test_generate.py 3.35 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
import weakref

6
7
import pytest

8
from vllm import LLM, RequestOutput, SamplingParams
9
from vllm.distributed import cleanup_dist_env_and_memory
10

11
MODEL_NAME = "distilbert/distilgpt2"
12
13
14
15
16
17
18

PROMPTS = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
19

20
21
22
23
24
25
TOKEN_IDS = [
    [0],
    [0, 1],
    [0, 2, 1],
    [0, 3, 1, 2],
]
26

27
28
29
30
31
32

@pytest.fixture(scope="module")
def llm():
    # pytest caches the fixture so we use weakref.proxy to
    # enable garbage collection
    llm = LLM(model=MODEL_NAME,
33
              max_num_batched_tokens=4096,
34
35
36
37
38
39
40
41
42
              tensor_parallel_size=1,
              gpu_memory_utilization=0.10,
              enforce_eager=True)

    with llm.deprecate_legacy_api():
        yield weakref.proxy(llm)

        del llm

43
    cleanup_dist_env_and_memory()
44
45


46
def assert_outputs_equal(o1: list[RequestOutput], o2: list[RequestOutput]):
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
    assert [o.outputs for o in o1] == [o.outputs for o in o2]


@pytest.mark.skip_global_cleanup
@pytest.mark.parametrize('prompt_token_ids', TOKEN_IDS)
def test_v1_v2_api_consistency_single_prompt_tokens(llm: LLM,
                                                    prompt_token_ids):
    sampling_params = SamplingParams(temperature=0.0, top_p=1.0)

    with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
        v1_output = llm.generate(prompt_token_ids=prompt_token_ids,
                                 sampling_params=sampling_params)

    v2_output = llm.generate({"prompt_token_ids": prompt_token_ids},
                             sampling_params=sampling_params)
    assert_outputs_equal(v1_output, v2_output)


@pytest.mark.skip_global_cleanup
def test_v1_v2_api_consistency_multi_prompt_tokens(llm: LLM):
    sampling_params = SamplingParams(temperature=0.0, top_p=1.0)

    with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
        v1_output = llm.generate(prompt_token_ids=TOKEN_IDS,
                                 sampling_params=sampling_params)

    v2_output = llm.generate(
        [{
            "prompt_token_ids": p
        } for p in TOKEN_IDS],
        sampling_params=sampling_params,
    )
    assert_outputs_equal(v1_output, v2_output)
80
81


82
83
@pytest.mark.skip_global_cleanup
def test_multiple_sampling_params(llm: LLM):
84
85
86
87
88
89
90
91
    sampling_params = [
        SamplingParams(temperature=0.01, top_p=0.95),
        SamplingParams(temperature=0.3, top_p=0.95),
        SamplingParams(temperature=0.7, top_p=0.95),
        SamplingParams(temperature=0.99, top_p=0.95),
    ]

    # Multiple SamplingParams should be matched with each prompt
92
93
    outputs = llm.generate(PROMPTS, sampling_params=sampling_params)
    assert len(PROMPTS) == len(outputs)
94
95
96

    # Exception raised, if the size of params does not match the size of prompts
    with pytest.raises(ValueError):
97
        outputs = llm.generate(PROMPTS, sampling_params=sampling_params[:3])
98
99
100

    # Single SamplingParams should be applied to every prompt
    single_sampling_params = SamplingParams(temperature=0.3, top_p=0.95)
101
102
    outputs = llm.generate(PROMPTS, sampling_params=single_sampling_params)
    assert len(PROMPTS) == len(outputs)
103
104

    # sampling_params is None, default params should be applied
105
106
    outputs = llm.generate(PROMPTS, sampling_params=None)
    assert len(PROMPTS) == len(outputs)