Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_pooling_params.py 3.22 KB
Newer Older
wang.yuqi's avatar
[Model] Pooling model activation supports per request control by PoolingParams (#20538)  
wang.yuqi committed
1
2
3
4
5
6
7
8
9
10 
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest

from tests.models.utils import EmbedModelInfo
from vllm import PoolingParams
from vllm.config import ModelConfig

EMBEDDING_MODELS = [
    EmbedModelInfo("intfloat/multilingual-e5-small", is_matryoshka=False),
Harry Mellor's avatar
Convert formatting to use `ruff` instead of `yapf` + `isort` (#26247)  
Harry Mellor committed
11
12
13
14
15 
    EmbedModelInfo(
        "Snowflake/snowflake-arctic-embed-m-v1.5",
        is_matryoshka=True,
        matryoshka_dimensions=[256],
    ),
wang.yuqi's avatar
[Model] Pooling model activation supports per request control by PoolingParams (#20538)  
wang.yuqi committed
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69 
]


def test_task():
    pooling_params = PoolingParams()
    pooling_params.verify(task="score")

    pooling_params = PoolingParams(task="score")
    pooling_params.verify(task="score")

    with pytest.raises(ValueError):
        pooling_params.verify(task="encode")


def test_embed():
    task = "embed"
    pooling_params = PoolingParams(normalize=None)
    pooling_params.verify(task=task)

    pooling_params = PoolingParams(normalize=True)
    pooling_params.verify(task=task)

    pooling_params = PoolingParams(normalize=False)
    pooling_params.verify(task=task)

    invalid_parameters = ["activation", "softmax"]
    for p in invalid_parameters:
        with pytest.raises(ValueError):
            pooling_params = PoolingParams(**{p: True})
            pooling_params.verify(task=task)


@pytest.mark.parametrize("model_info", EMBEDDING_MODELS)
def test_embed_dimensions(model_info: EmbedModelInfo):
    task = "embed"
    model_config = ModelConfig(
        model_info.name,
        task="auto",
        tokenizer=model_info.name,
        tokenizer_mode="auto",
        trust_remote_code=False,
        seed=0,
        dtype="float16",
    )

    pooling_params = PoolingParams(dimensions=None)
    pooling_params.verify(task=task, model_config=model_config)

    with pytest.raises(ValueError):
        pooling_params = PoolingParams(dimensions=1)
        pooling_params.verify(task=task, model_config=model_config)

    if model_info.is_matryoshka:
        assert model_info.matryoshka_dimensions is not None
Harry Mellor's avatar
Convert formatting to use `ruff` instead of `yapf` + `isort` (#26247)  
Harry Mellor committed
70 
        pooling_params = PoolingParams(dimensions=model_info.matryoshka_dimensions[0])
wang.yuqi's avatar
[Model] Pooling model activation supports per request control by PoolingParams (#20538)  
wang.yuqi committed
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107 
        pooling_params.verify(task=task, model_config=model_config)


@pytest.mark.parametrize("task", ["score", "classify"])
def test_classify(task):
    pooling_params = PoolingParams(activation=None)
    pooling_params.verify(task=task)

    pooling_params = PoolingParams(activation=True)
    pooling_params.verify(task=task)

    pooling_params = PoolingParams(activation=False)
    pooling_params.verify(task=task)

    invalid_parameters = ["dimensions", "normalize", "softmax"]
    for p in invalid_parameters:
        with pytest.raises(ValueError):
            pooling_params = PoolingParams(**{p: True})
            pooling_params.verify(task=task)


def test_encode():
    task = "encode"
    pooling_params = PoolingParams(softmax=None)
    pooling_params.verify(task=task)

    pooling_params = PoolingParams(softmax=True)
    pooling_params.verify(task=task)

    pooling_params = PoolingParams(softmax=False)
    pooling_params.verify(task=task)

    invalid_parameters = ["dimensions", "normalize", "activation"]
    for p in invalid_parameters:
        with pytest.raises(ValueError):
            pooling_params = PoolingParams(**{p: True})
            pooling_params.verify(task=task)