test_vision_openai_server_b.py 8.31 KB
Newer Older
1
2
import unittest

3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
from test_vision_openai_server_common import *

from sglang.test.test_utils import (
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
    popen_launch_server,
)


class TestPixtralServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "mistral-community/pixtral-12b"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--mem-fraction-static",
25
                "0.70",
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass


class TestMistral3_1Server(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "unsloth/Mistral-Small-3.1-24B-Instruct-2503"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--mem-fraction-static",
47
                "0.75",
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass


class TestDeepseekVL2Server(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "deepseek-ai/deepseek-vl2-small"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--context-length",
                "4096",
70
                "--disable-cuda-graph",
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass


class TestJanusProServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "deepseek-ai/Janus-Pro-7B"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--mem-fraction-static",
92
                "0.35",
93
94
95
96
            ],
        )
        cls.base_url += "/v1"

97
    def test_video_images_chat_completion(self):
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
        pass

    def test_single_image_chat_completion(self):
        # Skip this test because it is flaky
        pass


## Skip for ci test
# class TestLlama4Server(TestOpenAIVisionServer):
#     @classmethod
#     def setUpClass(cls):
#         cls.model = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
#         cls.base_url = DEFAULT_URL_FOR_TEST
#         cls.api_key = "sk-123456"
#         cls.process = popen_launch_server(
#             cls.model,
#             cls.base_url,
#             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
#             other_args=[
#                 "--chat-template",
#                 "llama-4",
#                 "--mem-fraction-static",
#                 "0.8",
#                 "--tp-size=8",
#                 "--context-length=8192",
#             ],
#         )
#         cls.base_url += "/v1"

#     def test_video_chat_completion(self):
#         pass


class TestGemma3itServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "google/gemma-3-4b-it"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--mem-fraction-static",
144
                "0.70",
145
146
147
148
149
150
151
152
153
                "--enable-multimodal",
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass


154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
class TestGemma3nServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "google/gemma-3n-E2B-it"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--mem-fraction-static",
                "0.70",
                "--cuda-graph-max-bs",
                "1",
            ],
        )
        cls.base_url += "/v1"

174
175
176
177
178
    def test_audio_chat_completion(self):
        self._test_audio_speech_completion()
        # This _test_audio_ambient_completion test is way too complicated to pass for a small LLM
        # self._test_audio_ambient_completion()

179

Stefan He's avatar
Stefan He committed
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
class TestKimiVLServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "moonshotai/Kimi-VL-A3B-Instruct"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--context-length",
                "4096",
                "--dtype",
                "bfloat16",
            ],
        )
        cls.base_url += "/v1"
199

Stefan He's avatar
Stefan He committed
200
201
    def test_video_images_chat_completion(self):
        pass
202
203


204
205
206
class TestPhi4MMServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
207
208
209
210
211
212
213
214
        # Manually download LoRA adapter_config.json as it's not downloaded by the model loader by default.
        from huggingface_hub import constants, snapshot_download

        snapshot_download(
            "microsoft/Phi-4-multimodal-instruct",
            allow_patterns=["**/adapter_config.json"],
        )

215
216
217
        cls.model = "microsoft/Phi-4-multimodal-instruct"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
218
219

        revision = "33e62acdd07cd7d6635badd529aa0a3467bb9c6a"
220
221
222
223
224
225
226
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--mem-fraction-static",
227
                "0.70",
228
229
                "--disable-radix-cache",
                "--max-loras-per-batch",
230
                "2",
231
232
233
234
                "--revision",
                revision,
                "--lora-paths",
                f"vision={constants.HF_HUB_CACHE}/models--microsoft--Phi-4-multimodal-instruct/snapshots/{revision}/vision-lora",
235
                f"speech={constants.HF_HUB_CACHE}/models--microsoft--Phi-4-multimodal-instruct/snapshots/{revision}/speech-lora",
236
237
238
239
            ],
        )
        cls.base_url += "/v1"

240
    def get_vision_request_kwargs(self):
241
242
243
244
245
246
247
        return {
            "extra_body": {
                "lora_path": "vision",
                "top_k": 1,
                "top_p": 1.0,
            }
        }
248

249
250
251
252
253
254
255
256
257
258
259
    def get_audio_request_kwargs(self):
        return {
            "extra_body": {
                "lora_path": "speech",
                "top_k": 1,
                "top_p": 1.0,
            }
        }

    def test_audio_chat_completion(self):
        self._test_audio_speech_completion()
260
        # This _test_audio_ambient_completion test is way too complicated to pass for a small LLM
261
        # self._test_audio_ambient_completion()
262
263


Zijian's avatar
Zijian committed
264
265
266
class TestVILAServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
267
        cls.model = "Efficient-Large-Model/NVILA-Lite-2B-hf-0626"
Zijian's avatar
Zijian committed
268
269
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
270
        cls.revision = "6bde1de5964b40e61c802b375fff419edc867506"
Zijian's avatar
Zijian committed
271
272
273
274
275
276
277
278
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            api_key=cls.api_key,
            other_args=[
                "--trust-remote-code",
                "--context-length=65536",
279
                f"--revision={cls.revision}",
Zijian's avatar
Zijian committed
280
281
282
283
284
            ],
        )
        cls.base_url += "/v1"


285
286
if __name__ == "__main__":
    unittest.main()