test_triton_attention_backend.py 1.58 KB
Newer Older
1
2
3
4
5
"""
Usage:
python3 -m unittest test_triton_attention_backend.TestTritonAttnBackend.test_mmlu
"""

Lianmin Zheng's avatar
Lianmin Zheng committed
6
7
8
9
10
11
12
import unittest
from types import SimpleNamespace

from sglang.srt.utils import kill_child_process
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
    DEFAULT_MODEL_NAME_FOR_TEST,
13
14
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
15
    is_in_ci,
Lianmin Zheng's avatar
Lianmin Zheng committed
16
    popen_launch_server,
17
    run_bench_one_batch,
Lianmin Zheng's avatar
Lianmin Zheng committed
18
19
20
21
)


class TestTritonAttnBackend(unittest.TestCase):
22
    def test_latency(self):
23
        output_throughput = run_bench_one_batch(
24
25
26
27
28
29
            DEFAULT_MODEL_NAME_FOR_TEST,
            [
                "--attention-backend",
                "triton",
                "--enable-torch-compile",
            ],
Lianmin Zheng's avatar
Lianmin Zheng committed
30
31
        )

32
        if is_in_ci():
33
            assert output_throughput > 153, f"{output_throughput=}"
Lianmin Zheng's avatar
Lianmin Zheng committed
34
35

    def test_mmlu(self):
36
37
38
39
40
41
42
        model = DEFAULT_MODEL_NAME_FOR_TEST
        base_url = DEFAULT_URL_FOR_TEST
        process = popen_launch_server(
            model,
            base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=["--attention-backend", "triton"],
Lianmin Zheng's avatar
Lianmin Zheng committed
43
44
        )

45
46
47
48
49
50
51
52
53
54
        try:
            args = SimpleNamespace(
                base_url=base_url,
                model=model,
                eval_name="mmlu",
                num_examples=64,
                num_threads=32,
            )

            metrics = run_eval(args)
55
            self.assertGreaterEqual(metrics["score"], 0.65)
56
        finally:
Lianmin Zheng's avatar
Lianmin Zheng committed
57
            kill_child_process(process.pid, include_self=True)
Lianmin Zheng's avatar
Lianmin Zheng committed
58
59
60
61


if __name__ == "__main__":
    unittest.main()