test_lazy_outlines.py 2.94 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import sys
youkaichao's avatar
youkaichao committed
5
from contextlib import nullcontext
6

youkaichao's avatar
youkaichao committed
7
from vllm_test_utils import BlameResult, blame
youkaichao's avatar
youkaichao committed
8

9
from vllm import LLM, SamplingParams
10
from vllm.distributed import cleanup_dist_env_and_memory
11
from vllm.sampling_params import GuidedDecodingParams
12

youkaichao's avatar
youkaichao committed
13

14
15
16
17
18
19
20
21
22
23
def run_normal():
    prompts = [
        "Hello, my name is",
        "The president of the United States is",
        "The capital of France is",
        "The future of AI is",
    ]
    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

    # Create an LLM without guided decoding as a baseline.
24
    llm = LLM(model="distilbert/distilgpt2",
25
26
27
28
29
30
31
32
33
34
35
36
37
              enforce_eager=True,
              gpu_memory_utilization=0.3)
    outputs = llm.generate(prompts, sampling_params)
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

    # Destroy the LLM object and free up the GPU memory.
    del llm
    cleanup_dist_env_and_memory()


38
def run_xgrammar(sample_regex):
39
    # Create an LLM with guided decoding enabled.
40
    llm = LLM(model="distilbert/distilgpt2",
41
              enforce_eager=True,
42
              guided_decoding_backend="xgrammar",
43
              gpu_memory_utilization=0.3)
44
45
46
47
48
    prompt = f"Give an example IPv4 address with this regex: {sample_regex}"
    guided_decoding = GuidedDecodingParams(regex=sample_regex)
    sampling_params = SamplingParams(temperature=0.8,
                                     top_p=0.95,
                                     guided_decoding=guided_decoding)
49
    outputs = llm.generate(
50
        prompts=[prompt] * 2,
51
52
        sampling_params=sampling_params,
        use_tqdm=True,
53
    )
54
55
56
57
58
59

    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

youkaichao's avatar
youkaichao committed
60
61
62
63

def test_lazy_outlines(sample_regex):
    """If users don't use guided decoding, outlines should not be imported.
    """
64
    # make sure outlines is not imported
youkaichao's avatar
youkaichao committed
65
    module_name = "outlines"
youkaichao's avatar
youkaichao committed
66
67
68
69
70
71
72
73
74
    # In CI, we only check finally if the module is imported.
    # If it is indeed imported, we can rerun the test with `use_blame=True`,
    # which will trace every function call to find the first import location,
    # and help find the root cause.
    # We don't run it in CI by default because it is slow.
    use_blame = False
    context = blame(
        lambda: module_name in sys.modules) if use_blame else nullcontext()
    with context as result:
youkaichao's avatar
youkaichao committed
75
        run_normal()
76
        run_xgrammar(sample_regex)
youkaichao's avatar
youkaichao committed
77
78
79
80
81
82
    if use_blame:
        assert isinstance(result, BlameResult)
        print(f"the first import location is:\n{result.trace_stack}")
    assert module_name not in sys.modules, (
        f"Module {module_name} is imported. To see the first"
        f" import location, run the test with `use_blame=True`.")