bench_sglang.py 2.82 KB
Newer Older
Lianmin Zheng's avatar
Lianmin Zheng committed
1
2
3
4
5
6
7
import argparse
import json
import os
import time
import uuid

import sglang as sgl
Liangsheng Yin's avatar
Liangsheng Yin committed
8
9
10
11
from sglang.test.test_utils import (
    add_common_sglang_args_and_parse,
    select_sglang_backend,
)
Lianmin Zheng's avatar
Lianmin Zheng committed
12
13
14
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


def load_questions(filename):
    questions = []
    with open(filename, "r") as fin:
        for line in fin:
            obj = json.loads(line)
            questions.append(obj)
    return questions


def write_answers(filename, model_id, questions, answers):
    with open(os.path.expanduser(filename), "w") as fout:
        for i in range(len(answers)):
            ans_json = {
                "question_id": questions[i]["question_id"],
                "answer_id": uuid.uuid4().hex,
                "model_id": model_id,
                "choices": {
                    "index": 0,
                    "turns": [answers[i][0], answers[i][1]],
                },
                "tstamp": time.time(),
            }
            fout.write(json.dumps(ans_json) + "\n")


@sgl.function
def answer_mt_bench(s, question_1, question_2):
    s += sgl.system()
    s += sgl.user(question_1)
    s += sgl.assistant(sgl.gen("answer_1"))
    s += sgl.user(question_2)
    s += sgl.assistant(sgl.gen("answer_2"))


def main(args):
    # Construct prompts
Liangsheng Yin's avatar
Liangsheng Yin committed
50
    questions = load_questions(args.question_file)[: args.num_questions]
Lianmin Zheng's avatar
Lianmin Zheng committed
51
    arguments = [
Liangsheng Yin's avatar
Liangsheng Yin committed
52
        {"question_1": q["turns"][0], "question_2": q["turns"][1]} for q in questions
Lianmin Zheng's avatar
Lianmin Zheng committed
53
54
55
56
57
58
59
60
61
62
63
64
    ]

    # Select backend
    backend = select_sglang_backend(args)
    sgl.set_default_backend(backend)

    # Run requests
    tic = time.time()
    rets = answer_mt_bench.run_batch(
        arguments,
        temperature=0,
        max_new_tokens=256,
Liangsheng Yin's avatar
Liangsheng Yin committed
65
66
67
        num_threads=args.parallel,
        progress_bar=True,
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
    answers = [[s["answer_1"], s["answer_2"]] for s in rets]
    latency = time.time() - tic

    print(f"#questions: {len(questions)}, Latency: {latency:.2f}")

    # Write results
    model_id = backend.model_info["model_path"]
    answer_file = args.answer_file or f"tmp_output_{args.backend}.txt"
    write_answers(answer_file, model_id, questions, answers)

    with open(args.result_file, "a") as fout:
        value = {
            "task": "mtbench",
            "backend": args.backend,
            "num_gpus": 1,
            "latency": round(latency, 3),
            "num_requests": args.num_questions,
            "other": {
                "num_questions": args.num_questions,
                "parallel": args.parallel,
Liangsheng Yin's avatar
Liangsheng Yin committed
88
            },
Lianmin Zheng's avatar
Lianmin Zheng committed
89
90
91
92
93
94
95
96
97
98
99
        }
        fout.write(json.dumps(value) + "\n")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--question-file", type=str, default="question.jsonl")
    parser.add_argument("--answer-file", type=str, default=None)
    parser.add_argument("--num-questions", type=int, default=80)
    args = add_common_sglang_args_and_parse(parser)
    main(args)