[Model] Add support for BERT-like Chinese ERNIE pooling models (#36385)

Signed-off-by: whyiug <whyiug@hotmail.com>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
whyiug
2026-03-13 03:23:53 +00:00
committed by GitHub
co-authored by wang.yuqi
parent 10f08dedfa
commit 1ce13cf992
7 changed files with 317 additions and 3 deletions
@@ -18,6 +18,7 @@ from vllm.platforms import current_platform
pytest.mark.slow_test,
],
),
pytest.param("Forrest20231206/ernie-3.0-base-zh-cls"),
],
)
@pytest.mark.parametrize("dtype", ["half"] if current_platform.is_rocm() else ["float"])
@@ -47,5 +48,6 @@ def test_models(
assert torch.allclose(
hf_output,
vllm_output,
atol=1e-3 if dtype == "float" else 1e-2,
rtol=2e-3 if dtype == "float" else 1e-2,
)
@@ -25,11 +25,17 @@ def seed_everything():
yield
@pytest.mark.parametrize("model", ["boltuix/NeuroBERT-NER"])
@pytest.mark.parametrize(
"model",
[
"boltuix/NeuroBERT-NER",
"gyr66/Ernie-3.0-base-chinese-finetuned-ner",
],
)
# The float32 is required for this tiny model to pass the test.
@pytest.mark.parametrize("dtype", ["float"])
@torch.inference_mode
def test_bert_models(
def test_bert_like_models(
hf_runner,
vllm_runner,
example_prompts,
@@ -0,0 +1,45 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from tests.models.language.pooling.embed_utils import correctness_test_embed_models
from tests.models.utils import EmbedModelInfo
from .mteb_embed_utils import mteb_test_embed_models
MODELS = [
EmbedModelInfo(
"shibing624/text2vec-base-chinese-sentence",
architecture="ErnieModel",
mteb_score=0.536523112,
seq_pooling_type="MEAN",
attn_type="encoder_only",
is_prefix_caching_supported=False,
is_chunked_prefill_supported=False,
enable_test=True,
),
]
@pytest.mark.parametrize("model_info", MODELS)
def test_embed_models_mteb(hf_runner, vllm_runner, model_info: EmbedModelInfo) -> None:
mteb_test_embed_models(
hf_runner,
vllm_runner,
model_info,
vllm_extra_kwargs={"gpu_memory_utilization": 0.2},
)
@pytest.mark.parametrize("model_info", MODELS)
def test_embed_models_correctness(
hf_runner, vllm_runner, model_info: EmbedModelInfo, example_prompts
) -> None:
correctness_test_embed_models(
hf_runner,
vllm_runner,
model_info,
example_prompts,
vllm_extra_kwargs={"gpu_memory_utilization": 0.2},
)
+7
View File
@@ -552,6 +552,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
_EMBEDDING_EXAMPLE_MODELS = {
# [Text-only]
"BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5"),
"ErnieModel": _HfExamplesInfo("shibing624/text2vec-base-chinese-sentence"),
"BertSpladeSparseEmbeddingModel": _HfExamplesInfo(
"naver/splade-v3",
hf_overrides={"architectures": ["BertSpladeSparseEmbeddingModel"]},
@@ -666,6 +667,9 @@ _REWARD_EXAMPLE_MODELS = {
_TOKEN_CLASSIFICATION_EXAMPLE_MODELS = {
"BertForTokenClassification": _HfExamplesInfo("boltuix/NeuroBERT-NER"),
"ErnieForTokenClassification": _HfExamplesInfo(
"gyr66/Ernie-3.0-base-chinese-finetuned-ner"
),
"ModernBertForTokenClassification": _HfExamplesInfo(
"disham993/electrical-ner-ModernBERT-base"
),
@@ -675,6 +679,9 @@ _SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
"BertForSequenceClassification": _HfExamplesInfo(
"cross-encoder/ms-marco-MiniLM-L-6-v2"
),
"ErnieForSequenceClassification": _HfExamplesInfo(
"Forrest20231206/ernie-3.0-base-zh-cls",
),
"GPT2ForSequenceClassification": _HfExamplesInfo(
"nie3e/sentiment-polish-gpt2-small"
),