Use level-1 (quiet) busy memory check in chunked-prefill and streaming tests (#27303)
This commit is contained in:
@@ -402,7 +402,7 @@ class StreamingSessionServerBase(CustomTestCase):
|
|||||||
- `env_overrides`: list of `(env_attr_name, value)` tuples; each is
|
- `env_overrides`: list of `(env_attr_name, value)` tuples; each is
|
||||||
pushed onto the `setUpClass` context stack so the env override is
|
pushed onto the `setUpClass` context stack so the env override is
|
||||||
live during `popen_launch_server` and torn down on
|
live during `popen_launch_server` and torn down on
|
||||||
`tearDownClass`-time. `SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY=2`
|
`tearDownClass`-time. `SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY=1`
|
||||||
is always applied on top of these.
|
is always applied on top of these.
|
||||||
"""
|
"""
|
||||||
|
|
||||||
@@ -417,7 +417,7 @@ class StreamingSessionServerBase(CustomTestCase):
|
|||||||
|
|
||||||
with contextlib.ExitStack() as stack:
|
with contextlib.ExitStack() as stack:
|
||||||
stack.enter_context(
|
stack.enter_context(
|
||||||
envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(2)
|
envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(1)
|
||||||
)
|
)
|
||||||
for name, val in cls.env_overrides:
|
for name, val in cls.env_overrides:
|
||||||
stack.enter_context(getattr(envs, name).override(val))
|
stack.enter_context(getattr(envs, name).override(val))
|
||||||
|
|||||||
@@ -29,7 +29,7 @@ class TestMixedChunkedPrefill(GSM8KMixin, CustomTestCase):
|
|||||||
|
|
||||||
@classmethod
|
@classmethod
|
||||||
def setUpClass(cls):
|
def setUpClass(cls):
|
||||||
with envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(2):
|
with envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(1):
|
||||||
cls.process = popen_launch_server(
|
cls.process = popen_launch_server(
|
||||||
cls.model,
|
cls.model,
|
||||||
cls.base_url,
|
cls.base_url,
|
||||||
|
|||||||
Reference in New Issue
Block a user