[AMD] skip deterministic inference for MLA FP8 test (#23382)
Co-authored-by: YC Yen-Ching Tseng <yctseng@amd.com>
This commit is contained in:
co-authored by
YC Yen-Ching Tseng
parent
1340d8b3bd
commit
fd88a1c562
@@ -8,6 +8,7 @@ from sglang.test.test_utils import (
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
CustomTestCase,
|
||||
is_in_amd_ci,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
@@ -23,11 +24,7 @@ class TestMLA(CustomTestCase, MGSMEnMixin):
|
||||
def setUpClass(cls):
|
||||
cls.model = DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
other_args=[
|
||||
other_args = [
|
||||
"--trust-remote-code",
|
||||
"--kv-cache-dtype",
|
||||
"fp8_e5m2",
|
||||
@@ -38,8 +35,15 @@ class TestMLA(CustomTestCase, MGSMEnMixin):
|
||||
# straddle the 0.8 threshold. With deterministic inference,
|
||||
# the score becomes a fixed function of (model, weights, CUDA
|
||||
# stack), so threshold-edge flakes stop being random noise.
|
||||
"--enable-deterministic-inference",
|
||||
],
|
||||
]
|
||||
if not is_in_amd_ci():
|
||||
# On AMD, the default attention backend (aiter) is not in the deterministic-inference allowlist, so the server fails to start, disable it.
|
||||
other_args.append("--enable-deterministic-inference")
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
other_args=other_args,
|
||||
)
|
||||
|
||||
@classmethod
|
||||
|
||||
Reference in New Issue
Block a user