[Minor] Remove unused code in attention (#2384)

vllm/model_executor/layers/attention.py

@@ -156,7 +156,6 @@ class PagedAttention(nn.Module):
             output = out.view_as(query)
         else:
             # Decoding run.
-            if key_cache is not None and value_cache is not None:
             output = _paged_attention(
                 query,
                 key_cache,
@@ -166,10 +165,6 @@ class PagedAttention(nn.Module):
                 self.scale,
                 self.alibi_slopes,
             )
-            else:
-                # This happens during the initial memory profiling run for
-                # CUDA graphs.
-                output = torch.zeros_like(query)
 
         # Reshape the output tensor.
         return output.view(batch_size, seq_len, hidden_size)