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

Double vision prefill throughput by defaulting to optimal vision attention backend (#8484)

Browse Source 
Co-authored-by: Xiang (Kevin) Li <lik@nvidia.com>
This commit is contained in:
Kevin Xiang Li
2025-08-13 02:08:30 -07:00
committed by GitHub
parent 35e6bc92e3
commit 3b3b3baf9f
3 changed files with 20 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
test/srt/test_vision_openai_server_common.py
View File

@@ -328,13 +328,14 @@ class TestOpenAIVisionServer(CustomTestCase):
or "person" in video_response
or "individual" in video_response
or "speaker" in video_response
or "presenter" in video_response
or "Steve" in video_response
or "hand" in video_response
), f"""
====================== video_response =====================
{video_response}
===========================================================
should contain 'man' or 'person' or 'individual' or 'speaker' or 'hand'
should contain 'man' or 'person' or 'individual' or 'speaker' or 'presenter' or 'Steve' or 'hand'
"""
assert (
"present" in video_response
@@ -347,7 +348,6 @@ class TestOpenAIVisionServer(CustomTestCase):
===========================================================
should contain 'present' or 'examine' or 'display' or 'hold'
"""
assert "black" in video_response or "dark" in video_response
self.assertIsNotNone(video_response)
self.assertGreater(len(video_response), 0)
@@ -385,8 +385,9 @@ class TestOpenAIVisionServer(CustomTestCase):
or "person" in video_response
or "individual" in video_response
or "speaker" in video_response
or "presenter" in video_response
or "hand" in video_response
), f"video_response: {video_response}, should either have 'man' in video_response, or 'person' in video_response, or 'individual' in video_response, or 'speaker' in video_response or 'hand' in video_response"
), f"video_response: {video_response}, should either have 'man' in video_response, or 'person' in video_response, or 'individual' in video_response or 'speaker' in video_response or 'presenter' or 'hand' in video_response"
assert (
"present" in video_response
or "examine" in video_response