From 5f2a473ff3245c0db065c62568a515b72007ac57 Mon Sep 17 00:00:00 2001 From: Andreas Karatzas Date: Thu, 8 Jan 2026 00:37:50 -0600 Subject: [PATCH] [ROCm][CI] v1 cpu offloading attention backend fix (#31833) Signed-off-by: Andreas Karatzas --- tests/v1/kv_offload/test_cpu_offloading.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/tests/v1/kv_offload/test_cpu_offloading.py b/tests/v1/kv_offload/test_cpu_offloading.py index 239f3b5d41..8aeeffcaea 100644 --- a/tests/v1/kv_offload/test_cpu_offloading.py +++ b/tests/v1/kv_offload/test_cpu_offloading.py @@ -15,10 +15,12 @@ from vllm.distributed.kv_events import BlockStored, KVEventBatch from vllm.platforms import current_platform CPU_BLOCK_SIZES = [48] -ATTN_BACKENDS = ["FLASH_ATTN", "TRITON_ATTN"] +ATTN_BACKENDS = [] if current_platform.is_cuda(): - ATTN_BACKENDS.append("FLASHINFER") + ATTN_BACKENDS = ["FLASH_ATTN", "FLASHINFER", "TRITON_ATTN"] +elif current_platform.is_rocm(): + ATTN_BACKENDS = ["TRITON_ATTN"] class MockSubscriber: