[ROCm] [Bugfix] `compute_attn_mask_seqlen` for qwen3 omni (#29974)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>

vllm/model_executor/models/qwen3_omni_moe_thinker.py

@@ -494,7 +494,10 @@ class Qwen3Omni_VisionTransformer(nn.Module):
         cu_seqlens: torch.Tensor,
     ) -> torch.Tensor:
         max_seqlen = torch.zeros([], device=cu_seqlens.device)
-        if self.attn_backend == AttentionBackendEnum.FLASH_ATTN:
+        if self.attn_backend in {
+            AttentionBackendEnum.FLASH_ATTN,
+            AttentionBackendEnum.ROCM_AITER_FA,
+        }:
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
         return max_seqlen