Fix attention

cacheflow/models/attention.py

-from typing import Optional, Tuple
+from typing import Optional
 
 import torch
 import torch.nn as nn
@@ -24,8 +24,12 @@ class OPTCacheFlowAttention(nn.Module):
         key: torch.Tensor,
         value: torch.Tensor,
     ) -> None:
+        query = query.unsqueeze(0)
+        key = key.unsqueeze(0)
+        value = value.unsqueeze(0)
         out = xops.memory_efficient_attention(
             query, key, value, attn_bias=self.attention_mask, scale=self.scale)
+        out = out.squeeze(0)
         # FIXME(woosuk): Directly write the attention output.
         output.copy_(out, non_blocking=True)