vllm_infer.py 6.44 KB
Newer Older
chenych's avatar
chenych committed
1
# Copyright 2025 the LlamaFactory team.
luopl's avatar
luopl committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import json
chenych's avatar
chenych committed
16
from typing import Optional
luopl's avatar
luopl committed
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

import fire
from transformers import Seq2SeqTrainingArguments

from llamafactory.data import get_dataset, get_template_and_fix_tokenizer
from llamafactory.extras.constants import IGNORE_INDEX
from llamafactory.extras.misc import check_version, get_device_count
from llamafactory.extras.packages import is_vllm_available
from llamafactory.hparams import get_infer_args
from llamafactory.model import load_tokenizer


if is_vllm_available():
    from vllm import LLM, SamplingParams
    from vllm.lora.request import LoRARequest


def vllm_infer(
    model_name_or_path: str,
    adapter_name_or_path: str = None,
    dataset: str = "alpaca_en_demo",
    dataset_dir: str = "data",
    template: str = "default",
    cutoff_len: int = 2048,
chenych's avatar
chenych committed
41
    max_samples: Optional[int] = None,
luopl's avatar
luopl committed
42
43
44
45
46
47
48
    vllm_config: str = "{}",
    save_name: str = "generated_predictions.jsonl",
    temperature: float = 0.95,
    top_p: float = 0.7,
    top_k: int = 50,
    max_new_tokens: int = 1024,
    repetition_penalty: float = 1.0,
chenych's avatar
chenych committed
49
    skip_special_tokens: bool = True,
chenych's avatar
chenych committed
50
    seed: Optional[int] = None,
luopl's avatar
luopl committed
51
    pipeline_parallel_size: int = 1,
chenych's avatar
chenych committed
52
53
    image_max_pixels: int = 768 * 768,
    image_min_pixels: int = 32 * 32,
luopl's avatar
luopl committed
54
):
chenych's avatar
chenych committed
55
56
    r"""Perform batch generation using vLLM engine, which supports tensor parallelism.

luopl's avatar
luopl committed
57
58
    Usage: python vllm_infer.py --model_name_or_path meta-llama/Llama-2-7b-hf --template llama --dataset alpaca_en_demo
    """
chenych's avatar
chenych committed
59
    check_version("vllm>=0.4.3,<=0.8.2")
luopl's avatar
luopl committed
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
    if pipeline_parallel_size > get_device_count():
        raise ValueError("Pipeline parallel size should be smaller than the number of gpus.")

    model_args, data_args, _, generating_args = get_infer_args(
        dict(
            model_name_or_path=model_name_or_path,
            adapter_name_or_path=adapter_name_or_path,
            dataset=dataset,
            dataset_dir=dataset_dir,
            template=template,
            cutoff_len=cutoff_len,
            max_samples=max_samples,
            preprocessing_num_workers=16,
            vllm_config=vllm_config,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            max_new_tokens=max_new_tokens,
            repetition_penalty=repetition_penalty,
        )
    )

    training_args = Seq2SeqTrainingArguments(output_dir="dummy_dir")
    tokenizer_module = load_tokenizer(model_args)
    tokenizer = tokenizer_module["tokenizer"]
    template_obj = get_template_and_fix_tokenizer(tokenizer, data_args)
    template_obj.mm_plugin.expand_mm_tokens = False  # for vllm generate
    dataset_module = get_dataset(template_obj, model_args, data_args, training_args, "ppo", **tokenizer_module)

    inputs, prompts, labels = [], [], []
    for sample in dataset_module["train_dataset"]:
        if sample["images"]:
            multi_modal_data = {
chenych's avatar
chenych committed
93
94
                "image": template_obj.mm_plugin._regularize_images(
                    sample["images"], image_max_pixels=image_max_pixels, image_min_pixels=image_min_pixels
chenych's avatar
chenych committed
95
96
97
98
99
100
101
                )["images"]
            }
        elif sample["videos"]:
            multi_modal_data = {
                "video": template_obj.mm_plugin._regularize_videos(
                    sample["videos"], image_max_pixels=image_max_pixels, image_min_pixels=image_min_pixels
                )["videos"]
luopl's avatar
luopl committed
102
            }
chenych's avatar
chenych committed
103
104
105
106
107
108
        elif sample["audios"]:
            audio_data = template_obj.mm_plugin._regularize_audios(
                sample["audios"],
                sampling_rate=16000,
            )
            multi_modal_data = {"audio": zip(audio_data["audios"], audio_data["sampling_rates"])}
luopl's avatar
luopl committed
109
110
111
112
        else:
            multi_modal_data = None

        inputs.append({"prompt_token_ids": sample["input_ids"], "multi_modal_data": multi_modal_data})
chenych's avatar
chenych committed
113
        prompts.append(tokenizer.decode(sample["input_ids"], skip_special_tokens=skip_special_tokens))
luopl's avatar
luopl committed
114
        labels.append(
chenych's avatar
chenych committed
115
116
117
            tokenizer.decode(
                list(filter(lambda x: x != IGNORE_INDEX, sample["labels"])), skip_special_tokens=skip_special_tokens
            )
luopl's avatar
luopl committed
118
119
120
121
122
123
        )

    sampling_params = SamplingParams(
        repetition_penalty=generating_args.repetition_penalty or 1.0,  # repetition_penalty must > 0
        temperature=generating_args.temperature,
        top_p=generating_args.top_p or 1.0,  # top_p must > 0
chenych's avatar
chenych committed
124
        top_k=generating_args.top_k or -1,  # top_k must > 0
luopl's avatar
luopl committed
125
126
        stop_token_ids=template_obj.get_stop_token_ids(tokenizer),
        max_tokens=generating_args.max_new_tokens,
chenych's avatar
chenych committed
127
        skip_special_tokens=skip_special_tokens,
chenych's avatar
chenych committed
128
        seed=seed,
luopl's avatar
luopl committed
129
130
131
132
133
134
135
136
137
138
    )
    if model_args.adapter_name_or_path is not None:
        lora_request = LoRARequest("default", 1, model_args.adapter_name_or_path[0])
    else:
        lora_request = None

    engine_args = {
        "model": model_args.model_name_or_path,
        "trust_remote_code": True,
        "dtype": model_args.infer_dtype,
chenych's avatar
chenych committed
139
        "max_model_len": cutoff_len + max_new_tokens,
luopl's avatar
luopl committed
140
141
142
143
144
145
        "tensor_parallel_size": (get_device_count() // pipeline_parallel_size) or 1,
        "pipeline_parallel_size": pipeline_parallel_size,
        "disable_log_stats": True,
        "enable_lora": model_args.adapter_name_or_path is not None,
    }
    if template_obj.mm_plugin.__class__.__name__ != "BasePlugin":
chenych's avatar
chenych committed
146
        engine_args["limit_mm_per_prompt"] = {"image": 4, "video": 2, "audio": 2}
luopl's avatar
luopl committed
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

    if isinstance(model_args.vllm_config, dict):
        engine_args.update(model_args.vllm_config)

    results = LLM(**engine_args).generate(inputs, sampling_params, lora_request=lora_request)
    preds = [result.outputs[0].text for result in results]
    with open(save_name, "w", encoding="utf-8") as f:
        for text, pred, label in zip(prompts, preds, labels):
            f.write(json.dumps({"prompt": text, "predict": pred, "label": label}, ensure_ascii=False) + "\n")

    print("*" * 70)
    print(f"{len(prompts)} generated results have been saved at {save_name}.")
    print("*" * 70)


if __name__ == "__main__":
    fire.Fire(vllm_infer)