EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

ci: simplify multi-modality tests by using mixins (#9006)

Browse Source
This commit is contained in:
Mick
2025-08-17 13:25:02 +08:00
committed by GitHub
parent 66d6be0874
commit 1df84ff414
6 changed files with 264 additions and 400 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
test/srt/test_vision_openai_server_a.py
View File

@@ -8,16 +8,28 @@ import unittest
from test_vision_openai_server_common import *
from sglang.srt.utils import kill_process_tree
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestQwen2VLServer(TestOpenAIVisionServer):
class TestLlava(ImageOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "lmms-lab/llava-onevision-qwen2-0.5b-ov"
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,
api_key=cls.api_key,
)
cls.base_url += "/v1"
class TestQwen2VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "Qwen/Qwen2-VL-7B-Instruct"
@@ -37,11 +49,8 @@ class TestQwen2VLServer(TestOpenAIVisionServer):
)
cls.base_url += "/v1"
def test_video_chat_completion(self):
self._test_video_chat_completion()
class TestQwen2_5_VLServer(TestOpenAIVisionServer):
class TestQwen2_5_VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "Qwen/Qwen2.5-VL-7B-Instruct"
@@ -61,9 +70,6 @@ class TestQwen2_5_VLServer(TestOpenAIVisionServer):
)
cls.base_url += "/v1"
def test_video_chat_completion(self):
self._test_video_chat_completion()
class TestVLMContextLengthIssue(CustomTestCase):
@classmethod
@@ -137,11 +143,8 @@ class TestVLMContextLengthIssue(CustomTestCase):
# )
# cls.base_url += "/v1"
# def test_video_chat_completion(self):
# pass
class TestMinicpmvServer(TestOpenAIVisionServer):
class TestMinicpmvServer(ImageOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "openbmb/MiniCPM-V-2_6"
@@ -162,7 +165,7 @@ class TestMinicpmvServer(TestOpenAIVisionServer):
cls.base_url += "/v1"
class TestInternVL2_5Server(TestOpenAIVisionServer):
class TestInternVL2_5Server(ImageOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "OpenGVLab/InternVL2_5-2B"
@@ -181,7 +184,7 @@ class TestInternVL2_5Server(TestOpenAIVisionServer):
cls.base_url += "/v1"
class TestMinicpmoServer(TestOpenAIVisionServer):
class TestMinicpmoServer(ImageOpenAITestMixin, AudioOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "openbmb/MiniCPM-o-2_6"
@@ -201,12 +204,8 @@ class TestMinicpmoServer(TestOpenAIVisionServer):
)
cls.base_url += "/v1"
def test_audio_chat_completion(self):
self._test_audio_speech_completion()
self._test_audio_ambient_completion()
class TestMimoVLServer(TestOpenAIVisionServer):
class TestMimoVLServer(ImageOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "XiaomiMiMo/MiMo-VL-7B-RL"
@@ -228,6 +227,95 @@ class TestMimoVLServer(TestOpenAIVisionServer):
cls.base_url += "/v1"
class TestVILAServer(ImageOpenAITestMixin):
@classmethod
def setUpClass(cls):
cls.model = "Efficient-Large-Model/NVILA-Lite-2B-hf-0626"
cls.base_url = DEFAULT_URL_FOR_TEST
cls.api_key = "sk-123456"
cls.revision = "6bde1de5964b40e61c802b375fff419edc867506"
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",
f"--revision={cls.revision}",
"--cuda-graph-max-bs",
"4",
],
)
cls.base_url += "/v1"
class TestPhi4MMServer(ImageOpenAITestMixin, AudioOpenAITestMixin):
@classmethod
def setUpClass(cls):
# 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"],
)
cls.model = "microsoft/Phi-4-multimodal-instruct"
cls.base_url = DEFAULT_URL_FOR_TEST
cls.api_key = "sk-123456"
revision = "33e62acdd07cd7d6635badd529aa0a3467bb9c6a"
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",
"--disable-radix-cache",
"--max-loras-per-batch",
"2",
"--revision",
revision,
"--lora-paths",
f"vision={constants.HF_HUB_CACHE}/models--microsoft--Phi-4-multimodal-instruct/snapshots/{revision}/vision-lora",
f"speech={constants.HF_HUB_CACHE}/models--microsoft--Phi-4-multimodal-instruct/snapshots/{revision}/speech-lora",
"--cuda-graph-max-bs",
"4",
],
)
cls.base_url += "/v1"
def get_vision_request_kwargs(self):
return {
"extra_body": {
"lora_path": "vision",
"top_k": 1,
"top_p": 1.0,
}
}
def get_audio_request_kwargs(self):
return {
"extra_body": {
"lora_path": "speech",
"top_k": 1,
"top_p": 1.0,
}
}
# This _test_audio_ambient_completion test is way too complicated to pass for a small LLM
def test_audio_ambient_completion(self):
pass
if __name__ == "__main__":
del TestOpenAIVisionServer
del (
TestOpenAIOmniServerBase,
ImageOpenAITestMixin,
VideoOpenAITestMixin,
AudioOpenAITestMixin,
)
unittest.main()