[CI][ROCm] Restrict MLA cross-layer KV cache test to supported backends on ROCm (#46401)
Signed-off-by: aarushjain29 <Aarushi.Jain2@amd.com>
Aarushi Jain committed
Jun 22, 2026 at 17:05 UTC
fbf9ff7cf4a466387126097f15c3f00da316e6cb
1 file changed
+7
tests/v1/kv_connector/unit/test_kv_cache_layout.py
+7
@@ -3,6 +3,8 @@
3
4
import pytest
5
6
+from vllm.platforms import current_platform
7
+
8
9
def test_mla_common_backend_rejects_cross_layer_kv_cache():
10
"""MLACommonBackend defaults to the identity permutation (layers dim
@@ -24,8 +26,13 @@ def test_mla_common_backend_rejects_cross_layer_kv_cache():
26
27
@pytest.mark.parametrize(
28
"backend_path",
29
+ # See: https://github.com/vllm-project/vllm/issues/46411
30
[
31
"vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend",
32
+ ]
33
+ if current_platform.is_rocm()
34
+ else [
35
+ "vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend",
36
"vllm.v1.attention.backends.mla.cutlass_mla.CutlassMLABackend",
37
"vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend",
38
"vllm.v1.attention.backends.mla.flashmla.FlashMLABackend",