A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI] Skip the pooling models that do not support transformers v4.55 (#22411)
Signed-off-by: wang.yuqi <noooop@126.com>
wang.yuqi committed
Aug 7, 2025 at 14:05 UTC
2a4c825523d5715068bf3ec373f662e113c66f45
4 files changed
+28
-1
tests/models/language/pooling/test_embedding.py
+4
-1
@@ -7,7 +7,7 @@ import pytest
7
from vllm.config import PoolerConfig
8
from vllm.platforms import current_platform
9
10
-from ...utils import check_embeddings_close
10
+from ...utils import check_embeddings_close, check_transformers_version
11
12
13
@pytest.fixture(autouse=True)
@@ -56,6 +56,9 @@ def test_models(
56
model,
57
monkeypatch,
58
) -> None:
59
+ if model == "Alibaba-NLP/gte-Qwen2-1.5B-instruct":
60
+ check_transformers_version(model, max_transformers_version="4.53.2")
61
+
62
if model == "BAAI/bge-multilingual-gemma2" and current_platform.is_rocm():
63
# ROCm Triton FA does not currently support sliding window attention
64
# switch to use ROCm CK FA backend
tests/models/language/pooling/test_gte.py
+9
@@ -4,6 +4,7 @@ from typing import Any
4
5
import pytest
6
7
+from ...utils import check_transformers_version
8
from .embed_utils import EmbedModelInfo, correctness_test_embed_models
9
from .mteb_utils import mteb_test_embed_models
10
@@ -60,6 +61,10 @@ MODELS = [
61
@pytest.mark.parametrize("model_info", MODELS)
62
def test_embed_models_mteb(hf_runner, vllm_runner,
63
model_info: EmbedModelInfo) -> None:
64
+ if model_info.name == "Alibaba-NLP/gte-Qwen2-1.5B-instruct":
65
+ check_transformers_version(model_info.name,
66
+ max_transformers_version="4.53.2")
67
+
68
vllm_extra_kwargs: dict[str, Any] = {}
69
if model_info.architecture == "GteNewModel":
70
vllm_extra_kwargs["hf_overrides"] = {"architectures": ["GteNewModel"]}
@@ -72,6 +77,10 @@ def test_embed_models_mteb(hf_runner, vllm_runner,
77
def test_embed_models_correctness(hf_runner, vllm_runner,
78
model_info: EmbedModelInfo,
79
example_prompts) -> None:
80
+ if model_info.name == "Alibaba-NLP/gte-Qwen2-1.5B-instruct":
81
+ check_transformers_version(model_info.name,
82
+ max_transformers_version="4.53.2")
83
+
84
vllm_extra_kwargs: dict[str, Any] = {}
85
if model_info.architecture == "GteNewModel":
86
vllm_extra_kwargs["hf_overrides"] = {"architectures": ["GteNewModel"]}
tests/models/language/pooling/test_reward.py
+4
@@ -10,6 +10,7 @@ from transformers import AutoModel
10
from vllm.platforms import current_platform
11
12
from ....conftest import HfRunner
13
+from ...utils import check_transformers_version
14
15
16
@pytest.fixture(autouse=True)
@@ -86,6 +87,9 @@ def test_prm_models(
87
dtype: str,
88
monkeypatch,
89
) -> None:
90
+ check_transformers_version("Qwen/Qwen2.5-Math-PRM-7B",
91
+ max_transformers_version="4.53.2")
92
+
93
if current_platform.is_cpu() and os.environ.get("VLLM_USE_V1", "0") == "0":
94
pytest.skip("CPU only supports V1")
95
tests/models/utils.py
+11
@@ -412,3 +412,14 @@ def dummy_hf_overrides(
412
})
413
414
return hf_config
415
+
416
+
417
+def check_transformers_version(model: str,
418
+ min_transformers_version: Optional[str] = None,
419
+ max_transformers_version: Optional[str] = None):
420
+ from .registry import _HfExamplesInfo
421
+
422
+ return _HfExamplesInfo(model,
423
+ min_transformers_version=min_transformers_version,
424
+ max_transformers_version=max_transformers_version
425
+ ).check_transformers_version(on_fail="skip")