bench_sglang.py 2.6 KB
Newer Older
Lianmin Zheng's avatar
Lianmin Zheng committed
1
2
3
4
5
import argparse
import json
import time

import sglang as sgl
Liangsheng Yin's avatar
Liangsheng Yin committed
6
7
8
9
10
from sglang.test.test_utils import (
    add_common_sglang_args_and_parse,
    select_sglang_backend,
)
from sglang.utils import dump_state_text, read_jsonl
Lianmin Zheng's avatar
Lianmin Zheng committed
11
12
13
14
15
16
17
18
19
20
21
22
23


@sgl.function
def multi_document_qa(s, docs, question):
    s += sgl.user_begin()
    s += "Pleaes answer a question according to given documents.\n"
    s += "Question:" + question + "Documents begin.\n"

    forks = s.fork(len(docs))
    forks += lambda i: docs[i]
    forks.join("concate_and_append")

    s += "\nDocuments end."
Liangsheng Yin's avatar
Liangsheng Yin committed
24
25
26
27
28
    s += (
        "\n\nBased on the above documents, please answer this question:\n"
        + question
        + "\nAnswer in three words or fewer."
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
29
30
31
32
33
34
35
36
37
    s += sgl.user_end()
    s += sgl.assistant(sgl.gen("answer", max_tokens=16))


def main(args):
    lines = read_jsonl(args.data_path)
    l = lines[0]
    arguments = []
    labels = []
Liangsheng Yin's avatar
Liangsheng Yin committed
38
39
40
41
42
43
44
    for i in range(len(l["questions"][: args.num_questions])):
        arguments.append(
            {
                "docs": l["documents"][:10],
                "question": l["questions"][i],
            }
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
45
46
47
48
49
50
51
52
53
        labels.append(l["answers"][i])

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

    # Run requests
    tic = time.time()
    states = multi_document_qa.run_batch(
Liangsheng Yin's avatar
Liangsheng Yin committed
54
55
        arguments, temperature=0, num_threads=args.parallel, progress_bar=True
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
56
57
    latency = time.time() - tic

Liangsheng Yin's avatar
Liangsheng Yin committed
58
    # Compute accuracy
Lianmin Zheng's avatar
Lianmin Zheng committed
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
    print([s["answer"] for s in states])
    correct = 0
    for s, label in zip(states, labels):
        answer = s["answer"].lower()
        if all(x in answer for x in label.lower().split(" ")):
            correct += 1
    accuracy = correct / len(labels)
    print(f"Accuracy: {accuracy:.3f}")
    print(f"Latency: {latency:.3f}")

    # Write results
    dump_state_text(f"tmp_output_{args.backend}.txt", states)

    with open(args.result_file, "a") as fout:
        value = {
            "task": "multi_document_qa",
            "backend": args.backend,
            "num_gpus": 1,
            "latency": round(latency, 3),
            "num_requests": args.num_questions,
            "accuracy": accuracy,
            "other": {
                "num_questions": args.num_questions,
                "parallel": args.parallel,
Liangsheng Yin's avatar
Liangsheng Yin committed
83
            },
Lianmin Zheng's avatar
Lianmin Zheng committed
84
85
86
87
88
89
90
91
92
93
        }
        fout.write(json.dumps(value) + "\n")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--data-path", type=str, default="questions.jsonl")
    parser.add_argument("--num-questions", type=int, default=100)
    args = add_common_sglang_args_and_parse(parser)
    main(args)