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

Use more general heuristics to set the default value of --mem-fraction-static (#10975)

Browse Source 
Co-authored-by: sglang-bot <sglangbot@gmail.com>
This commit is contained in:
Lianmin Zheng
2025-09-29 10:11:03 -07:00
committed by GitHub
parent 816b3a433a
commit a17e70f5cc
9 changed files with 167 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
test/srt/test_multi_instance_release_memory_occupation.py
View File

@@ -1,6 +1,6 @@
import multiprocessing
import os
import subprocess
import time
import traceback
import unittest
from multiprocessing import Process
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
TEST_SUITE = dict(
model_path=DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
mem_fraction_static=0.85,
mem_fraction_static=0.83,
dp_size=2,
tp_size=2,
)
@@ -214,6 +214,9 @@ def _run_sglang_subprocess(
_mem_usage = get_gpu_memory_gb(rank)
print(f"GPU{rank} Memory usage after resuming Sgl weights: {_mem_usage}")
del hf_model
hf_model = None
torch.cuda.empty_cache()
time.sleep(5)
torch.cuda.empty_cache()
_curr_usage = get_gpu_memory_gb(rank)
assert (