test_models.py 4.64 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
Woosuk Kwon's avatar
Woosuk Kwon committed
2
3
"""Compare the outputs of HF and vLLM when using greedy sampling.

4
Run `pytest tests/models/test_models.py`.
Woosuk Kwon's avatar
Woosuk Kwon committed
5
"""
6

Woosuk Kwon's avatar
Woosuk Kwon committed
7
import pytest
8
9
10
import torch

from vllm.platforms import current_platform
Woosuk Kwon's avatar
Woosuk Kwon committed
11

12
from ...utils import check_logprobs_close
13

14
15
16
17
18
19
# These have unsupported head_dim for FA. We do not
# not have a clean way to fall back, so we fail with
# a clear msg when it happens.
# https://github.com/vllm-project/vllm/issues/14524
REQUIRES_V0 = ["microsoft/phi-2", "stabilityai/stablelm-3b-4e1t"]

20
21
22
23
24
25
26
27
28
29
30
31
32
# This list contains the model that are using AITER kernel.
# Skip model that are not using AITER tests.
# When more AITER kernels are added, this list will not be
# needed as all the models will be calling AITER kernels
# in parts of the operators
AITER_MODEL_LIST = [
    "meta-llama/Llama-3.2-1B-Instruct",
    "openbmb/MiniCPM3-4B",
    "Qwen/Qwen-7B",
    "Qwen/Qwen2.5-0.5B-Instruct",
    "ehristoforu/Falcon3-MoE-2x7B-Insruct",
]

Woosuk Kwon's avatar
Woosuk Kwon committed
33

34
# @maybe_test_rocm_aiter
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
@pytest.mark.parametrize(
    "model",
    [
        pytest.param(
            "bigscience/bloom-560m",  # bloom - testing alibi slopes
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(
            "openai-community/gpt2",  # gpt2
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param("Milos/slovak-gpt-j-405M"),  # gptj
        pytest.param("bigcode/tiny_starcoder_py"),  # gpt_bigcode
        pytest.param("EleutherAI/pythia-70m"),  # gpt_neox
        pytest.param(
            "google/gemma-1.1-2b-it",  # gemma
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
53
        pytest.param(
54
            "THUDM/chatglm3-6b",  # chatglm (text-only)
55
        ),
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
        pytest.param(
            "meta-llama/Llama-3.2-1B-Instruct",  # llama
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(
            "openbmb/MiniCPM3-4B",
            # fused_moe not supported on CPU
            marks=[pytest.mark.core_model],
        ),
        pytest.param(
            "facebook/opt-125m",  # opt
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(
            "microsoft/phi-2",  # phi
            marks=[pytest.mark.core_model],
        ),
73
74
75
        pytest.param(
            "Qwen/Qwen-7B",  # qwen (text-only)
        ),
76
77
78
79
80
81
        pytest.param(
            "Qwen/Qwen2.5-0.5B-Instruct",  # qwen2
            marks=[pytest.mark.core_model],
        ),
        pytest.param("stabilityai/stablelm-3b-4e1t"),  # stablelm
        pytest.param("bigcode/starcoder2-3b"),  # starcoder2
82
83
84
85
        pytest.param(
            "ehristoforu/Falcon3-MoE-2x7B-Insruct",  # mixtral
            marks=[pytest.mark.cpu_model],
        )
86
87
    ])
@pytest.mark.parametrize("dtype", ["half"])
88
89
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [5])
90
91
92
93
94
95
@pytest.mark.parametrize(
    "use_rocm_aiter", [True, False] if current_platform.is_rocm() else [False])
def test_models(hf_runner, vllm_runner, example_prompts, model: str,
                dtype: str, max_tokens: int, num_logprobs: int,
                use_rocm_aiter: bool, monkeypatch) -> None:

96
97
    if model in REQUIRES_V0:
        monkeypatch.setenv("VLLM_USE_V1", "0")
98

99
100
101
102
103
104
105
106
107
    if use_rocm_aiter and (model in AITER_MODEL_LIST):
        monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
    elif use_rocm_aiter and model not in AITER_MODEL_LIST:
        # Skip model that are not using AITER tests.
        # When more AITER kernels are added, this list will not be
        # needed as all the models will be calling AITER kernels
        # in parts of the operators
        pytest.skip(f"Skipping '{model}' model test with AITER kernel.")

108
    with hf_runner(model, dtype=dtype) as hf_model:
109
110
111
112
        if model.startswith("THUDM/chatglm3"):
            hf_model.model.get_output_embeddings = lambda: \
                hf_model.model.transformer.output_layer

113
114
        hf_outputs = hf_model.generate_greedy_logprobs_limit(
            example_prompts, max_tokens, num_logprobs)
Woosuk Kwon's avatar
Woosuk Kwon committed
115

116
    with vllm_runner(model, dtype=dtype) as vllm_model:
117
118
        vllm_outputs = vllm_model.generate_greedy_logprobs(
            example_prompts, max_tokens, num_logprobs)
119

120
    check_logprobs_close(
121
122
123
124
125
        outputs_0_lst=hf_outputs,
        outputs_1_lst=vllm_outputs,
        name_0="hf",
        name_1="vllm",
    )
126
127
128
129
130
131
132
    if use_rocm_aiter:
        # this is to ensure that vllm engine
        # has deallocated the memory before running the next
        # unit tests. On ROCm, when using AITER
        # the memory might not be deallocated completely
        # before running the next test case
        torch.cuda.synchronize()