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

[Model] Re-add the implicit conversion feature for as_seq_cls_model (#21103)

Browse Source 
Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-07-18 15:15:07 +08:00
committed by GitHub
parent ba2dfbb0c2
commit ca4eb82bcb
11 changed files with 165 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
tests/models/registry.py
View File

@ -265,7 +265,6 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"Qwen2MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen1.5-MoE-A2.7B-Chat"),
"Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
"Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
"Qwen3ForSequenceClassification": _HfExamplesInfo("tomaarsen/Qwen3-Reranker-0.6B-seq-cls"), # noqa: E501
"RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
"StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b"), # noqa: E501
"StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
@ -292,7 +291,6 @@ _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
"GPT2ForSequenceClassification": _HfExamplesInfo("nie3e/sentiment-polish-gpt2-small"), # noqa: E501
"GritLM": _HfExamplesInfo("parasail-ai/GritLM-7B-vllm"),
"GteModel": _HfExamplesInfo("Snowflake/snowflake-arctic-embed-m-v2.0",
trust_remote_code=True),
@ -311,7 +309,6 @@ _EMBEDDING_EXAMPLE_MODELS = {
"Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
"Qwen2ForRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-RM-72B"),
"Qwen2ForProcessRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-PRM-7B"),
"Qwen2ForSequenceClassification": _HfExamplesInfo("jason9693/Qwen2.5-1.5B-apeach"), # 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
@ -324,20 +321,29 @@ _EMBEDDING_EXAMPLE_MODELS = {
is_available_online=False), # noqa: E501
}
_CROSS_ENCODER_EXAMPLE_MODELS = {
# [Text-only]
_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
# [Decoder-only]
"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
"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
"LlamaForSequenceClassification": _HfExamplesInfo("Skywork/Skywork-Reward-V2-Llama-3.2-1B"), # 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
}
_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
"LlamaForSequenceClassification": _HfExamplesInfo("Skywork/Skywork-Reward-V2-Llama-3.2-1B"), # noqa: E501
"Qwen2ForSequenceClassification": _HfExamplesInfo("jason9693/Qwen2.5-1.5B-apeach"), # noqa: E501
"Qwen3ForSequenceClassification": _HfExamplesInfo("tomaarsen/Qwen3-Reranker-0.6B-seq-cls"), # noqa: E501
}
_MULTIMODAL_EXAMPLE_MODELS = {
# [Decoder-only]
"AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
@ -449,6 +455,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"), # noqa: E501
}
_SPECULATIVE_DECODING_EXAMPLE_MODELS = {
"EAGLEModel": _HfExamplesInfo("JackFram/llama-68m",
speculative_model="abhigoyal/vllm-eagle-llama-68m-random"), # noqa: E501
@ -489,7 +496,7 @@ _TRANSFORMERS_MODELS = {
_EXAMPLE_MODELS = {
**_TEXT_GENERATION_EXAMPLE_MODELS,
**_EMBEDDING_EXAMPLE_MODELS,
**_CROSS_ENCODER_EXAMPLE_MODELS,
**_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS,
**_MULTIMODAL_EXAMPLE_MODELS,
**_SPECULATIVE_DECODING_EXAMPLE_MODELS,
**_TRANSFORMERS_MODELS,
@ -522,3 +529,4 @@ class HfExampleModels:
HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)
AUTO_EXAMPLE_MODELS = HfExampleModels(_AUTOMATIC_CONVERTED_MODELS)

29
tests/models/test_initialization.py
View File

@ -13,20 +13,21 @@ from vllm.v1.core.kv_cache_utils import get_kv_cache_config
from vllm.v1.engine.core import EngineCore as V1EngineCore
from ..utils import create_new_process_for_each_test
from .registry import HF_EXAMPLE_MODELS
from .registry import AUTO_EXAMPLE_MODELS, HF_EXAMPLE_MODELS, HfExampleModels
@pytest.mark.parametrize("model_arch", HF_EXAMPLE_MODELS.get_supported_archs())
@create_new_process_for_each_test()
def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
"""The reason for using create_new_process_for_each_test is to avoid
the WARNING:
"We must use the 'spawn' multiprocessing start method. Overriding
def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
EXAMPLE_MODELS: HfExampleModels):
"""The reason for using create_new_process_for_each_test is to avoid
the WARNING:
"We must use the 'spawn' multiprocessing start method. Overriding
VLLM_WORKER_MULTIPROC_METHOD to 'spawn'."
The spawn process causes the _initialize_kv_caches_v1 function below to
The spawn process causes the _initialize_kv_caches_v1 function below to
become ineffective.
"""
model_info = HF_EXAMPLE_MODELS.get_hf_info(model_arch)
model_info = EXAMPLE_MODELS.get_hf_info(model_arch)
model_info.check_available_online(on_fail="skip")
model_info.check_transformers_version(on_fail="skip")
@ -127,3 +128,15 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
load_format="dummy",
hf_overrides=hf_overrides,
)
@pytest.mark.parametrize("model_arch", HF_EXAMPLE_MODELS.get_supported_archs())
def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
can_initialize(model_arch, monkeypatch, HF_EXAMPLE_MODELS)
@pytest.mark.parametrize("model_arch",
AUTO_EXAMPLE_MODELS.get_supported_archs())
def test_implicit_converted_models(model_arch: str,
monkeypatch: pytest.MonkeyPatch):
can_initialize(model_arch, monkeypatch, AUTO_EXAMPLE_MODELS)

35
tests/models/test_transformers.py
View File

@ -138,3 +138,38 @@ def test_quantization(
name_0="transformers",
name_1="vllm",
)
@pytest.mark.parametrize(
"model",
["jason9693/Qwen2.5-1.5B-apeach"],
)
@pytest.mark.parametrize("dtype", ["half"])
def test_classify(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
monkeypatch,
) -> None:
import torch
from transformers import AutoModelForSequenceClassification
with vllm_runner(model,
max_model_len=512,
dtype=dtype,
model_impl="transformers") as vllm_model:
vllm_outputs = vllm_model.classify(example_prompts)
with hf_runner(model,
dtype=dtype,
auto_cls=AutoModelForSequenceClassification) as hf_model:
hf_outputs = hf_model.classify(example_prompts)
for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
hf_output = torch.tensor(hf_output)
vllm_output = torch.tensor(vllm_output)
assert torch.allclose(hf_output, vllm_output,
1e-3 if dtype == "float" else 1e-2)