test_mla_deepseek_v3.py 6.1 KB
Newer Older
1
2
3
4
5
6
import unittest
from types import SimpleNamespace

import requests
import torch

7
from sglang.srt.utils import is_cuda, is_hip, kill_process_tree
8
9
10
11
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
12
    CustomTestCase,
13
14
15
16
    popen_launch_server,
)


17
class TestMLADeepseekV3(CustomTestCase):
18
19
20
21
    @classmethod
    def setUpClass(cls):
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
22
        other_args = ["--trust-remote-code", "--chunked-prefill-size", "256"]
23
        if is_cuda():
24
25
26
27
28
29
30
31
32
            other_args.extend(["--enable-torch-compile", "--cuda-graph-max-bs", "2"])
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=other_args,
        )

    @classmethod
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)

    def test_gsm8k(self):
        args = SimpleNamespace(
            num_shots=5,
            data_path=None,
            num_questions=200,
            max_new_tokens=512,
            parallel=128,
            host="http://127.0.0.1",
            port=int(self.base_url.split(":")[-1]),
        )
        metrics = run_eval_few_shot_gsm8k(args)
        print(metrics)

        self.assertGreater(metrics["accuracy"], 0.62)


@unittest.skipIf(is_hip(), "FA is not available.")
class TestMLADeepseekV3Fa3Fp8Kvcache(CustomTestCase):
    @classmethod
    def setUpClass(cls):
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = [
            "--trust-remote-code",
            "--chunked-prefill-size",
            "256",
            "--kv-cache-dtype",
            "fp8_e4m3",
        ]
        if is_cuda():
            other_args.extend(["--attention-backend", "fa3"])
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=other_args,
        )

    @classmethod
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)

    def test_gsm8k(self):
        args = SimpleNamespace(
            num_shots=5,
            data_path=None,
            num_questions=200,
            max_new_tokens=512,
            parallel=128,
            host="http://127.0.0.1",
            port=int(self.base_url.split(":")[-1]),
        )
        metrics = run_eval_few_shot_gsm8k(args)
        print(metrics)

        self.assertGreater(metrics["accuracy"], 0.62)


94
class TestDeepseekV3MTP(CustomTestCase):
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
    @classmethod
    def setUpClass(cls):
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
        other_args = [
            "--trust-remote-code",
            "--cuda-graph-max-bs",
            "2",
            "--disable-radix",
            "--enable-torch-compile",
            "--torch-compile-max-bs",
            "1",
            "--speculative-algorithm",
            "EAGLE",
            "--speculative-num-steps",
            "2",
            "--speculative-eagle-topk",
            "4",
            "--speculative-num-draft-tokens",
            "4",
        ]
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=other_args,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)

    def test_gsm8k(self):
        requests.get(self.base_url + "/flush_cache")

        args = SimpleNamespace(
            num_shots=5,
            data_path=None,
            num_questions=200,
            max_new_tokens=512,
            parallel=128,
            host="http://127.0.0.1",
            port=int(self.base_url.split(":")[-1]),
        )
        metrics = run_eval_few_shot_gsm8k(args)
        print(metrics)

        self.assertGreater(metrics["accuracy"], 0.60)

        server_info = requests.get(self.base_url + "/get_server_info")
145
146
147
        avg_spec_accept_length = server_info.json()["internal_states"][0][
            "avg_spec_accept_length"
        ]
148
149
150
151
152
153
        print(f"{avg_spec_accept_length=}")
        self.assertGreater(avg_spec_accept_length, 2.5)


# compatible with old APIs
class TestDeepseekV3MTPWithDraft(CustomTestCase):
154
155
156
157
    @classmethod
    def setUpClass(cls):
        cls.model = "lmsys/sglang-ci-dsv3-test"
        cls.base_url = DEFAULT_URL_FOR_TEST
Lianmin Zheng's avatar
Lianmin Zheng committed
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
        other_args = [
            "--trust-remote-code",
            "--cuda-graph-max-bs",
            "2",
            "--disable-radix",
            "--enable-torch-compile",
            "--torch-compile-max-bs",
            "1",
            "--speculative-algorithm",
            "EAGLE",
            "--speculative-draft",
            "lmsys/sglang-ci-dsv3-test-NextN",
            "--speculative-num-steps",
            "2",
            "--speculative-eagle-topk",
            "4",
            "--speculative-num-draft-tokens",
            "4",
        ]
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=other_args,
        )

    @classmethod
    def tearDownClass(cls):
        kill_process_tree(cls.process.pid)

    def test_gsm8k(self):
        requests.get(self.base_url + "/flush_cache")

        args = SimpleNamespace(
            num_shots=5,
            data_path=None,
            num_questions=200,
            max_new_tokens=512,
            parallel=128,
            host="http://127.0.0.1",
            port=int(self.base_url.split(":")[-1]),
        )
        metrics = run_eval_few_shot_gsm8k(args)
        print(metrics)

        self.assertGreater(metrics["accuracy"], 0.60)

        server_info = requests.get(self.base_url + "/get_server_info")
206
207
208
        avg_spec_accept_length = server_info.json()["internal_states"][0][
            "avg_spec_accept_length"
        ]
209
210
211
212
213
214
        print(f"{avg_spec_accept_length=}")
        self.assertGreater(avg_spec_accept_length, 2.5)


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