diff --git a/vllm/v1/attention/backends/flash_attn.py b/vllm/v1/attention/backends/flash_attn.py
index 4c2a6c6b985b..f5cc7e9762de 100755
--- a/vllm/v1/attention/backends/flash_attn.py
+++ b/vllm/v1/attention/backends/flash_attn.py
@@ -759,6 +759,9 @@ def cascade_attention(
         if v_descale is not None else None,
     )
 
+    prefix_lse = prefix_lse.contiguous()
+    suffix_lse = suffix_lse.contiguous()
+
     # Merge prefix and suffix outputs, and store the result in output.
     merge_attn_states(output, prefix_output, prefix_lse, suffix_output,
                       suffix_lse)