[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
+6 -2
View File
@@ -118,6 +118,8 @@ class PPTestSettings:
multi_node_only: bool = False,
load_format: Optional[str] = None,
):
vllm_major_versions = ["1"] if runner == "pooling" else ["0"]
return PPTestSettings(
parallel_setups=[
ParallelSetup(tp_size=tp_base,
@@ -126,7 +128,7 @@ class PPTestSettings:
chunked_prefill=False),
],
distributed_backends=["mp"],
vllm_major_versions=["0"],
vllm_major_versions=vllm_major_versions,
runner=runner,
test_options=PPTestOptions(multi_node_only=multi_node_only,
load_format=load_format),
@@ -213,7 +215,9 @@ TEXT_GENERATION_MODELS = {
EMBEDDING_MODELS = { # type: ignore[var-annotated]
# [Text-only]
"intfloat/e5-mistral-7b-instruct": PPTestSettings.fast(runner="pooling"),
"BAAI/bge-multilingual-gemma2": PPTestSettings.fast(runner="pooling"),
# TODO: re-enable when https://github.com/vllm-project/vllm/issues/23883
# is fixed
#"BAAI/bge-multilingual-gemma2": PPTestSettings.fast(runner="pooling"),
"Qwen/Qwen2.5-Math-RM-72B": PPTestSettings.fast(
load_format="dummy", runner="pooling"
),
-8
View File
@@ -16,14 +16,6 @@ MODEL_NAME = "jason9693/Qwen2.5-1.5B-apeach"
prompts = ["The chef prepared a delicious meal."]
@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(scope="module")
def llm():
# pytest caches the fixture so we use weakref.proxy to
-8
View File
@@ -27,14 +27,6 @@ TOKEN_IDS = [
]
@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(scope="module")
def llm():
# pytest caches the fixture so we use weakref.proxy to
-8
View File
@@ -16,14 +16,6 @@ MODEL_NAME = "internlm/internlm2-1_8b-reward"
prompts = ["The chef prepared a delicious meal."]
@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(scope="module")
def llm():
# pytest caches the fixture so we use weakref.proxy to
-8
View File
@@ -14,14 +14,6 @@ from ...models.utils import softmax
MODEL_NAME = "tomaarsen/Qwen3-Reranker-0.6B-seq-cls"
@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(scope="module")
def llm():
# pytest caches the fixture so we use weakref.proxy to
@@ -32,15 +32,16 @@ MODEL_CONFIGS = [
"tensor_parallel_size": 1,
"tokenizer_mode": "mistral",
},
{
"model": "sentence-transformers/all-MiniLM-L12-v2",
"enforce_eager": True,
"gpu_memory_utilization": 0.20,
"max_model_len": 64,
"max_num_batched_tokens": 64,
"max_num_seqs": 64,
"tensor_parallel_size": 1,
},
# TODO: re-enable once these tests are run with V1
# {
# "model": "sentence-transformers/all-MiniLM-L12-v2",
# "enforce_eager": True,
# "gpu_memory_utilization": 0.20,
# "max_model_len": 64,
# "max_num_batched_tokens": 64,
# "max_num_seqs": 64,
# "tensor_parallel_size": 1,
# },
]
@@ -24,14 +24,6 @@ DUMMY_CHAT_TEMPLATE = """{% for message in messages %}{{message['role'] + ': ' +
DTYPE = "bfloat16"
@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(scope="module")
def server():
args = [
-8
View File
@@ -14,14 +14,6 @@ MODEL_NAME = "BAAI/bge-reranker-base"
DTYPE = "bfloat16"
@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(scope="module")
def server():
args = ["--enforce-eager", "--max-model-len", "100", "--dtype", DTYPE]
-9
View File
@@ -12,15 +12,6 @@ from vllm.entrypoints.openai.protocol import ScoreResponse
from ...utils import RemoteOpenAIServer
@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
MODELS = [
{
"name": "BAAI/bge-reranker-v2-m3",
@@ -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
-54
View File
@@ -9,10 +9,7 @@ from vllm.attention import AttentionMetadata, AttentionMetadataBuilder
from vllm.attention.backends.abstract import AttentionBackend
from vllm.attention.backends.utils import CommonAttentionState
from vllm.model_executor import SamplingMetadata
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
from vllm.worker.pooling_model_runner import (
ModelInputForGPUWithPoolingMetadata)
class MockAttentionBackend(AttentionBackend):
@@ -114,54 +111,3 @@ def test_model_runner_input():
assert (received_model_input.sampling_metadata.selected_token_indices ==
sampling_metadata.selected_token_indices)
assert received_model_input.sampling_metadata.seq_groups is None
def test_embedding_model_runner_input():
pooling_metadata = PoolingMetadata(
seq_groups=[[0]],
seq_data={},
prompt_lens=[1],
)
attn_metadata = AttentionMetadata(
num_prefills=1,
num_prefill_tokens=2,
num_decode_tokens=3,
slot_mapping=torch.zeros(1),
multi_modal_placeholder_index_maps=None,
enable_kv_scales_calculation=True,
)
model_input = ModelInputForGPUWithPoolingMetadata(
input_tokens=torch.ones(10),
input_positions=torch.ones(10),
pooling_metadata=pooling_metadata,
attn_metadata=attn_metadata)
assert isinstance(model_input, ModelInputForGPUWithPoolingMetadata)
# Test round trip serialization.
tensor_dict = model_input.as_broadcastable_tensor_dict()
attn_backend = MockAttentionBackend()
received_model_input = (
ModelInputForGPUWithPoolingMetadata.from_broadcasted_tensor_dict(
tensor_dict, attn_backend=attn_backend))
# Check that received copy has correct values.
assert isinstance(received_model_input,
ModelInputForGPUWithPoolingMetadata)
assert received_model_input.input_tokens is not None
assert (
received_model_input.input_tokens == model_input.input_tokens).all()
assert received_model_input.input_positions is not None
assert (received_model_input.input_positions == model_input.input_positions
).all()
assert received_model_input.multi_modal_kwargs is None
assert (received_model_input.multi_modal_kwargs ==
model_input.multi_modal_kwargs)
assert received_model_input.lora_requests is None
assert received_model_input.lora_requests == model_input.lora_requests
assert received_model_input.lora_mapping is None
assert received_model_input.lora_mapping == model_input.lora_mapping
for field in dataclasses.fields(AttentionMetadata):
assert getattr(received_model_input.attn_metadata, field.name,
None) == getattr(attn_metadata, field.name, None)
# Pooling metadata is not broadcast.
assert received_model_input.pooling_metadata is None