test_classification.py 2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

zhuwenwen's avatar
zhuwenwen committed
4
import os
5
6
7
import pytest
import torch
from transformers import AutoModelForSequenceClassification
zhuwenwen's avatar
zhuwenwen committed
8
from ....utils import models_path_prefix
9

10
11
from vllm.platforms import current_platform

12
13
14
15
16
17
18
19
# TODO: enable when float32 is supported by V1
# @pytest.fixture(autouse=True)
# def v1(run_with_both_engines):
#     # Simple autouse wrapper to run both engines for each test
#     # This can be promoted up to conftest.py to run for every
#     # test in a package
#     pass

20

21
22
23
@pytest.mark.parametrize(
    "model",
    [
zhuwenwen's avatar
zhuwenwen committed
24
        pytest.param(os.path.join(models_path_prefix, "jason9693/Qwen2.5-1.5B-apeach"),
25
26
27
                     marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
    ],
)
28
29
@pytest.mark.parametrize("dtype",
                         ["half"] if current_platform.is_rocm() else ["float"])
30
def test_models(
31
32
33
34
35
    hf_runner,
    vllm_runner,
    example_prompts,
    model: str,
    dtype: str,
36
    monkeypatch,
37
) -> None:
38
39
40
41
42
    if current_platform.is_rocm():
        # ROCm Triton FA does not currently support sliding window attention
        # switch to use ROCm CK FA backend
        monkeypatch.setenv("VLLM_USE_TRITON_FLASH_ATTN", "False")

43
    with vllm_runner(model, max_model_len=512, dtype=dtype) as vllm_model:
44
        vllm_outputs = vllm_model.classify(example_prompts)
45

46
47
48
49
50
51
52
53
54
55
    with hf_runner(model,
                   dtype=dtype,
                   auto_cls=AutoModelForSequenceClassification) as hf_model:
        hf_outputs = hf_model.classify(example_prompts)

    # check logits difference
    for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
        hf_output = torch.tensor(hf_output)
        vllm_output = torch.tensor(vllm_output)

56
57
        # the tolerance value of 1e-2 is selected based on the
        # half datatype tests in
58
        # tests/models/language/pooling/test_embedding.py
59
60
        assert torch.allclose(hf_output, vllm_output,
                              1e-3 if dtype == "float" else 1e-2)