[Fix] memory leak by overlap + retract (#11981)

Co-authored-by: Liangsheng Yin <lsyincs@gmail.com>
This commit is contained in:
cctry
2025-10-23 07:59:23 -07:00
committed by GitHub
parent 6c18addb6f
commit b0b4f71679
9 changed files with 132 additions and 25 deletions

View File

@@ -114,7 +114,6 @@ class Envs:
# Test & Debug
SGLANG_IS_IN_CI = EnvBool(False)
SGLANG_IS_IN_CI_AMD = EnvBool(False)
SGLANG_TEST_RETRACT = EnvBool(False)
SGLANG_SET_CPU_AFFINITY = EnvBool(False)
SGLANG_PROFILE_WITH_STACK = EnvBool(True)
SGLANG_RECORD_STEP_TIME = EnvBool(False)
@@ -128,6 +127,11 @@ class Envs:
SGLANG_SIMULATE_ACC_METHOD = EnvStr("multinomial")
SGLANG_TORCH_PROFILER_DIR = EnvStr("/tmp")
# Scheduler: memory leak test
SGLANG_TEST_RETRACT = EnvBool(False)
SGLANG_TEST_RETRACT_INTERVAL = EnvInt(3)
SGLANG_ENABLE_RUNTIME_MEM_LEAK_CHECK = EnvBool(False)
# Scheduler: new token ratio hyperparameters
SGLANG_INIT_NEW_TOKEN_RATIO = EnvFloat(0.7)
SGLANG_MIN_NEW_TOKEN_RATIO_FACTOR = EnvFloat(0.14)