/ms-marco-MiniLM-L-6-v2", v0_only=True), # noqa: E501
"BertForSequenceClassification": _HfExamplesInfo("cross-encoder/ms-marco-MiniLM-L-6-v2"), # noqa: E501
"GteNewForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-multilingual-reranker-base", # noqa: E501
trust_remote_code=True,
hf_overrides={
"architectures": ["GteNewForSequenceClassification"]}),# noqa: E501
"ModernBertForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-reranker-modernbert-base", v0_only=True), # noqa: E501
"RobertaForSequenceClassification": _HfExamplesInfo("cross-encoder/quora-roberta-base", v0_only=True), # noqa: E501
"XLMRobertaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-m3", v0_only=True), # noqa: E501
"ModernBertForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-reranker-modernbert-base"), # noqa: E501
"RobertaForSequenceClassification": _HfExamplesInfo("cross-encoder/quora-roberta-base"), # noqa: E501
"XLMRobertaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-m3"), # noqa: E501
}
_AUTOMATIC_CONVERTED_MODELS = {
# Use as_seq_cls_model for automatic conversion
"GemmaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-gemma", # noqa: E501
v0_only=True,
hf_overrides={"architectures": ["GemmaForSequenceClassification"], # noqa: E501
"classifier_from_token": ["Yes"], # noqa: E501
"method": "no_post_processing"}), # noqa: E501