[AMD] Skip shared-KV verify test on ROCm 7.0 CI (#36139)
Co-authored-by: chuyeh <chuyeh@users.noreply.github.com>
This commit is contained in:
@@ -11,6 +11,7 @@ from sglang.kernels.ops.attention.verify_mla import verify_shared_kv_fwd
|
|||||||
from sglang.srt.layers.attention.triton_backend import (
|
from sglang.srt.layers.attention.triton_backend import (
|
||||||
_should_use_verify_shared_kv,
|
_should_use_verify_shared_kv,
|
||||||
)
|
)
|
||||||
|
from sglang.srt.utils import get_hip_version
|
||||||
from sglang.test.ci.ci_register import register_amd_ci
|
from sglang.test.ci.ci_register import register_amd_ci
|
||||||
from sglang.test.test_utils import CustomTestCase
|
from sglang.test.test_utils import CustomTestCase
|
||||||
|
|
||||||
@@ -137,6 +138,11 @@ class TestVerifySharedKV(CustomTestCase):
|
|||||||
with self.subTest(l_ext=l_ext):
|
with self.subTest(l_ext=l_ext):
|
||||||
self._run_parity(head_dim=256, v_head_dim=256, l_ext=l_ext)
|
self._run_parity(head_dim=256, v_head_dim=256, l_ext=l_ext)
|
||||||
|
|
||||||
|
@unittest.skipIf(
|
||||||
|
get_hip_version()[:2] == (7, 0),
|
||||||
|
"Triton 3.4 on ROCm 7.0 aborts gfx950 fp8 KV tl.dot "
|
||||||
|
"(triton-lang/triton#8278). Remove once the image uses Triton >= 3.6.",
|
||||||
|
)
|
||||||
def test_qwen3_5_fp8_kv_cache(self):
|
def test_qwen3_5_fp8_kv_cache(self):
|
||||||
self._run_parity(
|
self._run_parity(
|
||||||
head_dim=256,
|
head_dim=256,
|
||||||
|
|||||||
Reference in New Issue
Block a user