diff --git a/test/registered/attention/test_verify_shared_kv.py b/test/registered/attention/test_verify_shared_kv.py index 330d4e37d..6a4549742 100644 --- a/test/registered/attention/test_verify_shared_kv.py +++ b/test/registered/attention/test_verify_shared_kv.py @@ -11,6 +11,7 @@ from sglang.kernels.ops.attention.verify_mla import verify_shared_kv_fwd from sglang.srt.layers.attention.triton_backend import ( _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.test_utils import CustomTestCase @@ -137,6 +138,11 @@ class TestVerifySharedKV(CustomTestCase): with self.subTest(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): self._run_parity( head_dim=256,