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

chore: tune mem fraction static for vlm (#6881)

Browse Source
This commit is contained in:
Mick
2025-07-19 08:19:27 +08:00
committed by GitHub
parent 9c7a46180c
commit 3964b352c3
4 changed files with 57 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
test/srt/test_vision_openai_server_a.py
View File

@@ -30,7 +30,7 @@ class TestQwen2VLServer(TestOpenAIVisionServer):
api_key=cls.api_key,
other_args=[
"--mem-fraction-static",
"0.4",
"0.35",
],
)
cls.base_url += "/v1"
@@ -52,7 +52,7 @@ class TestQwen2_5_VLServer(TestOpenAIVisionServer):
api_key=cls.api_key,
other_args=[
"--mem-fraction-static",
"0.4",
"0.35",
],
)
cls.base_url += "/v1"
@@ -75,7 +75,7 @@ class TestVLMContextLengthIssue(CustomTestCase):
other_args=[
"--context-length",
"300",
"--mem-fraction-static=0.80",
"--mem-fraction-static=0.75",
],
)
cls.base_url += "/v1"
@@ -147,7 +147,7 @@ class TestMinicpmvServer(TestOpenAIVisionServer):
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
"0.4",
"0.35",
],
)
cls.base_url += "/v1"
@@ -181,7 +181,7 @@ class TestMinicpmoServer(TestOpenAIVisionServer):
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
"0.7",
"0.65",
],
)
cls.base_url += "/v1"

8
test/srt/test_vision_openai_server_b.py
View File

@@ -22,7 +22,7 @@ class TestPixtralServer(TestOpenAIVisionServer):
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
"0.73",
"0.70",
],
)
cls.base_url += "/v1"
@@ -44,7 +44,7 @@ class TestMistral3_1Server(TestOpenAIVisionServer):
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
"0.8",
"0.75",
],
)
cls.base_url += "/v1"
@@ -88,7 +88,7 @@ class TestJanusProServer(TestOpenAIVisionServer):
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
"0.4",
"0.35",
],
)
cls.base_url += "/v1"
@@ -197,7 +197,7 @@ class TestPhi4MMServer(TestOpenAIVisionServer):
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
"0.75",
"0.70",
"--disable-radix-cache",
"--max-loras-per-batch",
"1",