From fbf9ff7cf4a466387126097f15c3f00da316e6cb Mon Sep 17 00:00:00 2001 From: Aarushi Jain <142941703+aarushjain29@users.noreply.github.com> Date: Mon, 22 Jun 2026 17:05:26 -0500 Subject: [PATCH] [CI][ROCm] Restrict MLA cross-layer KV cache test to supported backends on ROCm (#46401) Signed-off-by: aarushjain29 --- tests/v1/kv_connector/unit/test_kv_cache_layout.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/tests/v1/kv_connector/unit/test_kv_cache_layout.py b/tests/v1/kv_connector/unit/test_kv_cache_layout.py index 1da64e0b7e..313b06c6c6 100644 --- a/tests/v1/kv_connector/unit/test_kv_cache_layout.py +++ b/tests/v1/kv_connector/unit/test_kv_cache_layout.py @@ -3,6 +3,8 @@ import pytest +from vllm.platforms import current_platform + def test_mla_common_backend_rejects_cross_layer_kv_cache(): """MLACommonBackend defaults to the identity permutation (layers dim @@ -24,8 +26,13 @@ def test_mla_common_backend_rejects_cross_layer_kv_cache(): @pytest.mark.parametrize( "backend_path", + # See: https://github.com/vllm-project/vllm/issues/46411 [ "vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend", + ] + if current_platform.is_rocm() + else [ + "vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend", "vllm.v1.attention.backends.mla.cutlass_mla.CutlassMLABackend", "vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend", "vllm.v1.attention.backends.mla.flashmla.FlashMLABackend",