Merge pull request #96 from casper-hansen/fix_attention_mask

Only apply attention mask if seqlen is greater than 1

awq/modules/fused/attn.py

@@ -176,7 +176,8 @@ class QuantAttentionFused(nn.Module):
             if self.use_alibi:
                 scores = self.alibi.forward(scores, seqlen)
 
-            if attention_mask is not None:
+            # When seqlen is 1, there is nothing else to attend to
+            if attention_mask is not None and seqlen > 1:
                 scores = scores + attention_mask  # (bs, n_local_heads, slen, cache_len + slen)
                 
             scores = F.softmax(scores.float(), dim=-1).type_as(xq)