Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_stop_strings.py 3.61 KB
Newer Older
Nick Hill's avatar
[BugFix] Fix handling of stop strings and stop token ids (#3672)  
Nick Hill committed
1
2
3
4
5
6
7
8
9
10
11
12 
from typing import Any, List, Optional

import pytest

from vllm import CompletionOutput, LLMEngine, SamplingParams

MODEL = "meta-llama/llama-2-7b-hf"
MAX_TOKENS = 200


@pytest.fixture(scope="session")
def vllm_model(vllm_runner):
youkaichao's avatar
[CI/Test] improve robustness of test (vllm_runner) (#5357)  
youkaichao committed
13
14 
    with vllm_runner(MODEL) as vllm_model:
        yield vllm_model
Nick Hill's avatar
[BugFix] Fix handling of stop strings and stop token ids (#3672)  
Nick Hill committed
15
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
70
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
108
109
110
111
112 


@pytest.mark.skip_global_cleanup
def test_stop_basic(vllm_model):
    _test_stopping(vllm_model.model.llm_engine,
                   stop=["."],
                   include_in_output=False,
                   expected_output="VLLM is a 100% volunteer organization",
                   expected_reason=".")

    _test_stopping(vllm_model.model.llm_engine,
                   stop=["."],
                   include_in_output=True,
                   expected_output="VLLM is a 100% volunteer organization.",
                   expected_reason=".")


@pytest.mark.skip_global_cleanup
def test_stop_multi_tokens(vllm_model):
    _test_stopping(
        vllm_model.model.llm_engine,
        stop=["group of peo", "short"],
        include_in_output=False,
        expected_output="VLLM is a 100% volunteer organization. We are a ",
        expected_reason="group of peo")

    _test_stopping(
        vllm_model.model.llm_engine,
        stop=["group of peo", "short"],
        include_in_output=True,
        expected_output=
        "VLLM is a 100% volunteer organization. We are a group of peo",
        expected_reason="group of peo")


@pytest.mark.skip_global_cleanup
def test_stop_partial_token(vllm_model):
    _test_stopping(vllm_model.model.llm_engine,
                   stop=["gani"],
                   include_in_output=False,
                   expected_output="VLLM is a 100% volunteer or",
                   expected_reason="gani")

    _test_stopping(vllm_model.model.llm_engine,
                   stop=["gani"],
                   include_in_output=True,
                   expected_output="VLLM is a 100% volunteer organi",
                   expected_reason="gani")


@pytest.mark.skip_global_cleanup
def test_stop_token_id(vllm_model):
    # token id 13013 => " organization"

    _test_stopping(vllm_model.model.llm_engine,
                   stop_token_ids=[13013],
                   include_in_output=False,
                   expected_output="VLLM is a 100% volunteer",
                   expected_reason=13013)

    _test_stopping(vllm_model.model.llm_engine,
                   stop_token_ids=[13013],
                   include_in_output=True,
                   expected_output="VLLM is a 100% volunteer organization",
                   expected_reason=13013)


def _test_stopping(llm_engine: LLMEngine,
                   expected_output: str,
                   expected_reason: Any,
                   stop: Optional[List[str]] = None,
                   stop_token_ids: Optional[List[int]] = None,
                   include_in_output: bool = False) -> None:
    llm_engine.add_request(
        "id", "A story about vLLM:\n",
        SamplingParams(
            temperature=0.0,
            max_tokens=MAX_TOKENS,
            stop=stop,
            stop_token_ids=stop_token_ids,
            include_stop_str_in_output=include_in_output,
        ), None)

    output: Optional[CompletionOutput] = None
    output_text = ""
    stop_reason = None
    while llm_engine.has_unfinished_requests():
        (request_output, ) = llm_engine.step()
        (output, ) = request_output.outputs

        # Ensure we don't backtrack
        assert output.text.startswith(output_text)
        output_text = output.text
        stop_reason = output.stop_reason

    assert output is not None
    assert output_text == expected_output
    assert stop_reason == expected_reason