allow math kernel

modeling_bert.py

@@ -378,7 +378,7 @@ class JinaBertSelfAttention(nn.Module):
             b, _, s, _ = query_layer.shape
             new_bias = attention_mask + bias
             attn = scaled_dot_product_attention(query_layer, key_layer, value_layer, new_bias)
-            attn = attn.permute(0, 2, 1, 3)
+            attn = attn.permute(0, 2, 1, 3).contiguous()
             return (attn.view(b, s, self.all_head_size),)
 
         # Take the dot product between "query" and "key" to get the raw attention scores.