bench_sglang.py 3.88 KB
Newer Older
Lianmin Zheng's avatar
Lianmin Zheng committed
1
2
3
4
5
6
7
8
import argparse
import ast
import json
import re
import time

import numpy as np

Lianmin Zheng's avatar
Lianmin Zheng committed
9
from sglang.api import set_default_backend
Liangsheng Yin's avatar
Liangsheng Yin committed
10
11
12
13
from sglang.test.test_utils import (
    add_common_sglang_args_and_parse,
    select_sglang_backend,
)
Lianmin Zheng's avatar
Lianmin Zheng committed
14
from sglang.utils import download_and_cache_file, dump_state_text, read_jsonl
Lianmin Zheng's avatar
Lianmin Zheng committed
15
16
17
18
19

INVALID = -9999999


def get_one_example(lines, i, include_answer):
Liangsheng Yin's avatar
Liangsheng Yin committed
20
    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
Lianmin Zheng's avatar
Lianmin Zheng committed
21
22
23
24
25
26
27
28
29
30
31
32
33
34
    if include_answer:
        ret += " " + lines[i]["answer"]
    return ret


def get_few_shot_examples(lines, k):
    ret = ""
    for i in range(k):
        ret += get_one_example(lines, i, True) + "\n\n"
    return ret


def get_answer_value(answer_str):
    answer_str = answer_str.replace(",", "")
Liangsheng Yin's avatar
Liangsheng Yin committed
35
    numbers = re.findall(r"\d+", answer_str)
Lianmin Zheng's avatar
Lianmin Zheng committed
36
37
38
39
40
41
42
43
44
    if len(numbers) < 1:
        return INVALID
    try:
        return ast.literal_eval(numbers[-1])
    except SyntaxError:
        return INVALID


def main(args):
Lianmin Zheng's avatar
Lianmin Zheng committed
45
46
47
48
49
50
51
    # Select backend
    set_default_backend(select_sglang_backend(args))

    # Read data
    url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
    filename = download_and_cache_file(url)
    lines = list(read_jsonl(filename))
Lianmin Zheng's avatar
Lianmin Zheng committed
52
53

    # Construct prompts
Lianmin Zheng's avatar
Lianmin Zheng committed
54
55
56
    num_questions = args.num_questions
    num_shots = args.num_shots
    few_shot_examples = get_few_shot_examples(lines, num_shots)
Lianmin Zheng's avatar
Lianmin Zheng committed
57
58
59

    questions = []
    labels = []
Lianmin Zheng's avatar
Lianmin Zheng committed
60
    for i in range(len(lines[:num_questions])):
Lianmin Zheng's avatar
Lianmin Zheng committed
61
62
63
64
65
66
67
68
69
70
71
72
73
74
        questions.append(get_one_example(lines, i, False))
        labels.append(get_answer_value(lines[i]["answer"]))
    assert all(l != INVALID for l in labels)
    arguments = [{"question": q} for q in questions]

    #####################################
    ######### SGL Program Begin #########
    #####################################

    import sglang as sgl

    @sgl.function
    def few_shot_gsm8k(s, question):
        s += few_shot_examples + question
75
76
77
        s += sgl.gen(
            "answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
78
79
80
81
82
83
84
85

    #####################################
    ########## SGL Program End ##########
    #####################################

    # Run requests
    tic = time.time()
    states = few_shot_gsm8k.run_batch(
Liangsheng Yin's avatar
Liangsheng Yin committed
86
87
88
89
90
        arguments,
        temperature=0,
        num_threads=args.parallel,
        progress_bar=True,
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
91
92
93
94
95
96
    latency = time.time() - tic

    preds = []
    for i in range(len(states)):
        preds.append(get_answer_value(states[i]["answer"]))

Lianmin Zheng's avatar
Lianmin Zheng committed
97
98
99
    # print(f"{preds=}")
    # print(f"{labels=}")

Lianmin Zheng's avatar
Lianmin Zheng committed
100
101
102
    # Compute accuracy
    acc = np.mean(np.array(preds) == np.array(labels))
    invalid = np.mean(np.array(preds) == INVALID)
Lianmin Zheng's avatar
Lianmin Zheng committed
103
104
105
106
107
108
109
110

    # Compute speed
    num_output_tokens = sum(
        s.get_meta_info("answer")["completion_tokens"] for s in states
    )
    output_throughput = num_output_tokens / latency

    # Print results
Lianmin Zheng's avatar
Lianmin Zheng committed
111
    print(f"Accuracy: {acc:.3f}")
Lianmin Zheng's avatar
Lianmin Zheng committed
112
113
114
    print(f"Invalid: {invalid:.3f}")
    print(f"Latency: {latency:.3f} s")
    print(f"Output throughput: {output_throughput:.3f} token/s")
Lianmin Zheng's avatar
Lianmin Zheng committed
115

Lianmin Zheng's avatar
Lianmin Zheng committed
116
    # Dump results
Lianmin Zheng's avatar
Lianmin Zheng committed
117
118
119
120
121
122
123
124
125
126
127
128
129
    dump_state_text(f"tmp_output_{args.backend}.txt", states)

    with open(args.result_file, "a") as fout:
        value = {
            "task": "gsm8k",
            "backend": args.backend,
            "num_gpus": 1,
            "latency": round(latency, 3),
            "accuracy": round(acc, 3),
            "num_requests": args.num_questions,
            "other": {
                "num_questions": args.num_questions,
                "parallel": args.parallel,
Liangsheng Yin's avatar
Liangsheng Yin committed
130
            },
Lianmin Zheng's avatar
Lianmin Zheng committed
131
132
133
134
135
136
        }
        fout.write(json.dumps(value) + "\n")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
Lianmin Zheng's avatar
Lianmin Zheng committed
137
    parser.add_argument("--num-shots", type=int, default=5)
Lianmin Zheng's avatar
Lianmin Zheng committed
138
139
140
141
    parser.add_argument("--data-path", type=str, default="test.jsonl")
    parser.add_argument("--num-questions", type=int, default=200)
    args = add_common_sglang_args_and_parse(parser)
    main(args)