update fa backend on gfx926 and gfx928

vllm/platforms/rocm.py

@@ -154,7 +154,7 @@ def use_rocm_custom_paged_attention(
         sinks: Optional[torch.Tensor] = None) -> bool:
 
     GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
-    ON_GFX9 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950"])
+    ON_GFX9 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950", "gfx928", "gfx936"])
     ON_GFX11_GFX12 = any(arch in GPU_ARCH for arch in ["gfx11", "gfx12"])
 
     # custom paged attn always supported on V0. On V1, requires sliding window