"tests/vscode:/vscode.git/clone" did not exist on "5989f4684d62d5cb1852624ce0fd04fc08dd239b"
test_token_classification.py 4.38 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
import random

zhuwenwen's avatar
zhuwenwen committed
5
import os
6
import numpy as np
7
8
9
10
11
import pytest
import torch
from transformers import AutoModelForTokenClassification

from tests.models.utils import softmax
12
from vllm.platforms import current_platform
zhuwenwen's avatar
zhuwenwen committed
13
from ....utils import models_path_prefix
14
15


zhuwenwen's avatar
zhuwenwen committed
16
@pytest.mark.parametrize("model", [os.path.join(models_path_prefix, "boltuix/NeuroBERT-NER")])
17
18
19
# The float32 is required for this tiny model to pass the test.
@pytest.mark.parametrize("dtype", ["float"])
@torch.inference_mode
20
21
22
23
24
25
26
27
def test_bert_models(
    hf_runner,
    vllm_runner,
    example_prompts,
    model: str,
    dtype: str,
) -> None:
    with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
28
        vllm_outputs = vllm_model.token_classify(example_prompts)
29

30
31
32
33
34
35
    # Use eager attention on ROCm to avoid HF Transformers flash attention
    # accuracy issues: https://github.com/vllm-project/vllm/issues/30167
    hf_model_kwargs = {}
    if current_platform.is_rocm():
        hf_model_kwargs["attn_implementation"] = "eager"

36
    with hf_runner(
37
38
39
40
        model,
        dtype=dtype,
        auto_cls=AutoModelForTokenClassification,
        model_kwargs=hf_model_kwargs,
41
42
43
44
45
46
47
48
49
50
51
    ) as hf_model:
        tokenizer = hf_model.tokenizer
        hf_outputs = []
        for prompt in example_prompts:
            inputs = tokenizer([prompt], return_tensors="pt")
            inputs = hf_model.wrap_device(inputs)
            output = hf_model.model(**inputs)
            hf_outputs.append(softmax(output.logits[0]))

    # check logits difference
    for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
52
53
        hf_output = hf_output.detach().clone().cpu().float()
        vllm_output = vllm_output.detach().clone().cpu().float()
54
        torch.testing.assert_close(hf_output, vllm_output, atol=3.2e-2, rtol=1e-3)
55
56
57
58
59
60


@pytest.mark.parametrize("model", ["disham993/electrical-ner-ModernBERT-base"])
@pytest.mark.parametrize("dtype", ["float"])
@torch.inference_mode
def test_modernbert_models(
61
62
63
64
65
66
    hf_runner,
    vllm_runner,
    example_prompts,
    model: str,
    dtype: str,
) -> None:
67

68
    with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
69
        vllm_outputs = vllm_model.token_classify(example_prompts)
70

71
72
73
74
75
76
    # Use eager attention on ROCm to avoid HF Transformers flash attention
    # accuracy issues: https://github.com/vllm-project/vllm/issues/30167
    hf_model_kwargs = {}
    if current_platform.is_rocm():
        hf_model_kwargs["attn_implementation"] = "eager"

77
    with hf_runner(
78
79
80
81
        model,
        dtype=dtype,
        auto_cls=AutoModelForTokenClassification,
        model_kwargs=hf_model_kwargs,
82
    ) as hf_model:
83
84
85
86
87
88
89
90
91
92
        tokenizer = hf_model.tokenizer
        hf_outputs = []
        for prompt in example_prompts:
            inputs = tokenizer([prompt], return_tensors="pt")
            inputs = hf_model.wrap_device(inputs)
            output = hf_model.model(**inputs)
            hf_outputs.append(softmax(output.logits[0]))

    # check logits difference
    for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
93
94
        hf_output = hf_output.detach().clone().cpu().float()
        vllm_output = vllm_output.detach().clone().cpu().float()
95
        torch.testing.assert_close(hf_output, vllm_output, atol=3.2e-2, rtol=1e-3)
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123


@pytest.mark.parametrize("model", ["bd2lcco/Qwen3-0.6B-finetuned"])
@pytest.mark.parametrize("dtype", ["float"])
@torch.inference_mode
def test_auto_conversion(
    hf_runner,
    vllm_runner,
    example_prompts,
    model: str,
    dtype: str,
) -> None:
    with vllm_runner(model, max_model_len=1024, dtype=dtype) as vllm_model:
        vllm_outputs = vllm_model.token_classify(example_prompts)

    with hf_runner(
        model, dtype=dtype, auto_cls=AutoModelForTokenClassification
    ) as hf_model:
        tokenizer = hf_model.tokenizer
        hf_outputs = []
        for prompt in example_prompts:
            inputs = tokenizer([prompt], return_tensors="pt")
            inputs = hf_model.wrap_device(inputs)
            output = hf_model.model(**inputs)
            hf_outputs.append(softmax(output.logits[0]))

    # check logits difference
    for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
124
125
        hf_output = hf_output.detach().clone().cpu().float()
        vllm_output = vllm_output.detach().clone().cpu().float()
126
        assert torch.allclose(hf_output, vllm_output, atol=1e-2)