diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py
index 1d0714305cc8..0bedd80e5ecf 100644
--- a/vllm/platforms/rocm.py
+++ b/vllm/platforms/rocm.py
@@ -109,7 +109,8 @@ def use_rocm_custom_paged_attention(qtype: torch.dtype, head_size: int,
     ON_MI250_MI300 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942"])
 
     # rocm custom page attention not support on navi (gfx1*)
-    return (ON_MI250_MI300 and not ON_NAVI and (sliding_window == 0)
+    return (ON_MI250_MI300 and not ON_NAVI
+            and (sliding_window == 0 or sliding_window == (-1, -1))
             and (qtype == torch.half or qtype == torch.bfloat16)
             and (head_size == 64 or head_size == 128)
             and (block_size == 16 or block_size == 32)